Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

China Clears AI Firms to Buy Nvidia H200 Chips

China plans to allow domestic AI companies to purchase Nvidia’s H200 chips, according to The Information. The H200, a high-bandwidth memory GPU used for AI training and inference, sits below the H100 in export restriction tiers. The move would mark a policy shift after years of tightening U.S. controls that pushed Chinese firms toward domestically developed alternatives such as Huawei’s Ascend series. No official announcement has been made by either government. (Bloomberg Tech)

Vik: Guess Jensen got what he wanted, as he explained on the Dwarkesh Podcast, to sell chips to China to keep them hooked on American AI accelerators. It might be too little, too late though. China is already underway building their own domestic compute. Now everyone is worried that GLM 5.2 is smart enough to break through cybersecurity measures.

Austin: It’s never too late! Way easier to just buy and build with the best than bootstrap, and I think momentum will shift Nvidia’s way here. That said, Pandora’s box has been opened, and China’s government will fund the development and buildout of datacenters using domestic silicon, networking, power, and so on.

SK Hynix $28B ADR Bookbuild Oversubscribed, Closes Wednesday

SK Hynix will close the bookbuild for its $28 billion American depositary receipt offering on Wednesday after demand exceeded the available supply, according to a source familiar with the matter. The oversubscription signals strong investor appetite for the U.S. listing, which is part of a dual strategy that also includes a 1.4 trillion won commitment to expand its domestic AI supply chain in South Korea. (Reuters, The Chosun Daily)

Vik: Cerebras IPO was oversubscribed too, and look what happened to the stock price after. Guess its IPO hunting season after all.

Austin: Yeah SpaceX too. Never underestimate the power of get-rich-quick-schemes. Let’s be honest, thats why IPOs are oversubscribed.

Nvidia and d-Matrix Combine Hardware in Joint AI System

Nvidia and inference chip startup d-Matrix are integrating their respective hardware into a joint system for running AI models, per The Information. D-Matrix, founded in 2019, last raised $275 million at a $2 billion valuation in November and is reportedly in talks for a new round. The tie-up fits a pattern of Nvidia partnering with inference-focused competitors as a hedge, similar to earlier collaborations. (The Next Web)

Vik: No announcement from d-Matrix yet, but I am not sure what “integrating into a joint system” means. Is NVIDIA acquihiring? Are they going to just demonstrate how NVIDIA GPUs are used for prefill, and d-Matrix Corsair for decode? I’d bet on the latter, or some variation of it (like Attention-FeedForward Disaggregation maybe?).

Austin: Heterogenous compute. Maybe they saw my interview with Gimlet Labs and what Gimlet showed is possible with d-Matrix + Nvidia GPUs.

Chipstrat

An Interview with the Gimlet Labs Team About Heterogeneous Inference for AI Agents

I’ve been writing for a while about the shift from a one-size-fits-all GPU to multi-vendor, multi-silicon environments, so I wanted to talk to Gimlet directly about how cross-vendor orchestration actually works — and why most neoclouds, locked into a single-silicon vendor by equity terms, can’t compete with this model by design…

Read more

2 months ago · 21 likes · Austin Lyons

SambaNova Raises $1B Series F at $11B Valuation, Lands JPMorganChase as Customer

SambaNova closed a $1 billion Series F at an $11 billion valuation, five months after its prior funding round. General Atlantic led the oversubscribed raise, with Intel among existing backers. Separately, JPMorganChase signed on as a customer to deploy SambaNova’s hardware for on-premises AI inference. (Bloomberg, Reuters)

Vik: It’s always great for an inference startup when they get a new round of funding to keep the lights on, but what is even better is that they have a customer signed up to deploy their hardware on-prem. This is an interesting trend to me, where companies are starting to deploy their own compute hardware to avoid sensitive data from leaking back to the AI labs. Not all inference startups now need to be “acquired” like Groq by a big hardware company.

APS Acquires Korean Transformer Automation Leader UPI for 18B Won

APS is buying UPI, which holds a 95% share of South Korea’s cut-to-length machine market for transformer manufacturing, for 18 billion won in cash. UPI supplies core processing, winding, and E-stacking automation gear used to build transformers. APS is positioning the deal as an entry into AI data center power infrastructure, riding demand for 765kV ultra-high-voltage transformers and grid replacement projects across the U.S., Europe, and Middle East. (The Elec)

Vik: APS explains what UPI’s moat is:

“The transformer manufacturing equipment market has high barriers to entry because it requires sophisticated equipment design and control technologies, extensive customer-specific process expertise, and long-term supply references.”

Austin: I have a soft spot for heavy machinery / manufacturing companies. The industry dynamics are interesting and the technologies are usually surprisingly advanced and impressive. Would love a tour and an education, give me a call APS!

ZML Offers Software to Speed AI Inference Across Chips

ZML, a French AI startup, released ZML/LLMD, an LLM inference server designed to optimize inference across various AI chips, including those from Nvidia, AMD, Google, Apple, and Intel. The software aims to break vendor lock-in and allow enterprises and clouds to use a mix of chips for AI workloads. ZML has raised $20 million from venture firms. (TechCrunch)

Vik: I can imagine there are many uses for a hardware agnostic inference server. It will break all the ecosystem lock-ins, and treat compute as a fungible entity. According to founder Steeve Morin (that’s the first name, not a typo; also not to be confused with the serial killer), ZML is a competitor to vLLM and SGLang, but goes beyond them to co-design silicon.

Austin: I’ve been tracking ZML for quite some time. Interesting company. The recent Modular acquisition by Qualcomm was proof there’s something to the idea.

Chipstrat

ZML's Hardware-Agnostic Distributed Inference

I recently saw an interesting Linkedin post from French startup ZML demonstrating vendor-agnostic Llama 2 inference. The demo leveraged pipeline parallelism to distribute the inference across hardware from Nvidia, AMD, and Google. Each chip was housed in a separate location, and the inference results were streamed back to the Mac that kicked off the job. Check out the video on the…

Read more

2 years ago · 3 likes · Austin Lyons

Apple-Broadcom Chip Deal Tops $30 Billion

Apple has expanded its partnership with Broadcom into 2031, agreeing to spend at least $30 billion on chips under the multiyear supply agreement. Here is a quote from Apple’s IR page: (Apple IR)

“Broadcom is part of Apple’s American Manufacturing Program (AMP), launched last year to accelerate manufacturing in the U.S. This new agreement, which marks Apple’s largest AMP commitment to date, will enable Broadcom to expand and modernize its manufacturing facilities in Fort Collins, Colorado, with a $1.5 billion capital expenditure investment. Broadcom will produce advanced radio frequency components, including FBAR filters, and advanced wireless connectivity technologies at the Fort Collins facility.”

Vik: When this news broke a few days ago, I thought this has something to do with Apple making custom AI ASICs with Broadcom (for sure, I wasn’t thinking clearly; really wtf). As it turns out, the actual press release says nothing about AI chips. The whole thing is about Broadcom making RF chips and FBAR filters for phones and such. This whole thing reeks of a publicity stunt to ensure that Apple products are “made on American soil”.

Must Watch

This is sort of a Bloom Energy (Sc)andal.

Hunterbrook reports that one of their critical elements required for Bloom’s behind-the-meter energy sources is scandium, and their supply chain is from China. Reportedly, every shipment of scandium from China needs CCP approval. But the Bloom CEO has repeatedly denied that they are reliant on China. If Hunterbrook is right, then China has an OFF switch to US datacenters, as Hunterbrook put it. Spicy!

That said, Hunterbrook Capital is short $BE. So… yeah.

And Bloom says these fuel cells are “supplementing grid power” not the primary off-grid AI-datacenter use case the stock story is built on.

Oh yeah, and most of those datacenters aren’t actually stood up yet.

So… take it all with a grain of salt:

Sector Watch

Advanced Packaging

• Besi faces intensifying competition in hybrid bonding as rivals crowd into its once-uncontested “blue ocean” market. (bits-chips.com)

AI & Compute

• OpenAI will publicly launch its GPT-5.6 family of models on Thursday after regulatory clearance. (The Information)

Infrastructure

• Blue Owl unveils a new infrastructure investment venture focused on data center development to meet growing AI compute demand. (Bloomberg Tech)

Hiring & Layoffs

• Microsoft lays off approximately 4,800 employees as the company accelerates its AI infrastructure investment and reshapes its workforce accordingly. (The420.in)

• Eaton names Dan T. Simpson President of Global Energy Infrastructure Solutions, a business unit serving data center and power grid markets. (Financial Times)

EDA

• Arteris and IC-Link by imec announce a collaboration to accelerate next-generation AI and HPC chip design using Arteris’s network-on-chip IP technology. (Arteris)

  Thanks for reading! This post is public so feel free to share it.

  Share

Separately, Amazon launched a $25 billion bond offering, one of the largest corporate debt sales in US history, to fund its AI buildout. Wolfspeed sued Navitas over SiC patents, and Samsung delayed CXL 3.1 memory production while it waits on CXL-ready CPUs from Intel and AMD.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Apple Tests Chinese CXMT DRAM for China-Market Devices

Apple is testing DRAM chips from Chinese manufacturer CXMT for use in devices sold in China, according to the Financial Times. CXMT, which is on the Pentagon’s blacklist of companies with alleged ties to China’s military, produces chips that would be confined to the China market under this arrangement, avoiding direct conflict with US export controls. The testing covers potential supply for iPhones and other devices sold domestically in China.

Separately, Lexar announced its new DDR5-7600 desktop memory modules will use CXMT-sourced chips, marking another sign of the Chinese chipmaker gaining traction with consumer electronics brands outside the dominant Samsung, SK Hynix, and Micron triad. (CNBC, Financial Times)

Vik: “While CXMT’s capacity is expanding, it is unlikely to immediately flood the market with cheap chips, as its output is largely pre-committed, Ray Wang, a memory analyst at SemiAnalysis, told the FT.” — CXMT is #4 in the memory market after the big 3. Memory is like a broken Jevon’s paradox: The more expensive it gets, the more people will buy. 🤦🏽‍♂️

Austin: Ray got quoted in the Financial Times, nice job dude! Vik we should schedule Ray to come talk shop on our pod.

FuriosaAI Partners With Equinix to Deploy AI Inference in Europe

South Korean AI chip startup FuriosaAI has signed a partnership with Equinix to deploy its inference accelerators across Equinix’s European data centers. The deal gives FuriosaAI access to Equinix’s colocation footprint spanning multiple European markets, targeting customers seeking alternatives to Nvidia-based inference infrastructure. (The Register, Jon Peddie Research, Seoul Economic Daily)

Vik: I am guessing that a lot of inference accelerators with a TCO focus will find homes in sovereign AI deployments around the world, especially in cost sensitive places or those without too much energy surplus.

Austin: Europe allows the deployment of AI accelerators? 🙃

Amazon Raises $25B in Bonds for AI Infrastructure

Amazon launched a $25 billion bond offering, one of the largest corporate debt sales in U.S. history, to fund its AI infrastructure buildout. The multi-tranche deal spans maturities ranging from 3 to 40 years and follows the company’s previously announced plan to spend roughly $100 billion in capital expenditures in 2025, the bulk of which is directed at data centers and AI compute capacity. (SiliconANGLE, CNBC)

Vik: How do you feel about lending your money to Amazon to build AI infra? But lending to Amazon is not a bet on the AI spend. Bondholders get paid from all of Amazon (AWS, retail, ads), and the coupon is fixed, so if the bet turns into a gold mine you still get the same payment.

Austin: I would feel just fine. Amazon can monetize GPUs, whether via AWS or via e-commerce (ads or improving results etc).

Wolfspeed Sues Navitas for SiC Patent Infringement

Wolfspeed has filed a patent infringement lawsuit against Navitas Semiconductor, alleging that Navitas’s silicon carbide (SiC) products violate Wolfspeed’s intellectual property. The suit targets Navitas’s SiC technology, which Wolfspeed claims infringes on its patented processes and designs. Wolfspeed, one of the largest SiC manufacturers, is seeking damages and an injunction to halt sales of the accused products. (Wolfspeed)

Vik: The lawsuit encompasses both GaN and SiC products from Navitas — from the patent listing on the Wolfspeed announcement, it looks like they are going after the entire Navitas portfolio! Not a good look.

Samsung Begins Mass Production of NVMe Drives for Nvidia Vera Rubin

Samsung has begun mass production of NVMe storage drives designed for Nvidia’s Vera Rubin AI platform. The drives are built to meet Nvidia’s specifications for the Vera Rubin system, which is slated to succeed the current Blackwell architecture. (Bloomberg.com)

Vik: Cute looking drive

Austin: Samsung selling storage and memory into Vera Rubin. If only Samsung Foundry could participate in logic for Vera Rubin too, they’d be even richer.

Zhipu AI weighs custom chip due to demand, export controls

Zhipu AI, the Chinese AI lab behind the GLM series of open-source AI models, is considering designing its own AI chip. This move is driven by surging demand for its models and U.S. export controls, which are creating computing resource constraints. The company has made preliminary inquiries with Chinese chip design houses about developing a bespoke AI processor optimized for its models. (theinformation.com)

Vik: See the trend developing here? Each model maker is teaming up with a chip design house to make their own chips. OpenAI’s Jalapeno, Anthropic chip in the works, and now zAI? The hardware and model layer is coalescing. Pay attention.

Austin: The company formerly known as Zhipu AI, but now called z.ai but formally known as Knowledge Atlas Technology JSC Ltd?

DeepSeek Develops Custom AI Chip to Cut Nvidia Dependence

Chinese AI startup DeepSeek is developing its own AI chip, according to Reuters, in a move to reduce reliance on Nvidia hardware. The effort is separate from DeepSeek’s reported near-term use of Huawei’s Ascend chips as an interim solution while US export controls limit access to Nvidia’s H100 and H800 GPUs. DeepSeek, backed by Chinese quant fund High-Flyer, gained international attention earlier this year with its R1 model. No timeline or production partners for the in-house chip have been disclosed. (Reuters, Bloomberg.com)

Vik: What! Not DeepSeek too! Who will buy Nvidia chips now? jk, Nvidia does not sell in China after export controls. Everything is priced in because it is well known. Nothing to see here.

Austin: This is bearish Huawei, not Nvidia.

Samsung Delays CXL 3.1 Memory Production Due to CPU Delays

Samsung Electronics has postponed mass production of its CXL 3.1-based memory module (CMM-D 3.0) due to delays in the release of CXL 3.1-supporting CPUs from Intel and AMD. The company will continue producing CXL 2.0-based memory modules (CMM-D 2.0) for now. While Samsung initially aimed for CMM-D 3.0 customer sample production in June, full-scale mass production is now anticipated for next year. (thelec.kr)

Vik: Important dependency to note here: No point in making CXL 3.1 memory if there is no CPU to pair it with. Watch for Intel Diamond Rapids and AMD Venice launches.

Synopsys Exits Fab Manufacturing Software to Concentrate on AI Design

Synopsys is discontinuing its chip fab manufacturing control software products to redirect focus toward AI-driven chip design tools, according to people familiar with the matter. The EDA company is cutting select software lines used in semiconductor fab operations management as part of the strategic reallocation. (Reuters, grafa.com)

Vik: Wait what are these products anyway? From the article — “The ​affected products include the Equipment Engineering System (EES) and Fault Detection and Classification (FDC), a set of automation software that acts as the central nervous ​system of semiconductor fabrication plants to monitor and detect any anomalies before they cascade into costly defects” … dunno why, sounds important. Somebody comment and tell us what this means.

Analog Devices Closes Empower Semiconductor Acquisition

Analog Devices (ADI) has completed its acquisition of Empower Semiconductor, a power management chip company. The deal extends ADI’s power portfolio from grid to core, strengthening its position in AI infrastructure power delivery. (Analog Devices)

Vik: Empower has a whole slew of integrated voltage regulator products which fit next to the chip ensuring vertical power delivery. Now this IP belongs to Analog Devices. Could be a strong power player in the future — note to self: watch closely.

Austin: Vertical power is a big deal. I wrote about it here:

Chipstrat

Power Moves Into the Package. Empower, PowerLattice, and the IVR Socket

A 2.3 kW Vera Rubin pulls ~3,286 amps at the die. A Hopper H100 pulled ~1,000. Nearly 11x the conduction loss in three generations! Not good…

Read more

a month ago · 21 likes · Austin Lyons

Must Watch

Beautiful video from NVIDIA showing why a fast single core performance helps when running with GPUs.

Read the full blog post.

Quick Hits

• NAND Flash NAND prices surge to rival DRAM pricing as AI-driven memory demand tightens supply. (The Chosun Daily)

• Memory Wall Analysis argues cache miss tolerance, not raw compute throughput, now defines CPU performance amid the memory wall. (EDN)

• China smartphone sales fell 13% during the 618 shopping festival, with elevated memory component costs limiting device discounts and dampening consumer demand. (Reuters)

• Syntiant, an Intel-backed edge AI inference chip and software maker, has filed for a US IPO according to Bloomberg. (Bloomberg.com)

• Penguin Solutions releases Q3 fiscal 2026 financial results, covering revenue and profitability across its memory, networking, and AI server solutions portfolio. (Penguin Solutions) — Record net sales of $479 million, up 48% versus the year-ago quarter

  Thanks for reading! This post is public so feel free to share it.

  Share

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast playe

Nvidia denies Kyber AI server delay, affirms 2027 launch

Nvidia refuted a SemiAnalysis report claiming its next-generation Kyber AI server systems face a one-year delay. An Nvidia spokesperson stated the company’s roadmap “remains intact.” Nvidia plans to launch its Kyber server, which will feature a new vertical rack design to accommodate 144 GPUs, with the Vera Rubin Ultra platform in the second half of 2027. (finance.yahoo.com)

Vik: Not exactly a rebuttal to the SemiAnalysis article. If it is, Nvidia will probably say so at an earnings call, not as an off-cycle remark to a social media post. That’s how the big leagues play.

Kingboard hikes CCL prices again on rising material costs

Kingboard, a leading Chinese copper-clad laminate (CCL) manufacturer, announced price increases for its products, including FR-4 and PP prepreg by 15%, and CEM-1/22F by 10%. Copper foil processing fees will also rise by RMB 5 to 8 per kilogram, reflecting increased costs for glass fiber cloth and copper foil. This marks a continued trend of price adjustments in the CCL industry due to high demand from AI servers and rising raw material prices. (ctee.com.tw)

Vik: Although there was a temporary concern after SemiAnalysis saying that the midplane PCB in Kyber was delayed, AI PCB makers still face high demand. Strong signal.

Austin: Are CCL price increases going to impact consumers too? i.e. just like memory and recently CPUs?

PC Brands Adopt Chinese Memory to Reduce Costs

PC brands like Lenovo, ASUS, MSI, Gigabyte, and Acer are increasingly adopting Chinese memory and storage components from companies such as YMTC, CXMT, and BIWIN due to their price advantage. Lenovo has expanded its use of Chinese components, including YMTC SSDs in flagship laptops. MSI has completed tuning for CXMT DDR5 memory on AMD platforms, and Acer uses BIWIN-manufactured memory modules, some with CXMT standard particles. (ctee.com.tw)

Vik: Memory/storage pressures will be relieved in the consumer markets from Chinese suppliers. AI demand requires a different level of cutting edge performance from both memory and storage, which might not be Chinese suppliers’ strong suit (yet).

Austin: If I’m YMTC or CXMT, I’d use this opportunity of increased demand and increased profits to double-down on R&D. Just having supply might win you revenue today, but innovation drives differentiation and success tomorrow,

TeraWulf secures $19 billion Anthropic lease, sells Abernathy stake

TeraWulf announced a 20-year lease agreement with Anthropic at its Justified Data campus in Hawesville, Kentucky, expected to generate approximately $19 billion in contracted revenue. The campus will accommodate 401 MW of critical IT load, with initial capacity in service by H2 2027. Separately, TeraWulf sold its 50.1% ownership in the Abernathy Joint Venture to an investor group led by Fluidstack, monetizing its $450 million investment at a premium. (investors.terawulf.com)

Vik: 20 year agreement is a long time. Good on TeraWulf for landing an important client.

Austin: I would have titled this “Bitcoin-miner-turned-neocloud flips stake in Abernathy TX datacenter to even-more-neo-cloud”.

Broadcom Extends Apple Custom Chip Supply Deal to 2031

Broadcom has extended its custom chip supply agreement with Apple through 2031, expanding the scope of products and new custom chips it will deliver. The long-term commitment signals Apple’s increasing use of custom ASICs for its products and data center inference stack. Financial terms were not disclosed. (Bloomberg Tech, Reuters)

Vik: Google may be messing around with customer owned tooling (CoT) with MediaTek, but Apple is striking a long term deal with to ensure that Apple can deploy their own custom ASICs for AI.

Austin: The reports of my death are greatly exaggerated. —Broadcom

SK Hynix formally launches $28-29B US ADR listing

SK Hynix has formally launched a US American Depositary Receipt offering targeting a valuation of $28 to $29 billion, according to Bloomberg. The South Korean memory chipmaker, which currently trades on the Korea Exchange, is pursuing the US listing in part to secure a valuation premium that its domestic market has not fully reflected. The ADR structure would give US-based investors direct exposure to SK Hynix without routing trades through Seoul. (Bloomberg.com)

Vik: This basically means that you can buy SK Hynix shares in the US stock market. Supposedly, the domestic market has not fully reflected its valuation 👀 — Also kudos to on X for reading the original filing that shows both Coatue Management and Situational Awareness have indicated an interest in purchasing up to US $7B in these ADRs.

Austin: Get ready for some US retail nonsense! Volatility, on.

Huawei Brings 8,192-Chip Atlas SuperPod Cluster to South Korea

Huawei is entering the South Korean AI chip market with its Atlas SuperPod system, a cluster packing 8,192 Ascend 950 accelerators per deployment. The company claims the system delivers triple the inference performance of Nvidia’s H20 at one-quarter the cost, positioning it as a direct alternative to Nvidia hardware in a market where the H20 remains available following U.S. export controls that restrict more advanced Nvidia chips. (Tom’s Hardware)

Vik: Hahaha! 3x the performance of a nerfed dinosaur Nvidia chip. Why this weird comparison? So they can say 3x? How does it compare to a Blackwell? Hmm?

Austin: South Korea doesn’t have export controls? So this must be a “cheap” alternative—as in lower upfront CapEx?

Must Watch

Brilliant video from Anthropic — recommend!

