Daily Update - October 19th, 2026

CXMT hits 10% DRAM share on $16.7B Q3 revenue, Japan approves TSMC 3nm Kumamoto fab with $4.1B, Samsung routes Nvidia HBM4E base die to TSMC 3nm, Samsung Electro-Mechanics books $251M substrate order.

Hello world, it’s Monday, October 19th.

CXMT’s near-600% revenue surge has pushed its DRAM market share to roughly 10%, though its stock still fell 7.8% on the results. Elsewhere, Samsung is routing the logic base die for Nvidia’s next-generation custom HBM chips to rival TSMC’s 3nm process, Microsoft pulled out of Nscale’s Essex site days before its $3 billion IPO, and Japan is backing TSMC’s Kumamoto 3nm fab with up to $4.1 billion in subsidies.

Let’s get into it. — Austin & Vik

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

Presented by Crusoe

Crusoe builds the full stack for AI, from power and data centers to the cloud. It can also help you tune a small model on your own data for just a few bucks.

For repeatable work like reports, an open model trained on your terminology and formats can match a frontier model, and it’s cheaper and faster to run.

Crusoe’s Serverless Fine-Tuning handles the training: upload your dataset, pick an open model like Llama, DeepSeek, or GLM, and pay per million tokens used, not per GPU hour.

About $2 buys roughly 5M training tokens on smaller models.

Deploy on Crusoe Cloud in one click or download the weights and deploy on your own.

Get $5 in credit to start →

Samsung Sends Nvidia’s HBM4E Base Die Work to TSMC 3nm

Samsung is turning to rival TSMC’s 3nm process to manufacture the logic base die for Nvidia’s next-generation custom HBM chips, known internally as NVHBM, according to the Seoul Economic Daily. The base die, which handles memory control logic and sits at the bottom of an HBM stack, requires process sophistication that Samsung’s own foundry currently can’t match at the performance level Nvidia demands for HBM4E. The arrangement is notable precisely because Samsung operates its own foundry business and would ordinarily be the natural supplier. Reuters confirmed the broad outline of the report. TSMC’s 3nm capacity is already under heavy pressure from AI chip demand, with the company separately planning a new Taiwan fab targeting a 2028 production ramp.

Vik: This is the custom base die thesis playing out. Once the memory controller moves off the GPU onto the base die, bandwidth becomes a logic problem and the value accrues to whoever designs it (my note). Nvidia co-designing on TSMC 3nm, memory makers back to supplying commodity DRAM stacks?

Austin: Samsung built its HBM4 base die in-house on its own 4nm, so this is a reversal for HBM4E. And SK Hynix is reportedly looking at Intel for HBM4E base dies because TSMC’s cost 3-4x its own DRAM core dies.

CXMT’s Near-600% Revenue Surge Pushes Its DRAM Share to 10%

ChangXin Memory Technologies reported Q3 revenue as high as $16.7 billion, a near-600% year-on-year surge that lifted its share of the global DRAM market to roughly 10%. That number would have been unthinkable two years ago. The Hefei-based chipmaker also swung to a nine-month record profit, driven by a broad rally in commodity memory prices that amplified every unit it shipped. CXMT’s stock nonetheless fell 7.8% on Monday, closing at 41.52 yuan, well off its August peak of 61.8 yuan, as investors focused on margin sustainability and yield uncertainty rather than the headline figures. The drop came despite the company announcing a share buyback alongside its results. SK Hynix shares also slid, according to Timothy Sykes, as traders weighed what a credible tenth DRAM supplier means for commodity pricing.

Japan Greenlights TSMC Kumamoto 3nm Fab With $4.1 Billion in Subsidies

Japan has formally approved TSMC’s 3nm Kumamoto fab plan and committed up to $4.1 billion in government aid, clearing the biggest policy hurdle for advanced-node chip production outside Taiwan. The facility targets a 2028 production ramp and is designed to serve both domestic Japanese chipmakers and global AI chip demand, according to Taiwan News. TSMC is also adding a 3nm plant in Taiwan to meet surging AI demand, with that project sharing the same 2028 timeline. Kumamoto already hosts TSMC’s first Japan fab, producing on older nodes; the new plant would bring leading-edge capacity to the site for the first time.

Anthropic Offers Double-Digit Yields, Signs Nova Scotia Data Center Lease

Anthropic is offering Blackstone yields above 10% on subordinated chip-financing paper, a rate that puts frontier AI debt squarely in junk-bond territory. The deal, structured around AI compute infrastructure, comes as the company simultaneously signed its first Canadian data center lease with EverWind in Nova Scotia, taking a 95-megawatt phase of a planned $7 billion campus. Anthropic is also committing a 5% equity stake to a new public-benefit fund intended to fund donations addressing AI’s effect on employment. Three significant capital commitments in rapid succession, from a company still scaling revenue against steep liabilities, make for an unusually full week.

Microsoft Quits Nscale Essex Site Days Before $3 Billion IPO

Microsoft has pulled out of Nscale’s Loughton, Essex data centre, citing UK grid connection delays that, according to reporting by the Telegraph, stretch into the 2030s. The exit lands at the worst possible moment: Nscale was days from a planned $3 billion IPO, and Bloomberg reports investors were already pushing back on the valuation before the anchor tenant walked. The Essex site was meant to anchor Nscale’s AI cloud buildout, making the departure a direct hit to the capacity story the company intended to sell on its roadshow. UK grid constraints have been a known friction point for data centre developers, though the 2030s timeline makes the problem concrete in a way that’s hard to paper over in a prospectus.