Also here is the link —> full whitepaper — give it a read

Sector Watch

Memory

• CXMT ChangXin Memory Technology has ended its DRAM price war amid surging AI-driven demand. (chosun.com)

• Micron is investing to expand its Hiroshima, Japan chip fab as part of a multibillion-dollar capacity buildout. (hardwareluxx.de)

Advanced Packaging

• Samsung Electro-Mechanics is reportedly forming a glass-core substrate joint venture with a Sumitomo unit, targeting 2H27 production. (TrendForce)

Foundry & Logic

• TSMC is localizing equipment and materials suppliers into a resilient “second fleet” to secure CoWoS and CoPoS packaging supply chains. (DigiTimes)

• Tesla hired 17-year Intel fab veteran Gary Jiang, likely to oversee Terafab’s licensing of Intel’s 14A process. (Toms Hardware)

AI & Compute

• Nvidia longtime sales chief Jay Puri is retiring. (The Information)

Power & Infrastructure

• MLCCs AI-driven demand for high-end multilayer ceramic capacitors is pushing Japanese and Korean suppliers’ book-to-bill ratios to post-pandemic highs, raising 2H26 shortage risk. (TrendForce)

EDA & IP

• Intel 18A Synopsys detailed the foundation IP ecosystem underpinning Intel’s 18A process and why it matters for adoption. (SemiWiki)

Infrastructure

• SK Telecom announces a 15GW AI data center buildout plan for South Korea, targeting a position as the region’s leading AI infrastructure hub. (EE News Europe)

• KT Corp commits $12 billion to networks and AI data centers as South Korean telcos accelerate investment in domestic AI infrastructure capacity. (Light Reading)

• Csquare, Brookfield’s data center subsidiary, seeks $1.35 billion in an IPO to fund continued AI-driven data center construction and expansion. (Bloomberg.com)

Packaging

• LG Chem begins mass production supply of semiconductor process strippers to Amkor Technology, expanding into advanced packaging materials for outsourced assembly. (The Elec)

Hiring & Layoffs

• Celestica announces a leadership transition in its Connectivity and Cloud Solutions segment as the company sharpens its focus on AI infrastructure products. (Celestica)

• Microsoft cuts approximately 4,800 jobs in an AI-driven workforce restructuring, joining the broader wave of tech-sector headcount reductions. (The Detroit News)

• ASM International appoints KPN CFO Chris Figee as its new chief financial officer, filling the top finance role at the Dutch semiconductor equipment maker. (Reuters)

• Samsung appliance division workers plan a protest rally against the separate wage deal negotiated by the company’s semiconductor unit employees. (Reuters)

EDA

• Siemens EDA and Samsung Foundry deepen collaboration on silicon design enablement, covering process design kits and verification flows for advanced nodes. (dqindia.com)

By the Numbers

Closing moves, Jul 6:

• Up: IREN +13.1%, CRDO +9.8%, LCID +9.5%, ANET +8.3%, RIVN +8.1%, CLSK +7.1%

Why:

• IREN — Anthropic eyeing a massive Australian data center deal directly benefits IREN, Australia’s largest publicly listed AI data center operator.

• CRDO — Russell index reconstitution drove mechanical buying as CRDO was swept into index moves alongside a strong demand outlook.

• BE +9.8% — Brookfield expanded its AI power agreement with Bloom Energy to $25B, guaranteeing long-term fuel cell demand.

• ALAB +6.5% — Bank of America assigned Astera Labs a new street-high price target, triggering a broad re-rating.

• IMOS +15.1%, AEHR +3.4% — no clear public catalyst yet

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Nvidia Kyber NVL144 Delayed Over 12 Months to 2028

Nvidia’s Kyber NVL144, demoed at GTC three months prior, has been delayed by over 12 months, pushing its release to 2028. This delay also follows the cancellation of Nvidia’s NVL72x2 back-to-back rack architecture, which limits Rubin Ultra’s scale-up domain. (SemiAnalysis)

Vik: I had a wild conspiracy theory on X a month ago that this will happen! Lots of people seem to not believe that this is true though, and some say to wait for official announcements. The discussion on X is quite wide; lots of opinions and takes to process.

Austin: Favorable news for AMD, with Helios rack (MI455X) shipping 2H26 and MI500 series in 2027. AMD’s Advancing AI event is in two weeks, maybe we’ll hear more about the 2027+ roadmap and MI500 there?

Nvidia Introduces Revenue-Sharing Financing Model for AI Cloud Providers

Nvidia is deploying a revenue-sharing financing model to underwrite AI chip purchases by emerging cloud providers. In exchange for a cut of their rental revenues and a backstop for unused inventory, the model allows customers to acquire Hopper and Blackwell GPUs that their balance sheets might otherwise constrain. This strategy expands Nvidia’s addressable market by de-risking CapEx for customers. (The Information)

Vik: The recurring revenue model will smooth out financials for Nvidia instead of waiting for big purchase orders. Qualcomm’s % cut from handset provider sales for CDMA was a good source of money for decades for them. I wonder if this is the start of something similar.

Austin: Financing is an under appreciated competitive advantage Nvidia has against AI accelerator startups.

Global AI Startup Investment Reached $510B in H1 2026

Global startup investment hit $510 billion in the first half of 2026, with OpenAI and Anthropic capturing 43% of total venture capital. In parallel, AI CapEx from US hyperscalers is projected to reach 3.2% of GDP by 2027. The massive capital allocation confirms a structural industry shift from capability races to capacity deployment, with power constraints emerging as a tangible bottleneck. (Crunchbase)

Vik: Lion’s share of the startup investment is going to the top two, but still plenty of spending in AI overall. Looking forward to another half-tril in H2 2026.

Austin: It’s a good time to be raising money, even if you’re not OpenAI or Anthropic; the article says 16 companies raised billion-dollar rounds in the second quarter of 2026!

Intel Hikes Select CPU Prices, Citing Supply Costs and Demand

Intel confirmed price increases on some consumer and server CPUs, attributing the changes to market dynamics, rising costs, and high demand. Select enthusiast processors increased by $30 to $50, while data center-grade products saw increases of hundreds or thousands of dollars. For example, some Xeon 8000-series ‘Emerald Rapids’ processors now have higher Recommended Customer Prices than at their late 2023 release. (tomshardware.com)

Vik: If Intel can charge more for CPUs in an era of high demand, which AMD forecasts will continue till 2030/31, then CPU revenues 🚀… good for Intel’s financials for sure. Exciting to see how things will roll out at the next earnings call.

Austin: Inflation hits consumers again. First it was memory. Now CPUs. Folks are gonna have a love-hate relationship with AI. Love the usefulness, but hate the inflation.

Hon Hai Q2 Sales Beat Estimates on Record AI Server Revenue

Hon Hai Precision Industry reported second-quarter revenue that exceeded analyst estimates, with record sales in its AI server division as demand from hyperscale and enterprise customers stayed elevated. The Taiwan-based contract manufacturer, a key assembler in Nvidia’s supply chain, attributed the outperformance to sustained AI infrastructure spending; Oracle’s expanding capital commitments were cited among the demand drivers. (Bloomberg.com, Reuters)

Vik: Hon Hai, aka, Foxconn, assembles servers for Nvidia hardware. The article reports that AI servers have overtaken consumer electronics in terms of revenue. This is a proxy compute capacity really, and so far there is no slowdown even as there are fears of overcapacity.

South Korea Plans Growth Fund Backed by Chip Tax Revenue

South Korea is considering establishing a growth fund seeded by tax revenue from the semiconductor industry, according to Yonhap News. The move comes as domestic chipmakers Samsung Electronics and SK Hynix are committed to spending at least 1,350 trillion won ($880 billion) combined on chips and data centers, investments the government aims to leverage into a broader industrial financing vehicle. (Bloomberg Tech)

Vik: Feed the windfall from memory companies into a growth fund for other sectors. Good diversification strategy.

Austin: Heck yeah. Agree. Korea shouldn’t overinvest all its dollars in memory right now.

Tweet #1: Intel XBM

Intel XBM is good to see. Just don’t make it have the same fate as Optane. Too soon? Read the patent for yourself.

Tweet #2: Interconnects >> Memory

Amazing post by Big Boss on X, which explains from first principles why we should be designing AI accelerators with interconnects in mind.

CLUES

It is one of only three companies that dominate the world market for DRAM, and the only one of the three based in the United States. The other two are South Korean.

SPOILER ALERT: ANSWER BELOW

The company was Qualcomm, founded in 1985 by seven engineers who had left a San Diego outfit called Linkabit. Irwin Jacobs and Andrew Viterbi led the group, and they gave their new venture a name of almost defiant plainness: a portmanteau of “Quality Communications.” They had a big idea about how radio spectrum could work. They did not, yet, have a way to pay salaries with it.

So while that idea matured, they built something far less glamorous: OmniTRACS, a satellite terminal bolted to the roof of a truck cab, letting a dispatcher a thousand miles away see exactly where his fleet was and radio instructions straight to the driver. It was logistics software before logistics software was cool, and by 1989, one customer, Schneider National, accounted for half of Qualcomm’s $32 million in revenue.

That trucking money bought Qualcomm time, and it needed every bit of it, because the big idea was a hard sell.

It was called CDMA, and the concept struck most of the cellular industry as slightly unhinged: instead of making callers politely take turns on a channel, let everyone speak at once, with each voice wrapped in its own mathematical code and untangled at the receiving end.

The industry had already committed to a rival approach, TDMA, and when Qualcomm proposed CDMA in 1989, the response was a firm no. Engineers called it unworkable. The Wall Street Journal published warnings that a failure at scale could cost the industry billions. People in the trade still refer to the fight as the Holy Wars of Wireless, which tells you how seriously everyone took it.

Qualcomm won anyway.

The FCC ruled that carriers could deploy whatever standard they liked, so Qualcomm skipped the gatekeepers and sold CDMA straight to the operators. It became a formal standard in 1993 and went on to underpin the entire third generation of mobile networks. The industry has since moved on to newer methods for 4G and 5G, but Qualcomm’s patents and modems came along for the ride regardless.

The technology that respectable engineers had declared impossible became the thing an entire generation of phones was quietly built on, paid for in its infancy by keeping tabs on trucks.

As Qualcomm now reforms itself into an AI datacenter company, it could really use some of that early magic that made them the kingpin of communications technology for decades.

Could High Bandwidth Compute be it?

Vik's Newsletter

Qualcomm's High Bandwidth Compute and the Packaging Problem It Moved

At its 2026 Investor Day, Qualcomm introduced High Bandwidth Compute (HBC), which is DRAM stacked on top of a logic die with a dense interface that gives high bandwidth between the two. It lays the foundation for Qualcomm’s route into the data center using a memory architecture that does not use the traditional HBM+CoWoS approach…

Read more

8 days ago · 35 likes · Vikram Sekar

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast playe

CXMT Files $4.3B IPO, Secures $3B Tencent DRAM Deal

ChangXin Memory Technologies (CXMT), China’s largest DRAM maker, filed for a 29.5 billion yuan ($4.3 billion) listing on Shanghai’s Star Market and disclosed a $3 billion DRAM supply agreement with Tencent. The Hefei-based chipmaker is pursuing the offering during a global memory pricing upcycle driven by AI computing demand, with US export controls and competition in high-bandwidth memory flagged as key risks in its prospectus. (StorageNewsletter, SCMP)

Vik: What a great time to IPO! At the peak of the memory market. CXMT Q1 revenue was 700% YoY. They have inked deals with Tencent and ByteDance to provide memory. If the USG approves, who knows they might supply to Apple too!

SoftBank Plans 10-GW AI Cloud Platform in U.S.

SoftBank Corp. CEO Junichi Miyakawa announced plans to offer AI compute services in the United States at a scale of 10 gigawatts, positioning the company as an AI cloud provider to meet surging domestic demand. The initiative would place SoftBank among the largest AI infrastructure operators in the U.S., though specific timelines, capital commitments, and data center partners were not disclosed in the announcement. (Bloomberg)

Vik: More competition to neoclouds like CoreWeave and Nebius but I still don’t buy the overcapacity argument even as Meta is now selling compute too. We haven’t scratched the surface of what AI can do, and lots of compute still needed.

Nokia Wins Sole-Supplier Deal for Orange Belgium Optical Network Overhaul

Nokia has been selected by Orange Belgium as the sole supplier for a multi-year transformation of the operator’s transport infrastructure. The project will unify fixed and mobile networks into a single converged optical transport network across Belgium, with the goal of building a more resilient and scalable platform to handle growing bandwidth demand from consumer and business services. (Nokia)

Vik: Orange Belgium is upgrading their networks to support the next generation of communications infra. Nokia is solely entrusted to install optical networks, with AI-powered hardware, to spruce up performance.

SK Hynix Commits $64B to Memory Chip Expansion

SK Hynix announced a $64 billion investment plan to expand memory chip manufacturing capacity, framed as part of a broader push to meet AI-driven demand. The South Korean company plans to build new production facilities, with spending phased over multiple years across domestic and international sites. (Reuters)

Vik: This is part of South Korea’s broader push to double memory production capacity in 5 years. Michael Burry (of The Big Short fame) is concerned with Government level investments going into memory calling it “the beginning of the end.”

Oriole Networks, AMD Deploy First Commercial Photonic AI Network

UK startup Oriole Networks has deployed what it calls the world’s first large-scale commercial photonic AI network, built in partnership with AMD as part of the UK Advanced Research & Invention Agency’s (ARIA) Scaling Inference Lab. The system moves AI infrastructure optimization from processors to the optical interconnect fabric, connecting AMD accelerators via photonic links rather than conventional electrical interconnects. (EE News Europe)

Vik: Oriole has a great approach to optical networking, and its interesting that they are partnering with AMD. With their PRISM Ultra platform, they can hook up 1,000 nodes with zero hops! Clos networks be damned, this is radical! Look up their website, seriously.

Infineon Closes ams OSRAM Sensor Portfolio Deal

Infineon Technologies has completed its acquisition of ams OSRAM’s non-optical analog and mixed-signal sensor portfolio, a deal first announced in February 2026. The transaction, which cleared all required regulatory approvals, adds sensor technologies, engineering talent, and development sites to Infineon’s Edge Systems division, expanding its capabilities in automotive, industrial, and medical sensing. (EE News Europe)

Kioxia, Sandisk Begin Sampling 10th-Gen BiCS10 3D NAND

Kioxia and Sandisk have begun sample shipments of their 10th-generation BiCS10 3D NAND flash memory, targeting AI data center workloads. Sandisk’s BiCS10 1Tb TLC device delivers up to 4.8 Gb/s interface speed and a 59% bit density improvement over BiCS8, along with improved power efficiency. (Sandisk, Reuters, Bloomberg.com, TechPowerUp)

Vik: Some nice quotes from the Reuters article:

“Kioxia is two to four years ahead of rivals with NAND performance and power consumption due to strengths such as ​its wafer bonding technology, said Kazuyoshi ​Saito, an analyst at IwaiCosmo ⁠Securities.”

"They haven't been able to respond to the current NAND ​boom at all. That is why demand is now concentrated on Kioxia alone,"

Anthropic in talks with Samsung to produce custom AI chips

Anthropic is in discussions with Samsung to manufacture a custom AI chip, according to reports from TechCrunch and The Korea Herald. Financial terms, chip specifications, and a production timeline have not been disclosed.

--- (TechCrunch, The Korea Herald)

Vik: So Anthropic considering using Samsung’s 2nm process for their chip? I thought the major issue was yield. Or is TSMC's capacity so constrained that Anthropic will hedge elsewhere? There was some talk last month that Samsung could make Google’s chips too.

Global AI Sales Justify Data Center Costs in Q1, But Margins Remain Thin

Global AI sales excluding China hit $25B in Q1, narrowly exceeding the estimated $21B in data center and chip depreciation costs, but margins remain structurally thin as the industry confronts the unit economics of inference. This financial pressure is accelerating enterprise demand for cheaper open-source models, custom silicon, and aggressive inference optimization, evidenced by OpenAI’s reported success in halving inference costs. The shift from simple chatbots to multi-step agents has increased token consumption while simultaneously triggering cost discipline among buyers who are now routing work to cheaper models rather than paying premium rates. This dynamic marks a critical inflection point where revenue growth is justifying the physical buildout but failing to capture sufficient margin to sustain current capital expenditure levels without significant efficiency gains. The competitive landscape will increasingly favor companies that can deliver efficiency gains or own the software routing layer, rather than those relying on model capability alone to command premium pricing. (theinformation.com)

Vik: The era of mogging frontier tokens for the f*ck of it is done. The future belongs to those who can deploy AI “intelligently” while routing the right requests to the right models.

Hyperscaler CapEx Management and Investor Sentiment

Goldman Sachs projects AI infrastructure companies will account for 57% of S&P 500 EPS growth in Q2, signaling a structural pivot in market profitability toward the physical layer of AI. Concurrently, secondary signals regarding declining rental rates for Blackwell and Hopper platforms suggest hyperscalers are beginning to price in capacity normalization or increased merchant supply. Meta’s CapEx defense, arguing internal overbuild can be monetized externally, highlights the fragility of the current investment thesis as unit economics shift from capability to profitability. The market’s punitive reaction to Meta’s spending underscores growing investor sensitivity to the return on deployed capital. This marks a transition from pure capability expansion to a rigorous scrutiny of operating margins and capacity utilization. (theinformation.com)

Meta Monetization via AI Agents and WhatsApp

Meta is developing a cloud infrastructure business to sell AI compute and model hosting services, directly challenging Amazon, Microsoft, and Google, while simultaneously pushing internal teams to adopt its own development tools over external coding assistants. This marks Meta’s first major push into merchant AI infrastructure, shifting from a capital-heavy consumer model to a profit-generating compute provider. The move capitalizes on Meta’s excess GPU capacity and custom silicon efficiency to create a new high-margin revenue stream alongside its core advertising business. Watch for hyperscaler response and whether Meta can replicate AWS’s scale in the specialized AI cloud segment within 12-18 months. (bloomberg.com)

Vik: “Profit generating provider” is a good cash cow for Meta after all that coin they have spent in building out compute. Some will even argue its high time all those investments translate into revenue. Elon first ran this playbook. Seems like Zuck is on the train. Real question is — will others like Microsoft, Google, AWS also profit off their compute spend?

By the Numbers

Closing moves, Jul 2:

• Up: RIVN +8.4%, AAPL +4.8%, PLTR +2.8%

• Down: AEHR -17.1%, SITM -14.5%, SNDK -14.1%, AMKR -12.9%, KLAC -11.5%, IREN -10.4%

• Highs/lows: RIVN at a 90-day high; AEHR, SITM at 30-day lows

Why:

• AEHR -17.1% — Aehr Test’s deletion from the Russell index forced mandatory index-fund selling, compounded by concerns over stretched AI-orders valuation.

• AAOI -15.2% — Sector-wide optics valuation selloff dragged AAOI down alongside Coherent (-9%) and Lumentum (-8%) as investors questioned stretched multiples.

• VICR -19.2% — CEO and Chairman Patrizio Vinciarelli sold $7.5M (20,000 shares) via Rule 10b5-1, spooking investors despite strong recent guidance.

• UCTT -17.8% — Ultra Clean issued soft forward guidance that disappointed investors and drove the selloff.

• VECO -18.5%, ACLS -19.0%, ACMR -15.5%, MXL -17.1% — no clear public catalyst yet

Things we cover:

• Micron’s record-breaking revenue and gross margins

• Apple’s unprecedented mid-cycle price hikes and CXMT request

• South Korea’s half-trillion-dollar investment in memory capacity

• The impact of memory prices on consumer devices and companies like GoPro

• Technical reasons for HBM’s high wafer consumption

• The future of AI model training and inference costs

This podcast is lightly edited for clarity.

The Memory Crisis Hits the Streets

Vik: So check this out. Micron booked more revenue in a single quarter than in any full year of its 50-year history. They booked basically 41.5 billion dollars and its gross margin hit 85%. That’s more than Nvidia’s gross margin. And for a brief moment there, Micron as a memory company was worth more than Meta. And to add up to all this, the South Korean government is investing over half a trillion dollars in SK Hynix and Samsung just to increase memory capacity so that they can own the memory business entirely. And in the midst of all this, Apple is now increasing prices and they’re going to China CXMT to get memory capacity, so while the memory boom is creating winners on one side, it’s also causing a lot of trouble. This whole memory crisis is coming to a head now.

Austin: Welcome to another Semi Doped podcast. I’m Austin Lyons from Chipstrat and with me is Vik Sekar from Vik’s Newsletter. Hey, Vik, what’s up, man?

Vik: Not that much. Just been dealing with the memory crisis. I don’t know. I’m not dealing with it. What do I have to do with it? I’m not buying any devices. My Mac M3 Pro is doing amazing. So I don’t have to buy a laptop or anything. But I am also being more careful with it. I’m not going to drop it anytime because if I have to replace that thing, it’s terrible. The prices have gone up. Apple has raised prices and all that. Yeah, this was basically the entire discussion in the bar yesterday. I met a bunch of friends, and then we were having a few beers and this was the whole discussion: memory prices. What has the world come to? Everybody’s talking LLMs. And mind you, I’m not in the Bay Area, okay? This is, of course, Bangalore is a tech city, so it’s not unheard of, I suppose, but it was very fascinating. Everybody’s like, “Oh man, I can’t buy a laptop. This is Apple has increased prices.” Another friend of mine was like, “I can’t even buy musical instruments anymore because you can’t buy these synths and stuff because all of them have memory in it.” So, even musical instruments have gone up. This is serious memory crisis. It’s hit the streets.

Vik: Totally. Totally. Man, yeah, if it’s a conversation at the bar, then it’s truly impacting the consumer.

Austin: Exactly. When people start talking about the fact that memory prices from AI are affecting their day-to-day purchases, we are in trouble now because I don’t think the general population understood how much memory AI has been sucking up. We’ve been talking a lot about it on the podcast on and off, right? But that’s when it hits the streets like this, it gives you a visible reaction to people asking me, “Hey, you do the podcast, what’s happening with memory?” I’m like, “Yeah, yeah, let me tell you about HBM, let me tell you about DRAM.” And they’re all like, “Oh, wow, is it that much, huh?” And I’m like, “Wow, back to you.” To me, I’ve been thinking, “Oh, so much DRAM,” we’ve always been talking, it’s second nature, “Memory is short.” Yeah, yeah, we know this. Memory companies are making money. Sure, we get it, it’s in the news all the time for what we do. But for everybody else who is not on this train entirely like we are, it’s news actually. It’s like, “AI is killing my consumer devices.”

Vik: Yes, yeah, yes. So, the most simple definition of inflation is just sustained increased prices for a period of time, a year or two. And traditionally, when the consumer thinks about inflation, core inflation, it’s usually the price of food, but it’s also the price of oil, essentially, gas and whatnot. And so people, normal consumers have to tend to pay attention to various random industries based on the things that they buy that cause inflation for them. For example, in the United States, people, if you like to eat beef, you might know that there’s a shortage in the US beef herd or cattle herd. And it’s a very random thing, but you’re like, “Yeah, because when I go to buy steaks now, it costs way more than it used to.” And what’s interesting is because of AI, now the normal consumers are going to have to start paying attention, or at least right now they are, to the memory market. Normal people are going to have to learn, “What is memory? Why is it so expensive?” Because to your point, it’s just like, “Yo, I’m just a person that produces music with a synth and suddenly it’s expensive.” You just start pulling on that thread to find out why. And what’s super interesting is as much as I love AI, which I totally do, it’s interesting to say it as plainly as, “AI is causing inflation.” AI is going to make it so that your grandparents who are in the United States that are on a fixed income and using social security, for example, their social security dollars don’t go as far because the price of certain things have gone up because AI. So it’s crazy to think how it’s truly impacting just everyday people around the world too. A lot of the AI is produced in the United States, but it’s impacting people around the whole world, which is also crazy to think.

Austin: I think the whole sentiment in the market now because of these increased memory prices is that when it hits the consumer and when the consumer starts to feel the pinch, the sentiment seems to be that people are going to stop buying devices. And as much as we talk about AI and the deployment in data centers and how many GPUs people are buying and how much these data centers are spending, the consumer market is actually very significant. It has always been a very big driver of semiconductor business. So it’s not to be discounted or trifled with, right? So when the prices become this high and when everybody sees their electronics prices jumping up, the first tendency is what I said in the beginning, right? Look, I better not drop my laptop because I’m not going to buy anything in these prices. And that’s the thing that everybody is going to think about as well, at least in those lines, most people, I would think. So what that means is that if people stop buying consumer devices, now you’re going to have demand drop and then the supply will equalize better. And once people stop buying consumer devices, the market’s going to pull back or something. So, people anticipate this because they’ve seen this happen in the past. And so the market softened last week around this news. That’s my interpretation anyway.

Consumer vs. AI Demand

Vik: Hmm. Yeah, so definitely consumers are going to delay purchases or potentially, depending on what kind of consumer you are, you might say, “Well, I have to buy a new laptop.” My it’s at work, it’s time to refresh. I have to. They’re probably going to say, “Let’s just delay, delay a quarter or whatever.” Or they might say, “Well, we’ll get you that new MacBook still because it’s been three years. We promised you that, but we’re not going to go all out on memory, of course.” But yes, it’s definitely going to impact consumers. And then the interesting part is it’s not only going to impact the low-end consumers, but even high-end consumers. I think that’s where it’s really interesting is when you have people, Apple always has the best customers who are willing to pay the most for the premium experience of having an iPhone. But it’s pretty crazy when even the premium customers who are willing to spend three, four, five thousand dollars on a laptop, even they are saying, “Well, I guess I better wait or I’ll delay it. Let’s see.” The question I have is, so if we have a drop in demand from all the consumer devices of which there are many and it does consume a lot of memory, is that actually going to help even out the memory price increases? Because the question is, is that going to improve supply? Or is the supply that extra supply, if you will, that’s not coming from consumers, is just going to be sucked up by the AI and the hyperscalers, and all the accelerators who can’t get enough, right?

Austin: Yeah, the answer to that comes from looking at the supply commitments already that are awaiting these memory companies. It looks like people are booked out through 2027 at least. That’s what Micron said in their quarter at least that this is not going to let up in 2027 because they have these long-term agreements, and then they also have it as somehow a prepayment clause, which is basically that you pay for it whether you take it or not. So they’re going to continue to make money and AI book orders are in, they’re all booked up, they’ve been prepaid for. So even if this softens, I don’t see why it will let up in the near term. In the long term, I think supply capacity is being built up and when that comes online, maybe then we will see something letting up here. But up until that point in time, I don’t see anything changing in the near term. But let’s see how right I am in a year from now. We should revisit this episode.

Vik: Right. But, but, so you’re saying if it’s not going to let up in the near term because all the AI accelerator companies are already prepaying for it and they’re going to suck up any extra supply.

Austin: That’s possibly out there, whether the supply is from a new fab or the supply is from the lack of consumers buying it. That’s going to be really hard on consumer device companies because it means that they’re going to continue to have this pain. And so we, I know you and I were talking earlier about some companies that are having pain and trying to work around it. So GoPro, for example, tell us about GoPro.

Vik: GoPro was such an awesome company. It was such a lifestyle statement to have that GoPro stuff. When it came out, people were attaching it on their guitar headstocks and then you could see them playing a live show with the guitar moving around. It was so amazing. Obviously, there’s all these adventure sports that came out. I even considered it for some time mounting it on my surfboard when I lived in San Diego so that I could see how it looks as I surf. You see these surf videos of people going down the line and all that. So, it was such an amazing concept, but today what has happened is they don’t have money for memory. They know, how much will you pay for a GoPro anyway? It’s already something that you don’t really need. It’s a fun thing to have to do all these crazy things. But nobody really needs a GoPro, right? And now the company has become a penny stock, basically. There’s a good chance that they’ll actually go under. It’s just a sad thing because it’s a device that had a lifestyle statement at some point. And now the memory prices are driving companies like this out of business. It’s sad actually.

Austin: Yeah, it is sad. But yeah, you make a good point. Especially any discretionary spending, where people don’t definitely need it, those companies are going to definitely be hurting. What about Apple? How are they trying to deal with not passing on such crazy price hikes to customers?

Vik: So, Apple is actually the biggest giant of our times, right? Before Nvidia, they were first trillion dollar company, if I remember, right? This is a company that has been the gorilla in the room, so to speak, for decades because they had pricing power. They could always come in and say, “Hey, we’re going to sell a billion units, we need this price or we need this performance.” Now, “we need this capacity.” And there were companies before whose sole existence relied on Apple orders. I worked for some of them, actually. So, their whole existence is whether they make it into the next Apple design cycle or not. And Apple was in a sense brutal with their own business practices. It’s well known in the industry that they’re really tough to deal with and nobody would, if they can avoid it, if they didn’t have to make money, they’d rather avoid Apple. But Apple are a tough customer to work with. And this is the first time that Apple has increased product prices mid-cycle. This is not a next refresh, a next generation that comes in or whatever. They refresh prices in the middle of nowhere, saying, “Now the same product is now more expensive.” They’ve never really done that as far as I know.

Austin: Yeah, yeah, it’s pretty wild. Now, tell me, shouldn’t I feel like there’s a paradox here, which Apple is the big company, they make the premium products. They could probably afford to pass on the pricing to their customers the most because they’ve got the most premium customers, right? So if, let’s say you were making an Android phone and I was making an iPhone and memory’s going to get more expensive, it should be easier for me to pass it on, “Hey, things are going to this phone’s going to be an extra 100 bucks” than for you to pass it on as an Android phone maker because you’re like, “100 bucks, I can’t I can’t charge my customers 100 bucks.” But then paradoxically, isn’t it Apple who’s probably getting the best memory prices because they have such scale? So as me as Apple, wouldn’t my cost for memory actually be less than yours, Android maker, because I have such scale? Or what do you think there? I mean, I know Apple captures a lot of the margin, but I do know that there’s Samsung and other Android phone makers that still sell tons of units.

Vik: Yeah, so I think the distinction that we have to make here is premium versus non-premium device. The whole argument that I’ve been reading in recent times is that premium tier devices are actually okay because they have a high enough selling price that they can absorb the price increase in memory and they’ll be fine because of that. Although their gross margin will reduce. But in general, yes, it’s okay. Now, with the Apple price increases, I don’t know if Apple wants to hold on to their gross margins and therefore they increase the price of the device in lock step with what memory prices are increasing so that they can retain gross margins and look good in the next earning cycle. So it’s a gamble, right? Because people may not buy as many if you increase it. So now you lose on revenue anyway. So that’s I’m not sure which way they’re going. But there was news, we’ve mentioned this on this podcast too because we’ve spoken about memory so much. Apple was supposedly buying up DRAM like crazy. They were buying it up at premium prices. If at all, they could only lock out their competitors from DRAM, they were going to do it. And if there is a company that could do it, it’s Apple with their big cash reserves, right?

Austin: Right, right, right. Yeah, so interesting to think about, yes, there’s price elasticity here. So the more they increase the price, the fewer people that want to buy it. So of course, if they could absorb it all, then it wouldn’t affect the demand, but then of course, it’s going to affect their margins. And so, always as public companies, you have to play that game of how much of a margin hit do we want to take versus how much of fewer units do we want to sell. But then on top of it, it’s kind of like, well, are investors going to compare them to the previous quarter and say, “Hey, you sold less units than expected” or the previous year, year over year comparisons are probably best in a seasonal thing like smartphone sales. Or are they just going to compare them to everyone else and say, “Yeah, your margins went down, but they didn’t go down as far as everyone else.” So actually, we will still reward you for that.

Vik: Yeah, the expectation these days has become very AI-centric, right? Everybody wants a massive increase in earnings and massive boosts because the idea is, “Why would I invest in Apple when I can invest in Micron?” Because Micron is making money. “How about I invest in Micron?” You guys are losing gross margin, so why should I invest in Apple? So investors are not really tied to Apple as a company or anything. They’re just there for the money, right? So, we’re here to make money. So, there’s a logical thing is to go to the company that makes money. But so, I don’t know, ultimately, it seems like even the premium tier was not spared. And recently, over the weekend, there was also the discussion that now Apple is asking CXMT, which is a Chinese DRAM manufacturer, for DRAM supply. And CXMT is actually on the entity list, which is a list of no-no companies for US companies to do business with in China because those companies are also suppliers to the Chinese military. So, you don’t want to as a US company give money to a company that then supplies the Chinese military because that’s going to bite the US back in the future, right? They’ll have a stronger military or whatever. Anyway, so, they put CXMT on the entity list and now Apple is asking the US government, “Can we please get CXMT to give us DRAM because we need the supply?” This is another first, actually, because I don’t think there’s any company who’s gone to the US government and said, “Hey, you blacklisted this company, but now we actually need to buy from them.” Imagine Apple actually doing this. Actually, that’s a big deal.

Austin: Yes, the things not on my 2026 bingo card. Yes, Apple asking if CXMT could come off the entity list or any Chinese company could come off the entity list. It just goes to show the lengths that they’re willing to go to to prevent the inflation that we talked about, and obviously try to get more supply. Of course, the interesting thing is, it’s a temporary problem in that, as you said, we know more supply is coming online. Micron, Samsung, SK Hynix, these guys are building fabs. They’re going to come on in 2028, 2029 and so on. If you ask for CXMT to come off the entity list, that’s bringing supply on indefinitely, not just to solve this temporary problem, but indefinitely. Which of course, when you start to game theory this out, you wonder if the big three had it on their game theory bingo card that CXMT supply could get brought online. But it I’m sure it does feel like a slippery slope in that it’s like, yes, you would like more supply right now, but do you want CXMT DRAM in a year from now or two years from now? I mean, maybe at that point they just stop buying from them. They just say, “Hey, thank you for your service and for helping us in 2026.” Even though you’re not on the entity list, we don’t want to buy from you anymore. But it just yeah, it’s just so interesting to think through all the implications.

Vik: Yeah, so now the long and short of it is that the premium companies, premium tier companies like the Samsungs. Samsung is a bit different, right? Because they have DRAM supply by themselves. And they make phones. Think about that. Apple doesn’t have Apple doesn’t make DRAM. Samsung does. And they also make phones, which is a unique advantage if you come to think of it. But, apart from these premium tier companies, what about all the lower mid and low tier? They cannot absorb the costs. So, the one thing that will happen in this scenario is that those phones will stay at the same price, but it’ll get de-specced. So, a phone that was selling for 4 GB of RAM or whatever, some low tier, mid tier phone, will only go with 2 GB of RAM now. And so they get de-specced and that is one way to handle the lower tier. But nobody likes that. So now you’re like, “Okay, fine, I’ve spent the same amount of money, but I get a crappier phone,” right?

Austin: Right, totally, totally. Yeah, that’s a—what do they call that when in consumer foods where you pay the same, but your bag of chips is half as big?

Vik: Yeah, yeah, yeah. Isn’t it a form of inflation?

Austin: Yeah, yeah, yeah. There’s a funny term for it, but this is the same thing. Shrinkflation. That’s what it’s called. Shrinkflation.

Vik: Shrinkflation. Yeah, yeah.

Memory Market Dynamics

Austin: Shrinkflation hits low-end consumer phones. So you mentioned something interesting. It was kind of an aside, but you said, “Oh, Samsung, they make phones, they make memory.” They also have a foundry and can make logic chips, right?

Vik: Yeah.

Austin: So in theory, Samsung could presumably make quite a lot of phone, vertically integrated to some extent. And it would be interesting to think that through. Probably if you’re kind of doing the Taiwan, China war gaming of what happens if TSMC goes down, you could ask which smartphone makers are best poised to continue to be able to supply new smartphones and maybe it’s Samsung.

Vik: I think Samsung plays this game too because for profitability, they also have to make an internal decision as to where to direct their DRAM supply. Would you sell it for a nice hefty profit to AI companies or would you sell it into the consumer market and lose money? They will even refuse their own handset division if it comes to it because it’s all about getting the most out of the DRAM supply you have. So it’s very funny actually. And talking about Samsung and SK Hynix, maybe this is a good time to mention this. The government itself, right, the Korean government is investing north of a half a trillion dollars in DRAM capacity expansion in just these two companies. Because they realize that these two companies now hold the key to the AI semiconductor super cycle. Because yes, there’s TSMC and there’s Taiwan who has the key to manufacturing. Yes, there’s Nvidia in the US who have the dominance in training chips and also all the other hyperscalers, they’re all US-based. All this is good. But what do all of them need in common? What is the one breaking point is memory? And who has all the memory? Two-thirds it’s Korea, right? Micron is US, but two-thirds is a Korean-based company like SK Hynix and Samsung. So the Korean government is all in. They basically pushed all their chips in and said, “Here you go, we’re going to own this thing” and they’re spending half a trillion dollars on it. And then interestingly in other Korean memory company news, SK Hynix actually surpassed Samsung for the first time to become the most valuable company in Korea, ending a reign that Samsung has held for 25 years. So that’s the power of memory. So the other thing I wanted to tell you was about Kioxia because Kioxia is not really a DRAM memory company, it’s more of a NAND flash maker and they spun out of Toshiba. And they recently passed Toyota as the most valuable company in the country. So you can see what AI is doing. Toyota, it’s a big deal, it’s in every country, but now a NAND flash company that Toshiba actually spun off. What a bad idea. I mean, talk about a bad bet. So funny.

Austin: Yeah, pretty wild, pretty wild. Things are changing, that’s for sure.

Micron’s Profitability and Market Cycles

Vik: Yeah. And finally, we have to talk about Micron. We mentioned it in the cold open, obviously. But Micron is, let’s just say, for all the people who are suffering in the AI memory crunch, the GoPros going out of business. I also read about a small company that was making some routers or something. They were selling their routers for, I don’t know, $500, let’s say. And now they have to sell it for $2,000 to even stay profitable. And they’re like, the makers are like, “No, nobody’s going to buy this at $2,000. This is not a product like that.” Sometimes electronics has a certain intrinsic value. Nobody buys a laptop for $20,000. You buy a laptop for, I don’t know, $2,000, right? There’s a logical limit to certain class of devices that you can’t just break. So this is one of them. So for all the suffering that is going on in the memory industry, Micron is among other memory makers making a killing, right? Yes. They are at 84% or something, 85% profit margin, which means that you basically are selling it for four times more than you made it for. So if you make this widget for $10, you can sell it for $50, I think. Yeah. So anyway, that’s a lot of money to be made here.

Austin: Yes, totally. Yes. Micron’s gross margins, I think in the last, which I wrote about this on Chipstrat, you can check it out. In the last four quarters it went from 45% to 56% to 75% to now almost 85%. And they’re actually projecting another 85%, 86% next quarter.

Vik: That’s insane. Earlier you mentioned, maybe this isn’t price gouging, but I’m going to say it is. It is absolutely price gouging. And I think that a lot of people now have this inherent faith for memory companies because everybody’s thinking because of these three. And so basically it’s just like, I think when the down cycle comes, nobody’s going to like it. But everybody’s maybe secretly saying, “It had to happen. I’m glad it did.” Puts these people back in where they belong.

Austin: Yes, yes. I mean, to Micron’s credit, when I think price gouging, I think of someone trying to take advantage of people. And I think to their credit and other memory makers, it’s just that literally their customers are like, “No, I’ll pay more than that, dude, because I need it.” And that dude’s like, “No, I’ll pay more than them because I need it,” and it’s just, I mean, there’s probably literally almost to an effect a bidding war going on. So it is obviously very opportunistic, but if you have a customer saying, “Hey, you’ll sell that to Vik for $3, I’ll give you $5,” you can’t—it’s how do you say no to that?

Vik: But maybe that’s one way of looking at it that is, yeah, maybe they’re offering more to get the supply locked in and Micron is just like, “Sure, why not? You pay more, you get it.” And the next person comes by and like, “Sure, you can.” Or maybe they’re just going, “No, that’s not the price. This is our price. You want to take it or you want to leave it?” You can leave it because if you don’t take it, somebody else will. So there’s this whole FOMO thing. We’ll put up this another chart I have from the Wall Street Journal that shows basically Micron’s adjusted operating earnings every fiscal quarter from 2023 to 2024. If people are not aware of the history of this company, they had negative earnings, okay? They were losing money. It was pretty miserable. But then you can see what happened to this chart as you go from basically the beginning of 2024. They start making a turnaround and now they’re making, I don’t know, 41.5 billion in a single quarter. It’s insane. If you look at the period between 2024 and 2025, 26, the last two years, they haven’t made 41 billion dollars in a quarter put together in all those years. Now they do it in a single quarter. It’s interesting, I think.

Austin: It’s wild. Yes, it is so wild. I mean, we see this across AI, these companies where they’re spending more or making more in a quarter than they did in an entire year. So for the CAPEX guys, for the hyperscalers, it’s like, all of a sudden now they’re spending in a quarter in CAPEX what they used to spend in an entire year. And of course, on the other end, the people who the hyperscalers are buying from, they’re making more in a quarter than they used to make in an entire year. And it’s just pretty wild to have all these companies suddenly making more per quarter than they did in an entire year. And for some of them in several years added up. And for Micron, I mean, 2022 wasn’t that long ago. 2023 wasn’t that long ago. So pretty crazy how their fortunes have turned so quickly.

Vik: Yeah. I like the way you put it in your Substack, basically, you didn’t show the most recent earnings data in your Substack. It’s like, “You see, memory is a cyclical industry. You can see it goes up, it’s like a sine wave.” It’s almost like a sine wave if you draw a line joining all the peaks of these bar charts you have in your Substack article. By the way, whoever listening should go read the whole article. But then later you’re like, “Ah, but I sneakily did not put the most recent one. Here it is with that recent one in and then there’s this huge bar that goes up.” That’s totally, yes. You tell me, is this a cycle? Does it look like a cycle to you? Because that’s a peak. That’s a peak.

Vik: Right. Well, and what’s interesting and I linked to this nice chart from Doug O’Laughlin where he tries to show where different industries or submarkets are in terms of a cycle and I think of it as a merry-go-round. There’s these four quadrants of is there inventory increasing or decreasing and are your sales increasing or decreasing? And if you’re in a quadrant where your inventory is decreasing and your sales are increasing, that’s obviously very good. Prices are going to go up. But eventually it tips over and you start to build back up inventory. So even if prices are high and you’re still selling a lot, you’re building up inventory. And then eventually there’s going to be a point when you have so much inventory that you actually tip over and you don’t sell as much and then, you start to get in this place where it’s like, “Oh, now we’ve got more inventory than we need and prices are going down.” And it’s kind of like a merry-go-round. You just go all the way around. And so a lot of people are trying to make the argument that, “Oh, we have all these new demand drivers,” which is totally true, but they’re kind of arguing that that merry-go-round doesn’t exist. And I think that the merry-go-round still exists, but now I think there’s other drivers that are going to keep AI in a particular quadrant for a lot longer because, and we can talk about this, but actually, let’s definitely get into this really quick. When you buy HBM, you’re buying DRAM wafers essentially for stacked HBM. You can unpack this. And then also there’s people needing to buy DRAM. So, for example, I need lots of CPUs for agentic AI and those need DRAM. And so there’s these different drivers that want to take the same wafers. And it’s kind of like no matter where you slice it, it feels like the demand is going to continue be high and we know that supply is not going to increase until 2027, 2028, 2029. I’ve got it in the Substack as well. And so it it it there’s I’m a believer that the merry-go-round is still there, but there’s just so many drivers that things aren’t going to tip over. The supply is not going to catch up. The inventory is not going to decrease for a lot longer.

Technical Reasons for Memory Shortage

Austin: Yeah, so basically to the point of unpacking what you said, it’s that we are really short of memory because of AI. Okay, that’s one broad way of putting it, but essentially because we need high bandwidth memory, which requires stacking these DRAM chips one on top of each other. To make those DRAM chips that you have to actually stack, actually takes three times as many wafers to make the equivalent number of bits. You know, so I don’t know, to simplify that statement, let’s say you have to make a 1GB DRAM chip. Now, if I were just making a one single chip that I didn’t have to stack up into HBM, I would maybe make it in one wafer, and you can get one wafer’s worth of DRAM chips or whatever, 1GB chips. And let’s say you, I don’t know, maybe you get a thousand chips out of a wafer, each of which is 1GB, okay? Now, you want to make the same 1GB chip, but that 1GB DRAM is going to go into HBM. You’re going to need three times as many wafers than to make that same bit capacity if it were going to be used for HBM. Okay, there’s a technical reason for this. We won’t get into too much of it, but it’s basically that you have to drill holes through this thing and have through silicon vias, which is how you stack them. You actually stack them by connecting them through each DRAM die. Whenever you drill through it, your bit density drops. You can’t have as many bits now because you can’t put memory cells around these drilled holes for safety reasons or it won’t work. So, because of that, basically for every 1GB of HBM and for every 1GB of DRAM, the HBM takes three times as many wafers than the regular DRAM does. So, it’s a big suck up, it’s sucking up all the DRAM supply big time. That’s one big reason. The other reason is that we right now even need DRAM chips just as is. Forget about HBM. We need DRAM because apparently agentic AI needs CPUs and now CPUs need a lot of DRAM. Great. So now you have to get a whole lot of DRAM chips. Then it turns out that we want to run long agentic workloads, which means the AI has to remember stuff literally forever and then keep the context in its brain of what it is doing in a large code base or a very long context conversation. You want to keep asking it questions. You don’t want it to have amnesia ever and you want instant answers, right? You don’t want to wait or anything. So now all of that context that it holds in its brain is also stored on DRAM. It may not be stored in HBM. It could be stored in HBM, but it’s very expensive. So mostly it you can store it in DRAM, which is the next fastest memory. So all of that memory thing that AI has to remember is also stored in DRAM. And now people are looking at pooling DRAM. So what you do is you take all these DRAM sticks and put them together in an appliance and say, “Okay, look, this is my KV cache server,” so it could be CXL pooled DRAM or it could be CXL pooled NAND flash. So everything is going into AI. It doesn’t matter whether you’re looking at agentic CPUs, HBM or just KV cache storage. It’s all over. It’s all DRAM. It’s all memory. So this is this is why we are in this pickle. So in a long story short.

Vik: Totally. And that’s just talking about essentially the data center demand for DRAM or the enterprise demand if you will. There’s obviously still consumer demand for DRAM. The interesting thing is that the data center buyers basically are inelastic in their demand. So if the price goes up, they still want as much. Which and how much do they want? Every wafer you can make. Right? And it’s almost like the curve, the elasticity curve almost goes straight up. It’s like the price goes up, I don’t care, I want the same. But on the other side, the consumers, it’s very elastic. And so there’s this interesting sort of bifurcation into data center demand and consumer demand. And data center demand kind of bifurcates into HBM and DRAM. And DRAM kind of bifurcates into CPUs need it and, maybe some storage appliances to supply your KV cache needs it, right? And so you just have all these demand drivers. Supply is coming online. Meanwhile, of course, the consumers are hung out to dry and it’s impacting how far grandma’s dollar can go, and she can’t get you the new GoPro for Christmas because they’re going out of business, right? And so consumers are getting beat up. But it’s very difficult to see demand ever waning. So I don’t think people should think that data center demand wanes. I think consumer demand will wane, right? Because it’s phones, like we talked about, my laptop’s expensive, I’m going to put it off. But the data center, they’re just going to keep buying. And so that’s why, we kind of think it’s this AI super cycle in memory if you will, because like you said, there’s a little sine wave and then all of a sudden, boom, it’s like it is accelerating, it’s going a lot higher. So it’s going to be there for a while.