Vik: Power is the binding constraint on capacity now, and a grid connection slipping into the 2030s is about as hard a no as you get. Microsoft and Anthropic were around 85% of Nscale’s contracts per the filing.

Austin: Come on Europe!

Qualcomm Denies Snapdragon 8 Elite Extreme Gen 6 Overheating as Samsung Eyes Exynos

Reports from multiple outlets claim Samsung’s Snapdragon 8 Elite Extreme Gen 6 samples failed thermal tests during Galaxy S27 Ultra validation, prompting the company to weigh a return to its in-house Exynos chip for at least that model. Qualcomm pushed back directly, with an executive publicly denying any overheating problem with the chip. Samsung is reportedly considering Exynos for the Galaxy S27 Ultra specifically, a slot that has historically gone to Snapdragon in markets including the United States. The dispute arrives as Samsung is said to be accelerating the S27 production timeline, according to Nokiamob. Qualcomm’s denial doesn’t settle the underlying question of whether its silicon passed Samsung’s own internal benchmarks, and the two companies’ accounts remain publicly at odds.

Samsung Electro-Mechanics Books $251 Million Substrate Order as Capacity Lags Demand

Samsung Electro-Mechanics has signed a 360.5 billion won ($251 million) semiconductor package substrate supply agreement with an unnamed global customer, pushing its 2026 cumulative bookings to 4.3 trillion won according to 파이낸셜뉴스. The order, for memory substrates used in AI server applications, brings the year’s total contract value to a level that illustrates how persistently big-tech clients are leaning on the Korean supplier. That demand pressure sits awkwardly against a separate problem: Digitimes reports that Samsung Electro-Mechanics’ $5.1 billion capacity expansion plan is running into equipment supply constraints, a bottleneck that will slow the build-out of new substrate lines. More orders, slower factory ramp. The company has not publicly identified the customer or disclosed a delivery schedule for the latest contract.

Austin: Substrates are tight everywhere, which is probably why Nvidia is reportedly financing Unimicron’s Taiwan expansion (in Sector Watch below). Same thing we saw with the KYEC prepayment last week.

Sector Watch

Compute

• Intel unveiled Inference Microservice software to maximize AI inference on Xeon CPUs, expanding its software-defined compute stack for data-center deployments. (The Elec)

Memory

• Micron is developing Edge Inference Memory (EIM), a new memory category purpose-built for physical AI and humanoid-robot applications, targeting deployment in edge devices operating outside data centers. (The Elec)

• Phison nearly quadrupled quarterly revenue on AI storage demand and is now pursuing a flash-memory architecture capable of substituting for DRAM in inference workloads. (iTWire)

• Nvidia attributed consumer GPU shortages directly to insufficient memory supply, signaling that DRAM and HBM constraints are now limiting graphics card availability. (The Elec)

Foundry & Packaging

• Hanmi Semiconductor opened a Silicon Valley subsidiary, ‘Hanmi USA,’ to support U.S. chip plants with TC bonding equipment and plans to launch its Wide TC Bonder for 2.5D AI packaging in 2027. (digitimes)

• Nvidia is reportedly financing Unimicron‘s Taiwan capacity expansion, deepening hyperscaler-to-substrate-supplier direct investment as ABF substrate demand surges. (天下雜誌)

• Siemens and Nvidia expanded their computational lithography collaboration, delivering up to 12x speedup for next-generation AI chip mask synthesis workflows. (EE News Europe)

Power

• Lite-On demonstrated 800VDC power delivery technology at Nvidia’s MGX platform event, positioning its rack power systems for next-generation GPU infrastructure. (digitimes)

• Vicor published next-generation 48V AI power architecture specifications, advancing its high-density power delivery roadmap for GPU and accelerator rack deployments. (Tech Briefs)

Edge & Robotics

• Renesas launched a dedicated Physical AI team to expand its humanoid-robot chip business, signaling a formal organizational commitment to the emerging robotics silicon market. (THE ELEC, Korea Electronics Industry Media)

• Foxconn acquired a $727 million stake in a Mitsubishi Electric EV-parts unit, deepening the company’s vertical integration into electric-vehicle components alongside its AI server assembly growth. (Bloomberg.com)

• Nvidia and Mercedes-Benz extended their strategic autonomous driving partnership to jointly develop next-generation AI systems for upcoming vehicle platforms. (The Elec)

Data Centers

• Wistron board approved a NT$7.1 billion AI infrastructure expansion, committing fresh capital to scale its AI server assembly and data-center hardware capacity. (Wistron)

• HD Hyundai Electric and LG CNS formed a partnership to jointly expand power infrastructure supply for AI data centers in South Korea. (The Elec)

Policy & Trade

• Taiwan enacted strict controls on AI server exports to China, imposing penalties of up to five years imprisonment following a US tariff deal and a Super Micro enforcement action. (Tom’s Hardware)

• Germany plans to announce a national AI agreement with Nvidia next week, according to government sources, in a move to anchor sovereign AI computing capacity. (Reuters)

• A Chinese startup unveiled what it claims is the first domestically developed EUV laser source, a key subsystem previously monopolized by Cymer and Gigaphoton. (TechPowerUp)

Tell us how to make Semi Doped better! Give us your feedback! ❤️