The Future of AI Demand

Austin: Yeah, yeah. I think we should wrap this up. I have one question I think that will, maybe we can think on it because this is a bit of a macro question and I don’t know if anybody really knows the answer, but I just like to know what you think of it. So, what do you see happening from now to 2030? Okay, but that’s a broad question. So let me pin that down a little bit more. So, given that now leading frontier labs are not really being allowed to release leading edge models like Fable/Mythos was cut off by the US government, then OpenAI’s ChatGPT 5.6, what they code name Soul, was also not released for general public use. There is a lot of fear of distillation of these models from Chinese competition. So that is one front of it. So maybe the frontier won’t advance as much due to other reasons. It’s not actually scaling or technical reasons. It’s more political reasons or competitive reasons because now do you think that the next generation of leading edge models are going to be trained with the same gusto knowing that they will not be released, right? So that is my first question. What do you think will happen between now because the training seems to be having an inflection point of sorts. So what are your thoughts on that? Then I’ll ask you something else because I have more.

Vik: Okay, so if you assume that the government will always just prevent the leading edge thing from going out, then it does raise an interesting question, which is, well, how much money do we want to spend on advancing the leading edge if we can never let it go out? If that was all true, of course, there is an argument that the leading edge labs would continue to train even better models and just use them internally and just get continuously better and more efficient at doing what they do to improve their business. However, I’m not so sure that the government is going to prevent this forever. I think interestingly, this could be shooting oneself in the foot and that lately with the agentic AI stuff, it felt like the American labs had really run ahead of the Chinese labs. Now that it’s not just training the model, but it’s also training the harness. And so it’s this bigger system that you’re training and delivering. And Fable’s so awesome, that the government’s like, “Whoa, wait a minute, let’s not do that.” And but interestingly, that actually gives pause and gives a chance for Chinese labs to start figuring out how to train harnesses and sort of catch up again, if you will. So I won’t be surprised if there’s the pendulum swinging back and forth where it’s like, “Oh, never mind. Fable’s back on. We figured out how to feel comfortable about its safety,” because we feel like the rest of the world is catching up and we need to run ahead, which by the way, this is going to impact Anthropic and OpenAI’s IPOs. It’s going to impact their—there’s going to be revenue left on the table because they weren’t allowed to share the best technology with the consumers who would pay the most for it, right? So I think there’ll be all sorts of incentives for the government and for these companies to get these models back out there. But I definitely see your point, which is for as long as the government prevents them from doing it, it actually puts—it kind of disincentivizes maybe training the next model at the even almost scaling laws of, “Yeah, now we need a million accelerators. Now we need 5 million,” right? And so then, to your point, maybe that could actually flatten the growth of the demand curve a little bit for memory or for accelerators or whatnot.

Vik: Yeah, yeah. That’s the kind of point I wanted to make that if there is this giant rush towards AGI is being limited or governed by other competitive or political interferences, then it’s questionable how much more we will need. But that’s only the training side, right? The demand for inference is still enormous. We have just started with agents. We have not even scratched the surface of what is possible. I truly believe that, right? And it seems like it’s very, very helpful. The people I’ve spoken to have said, “I can’t believe this thing.” All companies now need to have an agentic AI approach to doing their job. Or their product, if it’s a software company, needs to have AI in it somewhere. Otherwise, as a startup, even a software startup, let’s say, they don’t have any valuation. You can’t go anywhere if there isn’t a component of AI. So it’s a strong driver everywhere. And it’s it is helping productivity gains. If you see the OpenAI jalapeno chip release with Broadcom, they say in the press release that it is accelerated by using ChatGPT. I’m sure it wrote some code to verify the chip or whatever. So, and there are a whole lot of EDA startups doing AI enabled this and that. And all of this are productivity gains. So, we have not even scratched the surface of what’s possible. And I truly believe that we’re going to only need more tokens from here. It’s not like we’re going to need less tokens. Totally. This leads me to my second question. I think now companies are becoming a little sensitive to what the token costs are. It’s not so much as to number of tokens, it’s about the cost of tokens, which is where things like the GLM 5.2 model comes in because it is a very capable model, but you can serve it at a fraction of the cost of what OpenAI can, right? Or you can talk about the DeepSeek V4 Pro and maybe that’s enough intelligence. You don’t have to have always the cutting edge intelligence for everything. So, basically my question is, do you think that people will stop using these frontier models that are so expensive in an attempt to cost optimize inference going forward?

Austin: My hot take is both are going to grow. I think companies are going to continue to just be thoughtful and say, “Guys, we don’t need Opus to do these little—I’m summarizing these 50 articles and giving you a daily thing on it. You don’t need Opus to do that. Just use Sonic or whatever.” So that kind of stuff is going to grow. And which by the way, someone else is going to say, “Don’t use Sonic, let’s buy a token generator and let’s just use an open source thing and just generate these tokens for free if you will, for the price of the CAPEX of buying the on premises token generator.” And let’s just use open source. So I think that’s definitely going to grow big time because lots of CFOs are going to go, “Wait a minute, we just went crazy, we have to.” But there’s people are going to always want the frontier and the frontier always is going to unlock more. And every time I touch the new frontier, I can do way more productive software building and automating and stuff. It was amazing with Opus. It was amazing with Fable when it lasted. We had a great 24-hour session where so much happened and I want Fable back. And so I definitely believe that frontier will always unlock new things. It’s going to be frontier video models. It’s going to just be continued frontier things. So I think both will grow. And of course, I do think that companies are going to start to fine tune their—and I think they’re doing this already—fine tune their own models for their own proprietary use cases because we can’t expect that Fable’s going to be amazing, but it’s not going to have the best biology data for your company as you’re doing RNA sequencing or whatever stuff. I don’t know anything about that, but right, you can tell that that’s not just going to be baked into Fable. You’re going to have data, you’re going to fine tune. That fine tune stuff may continue to be done on open source models, American or not, which we don’t seem to have a ton of American ones right now. But I so I see both growing. That’s my reaction. What about you?

Vik: Yeah, I think there’s one more thing that will become a standard going forward is basically model routing because you essentially have to put the right requests to the right model. That is going to be a very big cost optimizer for companies to make sure that the wrong request isn’t burning frontier model tokens. So that’s a piece of engineering somehow that’s going to become very important within a company.

Austin: Yes, yes. I’ve been wanting to do this on my own. I’ve got lots of agents that are doing different tasks and I want to get better at saying, “Oh, these tasks should only use these models,” which right now how I solve it is I just use OpenRouter and I’m like, “Yeah, use Gemini Flash light or whatever for these tasks.” But it’s still hard to, there’s not a good user interface for visualizing all this stuff, but I’d like a better way to know, here’s all the different agents, here’s what they’re doing, here’s the models they’re using and make sure that it is cost optimized, which, you know, I’m sure that there are startups probably out there that are solving this problem already and I just haven’t learned about it. So if you’re listening and you know how to help me here, feel free to send me an email or comment down below.

Vik: Let us know. We really only do hardware stuff. We don’t know much about this stuff otherwise. We play around with models, right? What do we know about? But then the other thing is somebody mentioned that, “No, no, no, I don’t think your companies can deploy on premise models at scale or whatever,” to which I actually replied that, no, actually currently if you see how big companies do simulation workloads because they do run a lot of workloads on premises and they don’t want all your simulation data to leave the company. So basically these load share facilities or LSF farms are basically data centers run by companies and they are only within the company network. You can’t reach them otherwise, right? So that’s how I ran simulations, for example, all the big simulations were always dispatched to a server somewhere in the world and I got my results back. That’s how it always works because all the Linux terminals I was given in many companies in the past have always been a virtual machine. It’s just a 2GB RAM virtual machine. It’s not nearly enough to do the engineering work I had to do. So it was always sent out on LSF. The same way companies can develop AI farms, they can put in hardware locally and they can deploy local models, even pretty big models, that does exactly what they have to do. So they don’t actually have to do it. Companies, I just read this piece of news too, Meta are now worrying that using Claude on the cloud or whatever is somehow leaking data to the model and then it will be used to distill future models. So they’re issuing warnings to their people saying, “Be careful, don’t give away all the information or whatever.” So yeah, there are all kinds of concerns coming in as AI becomes more mature in its use in the industry. So it’s interesting to see how all this will have an impact. So it’s not only about, “Look, look, we need more AI, more AI,” it’s about noticing that frontier models seem to face some kind of headwind at the moment. As you say, it may not be forever. And companies are getting smarter about what they actually use because they need to show revenue. You can use all the tokens you want, but at the end of the day, your CFO is going to come knocking, “Okay, you blew, I don’t know, $100 million worth of tokens this year. What revenue, but my company revenue hasn’t proportionally skyrocketed. I want to see multiples of that $100 million spent on tokens because we even laid off people for these tokens, remember?” So I need to see that revenue. So this is going to become a big driving factor, I think.

Austin: Yeah, totally. It’s always going to be at what cost. So obviously at what monetary cost, but also at what security cost, at what latency cost and so on. So all right folks, we hope you like our AI’s eating memory deep dive. Thanks for listening. Thanks for our YouTube commenters. We love you. There’s a core group of you. Thank you for that. Thank you for our podcast listeners. Thank you to the person out there that was like, “Yo, where’s the podcast? It’s late this week.” We don’t really stick to a schedule per se, but we try to get it out weekly, but we had some travel, but I love that some of you love the podcast so much that you’re telling us, “Guys, get it out there. I’m ready for it.” So, that’s awesome. Keep listening and we’ll keep bringing this to you. Feel free also to check out our Substacks, ViksNewsletter.com, Chipstrat.com, and share this with a friend. Thank you.

Vik: And check out semidoped.com where you get all these daily news picks but with our little words in it. Think of it as a little readable version of this podcast brought to you every day. So, don’t forget to check out.

Austin: Yes, yes. If you like our podcasts and we can’t bring our podcasts to you fast enough, check out semidoped.com to get Vik and Austin every single day.

Vik: Yes, and also leave us a five-star review on Apple. I’ve been told it’s very important. Yeah.

Austin: All right, that’s a wrap.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player

Meta Plans Cloud Business, Selling AI Compute and Models

Meta Platforms Inc. is developing plans for a cloud infrastructure business to sell access to AI computing power and models. This initiative aims to generate revenue from excess computing power, potentially offering access to AI models hosted on Meta’s infrastructure, similar to AWS Bedrock, or selling raw computing capacity like CoreWeave Inc. The plans are part of Meta Compute, an internal initiative led by Santosh Janardhan, Daniel Gross, and Dina Powell McCormick. (bloomberg.com)

Vik: This piece of news has led people into thinking that compute has peaked. In reality, it is a pathway for Meta to be more profitable going forward, in spite of all their capital investments. Meta has historically been a customer of CoreWeave, and with them selling compute, the market views this as eating neocloud lunch, causing their stocks to drop.

Micron Becomes Most-Traded US Stock with $70B in Daily Turnover

Micron recently became the most-traded stock in the US, reaching $70 billion in daily turnover. The trading volume reflects investor sentiment that memory chips are a critical and constrained component for AI compute infrastructure. This has led to a re-evaluation of memory from a cyclical commodity to a strategic asset with significant pricing power, particularly for domestically manufactured chips. (DigiTimes)

Vik: Yeah, we know that memory is important. More people buying/selling memory stock than ever. People are so skittish on memory right now. Everybody is frantically looking for the peak that even the slightest gust of wind, like the tweet below sets off a market sell-off.

For what it’s worth, I think Andrew is referring to Core Automation, an openAI spinoff that is rethinking how AI works.

Weave Robotics Launches Home Robot Isaac 1

From weaverobotics:

Vik: Watch the video. It’s a nice homely robot folding clothes and doing laundry. If it can clean up after young kids, its worth the $8K. Anybody objecting to the price tag needs to understand how messy life with young kids can be, and how a well to do family would buy one. Also, these things get cheaper over time.

Micron, GM Sign Long-Term Semiconductor Supply Agreement

Micron Technology and General Motors have signed a Strategic Customer Agreement to secure long-term supply of memory and storage semiconductors for GM vehicle platforms. The deal covers high-performance DRAM and NAND components used in next-generation GM vehicles, giving GM committed supply access while Micron locks in an automotive customer relationship. (Micron, Reuters)

Vik: Geez Louise! Auto industry is locking up memory supply now too?!

Nvidia, Partners Pledge US Manufacturing Investment

Nvidia and an unspecified group of domestic partners announced commitments to build semiconductor and AI infrastructure in the United States, targeting manufacturing capacity, supply chains, energy infrastructure, and workforce development. The company framed the initiative around applications in healthcare, scientific research, and industrial productivity, though no dollar figures, partner names, or production timelines were disclosed in the announcement. (Nvidia News)

Vik: Here is Nvidia’s “made in America” map. Montana, South Dakota, Louisiana, West Virginia, and Vermont have no hope. No NVIDIA green for you!

Oracle Questions Returns on AI Data Center Spending

Oracle warned that its investment in AI data centers may not pay off, raising concerns about the economics of large-scale AI infrastructure buildout. The caution comes as Oracle is a principal partner in Stargate, a joint venture with OpenAI and SoftBank to build AI data centers across the U.S., including a facility in Abilene, Texas. Stargate has received promotional backing from President Trump and represents one of the largest committed AI infrastructure programs in the country. (Bloomberg Tech)

Vik: Oracle has a $300B Stargate contract with OpenAI and the Oracle’s company filing lists nonpayment as one of the risks. Stock has been dropping pretty steadily this month too.

Oxmiq Raises $35M Series A, Expands to Data Center Design

Oxmiq Labs, the GPU IP startup founded by former Intel and AMD executive Raja Koduri, has raised $35M in a Series A round, bringing total funding to $60M. The company, which develops GPU hardware and software intellectual property, will use the capital to expand beyond GPU IP into broader data center design. (EE Times)

Vik: Oxmiq basically builds silicon IP and software for the entire AI stack that you can just license and build. Raja Koduri is a highly respected individual and an industry veteran who also posts great insights on X.

ASE Reportedly Raises Advanced Packaging Prices by Over 20%

ASE is reportedly raising its quotes for advanced packaging by more than 20 percent, driven by high demand for AI components. The price increase highlights acute supply constraints and the critical role of packaging in AI compute scaling. (TrendForce)

VikL Prices only go one way these days. Up.

TSMC Q2 Gross Margins Near 70% on AI Chip Demand

TrendForce reports that TSMC’s gross margins for the second quarter are approaching 70 percent, driven by strong demand for its 3nm and 5nm nodes used in AI accelerators. The company’s third-quarter revenue is projected to grow over 10 percent sequentially, reflecting its pricing power amid tight CoWoS packaging supply. (DigiTimes)

China’s DeepSeek Raises $7.4B to Compete with Frontier AI Models

Chinese AI firm DeepSeek has raised $7.4 billion to develop models competitive with Anthropic’s Mythos, while Zhipu’s GLM-5.2 has shown strong performance in cybersecurity benchmarks. The developments signal a narrowing capability gap between US and Chinese frontier AI models, as ByteDance also pursues in-house CPU designs for 2027. (South China Morning Post)

Important Tweet

Amazon is shifting away from externally sourcing silicon for its products, and instead plans to adopt a customer owned tooling (CoT) approach with Alchip to develop its own chips for kindle, fire TV, Alexa, etc. Expected to begin in 2027. Annual shipments expected around 40M units.

By the Numbers

Closing moves, 2026-07-01:

• Up: META +8.8%, PLTR +7.8%, CRM +4.2%, AUR +3.7%, MSFT +3.0%, SNOW +2.6%

• Down: NBIS -17.0%, CRWV -13.9%, AEHR -12.1%, KLAC -11.8%, ALAB -10.8%, MU -10.6%

• Extremes: CRWV at a 30-day low

  Thanks for reading! This post is public so feel free to share it.

  Share

Lighter on takes today since Austin is on vacation (he still got a take in!) and Vik is battling a cold.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Etched Emerges from Stealth with $1B+ Customer Contracts

Etched came out of stealth with a successful A0 tapeout, more than $1 billion in customer contracts, and $800 million raised, and says its first racks ship this summer after early tests showed state-of-the-art throughput, latency, and power efficiency. The round drew Jane Street (over $100 million on its own), HRT, Two Sigma, Jump, and a strategic investment from TSMC’s VentureTech Alliance, alongside a 400+ team pulled from Nvidia, Google’s TPU group, Broadcom, SK Hynix, and TSMC and angels including Geoffrey Hinton, Peter Thiel, Andrej Karpathy, and Fei-Fei Li.

To push the pareto curve on many-trillion-parameter MoEs, long context, and agentic workloads, Etched co-designed the whole cluster “from the transistor to the token” around two ideas. Low-Voltage Inference (LVI) runs its math blocks at under half the voltage of most AI chips, which today downclock as FLOPs utilization rises and often sustain under half their peak FLOPs; Etched claims 80%+ of peak on trillion-parameter sparse MoEs without thermal throttling. Cluster-Scale Memory (CSM) pools low-latency memory across the entire scale-up domain over a proprietary interconnect, pairing HBM and SRAM to reach SRAM-level decode speed without the capacity, cost, and yield tradeoffs of SRAM-only or 3D DRAM designs. Etched also built a 2MW datacenter in its office and opened a Taiwan factory, with performance and roadmap details due this summer. (@Etched, Bloomberg)

Vik: Etched seems to have designed a lot more than just the silicon. They seem to have designed the whole rack, including PCBs, liquid cooling, and network “connections” (whatever that means). Underclocking compute seems to help thermals, and they appear to treat all memory in the Etched cluster as “one.” Pretty smart stuff. Excited for more!

Austin: A lot has been announced recently in the AI ASIC startup space. I’ll write about it soon, probably right after the 4th of July holiday.

ASML’s $400M High-NA EUV Machine Achieves 8nm Resolution

ASML started shipping its high-NA EUV lithography machine, a $400 million, 150-ton system the size of a double-decker bus that patterns chip features down to 8 nanometers, versus 13nm on the prior EUV generation. High-NA is evolutionary rather than a new source of light: ASML raised the numerical aperture from 0.33 to 0.55, which cuts transistor size by nearly half and roughly triples density, and lets chipmakers single-pattern layers that today need slower, costlier multi-patterning.

Intel bought the very first high-NA machine and is testing it in Oregon, hoping first-mover access gives its foundry an edge on TSMC and Samsung. TSMC is holding back, calling high-NA only 30 to 50% better and not obviously worth $400 million, and may not run it in volume until the 2030s. The duopoly is also drawing challengers doing end runs around EUV: China is pouring billions into its own extreme-ultraviolet program, San Francisco’s Substrate is chasing x-ray lithography from a compact particle accelerator (targeting $10,000 wafers and its own fab by 2030), and Norway’s Lace Lithography is firing helium-atom beams at 0.1nm precision (2029 to 2030). ASML’s Jos Benschop doubts any of them scale to volume, and is already sketching a hyper-NA successor (0.55 to 0.75, roughly 6nm) for the second half of the 2030s. (technologyreview.com)

Vik: Where lithography goes from here is anyone’s guess, but its great to see so many options emerging.

Meta develops ultra-narrow steel-can batteries for AI glasses

Meta engineers developed steel-can battery technology with widths as narrow as 7mm for AI glasses like Ray-Ban Meta and Oakley Meta Vanguards. This involved replacing traditional wound “jelly roll” electrode material with die-cut stacked layers to achieve lower impedance. The Gen 1 Meta Ray-Ban’s cell capacity grew from 160 mAh to 210 mAh, and the Meta Ray-Ban Display glasses feature a 248 mAh steel-can cell. (engineering.fb.com)

Vik: Like a AAA battery but way smaller?

TSMC Accelerates CoPoS Optical Packaging Under Supply Chain Silence

TSMC is fast-tracking development of Co-Packaged Optics on Substrate (CoPoS), an advanced optical packaging technology that integrates photonics directly with chips at the substrate level. The company has placed its entire supply chain under a non-disclosure agreement, barring partners and vendors from discussing the program publicly. CoPoS is aimed at data center interconnect applications where electrical signaling faces bandwidth and power limits at scale. (digitimes)

Vik: CoPoS is an open secret now, and lots of people can guess who Taiwanese suppliers are.

From Pres. Bill Chiu: “TSMC will need to expand domestic procurement, since only local suppliers can respond quickly enough to keep pace with R&D needs — meaning Taiwan-based equipment and materials makers in the first wave of CoPoS suppliers stand to benefit.”

Sector Watch

• TSMC Q2 gross margins near 70% on insatiable 3nm and 5nm demand; Q3 revenue projected to grow over 10% sequentially. (TrendForce)

• Anthropic receives US government approval to expand Mythos 5 model access, removing regulatory bottleneck for commercial deployment. (Bloomberg Tech)

• ByteDance advances in-house CPU design for 2027 deployment, reinforcing vertical AI stack trend among Chinese tech giants. (SCMP Tech)

• Nvidia Research publishes a self-evolving AI agent framework for hardware design that treats RTL development as repository-level code evolution. (Semiconductor Engineering)

By the Numbers

Closing moves, 2026-06-30:

• AMBA ▲ +28.0% — Rosenblatt labeled Ambarella a ‘Physical AI Pure Play’ and set a $120 price target, sparking the 28% surge.

• MXL ▲ +18.0% — Stifel raised MaxLinear’s price target to $110, maintaining Buy on AI infrastructure end-market strength.

• FORM ▲ +10.9% — no clear catalyst from public sources

• SNDK ▲ +10.9% — Analysts flagged a record-breaking quarter and projected another large jump, driving Sandisk’s 10.9% gain.

• CRDO ▲ +10.7% — Strong Q4 results and analyst upgrades drove Credo Technology shares up ~11%.

• IMOS ▲ +10.2% — Post-dividend selling pressure faded for ChipMOS, allowing shares to recover ~10% as the dividend hangover cleared.

Down: CLSK -5.3%, HPQ -3.1%, SKYT -2.2%, QCOM -2.1%, WDC -2.0%

Or did we just miss something important? FOMO is real. Let us know in the comments.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast playe

Anthropic Claude Reaches General Availability on Azure

Anthropic’s Claude models are now generally available on Microsoft Azure via Microsoft Foundry, running on NVIDIA GB300 Blackwell Ultra GPUs. The deployment targets enterprise customers building agentic AI applications and domain-specific agents within Azure-native environments. (Nvidia News)

Austin: Anthropic is a new TAM for Nvidia here… used to only be on Trainium.

Taiwan Raids Super Micro in Nvidia GPU Export Probe

Taiwan authorities raided Super Micro’s offices as part of an investigation into alleged smuggling of Nvidia GPUs to China in violation of U.S. export controls. The probe is part of a broader Taiwan crackdown on chip export circumvention. Super Micro said it is cooperating with authorities. SMCI shares fell roughly 7% on the news. (Bloomberg.com, Stocktwits)

Vik: This was a whole thing in late March of this year, where NVIDIA GPUs were smuggled into China by swapping shipping labels on rack servers. The CEO later issued a letter saying that this was not Super Micro’s intention, and that the “three individuals” responsible for this violation have been “taken action against.”

Meituan Open-Sources Trillion-Parameter Model Trained on Chinese Chips

Meituan has open-sourced LongCat-2.0, a 1.6 trillion-parameter large language model with a 1 million-token context window, trained entirely on domestic Chinese chips. The Beijing-based food delivery company claims it is China’s first trillion-parameter model trained on home-grown hardware rather than imported GPUs, placing it on par with DeepSeek’s latest flagship by parameter count. (SCMP)

Vik: Food delivery companies in China are now training AI models? Who is delivering food then? Apparently its near frontier performance from 50k Chinese accelerators.

Austin: Food delivery used to be the side hustle. Now training AI models is.

Palantir, NVIDIA Integrate Nemotron Models Into Sovereign AI Platform for US Agencies

Palantir and NVIDIA have expanded their AI partnership, integrating NVIDIA’s Nemotron open models into Palantir’s Sovereign AI Operating System. The system targets US government agencies and critical infrastructure operators that require on-premises deployment, allowing them to train and run models without transferring sensitive data to public cloud environments. (EE News Europe)

Vik: Keeping sensitive data close to the chest is going to become more and more important in the world of AI. There is too much risk of info leak or distillation if AI is used by high profile organizations on the open internet.

Austin: My ears perk up anytime I see mention of Nemotron. As a merchant silicon vendor making training systems, it makes sense to train your own models to deeply understand your customers and test your systems. But deploying them? After you’ve invested all the depreciation and OpEx for training Nemotron, might as well get some use out of it! Not sure if Nvidia makes any money directly sharing Nemotron, but surely brand goodwill at a minimum.

Samsung Electro-Mechanics Expands AI Server Substrate Capacity in Busan

Samsung Electro-Mechanics will expand flip chip ball grid array (FC-BGA) substrate production at its Busan facility to meet AI server demand, with a further capacity increase at its Sejong plant also expected. Samsung Electronics Executive Chairman

Digital Realty Acquires Blackstone’s Stake in Three Northern Virginia Data Centers

Digital Realty (NYSE: DLR) agreed to purchase Blackstone’s interest in three fully-leased hyperscale data centers in Northern Virginia, increasing its ownership in assets in the top U.S. data center market. To fund the acquisition, Digital Realty priced an underwritten secondary offering of approximately 12.31 million shares of common stock. (Digital Realty)

Austin: Stock down on the news.

Worth a Watch

Bolt Graphics’ GPU Architecture and Plan to Challenge Nvidia with Founder Darwesh Singh — Gamers Nexus

Gamers Nexus sat down with Bolt Graphics founder Darwesh Singh, a US startup taking a genuinely strange swing at the GPU market. Bolt’s card is RISC-V and chiplet-based, with user-expandable memory (two standard DDR5 SODIMM slots, so you choose your own 8 to 48GB DIMMs), dual PCIe, and an onboard Ethernet jack.

The plan is to go where Nvidia is pulling back, winning over creative pros first and gamers later, and to prioritize path tracing over rasterization instead of chasing raw frame rates. Singh walks the roadmap (FPGA prototype today, 5nm tapeout by year end, mass production end of next year) and is candid that the hardest part is the same one Intel is still fighting: drivers and software.

Austin: It’s natural for the world’s largest accelerator companies to focus on datacenter AI and toward the future Physical AI S-curve. It’s also expected for startups to move in for a foothold where attention has waned. Bolt is bringing more acceleration to the workstation for creative professionals, simulations, high-performance, etc. You like FP64?

Transcript on Chipstrat

Sector Watch

AI & Compute

• Meta deployed a custom CXL ASIC to pool memory from decommissioned servers, bypassing HBM supply constraints for inference workloads. (The Register)

• Global AI sales reached $25B in Q1 2026, exceeding $21B in data center depreciation costs but leaving margins critically thin. (TechMeme)

Memory

• SEMI projects 300mm memory equipment spending to exceed $50B in 2026, driven by HBM and DRAM capacity expansion for AI workloads. (semi.org)

China & Policy

• CXMT secured ~$3B three-year server DRAM supply agreement with Tencent, with SemiAnalysis projecting $55B revenue this year at 75%+ gross margins ahead of IPO. (TechMeme)

• CXMT signed 20B yuan supply deal with Tencent while in talks with Alibaba Cloud, ByteDance, and Xiaomi; Apple sources CXMT DRAM exclusively for Chinese market. (chosun.com)

• Baidu’s Kunlunxin seeks $50B Hong Kong IPO valuation while requiring investors to commit to purchasing its AI chips. (DigiTimes)

Foundry & Logic

• South Korea unveiled $880B AI and semiconductor investment megaproject led by Samsung and SK Hynix to cement memory and logic market dominance. (Bloomberg Tech)

• Apple to launch first touch-screen MacBook with M5 Pro and M5 Max silicon, skipping M6 generation entirely and jumping to M7 for 2027 models. (co.uk)Jay Y.

Stock Movers

At close on June 29th, 2026.

• ALAB ▲ +16.4% — Astera Labs debuted in the Nasdaq-100 index, triggering mandatory buying from passive funds benchmarked to the index.

• ALGM ▲ +14.7% — Mizuho raised Allegro MicroSystems’ price target citing AI data center demand, pushing shares to a new all-time high.

• ACMR ▲ +13.8% — Morgan Stanley maintained its Overweight rating on ACM Research, catalyzing a sharp single-session surge in the semiconductor equipment maker.

• MXL ▲ +12.3% — no clear catalyst from public sources

• KLAC ▲ +12.0% — no clear catalyst from public sources

• WDC ▲ +11.2% — Micron fell 5% and SanDisk dropped 7% while Western Digital outperformed, as investors favored its diversified storage business over pure-play NAND peers.

Press Releases

• Google Gemini can now take notes in Google Meet for Google AI Pro and Ultra subscribers. (Google IR, June 28)

Things we cover:

• High Bandwidth Compute: stacking LPDDR on logic instead of HBM on the side

• Why disaggregated inference opens the door for latecomers

• The AI 200 / 250 / 300 accelerator roadmap

• Alphawave as Qualcomm’s Mellanox, plus the Modular software stack

• The C1000 data center CPU, with Meta as a customer

• Edge AI, AI-defined vehicles, and the robotics endgame

This podcast is lightly edited for clarity.

Cold Open and Catching Up

Austin: Qualcomm’s main point is that they’re diversifying their businesses, and the majority of their business in the long run will be non-communications. I thought this was interesting because the name of the company and the legacy of the company is communications. Quality communications. Qualcomm. But the crazy thing is the data center business could inflect so much that what Qualcomm makes from their new businesses, data center and automotive, could ultimately dwarf what they’ve made in the aggregate of the company’s history as a communications player.

So that’s something big to wrap your head around: Qualcomm being a communications player could just be the start of their story.

Austin: All right, hello, welcome listeners. Welcome to another Semi Doped podcast. I’m Austin Lyons with Chipstrat, and with me is Vik from Vik’s Newsletter. Hey Vik, what’s up, man? It’s been a while since we chatted.

Vik: Yeah, it’s been a while. I’ve been traveling, you’ve been traveling. Finally we’re back at the home studio, so it’s time to do something now.

Austin: Totally. So you were in the UK and it was hot?

Vik: It was crazy. I thought I ran away from India to get away from the heat and feel some cool air in the UK, because everybody says, oh, it rains all the time. So I was hoping to get rained upon a little bit. But anybody listening from the UK will be like, no, no, you never wish for that stuff. We like a little bit of sun once in a while. I know it’s a lot of sun right now in Europe, but it’s okay, maybe for a week or two.

Austin: Yeah, totally. I saw it was like 110 degrees Fahrenheit or something in France. When you said it was hot, I was like, oh yeah, it’s UK hot. But no, dude, it was hot, hot.

Vik: It was hot, hot. And the other thing is, in the UK they’re just not equipped to deal with this stuff. In India there’s AC in some places. If not, if you go by buses, for example, there are windows, sometimes there are no windows. So you always feel the breeze. It’s not cool breeze, but at least there’s movement of air. In the buses in the UK, it’s just sealed. There’s no air movement, the windows don’t open out or slide out or anything like that. And you’re just trapped, with a bus full of people just sweating. It’s a sweat lodge.

Austin: Just suffocating. Brutal. Well, I was in New York City. Qualcomm flew me out for their investor day event. And the weather was really nice in New York City. It was very beautiful. I woke up early, the sun’s out for a long time right now, and I ran in Central Park. I ran one day to the west side along the Hudson River. I saw crazy huge yachts. I have never seen such big boats in my life.

I don’t know if it’s someone who lives there’s yacht, or if it had to do with the World Cup, because there were a bunch of people in town for the World Cup. But I was like, dude, you need a whole staff to run this thing. This is crazy. And here it is, and I’m just right by it. I was going to take a selfie in front of it, but I thought, I don’t know, that seems kind of touristy.

Vik: I mean, all the investment bankers live there, right? Half the people who listen to our show must be on their yachts. Maybe they should let us know. Hey Austin, you saw our yacht.

Austin: That’s true. Hey, live Semi Doped podcast from your yacht. Send me an email and we can chat.

Vik: Okay, okay. So wait, wait, wait. You were in New York for the Qualcomm investor day then.

Austin: Yes, Qualcomm investor day. I was there. They hosted it in New York City because the audience is the sell-side analysts and the financial community, but they brought out industry analysts as well, because they had a lot to talk about.

Qualcomm’s Diversification Bet

Austin: Really, the story we’ll walk through, and I want to talk through the technical bits and hear your take on some of it, the story that Cristiano Amon and Qualcomm were trying to communicate was all about their diversification away from handsets. Qualcomm has always been a communications company, a big smartphone business.

Everyone is seeing the writing on the wall with smartphones: we’re at the top of the S-curve. Everyone has a smartphone. There’s lots of competition in the space. Qualcomm has been moving into automotive, and that business has been growing, and it’s a proof point that they can use M&A to get into a new business, build the technology, use their channel and their manufacturing scale to grow that business and take it to market.

And the whole point of this was Qualcomm trying to say, we are going to do that with data center as well, and here’s our plan. There was, and we’ll include the image, but they had this one money slide, I think from the CFO’s presentation, where there were three circles. It was a donut ring slide, and they were showing: okay, in fiscal year 25, handsets were two-thirds of our business, automotive and IoT is one-third. Our goal for 2027 is roughly half and half, half handsets, half other, which now starts to include data center.

And then fiscal year 29, the goal is actually two-thirds of the business is automotive, IoT, and data center, and only one-third is handsets. So they’re trying to move beyond their legacy communications-only business into these other new businesses. Last point and then I’ll let you chime in: obviously communications is part of all of these businesses. The IP and the technology is still important to the other businesses. It’s just that they’re going to have to build more and do more than just handsets.

Vik: I saw the whole investor day talk. I wasn’t there, but I saw it on YouTube. It was interesting because Cristiano Amon said that for the 40th anniversary of Qualcomm they had the founder over, Irwin Jacobs. And during that talk he mentioned that he made one mistake in naming the company Quality Communications. He said, I should have just stopped with one M in the com, because now it could be quality compute.

But now it’s a problem, so they couldn’t. In my opinion, they should rebrand it and drop one M.

Austin: Totally. Oh yeah, if only he had the foresight to think 40 years ahead, right? That’s so funny. So true.

Vik: Yeah. So that’s the big deal. What you say is actually a big deal, because this company from the get-go has always been about communications. They have a rich history. It started with satellite communications. When I joined there in 2018, they had a tour of the museum within the campus that showed the entire history of all their devices made from the beginning. It’s really nice, it’s part of the orientation, they take you to the museum. It’s a very inspiring way to start working there.

But that’s the legacy. They’ve been through 3G and 4G and 5G. They’re also on the 6G train, we’ll talk a little bit about that. But now they’re saying that only one-third of the business is going to be handset-based, which is a way of saying that’s the communications business. Two-thirds will be IoT and data centers. It’s a big shift.

Austin: Indeed. It is a big shift. So let’s start with maybe the confidence question: do you think Qualcomm can get into new businesses? Maybe we’ll start there, strategically, high level. Do you think they can build a real data center business? I’m going to let you have your take, and then maybe I’ll provide some color on what they said as well.

Vik: They can, because there’s no reason a company like Qualcomm cannot get into the data center business. They have compute. They have the CPU business, they have NPUs, which they use for graphics acceleration on their SoCs and mobile handsets anyway. They have all kinds of IP now through the Alphawave acquisition. I was looking at what all they got through that, and it’s quite a lot, because they have SerDes IP in copper and optical.

They have PCIe Gen 6, CXL for servers and storage. They have Ethernet IP, 800 gig, 1.6T, for switches, routers, DPUs, NICs. And they have HBM and DRAM IP for GPUs, which is very important. We’ll talk about that, because that’s one of their key technological innovations in the era of AI and the way they’re doing inferencing chips. And they also have chiplet IP through the Alphawave acquisition, through UCIe and a bunch of other standards.

So they’re very well positioned as a company that has also shipped billions of devices into the handset world. It’s not like they’re new to doing this. I could argue that if OpenAI says, I’m now a chip company, or Anthropic wants to be a chip company, they’ve never been a chip company. It takes experience to be one, which Qualcomm has. So yeah, they’re well positioned to do so, although one could argue that they’re a bit late to the scene.

Austin: Yes. Okay, so let’s get into that. In the grand arc of things, late and timing is interesting, because obviously this has been going on since late 2022. It’s only 2026, and this is the type of technology that will be here for the next 30, 40 years, whatever, forever. So being late now feels like a big deal in the grand arc of time, but it won’t be. It’ll just be a little blip.

But the question is, what does it take to succeed? Which I think you hit on nicely: you need IP, you need talent, you need manufacturing prowess, and you need technological differentiation. It can’t just be a me-too fast follower, like, oh great, a company did this, I’ll build the same thing and sell it to a different market. That’s not going to work. There has to be a reason for a customer to choose you. And to the point of being late, you especially have to differentiate on some vector.

I think of Nvidia’s Hopper-era GPUs, and then eventually AMD came in with Instinct, and what was nice was they made a different decision about the amount of HBM capacity on a single chip. That small decision on one vector opened up some particular workloads that could run on eight Instinct 300s, I think, or maybe it was 350, I can’t remember off the top of my head, but it took 16 Nvidia GPUs, because they just didn’t have enough memory capacity on each GPU. That illustrated for me that even if you come in late, if you make some sort of different technological bet, it might unlock a certain set of customers, which can help you get your foot in the door and grow from there.

So to your point, Qualcomm checks a lot of the boxes: they’re a big company, they have experience, they have IP. And by the way, they were able to use M&A to get IP they didn’t have, get talent they didn’t have, and fill those gaps, which is something we’ve seen them do before in their automotive business. But there still needs to be something different, so that at the end of the day a customer like a hyperscaler or an AI lab will know why they want to choose Qualcomm. And it’s not just because it’s cheaper, or because everyone’s out of supply and Qualcomm happens to have some. None of those are sustainable, those are temporary. There has to be technological differentiation that really impacts TCO for a particular workload. This is even why we saw the Groqs and the Cerebras have success: they made design decisions with lots of negative consequences, but for a particular workload, high interactivity, it was much better than GPUs could do. So obviously Qualcomm had to come out swinging and say, we think we can do something different here.

Disaggregated Inference

Vik: Yeah. Their key differentiator here, apart from their memory architecture, which we’ll discuss, is basically the fact that inference now is getting disaggregated. This was their key theme that went through the whole talk, they kept saying this is now disaggregated, which means you can actually develop hardware to do a particular task. So you could put in a Qualcomm rack along with some other racks. It doesn’t mean you have to fill Nvidia racks with all these Blackwell GPUs along with their Vera CPUs. You don’t have to do that. Now you can put a rack full of CPUs separately. You can put a rack full of low-latency decode, like you were saying, the Cerebras or the Groq LPUs. And now you can put maybe a rack of Qualcomm inference chips just to do that one portion of it. If they have high memory bandwidth, you could just do decode on that, and still continue to use Nvidia GPUs for the prefill part. So the disaggregation across the whole inference space is becoming more and more a theme, because one size does not fit all. You’ve spoken about this in your Substack as well, there’s a right-sized hardware for the right-sized workload. So that’s increasingly becoming a theme, which is why you can run different kinds of hardware along a common software platform, which Qualcomm also wants to have, open and developer-friendly for everybody to work with.

Austin: Yes, you’re totally right: disaggregation has really opened the door to everyone. Once we started to break the workload down for inference, today LLM inference is obviously the defining workload, and the value is actually all being created by agentic inference workloads, reasoning models, but in breaking up that workload into prefill and decode, then people could say, okay, I’m late to the market, I’m just going to focus on decode. I’m going to do something very decode-specific and very different, and that will help me get my foot in the door. Again, it worked for the AI ASIC startups, and this will have to be the mindset that Qualcomm has.

Because ultimately, yes, would they like to sell data centers full of Qualcomm Dragonfly, which, by the way, that’s what they branded it, for anyone who’s not watching. Dragonfly is their data center brand. Would they ultimately like to sell, and the branding was white and gold, by the way, so it looks very nice, would they like a data center full of white racks, so that whenever on FinTwit you see a picture of someone saying, look, I built a data center and all the racks are white, everyone can go, whoa, Qualcomm to the moon? Yes, they want to do that. But the reality, which I think they’re fully aware of, is they need to sell a few racks into the existing data center that’s full of Nvidia GPUs. So focusing in on decode, for example, would be one way to get there. But okay, take us further. Do you want to talk about the memory, the HBC? That was sort of the big star of the show.

High Bandwidth Compute

Vik: Yeah, let’s go to HBC, because that’s an underlying technology that’s very useful regardless of whether they want to use it in data center hardware or IoT or automotive or robotics, because this is going to be the platform that lays the foundation for all of their businesses going forward, in my opinion. At least that’s the way it seemed from what I heard.

So high bandwidth compute is their way of getting better memory-bandwidth hardware packaged right next to an XPU, so that you can get a really fast inference bandwidth like you’d get with Cerebras or Groq LPUs. Those things can do, a SRAM bandwidth is in the range of 100 terabytes per second, and right now HBM bandwidth is about a tenth of that, about eight terabytes per second. So they want to break through this eight-terabyte barrier, which currently involves an HBM stack sitting next to a GPU or XPU and connected with maybe 2,000 lanes of interconnect between the memory and the GPU chips, packaged with an advanced packaging substrate like TSMC CoWoS.

So their idea with high bandwidth compute is that you take the memory and you put it on top of the XPU die. Instead of connecting the memory through the side, using shoreline density as it’s called, on one edge of the GPU die you can only put so many lanes before you run into advanced packaging limits, if you put the memory on top of the XPU, you expose the entire face of the chip between the XPU compute and the memory to have interconnects. So now, instead of 2,000 lanes going between these chips, you can have, I don’t know, tens of thousands or even 100,000 different lanes. What that does is increase your bandwidth, if you increase your number of lanes by 100x, you’ll get 100x more bandwidth. So this is their main idea, what they call high bandwidth compute.

Austin: Yes. Okay, there’s so much to unpack here. I’m going to walk through it again for people at a high level in case they didn’t catch it the first time, because, I’ll be honest, one, the naming might not do it justice, although I don’t have a better name, because HBC, HBM, it feels very similar, but there’s actually a lot of differences going on. And two, there are some nuances that weren’t obvious to me the very first time it came across, but as I looked back at it again preparing for this episode, some things stood out.

So, one of the problems: everyone knows the way GPUs and HBM work today. You’ve got the HBM, it sits next to the GPU, and you’ve got things like weights or activations, KV cache, sitting in this HBM. During decode especially, that stuff just has to fly back and forth. It’s like if you separated your kitchen and put your fridge in the garage, and you just had to walk to the garage and get something, bring it back, chop it up, walk to the garage, get the next thing, bring it back, chop it up, walk to the garage. Obviously it’s like, dude, what are you doing? Why are you spending all this time walking back and forth to your garage? Just put your ingredients right next to your chopping block. Because of course there’s lots of power loss, lots of latency, that kind of thing.

And to your point, the analogy breaks down, but you can’t get as many lanes and as high bandwidth. So one point is: how do we bring the compute as close to the memory as possible? Other people have called this near-memory compute, processing in memory, stuff like this. There are lots of startups doing this. I think it was d-Matrix doing something like this?

Vik: Yeah, yeah. This is d-Matrix’s approach, what they call in-memory compute.

Austin: In-memory.

Vik: So d-Matrix also puts memory on top of logic. So there’s a question of, really, is the Qualcomm innovation all that different? That’s another question to ask.

Austin: Totally. Okay, so that’s the first problem, how can we bring the compute as close to the memory as possible? And what you said is, well, the closest thing is to stack the memory on top of the compute and literally have it as close as possible. Now, there are obviously thermal concerns there. These GPUs, XPUs are crazy hot. So can you actually put memory on top of the XPU?

As I looked closer, and we’ll try to include some pictures as well, my understanding is they’re not actually stacking the memory on the XPU, but they’re putting a logic chip under it, and putting it close to the XPU, and they’re actually offloading some of the workload to live on that logic chip. I don’t even think it’s necessarily a programmable logic chip. I have a feeling it’s almost more of a true accelerator, in the sense of accelerator. So for example, maybe you have during inference some sort of primitive, some function that’s run all the time, like softmax, or something in attention. Can you actually take that off the XPU, put it in the logic that’s right under the memory, and the idea is: fetch what you need from the memory, do those primitives right there, and then send the result back? That would help you solve the thermal issue, you’re not actually putting the memory on top of this crazy hot XPU, but you’re still using advanced-node logic and putting the memory on top of it, and accelerating an even finer-grained piece of the workload.

This would be the name we’ve always used before AI, which is an accelerator. Like a camera, an image-sensing processor, or any other little acceleration we’ve done in the past, like with graphics, you’re like, dude, we just keep doing this particular loop over and over on a CPU, why not build a little thing to do it? Qualcomm does this all the time in their Snapdragon SoCs, build a little thing to do some image processing that’s dedicated for that. My understanding is it felt conceptually like they’re saying, take a tiny piece of the inference primitives, move it off, but also it’s very important that it’s very close to memory, because it’s very memory-bandwidth intensive. So I’ll pause there. What’s your reaction?

Vik: When I saw this announced on the talk, it was very confusing the way it was drawn out. In the first example, where they show how the current solution works, they had a GPU and an HBM, and there was this gold-colored back-and-forth happening. It was very beautiful to look at, you could see the gold data going up and down the HBM stack and going around into the GPU. So that’s how it works today. But then when they put in their solution, their high bandwidth compute, they still showed a stack of LPDDR. And then they said something about an XPU. And sitting next to it was an SoC. And then they were showing the same gold thing, presumably data, going between the SoC and the LPDDR stack.

So for somebody looking at this, it looks very similar to the HBM one. I’m like, what’s different? And then they go on to mention that, no, you don’t need advanced packaging anymore, you could do standard packaging, because you don’t need that bandwidth anymore. So my interpretation is that the logic die under this memory is not any simplistic logic die of any kind. It is a full-up XPU. Because that’s the only way you will not need advanced packaging to make this work, you have to do the full thing there. And that’s what d-Matrix is doing: they do all their compute right under memory.

Austin: Okay, so when you say XPU, what’s your definition of XPU in this particular use case?

Vik: This is the thing that does matrix multiplications. That’s what I call an XPU.

Austin: Yeah. So do you think they’ll do any matrix multiplication anywhere in the workload right there under the LPDDR?

Vik: Yes, I think so. To me it looks very similar to what d-Matrix is doing, because d-Matrix, digital in-memory compute, and their Raptor, which is the next generation of 3D DRAM, those slides are out there, they previously announced this stuff. So it’s nothing unannounced. They actually have a full-up XPU chip that does all the compute under the DRAM die. That’s their Raptor chip. So this has to be something like that. This SoC sitting next to it is doing something else. It could be a mobile SoC, if you’re using AI on a phone. It could be an automotive SoC, if you’re using AI in a car. But it’s not doing matrix multiplications, it’s doing SoC stuff.

Austin: Sure. I agree with you. Conceptually I like the idea of essentially a tensor core coming over and doing all the matrix multiplication. I totally agree that’s what makes sense, get all of this multi-dimensional data, put it right down into some sort of tensor-core-type thing, and have it do all the matmuls. It would be nice to understand what else is in this SoC. Obviously there’s die-to-die communication, there’s orchestration of the whole workload, there’s communication with CPUs, all these other blocks that don’t need to live right there. But I do agree that conceptually, anything in the workload related to the specific matrix multiplication would live under the memory.

Vik: Yes.

Austin: And you mentioned earlier that this technology could come to the edge, to phones, to auto, wherever. I think that’s also a key point: if people are thinking GPU with LPDDR stacked on top, that might be confusing, because it’s like, oh, are we putting big GPUs into a car now? But the point is the concept of the technology: if your car’s going to run neural nets, wouldn’t it be nice to have the matrix multiplication under a big stack of, not HBM, but LPDDR, something lower power, cheaper, higher capacity, that would still let you do generative AI at the edge, presumably with even bigger models?

The Hard Part: Thermals, Stacking, and TSV Density

Vik: Yeah. I want to talk about this memory-stacking part first, because we’ll definitely talk about the automotive, IoT, and robotics aspects, which they really want to get into in the future, and I think they’re well positioned to do it. But there are several issues with the way this memory works to begin with.

First of all, you already mentioned the thermal aspect. It is pretty challenging to stack logic on top of GPUs or XPUs like this. The second thing is that there’s a fundamental size mismatch. These XPUs that sit under any form of memory, even when d-Matrix does their Raptor, they’re pretty big dies, because they’re reticle-sized GPUs that sit there. Nvidia even packages multiple GPUs in a single Rubin or whatever nowadays. So it’s easy and okay to assume the XPU will be a reticle-sized compute unit, which means it’s like 850 square millimeters, anyway, above 800 mm². Now, if you have to stack DRAM on top of that, can you imagine how big a DRAM die you need? And can you imagine the planarity requirements of actually hooking this up and keeping it planar in a thermal environment? Stuff is heating up, and the planarity between the top die and the bottom die needs to be maintained without having the bumps rip off each other. This is a challenging problem.

Austin: Yes. You’re saying that things move and warp when it’s hot, but obviously you need physical connections, wires, pipes, so to speak, going up and down. So you can’t have all this moving around. It needs to be rigid, stable, heat up together, cool off together, or dissipate all the heat so it doesn’t move around.

Vik: Yes. So it’s a challenging problem. And as much as they say, oh, we don’t need advanced CoWoS packaging anymore, yay, we solved packaging, it’s enough to use a standard-pitch package for this, they have not really solved the packaging problem. They moved it to a different place.

Austin: From next to the XPU to on top of it, which is even harder. And one more thing I want to point out: the picture deceivingly shows multiple LPDDR stacks. It’s not that easy, because even the Raptor first generation is probably going to have one die layer. The expansion is going to happen later, people want to stack more DDR die on top of each other, but it’s not simple by any means, because as it is, it’s hard enough to stack one memory chip on top of logic like this. This is logic-on-logic stacking, in a sense, if you think of it that way. It’s like stacking SRAM on top of compute, like AMD did with its V-Cache. Similar stuff.

Vik: Right. Well, talk to me about this then. Obviously Qualcomm’s going to need to work with the memory vendor, but then there’s a logic accelerator under it. So they obviously have to work with the logic foundry as well, TSMC.

Austin: TSMC, yeah.

Vik: TSMC, exactly. Then, as far as the thermals and the stacking and making the logic and the memory behave nicely, ultimately, who does that fall on? HBM4 has, or maybe it’s HBM4E, memory stacked on top of a custom logic die now. So I’m wondering if some of these thermal problems are already being solved for HBM4, the HBM stacked onto a custom base logic die, and if there are any learnings that can be applied here. So it’s like, oh, the industry has already solved this, or is it, no, no, we’re talking about different memories and a different level of compute under the memory?

Austin: I don’t think the logic die and the HBM die sizes are as big as a GPU die. They’re smaller.

Vik: True. Yes.

Austin: So that’s what I was pointing out earlier, it’s not as easy as that. If you want to stack something the size of a GPU die, that’s difficult. The logic dies under HBM4 are smaller, and they don’t nearly do as much work as something doing matrix multiplications to infinity, which is what XPUs really do. They do a lot of matrix multiplications. So it’s a computationally much heavier workload when you try to do it under.

Vik: So presumably running hotter, more often.

Austin: Yeah. There’s another catch to this. When you do DDR7 on top, let’s say LPDDR on top of LPDDR, the whole idea was that DDR has more capacity than HBM per layer, and therefore you’re getting more capacity. But now, when you stack DDR on top of DDR, that’s pretty much HBM. Now what you’re doing is stacking HBM on top of compute, in my opinion. And the reason HBM per layer does not have that density, that memory density, is because there are through-silicon vias. When you have through-silicon vias, you can’t put as many memory cells around them. There’s a keep-out zone: if you put a through-silicon via here, you can’t put memory cells around it for some distance. So per die, your memory capacity drops.

You get high density if you stack only one layer, because you don’t have to put vias through that memory layer. The moment you start stacking two or more, you have to put vias through the layer, which means the capacity per layer starts to drop, let’s say by half. So you need to stack four to make it even useful. You can’t stack two, because you lose it on density and keep-out regions. So you see the complexities here that nobody ever talks about. They just say, look at this, yay, high bandwidth compute, let’s go. And that’s it.

Vik: Totally. On that point, this was obviously an investor day, so they teased some technology, but they didn’t get into the technology weeds. I would love to see Qualcomm have a data center day or something focused on the technical architecture and all the technical details. Or ultimately just publish some papers on this, show it at Hot Chips, whatever, because there are lots of technical questions. But if they can answer them confidently, then it gives a lot more credence to what they tried to say at investor day, which is: we’re late, we’ve got a new approach, we think it can ultimately be competitive, it’s HBC, and we’ve shown you all the technical bits and answered all your questions with satisfaction. So you can believe that it’s possible, it’ll work, and it’ll scale.

The Roadmap: Accelerators, Alphawave, and the C1000 CPU

Austin: Yeah. So I think we’ve spoken about the technical complexities of HBC. We don’t know any more details, like you say, unless they publish something. But the first HBC-based chip is expected in 2027, that’s what they said. And then they showed a chart that they’ll do the AI 300 in 2028, and they’ll use UALink and E-sun for scale-up fabrics. And then somebody in the audience asked, when are you doing scale-up CPO? They said, yeah, it’s going to be after that. So maybe a 2029 thing would be CPO. It’s interesting that this is in the works, but we really have to see the silicon show up to understand anything more than what’s already been announced here.

Vik: Totally. So on the accelerator roadmap, they said, yes, we’ve got a multi-generation accelerator roadmap. Historically they had the AI 100, that was a long time ago. I’m sure they learned some lessons, but you could argue it doesn’t really count. The AI 200 is sampling in 2026, and it does not have this HBC we’re talking about. It is the AI 250 that has the first generation of HBC technology, and that’s in 2027. And then, to your point, the AI 300, which isn’t until fiscal year 28. When does Qualcomm’s fiscal year start? You would know.

Austin: That’s a good question. I never followed the financial side of things when I was at Qualcomm, I was doing engineering work. But I think it starts in 2027 actually. The reason I say that is, at the end, when the CFO was talking about FY 2027, he mentioned, hey, FYI, that’s actually in calendar year 26.

Vik: Yeah, right. Exactly. Well, I don’t know if it’s a full year, but it probably starts in June or something.

Austin: Which is actually one year earlier, yeah.

Vik: Qualcomm fiscal year start, I’m Googling it, audience. According to the AI overview, Qualcomm’s fiscal year begins on the last Sunday of September. So they pull it forward by a quarter, essentially.

Austin: Right. So I did pay attention a little bit.

Vik: So the AI 300, that’s when they get the second generation of this HBC. That’s also when they add scale-up, UALink or E-sun, and copper and optical scale-out. So they started to have the buzzwords. I think this is a nice transition to talk a little bit about Alphawave Semi. One of the things nowadays, if you’re coming to compete in the data center accelerator market, you can no longer say, I have a chip. You have to say, I have a rack-scale solution, which not only means the accelerator chip, but the interconnect.

Going all the way back, Nvidia bought Mellanox, and that was their way of using M&A to say, we want to be more than just the GPU, we want to do the whole system, and Mellanox gives us all the networking, scale-up, scale-out, switches, all the good stuff. So the question is, what is Qualcomm’s Mellanox? And Alphawave is an acquisition they made, I think about a year ago, maybe a little less. Alphawave historically is well regarded for their IP, a lot of communication IP, from SerDes and essentially taking an accelerator and making it able to talk at high enough bandwidth, getting data in and out, communicating with industry-standard networks, talking to a Broadcom switch on the other end, or whatever.

I’ll also mention that Alphawave Semi brings custom silicon customers to Qualcomm, because Alphawave was pretty smart: hey, if we’re good at the hard part, SerDes and IO, we can help people do custom XPUs. That’s essentially what Marvell does, what Broadcom does, we could license you the hard part, or we could also do the so-called easier part, like the front-end RTL stuff. So Alphawave Semi was actually also in the game of custom XPUs. Not only did Qualcomm acquire the IP and the talent, but they also acquired customer relationships and roadmaps that Alphawave Semi had with these huge hyperscalers. They didn’t name them, they said there are two. And then separately on the call, Satya Nadella spoke and Mark Zuckerberg spoke, so that implies Meta and Microsoft are working with Qualcomm on data center. But they specifically called out two custom silicon customers, which are presumably heritage, legacy Alphawave ones. I didn’t think about this until just now, but Alphawave was public, so a person could probably go back and figure out who those customers are.

The CFO also said they expect $1 billion of revenue in 2028, I think, where total revenue would be $5 billion. But $1 billion will come from each of these hyperscalers, so $2 billion is already spoken for. I don’t think it’s actually Alphawave’s customers, because he said that on top of this, we have the Alphawave business as well that will add on top. So he says things are looking very good, that was the CFO’s whole thing.

But in terms of acquisitions, their Modular acquisition that was recently announced is very interesting too, because you can’t build a rack-scale solution without providing a software layer to go with it. That’s exactly what Modular brings to the table. They have something called Mojo, which is somewhat equivalent to Nvidia CUDA in the programming layer. They also have something called MAX, which is the equivalent of Triton or TRT-LLM, for model serving. And they have the cloud product, which is distributed, it’s like Nvidia’s Dynamo, it handles KV cache offloading and data movements and all of that. So Modular basically lets you use this software layer to run what they call multi-silicon token factories, which means this software layer will tie together hardware from multiple vendors in a single solution, so you can mix and match. That’s a powerful thing, a Qualcomm solution can be slotted in with others and still work.

Austin: Nice. Right. So this is the unicorn that everyone’s always talked about: write it once and run it everywhere. But when you peel it back, there are always trade-offs, you’re writing at some high-level abstraction, and then it’s trying to translate it down to something that runs, and usually what happens is things are unoptimized. There’s a translation layer higher up, and it’s unoptimized. So the natural question is, is this different?

This was a very surprising and strategically good decision by Qualcomm, because Modular is started by Chris Lattner, he’s one of the co-founders, and he’s a legend in the space of compilers and programming languages. In his PhD, I think he created the Clang compiler and LLVM. At Apple, he invented the Swift programming language. And he also has history with MLIR. At the end of the day, what Chris and team did is make it so you can build in, I think if you use their language, it’s like a superset of Python, but they have the hooks to go all the way down to MLIR, multi-level intermediate representation, to go way down the stack and hook in there.

So at the highest conceptual level, I’d think of it as: other approaches would say, yeah, write using our software and it’ll run on Nvidia’s and ours, but they’d have to do some translation at a much higher level, and it was unoptimized. Lattner’s team is able to go much further down the stack. So in the data center, this would be great, because you could say, hey, use Mojo, use Modular’s tools, write it for Qualcomm, and it will also run on Nvidia. Chris has said before that they can actually squeeze more performance out of Nvidia’s chips than Nvidia. So it’s like, we can also make sure you’re getting as much as possible out of your existing Nvidia infrastructure, potentially even more, that’s compelling.

Another really compelling angle is just Qualcomm infrastructure: what you deploy in the cloud on Qualcomm hardware or someone else’s hardware, you could also deploy on premises, on your desk, on your phone. That’s interesting too. It’s not just cloud multi-vendor silicon you could deploy across, in theory, you could easily deploy from cloud all the way to edge. Nvidia’s doing that too, with their laptop and stuff they’ve come out with recently, where you can write CUDA and it can run in the cloud, or on the DGX little box on your desk, or even on your laptop. But that’s still a proprietary environment. If you use Mojo and Modular, you’d get all those same benefits, but in a true open, modular, hence the name, environment.

Vik: Okay, so that’s good. We should move on and quickly mention what they called their CPUs for data centers, the C1000. It’s supposedly 5 GHz per core and 250-plus core count, runs a lot of PCIe Gen 7 with LPDDR and all of this. I think they have three different versions: one for agentic CPUs, one for general purpose, and one for AI head nodes, which is nice. And it looks like they had Mark Zuckerberg come in and do a video clip saying Meta is planning to deploy the C1000 in the data center, and that they have a multi-generational agreement to supply to Meta. So they already have a business customer there in Meta.

But the one question from the audience at the end on this topic was really funny. That analyst asks, hey, your CPU is for 2028 or something, it’s not this year. So the analyst asks, you say 5 GHz, but is your CPU any good in 2028? Now, everybody knows there’s an agentic AI CPU shortage now, but you don’t have a product for two years, what’s with that? And for that, the Alphawave CEO, Tony Pialis, right, his name was?

Austin: Yes, yeah.

Vik: Yeah, he was like, oh yeah, Qualcomm has the best engineers, and when they design stuff from the ground up, it doesn’t matter when it hits the market, it’s still going to be the best. I’m like, dude, I thought Amon was the bullish Qualcomm guy, which makes sense, because he’s been a Qualcomm guy since he was an engineer and now he’s CEO. But Tony Pialis showed up six months ago with the Alphawave acquisition, and I’m like, what’s with the bullishness about CPUs? I can’t be that bullish. I’m going to take that with a grain of salt. It’s good, maybe we still need CPUs. I don’t think we’re anywhere near the peak of what hardware is required for AI, agentic AI, and all that. Maybe CPUs will be required for another five years, because if you see AMD’s forecasts, there’s a massive growth of the CPU industry all the way to 2030, 2031. So in a sense, yeah, I buy that Qualcomm is not really late to the CPU game. But it’s not that easy to just say, yeah, we’re going to be the best. It’ll be one of the CPUs. And if they can provide it at scale and volume, with sufficient supply chain capacity, then people will buy it. So that’s the level take on it.

Austin: Yeah, to your point, there is a wave happening right now, and they’re missing that wave. It’s just like, we need all the CPUs we can get, they’re missing it. So what happens when they come back in? Well, if CPU capacity is a lot higher, then you’re going to have to compete more on performance or power or cost. So it’s good, if you know you’re going to be late, make sure you feel very confident that you know where it slots in. Which, by the way, I appreciate them calling out the difference in the type of workload and requirements you need from the head node versus general purpose, which is just running more of your SQL databases, because now your agents are pummeling them, versus the agentic rack running a lot of VMs with little agents spinning up tools, some making web requests, some compiling code. So I appreciate that they spelled that out. But if they’re going to compete at that agentic layer, they need to think really hard about the specs they want and make sure they get it right. It’s too bad they couldn’t have brought it in a year earlier, but it is what it is.

Edge AI and Robotics

Vik: There are reasons for that, which we won’t get into. But I want to quickly mention their other big angle, which is the edge AI thing, because Qualcomm is one of the biggest proponents of edge AI, other than maybe NXP, since it’s very important to their whole IoT and automotive business, which they’ve been doing for the longest time. They expect that what they call the software-defined vehicle is now going to be called the AI-defined vehicle. They gave a really interesting example: a car drives into a parking lot, sees a QR code, looks at it, identifies what it is, and then pays for it. You don’t get parking tickets, if time runs out, it can renew it, or get a notification on your phone, and take care of it all by itself. I thought that was a nice example, and apparently that’s already in deployment.

So this is one use case they’re looking at, and they want to couple these high bandwidth compute units we spoke about into automotive SoCs, which means you can put AI into a car so it can think for itself and look at QR codes and identify this and that.

And they want to put vision as a primary driver for their whole edge inference play, because they say vision is the biggest way to get information, if humans can do most things with vision, AI should be able to. So they’re heavily centered on vision as a major unlock for industrial AI.

All of this was really cool, because they’re talking about converting cars into basically token generators. If you have this high bandwidth compute unit in a car, it’s actually using local AI inference, connectivity may exist, yes, but you can use local AI to do a lot of stuff, and it’s really fun. I think this concept is really nice. I don’t know whether it’ll materialize, but it’s fun to hear about all these applications.

Austin: It’s how do they. And the nuance I’ll add is that many cars are already token generators, but they’re captive. It’s for ADAS and autonomy. So your Rivian, your Tesla, your whoever, you’re running end-to-end neural nets these days on potentially Nvidia chips, or custom chips in the case of Rivian. But it’s captive, that is for essentially driving the car.

So what’s cool is, if you can enable the cockpit, the user experience, the user interface, to have AI and enough compute headroom to go experiment and do very interesting things, whether it’s just, hey, next song please, and you don’t have to take your hands off and it actually works, or the car itself doing interesting things for you.

Vik: Yeah. A lot of these edge IoT things are really fun, because, if you remember, they recently acquired Arduino too, and they have the Dragonwing platform, which I believe they’re launching later this summer. So what you can actually do is buy an AI card off of Amazon and use that to run Claude locally.

Those are cool. I don’t know what kind of model sizes, but they said you can run Claude Code locally, that’s what Qualcomm said in their investor day presentation. So I assume it’s a capable model. But I think it’s nice. They showed all these applications in retail, where you can stock shelves and it automatically restocks, and you have all this intelligence within the department store. And how they use it in energy, oil, gas, everything.

But their long-term play at this moment is basically robotics, because that’s what they view as their absolute future opportunity, which they mentioned to be a $1 trillion opportunity, in 2040, though, if you see it.

Austin: Always 2040. Yeah.

Vik: It’s 2040.

Austin: So it’s like, okay, whatever, we’ll retire by then. But what do you need to do interesting robotics? You need a CPU that’s constrained by power, because it’s probably battery powered. You need accelerators, you need memory. And one of the things we don’t have today is cheap enough, high-capacity, high-bandwidth enough memory at the edge.

So we have to run these crappy little models with very little context. If HBM or other innovations can bring us much more memory and much more bandwidth, and still do it in a way that makes sense from a power-envelope perspective, that’s really compelling. What if today’s frontier-level model, not even Fable... Fable, I miss you, come back, give me a call, but, you know, Opus-level, three or four years from now, if you could get that at the edge and fit it in, I think the opportunities are very compelling.

Oh yeah, and then what else do you need? You need connectivity. So they kind of have the whole portfolio: CPU, accelerator, some memory innovations, connectivity. It’s very promising for Qualcomm. Oh, and you need the software stack, they’re acquiring that with Mojo, Modular. Wouldn’t it be interesting if they acquired it for the data center, but actually the best place for it to capture the most value was robotics, seven years from now?

Vik: That’s my whole outlook. I think we’re coming up on time anyway. So I just wanted to mention that my whole outlook on what the Qualcomm investor day came away with: their data center entry point is a little bit late, yes, but we are nowhere near the peak of what compute and AI is anyway. So maybe, like you say, it’s just a blip in the long time horizon here.

But that’s a short-to-medium-term play. I think if edge AI really becomes something widely adopted, Qualcomm has a very, very strong portfolio, and they could come up and dominate edge AI like they did communications for a good decade. So I think they have a very strong edge AI play, but whether that’s going to come next year or whatever is the question.

Austin: It’s going to take a while.

MOAR Memory

Vik: So I remember you mentioned in the beginning that you got to ask Cristiano a question. I don’t know if it was on the record or off the record, but it doesn’t matter, now I’m curious, what did you ask him?

Austin: Yes. So it was not on the recording. I was in a room with Akash, the CFO, and Cristiano, the CEO. Industry analysts had a chance to ask questions, and this one was on the record, so I can talk about it. Sometimes it’s on background, so you can’t talk about it.

I basically said, hey, you introduced HBC, which needs less HBM but more DDR. DDR uses fewer memory wafers than HBM. So on the one hand, that could shift the mix from HBM back into DDR’s favor. On the other hand, this actually opens up the opportunity to bring much more memory to the edge. So, Cristiano, from your perspective, how should we think about the memory market in 2028, 2029, and 2030, when more supply is coming on, but new innovations like this are going to shift the mix, but also just grow the pie, because the edge is going to need more? How should we think about it?

And he was like, that was very many questions, that was a big question, but it’s very good and very interesting. I think he was basically like, HBM’s not going anywhere. Don’t think HBM is dead, we’re going to need HBM. But he said what’s interesting is the amount of increased demand there will be from on-premise infrastructure, from the workstation on your desk to your laptop to your car, for LPDDR and more memory. So he was also giving the MOAR memory vibes, more memory.

Vik: Amazing. And in general, was everybody really excited about all the Qualcomm announcements, from what you could see around you?

Austin: Oh yeah. I think all the other analysts in the room were excited. Investors were excited. There are always the technical questions, some people are naturally more skeptical than others, so it’s like, okay, you introduced a lot, you talked about it, we want to see the technical proof. And some of that is just, you have to know the context of this being an investor day. What are we going to talk about, how in the weeds are we going to get? And these things take time. So I think you can give Qualcomm the benefit of the doubt, but they do eventually need to follow up with technical details, especially if stuff’s not shipping for a few years. The best thing they could do is write a paper, give a presentation, whatever, to prove that at least in the lab, it’s real and it works.

Vik: Yeah. So we still need more memory. Good to know.

Austin: Yes, totally. Okay, maybe here’s one last bonus thought, we won’t even talk about it, we’ll talk about it more some other time. If HBC can bring much more memory close to the accelerator, let’s not call it HBC, let’s just say, industry-wide, near-memory compute, what are the implications on the memory market? Obviously we’re still going to want HBM, but now all of a sudden, maybe even more DRAM. And not even from a data center perspective, but when we’re talking about bringing more data to the edge, robotics, cars, and stuff, it just feels like even more wafers are going to be needed.

Vik: More memory. MOAR. Yes.

Austin: MOAR. Exactly. Don’t let anyone DeepSeek you and tell you, oh, HBM’s dead because near-memory compute. No, MOAR, more memory.

Vik: All right, we should call it with that. Yeah.

Austin: Yeah, let’s call it there. Everyone, thanks for watching. We appreciate you curious, intelligent, interesting people who love semiconductors and listen all the way to the end. Check us out on YouTube, leave comments, check us out on Spotify, send us emails, whatever. We have a daily, if you like our takes, you can get them daily for free at semidope.com. Check it out, and thanks for listening.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Apple Seeks US Waiver to Buy Chips From Blacklisted CXMT

Apple has applied to the US government for permission to purchase memory chips from ChangXin Memory Technologies (CXMT), China’s largest DRAM maker, according to the Financial Times. CXMT was added to the US Entity List, which bars American companies from transacting with it without an explicit government waiver. The move marks a rare instance of a major US consumer electronics company seeking to source from a blacklisted Chinese chipmaker.

Sources: Bloomberg Tech

Vik: Memory makes companies do strange things. This week Apple raised prices on its products, and in what may be the first time for many outside the semiconductor sphere, people realized that memory is now a global problem. For the last two years, folks in the semi space have seen the rising demand for memory. Now the pot has boiled over.

For context, CXMT is in the Pentagon’s 1260H list of companies with alleged connections to the Chinese military. As a result, this company is on the Entity list which bars US companies from sourcing from them.

Many view this move as a publicity stunt, where Apple allegedly leaked this information to push the US government to open up trade with CXMT by taking them off the Entity list. The markets viewed Apple’s price increases as a sign that demand will finally show and start pulling back.

Austin: Interesting to see how AI is directly causing inflation.

“Inflation measures how much more expensive a set of goods and services has become over a certain period, usually a year” — IMF

Agility Robotics to go public via SPAC at $2.5 billion valuation

Agility Robotics will merge with Churchill Capital Corp XI, a SPAC, in a deal valuing Agility at $2.5 billion pre-money equity value. The transaction is expected to generate over $620 million in gross proceeds, including a $200 million PIPE at $10 per share. Agility has secured over $300 million in multi-year orders for its Digit v5 humanoid robot. (agilityrobotics.com)

Austin: Commercial deployments of humanoids, going public. Check out the video at Agility’s website.

OpenAI restricts GPT-5.6 access, citing Trump administration security concerns

OpenAI is limiting access to its new GPT-5.6 models after discussions with the Trump administration, making them initially available to a small, government-approved customer group. This follows a similar intervention with Anthropic’s Fable 5 and Mythos 5 models due to their ability to discover software vulnerabilities. OpenAI hopes for general availability in weeks, viewing the current approval process as a temporary measure during the implementation of President Trump’s executive order on model oversight. (wsj.com)

Austin: Hey, look, we should all be outside splashing in the summer sun. That’s what I keep telling myself after they took my Fable away…

Will be interesting to see if the government’s export controls delay Anthropic and OpenAI IPOs, and if that ends up actually being a good thing. Maybe you can have too many massive IPOs in too short of a time right? (Not looking at you Cerebras, I’m talking about SpaceX).

Samsung, SK Hynix Pledge $880 Billion for Two New Chip Fabs

Samsung Electronics and SK Hynix have committed a combined $880 billion to build two new chip fabrication facilities in South Korea, the companies announced. The investment represents record capital spending for the two Korean chipmakers, which together account for the majority of global memory chip production. The pledge covers construction of the new fabs as part of an effort to expand domestic semiconductor manufacturing capacity.

Sources: Bloomberg Tech

Vik: This is Korea’s way of retaining memory dominance. This represents 5% of Korea’s 2024 GDP, just to make memory. That’s how important it is. President calls SK Hynix and Samsung “national heroes” and says “Speed is the only way to survive.”

ShunSin Confirms TSMC COUPE Optical Packaging Partnership

Foxconn subsidiary ShunSin has confirmed a deal with TSMC under the COUPE (Co-Packaged Optics Using Photonics Ecosystem) program to provide co-packaged optics (CPO) and optical circuit switching (OCS) packaging services. ShunSin plans to commit NT$5 billion in capital expenditure to support the arrangement.

Sources: digitimes

Vik: Shunsun (6451.TW) — according dnystedt on X, sees double digit revenue growth in 2026, and more in 2027. 51.2T CPO currently in production, 102.4T CPO is sampling, and 12.8T NPO / OCS in development.

DeepSeek Cuts AI Inference Latency 85% With New DSpark Framework

DeepSeek has released DSpark, a speculative decoding framework for its V4 model that increases per-user response speeds by up to 85%, reducing the load on inference chips. The upgrade targets serving costs and user experience as Chinese AI developers compete on operational efficiency rather than model capability alone.

Sources: SCMP

Vik: Speculative decoding is a technique where there is a draft model that predicts tokens and the verifier model confirms that they are correct. This is a much faster way of token generation, and its usefulness depends on the accuracy of the draft model, and the acceptance rate of the verifier. This is not really a new technique, but good to see it in DeepSeek V4.

Austin: d-Matrix and Gimlet have a nice demo showing how speculative decoding on right-sized hardware can further improve the Pareto frontier.

Key Data

A large part of our reader/listener base on Semi Doped are professional investors, and on Austin’s and Vik’s substack too! We think this chart might be interesting. (h/t to The Kobeissi Letter on X)

Sector Watch

Memory

• SK Hynix listing ADRs on Nasdaq to raise $29B, providing liquidity for its $194B ten-year capex program targeting HBM4 and advanced packaging supply constraints. (chosun.com)

• CXMT projected to generate over $130B in gross profit over next 18 months with nearly all capital flowing back into capacity expansion, targeting 600k wafers-per-month by 2028. (substack.com) (DigiTimes)

• Apple raised Mac and iPad prices citing unprecedented 40-year memory cost spike driven by AI infrastructure buildout, with CEO Tim Cook noting memory quotes nearing historic highs through 2027. (Bloomberg Tech)

Foundry & Logic

• TSMC A14 node driving demand for Vera Rubin architectures with performance gains exceeding 20% while attracting orders from Nvidia and Apple, intensifying capacity constraints. (substack.com)

• Intel promised SpaceX and Apple a toolkit this fall to test its 14A node before final commitments, a critical validation step for Intel Foundry’s process between 18A and next-gen nodes. (TechMeme)

• Apple public roadmap for A-series SoC on 1.4nm node in 2027 confirms it will remain critical lighthouse customer for TSMC’s process technology leadership beyond immediate AI training boom. (DigiTimes)

Advanced Packaging

• Applied Materials unveiled six new tools targeting epitaxy, CMP, deposition, and eBeam metrology to address yield bottlenecks in HBM and 3D stacking for AI accelerators. (Semiconductor Digest)

• LG Chem planning to expand Copper Core Laminate (CCL) capacity to address supply constraints driven by AI chip demand, highlighting tightening bottleneck in backend packaging supply chain. (DigiTimes)

Earnings & Capital

• U.S. equity indices fell across the board this week with S&P 500 and Nasdaq declining in every session, driven by mounting anxiety over sustainability of AI compute buildout and unit economics. (chosun.com)

• Apple skipping high-end M6 Pro and Max chips, redirecting roadmap to launch base M6 this year followed by AI-focused M7 and M7 Pro/Max chips in 2027. (DigiTimes) (Toms Hardware)

When its foundry executives marked that nanosheet breakthrough, they lined up for the cameras holding three fingers aloft, one for each nanometer. Their founder might have approved of the count: the single word he picked for the company in 1938 sets three points of light in the night sky.

This tech powerhouse also has stakes in consumer electronics, insurance, food processing, and even led the joint venture that built the world's tallest skyscraper in Dubai.

Name the company.

CLUES

Founder Lee Byung-chul chose the name in 1938.

Its foundry built Apple’s A-series chips during the Galaxy-versus-iPhone wars

It is the largest maker of both DRAM and NAND flash.

It’s one of the family-run conglomerates, the chaebol, that dominate a single East Asian economy. Narrow it to the Korean ones and the field is already down to a handful.

SPOILER ALERT: ANSWER BELOW

When engineers in the early 1960s cracked open the sealed chips and looked through a microscope, the tiny wires inside had not stayed silver and gold. The contact points had turned a deep, sickly violet. The industry called it the “purple plague,” and for a few years it was one of the most feared failures in electronics.

The setup looked innocent. To wire up a chip you bonded a hair-thin gold wire onto an aluminum pad on the silicon. On the bench the joints were perfect, the signals crisp, the chips passing every test. Then they ran hot - in a missile, in a spacecraft, in any sealed box that warmed up in use - and something quiet and molecular went wrong. Gold and aluminum are bad roommates. Heat them together and they diffuse into one another, forming a family of brittle gold-aluminum compounds. One of them, gold aluminide (AuAl₂), happens to be a vivid purple. That purple was what engineers saw when they opened a dead chip, so the purple got the name, and the blame.

Here is the part the name gets wrong. The purple compound was mostly a bystander. AuAl₂ is so stable that jewelers still sell it as “purple gold,” and it turns up in bonds that stay perfectly strong. This is because gold diffuses into the aluminum faster than the aluminum moves back, the reaction leaves rows of microscopic empty pockets along the joint — a hollowing-out called Kirkendall voiding. Leftover manufacturing contaminants like fluorine and chlorine made it worse, drilling their own voids into the interface. The bond slowly turned into a purple honeycomb, and the first real vibration — a truck ride, a test firing — snapped it. The color was just the witness standing over the body.

The timing could not have been worse, because the plague hit the two most important chip projects in the country at once. The Apollo Guidance Computer, the machine meant to steer astronauts to the moon, ran entirely on early Fairchild integrated circuits; its lead hardware designer, Eldon Hall, later counted the purple plague among the team’s major problems, a metal that turned brittle and cracked. At the same time the Air Force’s Minuteman II missile was being built around custom Texas Instruments chips. A single snapped wire meant a dead guidance computer, whether it was bound for the Sea of Tranquility or sitting in a silo in North Dakota.

So the industry treated metallurgy with the same fear it had reserved for physics. Engineers deliberately baked wafers in hot ovens to force the reaction and find the weak joints early. They screened by the lot, and if one chip failed, the whole batch was thrown out. They scrubbed contamination out of their fabrication lines, and they redesigned the joint itself, doping the gold wire, switching to aluminum-on-aluminum, or slipping a barrier metal between the two so gold and aluminum never touched.

It is a strange thing to find at the foundation of the digital age: not a theory, not a transistor, but two common metals that could not be trusted to sit quietly together in the dark. And the most notorious villain in early chip history turned out to be named after a color that was mostly just standing at the scene.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player.

Marvell Explores Plasmonics for Higher Bandwidth Optical AI Networks

Marvell is researching plasmonics to enhance optical networks for AI infrastructure. Plasmonic-based silicon photonic (SiPho) light engines could enable 3.2T modules and beyond, reducing space and power per bit compared to existing technologies. Marvell has developed plasmonic modulator prototypes that are approximately 10 microns in length, 300x to 500x shorter than existing SiPho devices, and operate at 1THz, over 10x faster. (marvell.com)

Vik: This is from their acquisition of Polariton I am assuming. Plasmonics involves sandwiching an eletro-optic material between two metal electrodes and applying voltage to the electrodes based on data. The voltage interacts with light in the electro-optic material to cause surface plasma polaritons, creating high speed, tiny, high effective phase shift modulators. Watch out for these things in the 3.2T generation.

Austin: Well done Marvell social media team. These images are eye candy.

Anthropic Joins RAISE US Workforce Coalition as Founding Partner

Anthropic joined RAISE US, a new nonprofit coalition, as a founding partner. RAISE US works to strengthen the American workforce through employer-led action, AI-enabled training, and policy innovation to support the transition to transformative AI. (@AnthropicAI)

Austin: Making America’s workforce AI ready is a very important goal. See Jeff Ding’s book “Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition” which ultimately argues that the diffusion of technology is more important to a nation than whether it invented the technology or not.

Princeton professor Helen Milner put it well:

“Ding’s Technology and the Rise of Great Powers provides a powerful new argument about how technology can change world politics. He shows that creating an infrastructure of education and training systems for developing new general-purpose technology skills is critical to global leadership. Power transitions in global politics may depend more on ordinary engineers than on heroic inventors.”

EU Joins US-Led Pax Silica Chip Alliance, Now 24 Nations

The EU, Netherlands, Germany, and Greece joined Pax Silica, an American-led initiative to strengthen AI-related tech supply chains. Argentina, Chile, Costa Rica, Kazakhstan, and Panama are also joining this week, bringing the total to 24 countries. Jacob Helberg, Under-Secretary of State for Economic Affairs, stated Pax Silica aims to promote AI innovation and create a US alternative to initiatives emphasizing “digital sovereignty.” (ft.com, state.gov)

Austin: ASML is in the Netherlands. I wonder how they truly feel about this. I saw a quote from Reuters that it’s not all hunky dory:

“Speaking to reporters on Tuesday in Washington, Dutch ​Trade Minister Sjoerd Sjoerdsma said the U.S. and the Netherlands share common goals over preventing sensitive ​technology from ending up in dangerous hands.

However, “elements in that Act seem to suggest that the United States might take control over some of these decisions that affect our national security and the way our companies ​operate,” he said.”

SambaNova raising at $10B valuation

AI chip startup SambaNova aims to raise between $800 million and $1 billion, potentially valuing the company at $10 billion, a five-fold increase from its last funding round four months prior. General Atlantic is reportedly in talks to lead the round. Intel, a current customer and 9% stakeholder, previously considered acquiring SambaNova for $1.6 billion. (theinformation.com)

Vik: SambaNova has a useful approach to inference that focuses on low TCO by using reconfigurable data units (RDUs) to chart their way through multiple tiers on memory in a predetermined fashion. I have a whole article on this if you want a deep dive. Its good to see them raising a new round of funding.

Samsung, SK Hynix Ready Large AI Spending Commitments

Samsung Electronics is preparing a $648 billion investment plan, according to reports, as South Korea’s two largest memory chipmakers ready significant AI-driven capital commitments. SK Hynix is also preparing a major spending push. Both companies are responding to accelerating demand for AI infrastructure components, with the buildout expected to reshape South Korea’s broader technology sector. (Bloomberg Tech, Reuters)

Vik: The memory trade does not seem to be going anywhere but up. After Qualcomm’s investor day yesterday, its clear that memory needs will only increase when edge AI is here for good. Not financial advice, but the demand for memory is only increasing. Real question: for how long?

Austin: From Bloomberg, this caught my eye:

President Lee Jae Myung is set to host a “National Briefing on the Three Major Mega-Projects for the Great Leap Forward of the Republic of Korea” on Monday, a senior Blue House spokesperson said Thursday, adding that further details will be provided that day.

The Great Leap Forward! It kind of reminds me of the space race of the 60s and how much the general public paid attention to deep tech R&D and industrial mobilization.

Onsemi Acquires Synaptics to Expand Into Intelligent Systems

Onsemi has agreed to acquire Synaptics in a deal that expands its addressable market by $30 billion, bringing the total to $243 billion by 2030. The acquisition adds Synaptics’ edge AI and human-machine interface capabilities to onsemi’s existing power and sensing portfolio, positioning the combined company at the intersection of AI data centers and physical AI applications. (onsemi)

Vik: Power, analog, and sensing will become increasingly important themes as AI proliferates outside the datacenter eventually. In my opinion, the ChatGPT moment for robotics is not yet here. It will become apparent when robots can solve real human problems like laundry and dishes, outside of industrial automation. Current dancing robots don’t make the cut. Sorry.

Austin: There was an M&A investor update. I need to listen to that. Seems maybe analog companies are thinking they need to acquire compute and software to have a foothold in robotics.

MKS Instruments Invests $25M to Expand Guangzhou Manufacturing

MKS Instruments (NASDAQ: MKSI) is investing $25 million to expand its Atotech equipment manufacturing site in Guangzhou, China, adding capacity aimed at supporting AI infrastructure buildout. The Andover, Mass.-based company said the expansion will increase production capability at the facility, which makes equipment used in semiconductor and advanced packaging manufacturing processes. (MKS Instruments)

Vik: This seemingly innocuous piece of news is actually a panel-level packaging play, for things like glass substrate based advanced packaging — think CoPoS. MKSI-Atotech have tech that can form high aspect ratio vias in glass in panel level formats.

Key Data

Newly released ChatGPT 5.6 Sol beats Claude Mythos 5. May not get banned though. Let’s see.

Quick Hits

• ASE Group projects AI-driven packaging demand will constrain capacity through 2030, extending previous consensus and confirming advanced packaging as binding bottleneck over transistor scaling. (DigiTimes)

• Arm architecture now powers over 50% of hyperscale cloud computing capacity driven by AI workload demand, confirming structural shift from x86 to Arm-based custom silicon. (TechMeme)

• Bosch pushes third-generation SiC MOSFETs with 20% lower resistance and 10% lower switching losses for EV drivetrains as high-voltage infrastructure converges across AI and e-mobility. (EE Power)

By the Numbers

Closing moves, 2026-06-25:

• Up: SNDK +22.0%, MU +15.7%, AMAT +13.4%, KLAC +7.6%, LRCX +7.2%, WDC +4.9%

• Down: AAPL -6.1%, DELL -5.7%, PLTR -5.5%, IREN -5.1%, HPE -4.2%, MSFT -3.5%

• Extremes: SNDK at a 30-day high; MU at a 30-day high; AMAT at a 30-day high; AAPL at a 30-day low; PLTR at a 30-day low

  Thanks for reading! This post is public so feel free to share it.

  Share

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!

Qualcomm unveils datacenter CPU, AI accelerator, and custom silicon roadmap

Qualcomm announced its re-entry into the datacenter market at its 2026 Investor Day, detailing a new CPU, AI accelerator, and custom silicon offerings. The company introduced the Qualcomm Dragonfly C1000 CPU, a chiplet design with over 250 cores, PCIe Gen7, CXL, and optional High Bandwidth Compute (HBC) attach. Qualcomm also acquired Modular to bolster its AI software stack and announced a multi-generational agreement with Meta for its processors. (servethehome.com)

Vik: Qualcomm is actually a good AI CPU play because they likely have a ton of capacity from TSMC due to mobile chips allocation (if its even fungible). But hearing that they are “re-entering” the datacenter space (again) is like nails on a chalkboard. Are they going to stay in it this time?

Austin: I was there. A lot to digest. We’ll talk about it on an upcoming pod.

The Self-Driving Future of Chip Design

Intel’s CTO Pushkar Ranade out with a long essay on X on AI-driven chip design.

“Four years ago, the idea that models could write code, and that code could design chips, was a provocative thesis. I first wrote about it eight months before the emergence of ChatGPT.

This idea is no longer a thesis. It is now practice.

Large language models now author Register-Transfer-Level (RTL) code; Reinforcement Learning (RL) agents now place macros on silicon that is commercially shipping; commercial Electronic Design Automation (EDA) suites can now embed generative copilots across the flow…

This essay traces the state of the art in AI driven design, examines where human authority still holds (in intent, sign off and accountability) and why that ground is shrinking, and offers a forward-looking sketch of an eventual state in which humans exit the design loop entirely.”

(@magicsilicon)

Vik: The wonders of the modern era is not AI, its that you can hear from thought leaders like Pushkar directly.

Austin: I love how Pushkar goes direct and writes on X and Substack.

XLight, chaired by ex-Intel CEO, seeks $350M for chip lasers

XLight, a startup developing advanced lasers for semiconductor manufacturing, is in talks to raise $350 million from investment firms. This follows a $150 million investment from the U.S. Department of Commerce via the CHIPS and Science Act. Former Intel CEO Pat Gelsinger is the executive chair of XLight, which aims to develop an alternative to extreme ultraviolet lithography to reduce AI server chip manufacturing costs. (theinformation.com)

Vik: Their photons-as-a-service idea is fascinating. We need to continue to innovate on lithography.

Austin: xLight is a very interesting company with a pragmatic approach to an important problem. I wrote about it at Chipstrat, check it out if you haven’t:

Chipstrat

xLight. Photons as a Service.

In January, I wrote about the worsening cost curve of EUV lithography and two startups trying to bend it…

Read more

4 months ago · 16 likes · Austin Lyons

Samsung details 3D Stacked FET with 42nm gate pitch

Samsung Electronics’ Semiconductor Research Center presented a paper at the 2026 VLSI Symposium detailing its 3D Stacked FET technology. The research, which received a Best Paper award, describes a 3D Stacked FET with a 42 nm gate pitch featuring triple-stacked nanosheet channels. This approach vertically stacks n-type and p-type transistors, enabling higher transistor density within the same footprint. (semiconductor.samsung.com)

Vik: I saw TMSC’s CFET topology at IEDM last year, and they even demonstrated a ring oscillator with their stacked transistors. A key figure at the company told me that we have 20+ years of innovations in transistors still to come. Lots of excitement up ahead.

Unitree Robotics Teases New Robot, Details Forthcoming

Unitree Robotics posted a short video clip on X teasing a new robot. The company stated that more details about the robot are coming soon. (@unitreerobotics)

Vik: China is really quite ahead on robotics, but their marketing team needs work.

OpenAI unveils custom Jalapeño AI chip with Broadcom

OpenAI has designed and built its first AI chip, “Jalapeño,” in collaboration with Broadcom. The chip is purpose-built for large language model workloads powering ChatGPT, Codex, the API, and future agentic products. OpenAI states that building its own chips expands its full-stack platform from products to models to infrastructure.

Sources: @openai

Austin: We have to get OpenAI’s head of hardware Richard Ho to come chat sometime! And excerpt from his thoughts on Jalapeño via LinkedIn:

”Jalapeño is OpenAI’s first Intelligence Processor. We designed it from scratch around the kernels, memory movement, networking, scheduling, serving patterns, and latency requirements that matter for LLM inference. It is informed by the systems that serve ChatGPT, Codex, the API, and future agentic products, and built to remain flexible for current and future LLMs across the industry.

For this kind of system, peak performance is only one part of the problem. The harder and more important question is how much useful work you can get from the hardware in production. Jalapeño is designed to reduce data movement and balance compute, memory, and networking so our most important workloads can run much closer to the hardware’s theoretical limits.”

Qualcomm to Acquire Modular for AI Software Foundation

Qualcomm announced an agreement to acquire Modular Inc. to strengthen Qualcomm Technologies’ software foundation for generative and agentic AI across data center and edge environments. Modular provides an AI-native software stack that enables AI to run efficiently across hardware architectures, including CPU, GPU, NPU, and custom ASIC architectures. The acquisition is expected to close in the second half of 2026, subject to customary closing conditions and regulatory approvals.

Sources: modular.com

Austin: “Write once, run anywhere” can be a pipe dream, but Modular has made it real. There’s the obvious heterogeneous data center application here (e.g. run a single codebase across a datacenter with both QCOM accelerators and existing Nvidia chips) but there’s also a very interesting “deploy inference from edge to on-prem AI to cloud” angle too

Key Data

Nice map of ASMPT CoWoW solutions. h/t to on X.

Sector Watch

• Qualcomm and ByteDance in talks for custom chip design services for TikTok operations, highlighting China’s shift toward hyperscaler-owned ASIC capabilities amid export controls. (DigiTimes)

• Samsung and SK Hynix report profit surges driven by hyperscaler demand for HBM, as contract structures shift toward direct deals bypassing traditional intermediaries. (chosun.com)

• Apple iPhone 17 architecture teardown by TechInsights reveals engineering trade-offs in custom silicon to support next-generation mobile AI workloads. (techinsights.com)

• Blue-X exploring sub-13.5nm lithography as alternative to EUV systems, potentially bypassing ASML’s monopoly if yield parity is achieved. (bits-chips.com)

Press Releases

• Qualcomm accelerates diversification with a comprehensive strategy for the data center, anticipating multiple inflection points over the next 3 to 5 years. (Qualcomm IR, June 23)

• Micron Technology, Inc. reports record results for the third quarter of fiscal 2026. (Micron IR, June 24)

Here are our quick briefing of the news — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast playe

Cerebras 2026 Sales Forecast Disappoints AI Market Expectations

Cerebras projects 2026 revenue between $855 million and $865 million, exceeding analysts’ average estimate of $824.8 million. Despite this, shares fell approximately 10% in late trading as investors anticipated a larger market share in AI data centers. The company reported first-quarter sales of $193.4 million, a 94% increase, with a net loss of $14 million. (bloomberg.com)

Vik: Investors have high expectations from semiconductors these days. Just exceeding analyst estimates seems to disappoint everybody. Nevertheless, there are issues with Cerebras at scale, and I’ve written about it. In any case, I think my tweet below sums up some of it.

posts his observations from the earnings call transcript.

Irrational Analysis

Cerebras June 2026 Earnings

Irrational Analysis is heavily invested in the semiconductor industry…

Read more

14 days ago · 29 likes · 6 comments · Irrational Analysis

AI Customers Cut Bills, Shift to Cheaper Models

Large AI customers are reducing spending on Anthropic and OpenAI by using less expensive models, including open-source options. Ensemble Health Partners expects to save nearly $700,000 annually by switching to a lower-cost OpenAI model. Open-source model usage surged to 65% of OpenRouter tokens processed in June, up from 34% in January. (theinformation.com)

Vik: If this report is true, then we can expect the crazy ARR growth of Anthropic to slow down, and see inference shift more to open source models where cost per unit intelligence is much more aligned to actually be useful. Frontier models are not needed for everything. Several chip design folks I know say Claude Sonnet does the job well enough. No need for Opus. That tells you something. Watch closely.

Austin: Although a few folks have “reduced” current spend on frontier, I tend to think frontier continues to grow in aggregate for frontier tasks, especially agentic coding. But open-source (+fine-tuning) will grow a ton too as companies deploy LLM- and need to keep an eye on costs.

SK Hynix files $29B US ADR listing

SK Hynix said it will raise up to $29 billion through a US listing of American depositary receipts (ADR), with proceeds earmarked for AI-related investment. This filing would rank among the largest ADR offerings by an Asian issuer, and is expected on July 10th. (Bloomberg Tech)

Vik: This makes it easier for US investors to get in on the SK Hynix action, although retail investors have probably been buy in via Interactive Brokers or similar.

Samsung unveils first UFS 5.0 mobile storage

Samsung Electronics announced on June 23 the development of the industry’s first Universal Flash Storage (UFS) 5.0 product family, designed for on-device artificial intelligence workloads in smartphones and other mobile devices. The new lineup adopts the latest UFS 5.0 standard, which Samsung said is optimized to handle AI processing directly on devices rather than routing requests to the cloud. The South Korean memory maker did not disclose pricing, capacity tiers, or commercial availability for the new product family in its initial announcement. (The Elec)

Vik: NAND is important for AI now, and KV cache storage becomes an issue with just DRAM-based memory. Even on edge devices will require a different class of NAND storage that is suited to AI applications.

LG Chem commits $11B to chip, robotics materials

LG Chem plans to invest 15 trillion won (roughly $11 billion) in research and development through 2035, shifting its focus from conventional petrochemicals toward higher-margin materials for semiconductors, mobility and robotics. President Kim Dong-chun announced the plan at a town hall meeting, framing it as part of the South Korean company’s transformation into a high-value-added materials supplier. The Elec reported the commitment, which spans the next decade of R&D spending. (The Elec)

Vik: Signs of long term investments are a good sign but people from all industries have some kind of AI angle these days which makes supply chain research SO MUCH more interesting!

Key Data

A nice chart from TrendForce.

Quick Hits

• YEST is expected to see earnings growth as high-pressure hydrogen annealing equipment is considered for next-generation 1d-nanometer DRAM production. (The Elec)

• NVIDIA and AWS collaborate to bring AI to production at scale, addressing low-latency inference, vector search and GPU price-performance for enterprise workloads. (Nvidia News)

• Penguin Solutions becomes an NVIDIA AI Factory Specialized Partner, joining the program for delivering full-stack AI infrastructure to enterprise customers. (Penguin Solutions)

• Ciena’s Blue Planet and Telefónica Deutschland demonstrate how AI agents can accelerate 5G network slicing service design across operator infrastructure. (Ciena)

• Kioxia executive pay jumps sharply after AI-driven demand boosted NAND flash sales and the company’s stock price following its Tokyo listing. (Bloomberg.com)

Shorter edition today since we’re both traveling. Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast pl

SpaceX Erases $600 Billion in Market Value

SpaceX shares dropped 23% over three days, erasing over $600 billion in market value. The stock now sits at $154.60, though it remains about 15% above its $135 IPO price. The sell-off followed news that SpaceX plans to raise at least $20 billion in bonds to fund AI investments, including a new computing deal with AI startup Reflection AI. Thin trading volume and heavy retail interest amplified the volatility. Investors are also watching SpaceX as a bellwether for the AI race, ahead of expected IPOs from Anthropic and OpenAI, both targeting ~$1 trillion valuations.

MediaTek Secures Google TPU v9 Orders with 336G SerDes Solution

MediaTek has reportedly secured the primary orders for Google’s TPU v9, codenamed Triggerfish, leveraging its 336G SerDes solution. The enhanced design carries roughly three times the memory of the baseline TPU v9 and represents a continuation of Google’s multi-year custom silicon partnership with MediaTek. This win comes as Broadcom, which focused on 448G SerDes, faced delays due to technical maturity and product timeline pressures. Analysts project MediaTek’s AI accelerator ASIC revenue to reach $2 billion in 2026, with further growth anticipated in 2027.(ctee.com.tw)

Vik: There was quite some discussion on X whether MediaTek SerDes was any good due to the choice of their circuit architectures some time back. Even if its not, 300G/lane or 2.4T networking is going to be a reality for Google. If MediaTek now has TPU v9 volume, then that indicates SerDes is just fine.

Austin: 336G per lane instead of 448G. Interesting. 448G is hard:

Memory Pricing to Surge 40-50% in Q3 2026

Jefferies predicts memory pricing will increase 40-50% quarter-over-quarter in Q3 2026 and another 30-40% in Q4 2026, with continued price hikes into 2027 due to zero wafer capacity growth. Cloud Service Providers are securing 50% of total capacity through two-year long-term agreements requiring 40% prepayment, leaving consumer electronics players facing severe supply pressure. Samsung is expected to gain a competitive edge in HBM4 with its 4nm base die and hybrid bonding transition. (@pequityresearch)

Austin: Long-term agreements and 40% prepayment 👀

Qualcomm Nears Deal for AI Chip Startup Modular

Qualcomm is close to acquiring Modular, an AI chip software startup, according to a Bloomberg News report cited by Reuters. The transaction values ​the AI chip company at about $4B. Earlier, the Information reported that Qualcomm was in a deal to buy Tenstorrent. (Reuters)

Vik: Why is Qualcomm doing all this M&A stuff? What happened to the inference chip they were building? This week’s investor day will reveal all, we hope.

Austin: We’ll find out tomorrow!

BTW, adding Chris Lattner and Modular would be a slam dunk imo for QCOM.

And previous rumors said Jim Keller’s Tenstorrent is on the table too.

Micron and Anthropic sign AI memory supply pact

Micron and Anthropic announced a strategic agreement on June 22 covering joint design of memory and storage architectures for next-generation AI infrastructure, a long-term supply agreement, and enterprise deployment of Anthropic’s Claude models inside Micron. Micron is also taking part in Anthropic’s latest funding round as a strategic investor. Financial terms and supply volumes were not disclosed. (Micron, EE News Europe, Reuters)

Vik: hit an all time high for a moment there. If a company with a steeply increasing ARR is going down the supply chain trying to secure memory, what’s next?

Is Anthropic going to acquire InP laser supply too? This is going to happen….

I’ve argued before that there is a pecking order here in the form of a token value chain: supplier → provider → consumer. As computing resource becomes scarce for the provider (Anthropic), they will acquire a larger portion of the supplier universe.

Austin: Memory vendor is a strategic investor in major AI lab. Reminiscent of a time when the news was all about an accelerator vendor strategically investing in major AI lab. Is memory is the belle of the ball right now?

Upscale AI raises $190 million, reaches $500 million total funding

Upscale AI, an AI networking infrastructure company, announced it raised $190 million in Series A-1 financing, bringing its total funding to $500 million and its valuation to $2 billion. Premji Invest led the round, with new investors including Nvidia, Salesforce Ventures, Seligman Ventures, and Temasek. The company is engaged with hyperscalers and neocloud infrastructure providers for customer evaluations and deployments. (businesswire.com)

Austin: Upscale’s SkyHammer switch can support UAL, UEC, and SUE. Which enabled them to give a nice talk back at OCP where they tested and discussed the merits of UAL vs UALoE vs SUE:

Microsoft, Chevron ink Texas AI datacenter power deal

Microsoft and Chevron signed a power supply agreement for a new Microsoft AI datacenter in Pecos, Texas, with Chevron supplying electricity from on-site natural gas turbines under a long-term contract spanning roughly 20 years. The Pecos facility is part of Microsoft’s Azure capacity expansion for AI workloads and is being built with a closed-loop cooling design that Microsoft says uses less water than a fast food restaurant. Reuters reported the arrangement gives Microsoft dedicated, behind-the-meter generation in the Permian Basin, bypassing constrained grid interconnection queues in ERCOT. (The Official Microsoft Blog)

Vik: Seems like the comparison between fast food joints and datacenter water usage as perpetuated by SemiAnalysis is catching on. I suppose everyone can relate to burgers. I’ve argued in the past that the comparison is not as simple as that.

US Accuses ASML of EUV Shipments to China

The US Commerce Department accused ASML of shipping EUV-compatible components to Chinese customers, alleging the transfers may have breached export controls on advanced chipmaking gear. The Bureau of Industry and Security is reviewing the evidence and weighing enforcement action against the Dutch lithography maker, which supplies the only extreme ultraviolet scanners capable of producing leading-edge logic and memory chips. ASML rejected the allegation, saying it complies with applicable Dutch and US restrictions and has not delivered full EUV systems to China. The company has been barred from exporting its most advanced EUV tools to Chinese fabs since 2019, with curbs later extended to some immersion DUV systems. (Tech Times, AI Insider, Electronics Weekly, verdict.co.uk)

Nvidia unveils Vera Rubin platform for science

Nvidia introduced its Vera Rubin accelerated computing platform at ISC High Performance 2026, pairing new Vera CPUs with Rubin GPUs in a rack-scale, liquid-cooled architecture that combines double-precision HPC with large-scale AI workloads. Los Alamos National Laboratory will deploy the platform in three supercomputers — Mission, Vision and Veritas — built with Hewlett Packard Enterprise on the HPE Cray Supercomputing GX5000 architecture. Nvidia is positioning Vera Rubin for scientific research, industrial simulation and agentic AI applications, marking a shift toward unified systems that merge traditional HPC with AI acceleration. (Nvidia News, EE News Europe)

Austin: Hmm. Where does the double-precision (FP64) compute come from in these setups? Rubin can support FP64 no doubt, but has been prioritizing lower precision for AI. Are they running like agentic inference on Vera Rubin and doing all the HPC math on the Vera CPUs? Or using Rubin for the HPC oomph too?

Seems AMD’s MI430X would be a good candidate here too.

Key Data

Translated from Chinese using Claude. Errors are Dario’s fault. From Commercial Times.

Source: Commercial Times

Sector Watch

AI & Compute

• SpaceX and Reflection AI sign $150M/month compute deal for Nvidia GB300 systems at Colossus 2 datacenter, payments start July 2026. (TechCrunch)

• Groq confirms $650M raise and re-staffing after Nvidia deal. (TechCrunch)

Memory

• YMTC NAND market share rises to 13%, alarming Samsung, SK hynix, and Micron. (DigiTimes)

• Motherboard ASPs collapse while RAM prices surge as hyperscaler HBM demand crowds out standard memory production. (Tom’s Hardware)

Foundry & Logic

• TSMC cuts 28nm output by over 25% since early 2026 to reallocate resources toward advanced node and packaging capacity. (TrendForce)

• Intel CEO bets on advanced packaging as former SK hynix CEO returns to company. (DigiTimes)

• Samsung System LSI sees ongoing SoC losses drag down performance. (DigiTimes)

Optics & CPO

• MSScorp expands Taiwan investment to scale silicon photonics manufacturing capacity for optical interconnect. (DigiTimes)

Equipment

• Top five Japanese chip toolmakers report 10% China sales drop for FY25, driven by front-end fabrication tool weakness. (TrendForce)

• Nvidia unveils new datacenter water cooling system. (TechCrunch)

Press Releases

• Dell Technologies announced the Dell AI Factory with NVIDIA, advancing supercomputing-class infrastructure for HPC and AI. (Dell Technologies IR, June 22)

• GlobalFoundries qualified SLATE™ advanced packaging technology on its 9SW platform for next-generation radio frequency applications. (GlobalFoundries IR, June 23)

• Supermicro broadened its AI at the Edge Solutions portfolio with Intel-powered platforms optimized for low-latency inference and industrial deployments. (Super Micro Computer IR, June 22)

• Supermicro delivered the NVIDIA Vera Rubin NVL4 End-to-End DCBBS Blueprint with native FP64 performance for converged HPC and AI infrastructure. (Super Micro Computer IR, June 21)

• NVIDIA announced Halos for Robotics, the industry’s first full-stack safety system for Physical AI. (NVIDIA IR, June 22)

• NVIDIA announced the BioNeMo Agent Toolkit, providing tools for agents to accelerate scientific discovery. (NVIDIA IR, June 23)

Also today: HBM test bottlenecks, floating datacenters, Hyundai absorbing Boston Dynamics, and really interesting data about the optics market.

Let’s get into it. — Austin & Vik

Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player.

Amazon to sell Trainium chips externally

Amazon plans to sell its in-house Trainium AI chips directly to data center operators, joining Google, which announced a similar move in April, in offering custom silicon outside their own cloud platforms. The shift positions both hyperscalers as direct alternatives to Nvidia, whose chip business is valued at roughly $5 trillion and which counts Amazon and Google among its largest customers. EE Times reported the development, noting it marks a departure from the prior model in which hyperscaler-designed accelerators were available only through AWS and Google Cloud rentals. (EE Times)

Vik: First Google started selling externally, and now Amazon. Their CUDA equivalent is called Neuron, and the software ecosystem is just as important. It is inevitable that companies making AI chips will feed themselves first, and then start selling to others. Competition is good.

Austin: Who will be buying these TPUs and Trainiums? Surely neoclouds. And the obvious end customers renting those chips will be AI labs. But I’ll be super interested to understand which enterprise end customers sign up. According to the WSJ,

“Among them is Citadel Securities, a longtime Google Cloud customer that recently began using TPUs for some of its research software workloads. Josh Woods, the firm’s chief technology officer, said the company can run key workloads at a 30% lower cost and up to four times as fast with TPUs.”

Relatedly, neocloud originally referred to “new GPU-only cloud” players like CoreWeave, and was meant to differentiate from traditional cloud service providers that also rent CPU, storage, etc like Amazon Web Services, Microsoft Azure, and Google Cloud. Will the term neocloud now also encompass XPUs or more broadly, any AI accelerator? I’d expect said neoclouds to eventually also stand up racks of AI accelerators from startups.

Of course, if customers are demanding XPUs, one has to stop and ask why AWS and GCP don’t just expand their own fleet of XPU rentals and are instead selling them? Some reasons come to mind. The hardware rental business has a longer time-to-payback for the CapEx investment, plus it requires more datacenter buildouts and access to power. Selling the XPUs is a lot simpler and banks a quick ROI…

MangoBoost moves from DPU chips to server racks

MangoBoost, an AI data center DPU startup, is expanding from chip sales into complete server rack systems, CEO of MangoBoost Korea Kim Jang-woo said on June 19. The company is betting that customers increasingly prefer turnkey AI infrastructure over assembling servers themselves. (The Elec)

Bold claim by MangoBoost: “Outside of Nvidia, MangoBoost is the only company capable of delivering everything from chips to software as a unified offering.”

Vik: Delivering optimized hardware is a much bigger deal than people think. MangoBoost’s software platform is called LLM Boost, and they integrate networking, storage, CPUs, and GPUs (from AMD) into racks providing turnkey solutions to deploy AI hardware.

Austin: SemiAnalysis had a nice post discussing the idea of “goodput”, namely:

“In the context of training, goodput is defined as the amount of useful work users can perform on their cluster. Goodput plays on the term throughput to mean that not all throughput is “good”. Lots of training throughput can be “bad” if a GPU fell of the bus, NCCL is stalling, or there is an OOM hiding around the corner during the next checkpoint save.”

While that definition discusses training, obviously there’s tons of money on the table if inference isn’t fully optimized either. DPUs play an underappreciated role in offloading networking tasks from host CPUs and enabling high-throughput data access for KV caches, for example in Nvidia’s STX platform:

If MangoBoost is good at leveraging DPUs, they should be able squeeze more tokens out of the same hardware.

I wonder what they can do with AMD’s MI455X Helios platform, which also has a Pensando DPU in it:

If MangoBoost can squeeze more tokens out of Helios than AMD… well then, they ought to be a great acquisition target.

TSE builds HBM test handler doubling throughput

South Korean semiconductor test component maker TSE is developing a next-generation test handler aimed at doubling inspection throughput for high-bandwidth memory devices, the company said. The handler targets HBM production lines, where rising stack heights and tighter quality requirements have stretched test times and created a bottleneck for memory makers including SK Hynix, Samsung, and Micron. TSE, which supplies probe cards and test sockets to the major DRAM producers, said the new equipment is intended to improve manufacturing efficiency as HBM volumes scale to meet AI accelerator demand. (The Elec)

Austin: The article says the engineering development will be done in 9 months from now or so:

“The companies plan to complete development of both the handler and die socket by March 2027. Once commercialized, the system is expected to alleviate testing bottlenecks that currently constrain throughput in HBM manufacturing lines.”

But when will it actually be deployed at the big three and actually move the needle on amount of tested wafers per hour? Not until 2028 or later?

Imec details ferroelectric memory advances at VLSI Symposium

Belgian research institute imec presented two ferroelectric memory developments at the 2026 IEEE/JSAP Symposium on VLSI Technology & Circuits, targeting capacity, bandwidth, and energy-efficiency constraints in AI memory systems. The work covers low-voltage ferroelectric capacitors and vertically stacked ferroelectric field-effect transistors (FeFETs), positioned as alternatives to conventional DRAM and SRAM for high-density applications. Imec framed the research as a response to AI workloads pushing existing memory architectures beyond their limits. (EE News Europe)

Vik: New memory technologies are one of the long term escape hatches from our current memory shortage predicament. We need a breakthrough now, more than ever.

Austin: imec is so cool. Vik we need to get there and talk to folks. It’s one of the most interesting R&D labs in the world.

Phison Demos 28GB/s Pascari Gen6 SSD

Phison demonstrated a PCIe Gen6 Pascari SSD reaching 28GB/s sequential throughput at Computex 2026, alongside new additions to its Pascari enterprise lineup. The demo doubles the bandwidth of current Gen5 enterprise drives and positions Phison among the first controller vendors to publicly show working Gen6 silicon. The expanded Pascari portfolio targets AI training, inference, and high-capacity storage tiers in data center deployments. (The SSD Review)

Samsung Heavy, Supermicro plan floating AI data centers

Samsung Heavy Industries has partnered with Supermicro and a Greek shipowner to develop 50MW floating AI data centers, according to Hydrogen Central. The vessels are designed to be powered by solid oxide fuel cells running on liquefied natural gas. Samsung Heavy will lead the shipbuilding, while Supermicro supplies the compute infrastructure housed onboard. (Hydrogen Central)

Vik: Ha! First we want to chuck data centers in space, now we want to bob them on the oceans. I bet they could double as AI powered weather stations making sea-faring even safer. Also, for floating data centers?

Hyundai acquires Boston Dynamics, plans 25,000 Atlas robots for factories

Hyundai Motor Group is paying $325 million to acquire SoftBank’s remaining stake in Boston Dynamics, bringing its ownership to 100%. Hyundai plans to deploy over 25,000 Atlas robots across its global plants, starting with its Metaplant near Savannah, Georgia, by 2028. A dedicated humanoid factory is targeting 10,000 to 30,000 units annually by around 2030. (finance.yahoo.com)

Austin: Well, we lost Boston Dynamics from America’s grip, but at least we’ll see some deployment and operation know-how reside in America. Savannah Bananas home game in 2028 + a Metaplant visit? Hyundai call us!

SK Hynix surpasses Samsung as South Korea’s most valuable company

SK Hynix’s market capitalization reached $1.362 trillion, overtaking Samsung Electronics. This occurred as SK Hynix shares closed 5.6% higher on Monday, lifting its market capitalization to 2.080 quadrillion won. Samsung shares ended 0.1% lower, taking its market value to 2.067 quadrillion won. (wsj.com)

Key Data #1

A single AI training cluster now requires more optical and photonic components than the entire worldwide volume of such components in 2008. (via jwt0625)

Key Data #2

According to Goldman Sachs, here is the breakdown of the BoM of a CPO switch. (via omercheema)

Supply Chain Insight

So. Many. Companies.

Worth a Watch

How AI interconnects are built — a nice video by Lightmatter.

Worth a Read

FADU’s blog post on CMX using SSDs is highly informative. The table below explains how the needs for Agentic AI differs from traditional storage.

Sector Watch

Foundry & Logic

• Intel and Apple Intel shares rose ~9% after Trump announced Apple agreed to design and manufacture chips in the U.S. with Intel, shifting significant volume from TSMC and validating Intel’s foundry strategy with a lighthouse customer. (TechMeme)

• Samsung and AMD Samsung leveraging excess fab capacity to secure AMD as advanced node customer, carving supply share from TSMC and testing pricing power in the foundry oligopoly. (com.tw)

Advanced Packaging

• TSMC Advancing CoPoS panel-level packaging with glass substrates to replace CoWoS silicon interposers, targeting mass production Q4 2028 or Q1 2029 to align with Nvidia’s next-gen accelerators. (com.tw)

• MPI Corporation Weighing prepayment mechanism to secure probe card capacity amid AI-driven supply constraints, highlighting test infrastructure as binding constraint on accelerator throughput. (DigiTimes)

Networking

• Nvidia and optical interconnects Acceleration to 1.6T and 3.2T standards exposing copper limits, driving structural shift toward silicon photonics and MicroLED in Taiwan supply chain. (DigiTimes)

• Arista Networks Deploying 1.6T Ethernet switches for AI cluster interconnects, challenging Nvidia’s NVLink dominance with open, disaggregated alternative for hyperscale datacenters. (eenewseurope.com)

Optics & CPO

• JX Nippon Expanding Indium Phosphide substrate capacity to meet rising demand for optical interconnects in datacenters, supporting transition toward optical I/O as critical bottleneck for hyperscale AI. (Semiconductor Today)

Power

• BWX Technologies Licensed small modular reactor design following activist pressure; NRC reforming licensing framework to accelerate nuclear plant approvals for AI datacenter energy demand. (Bloomberg Tech)

Components

• MLCC shortage AI accelerator density driving structural shortage in multi-layer ceramic capacitors, forcing hyperscalers to secure long-term supply contracts before chip deliveries. (South China Morning Post)