Daily Update - July 31st, 2026
TSMC partners Kinsus on EMIB-like packaging, Samsung Q2 chip profit surged 250x on HBM demand, CXMT debuted at $590B after $4.4B IPO, Crusoe-Aalo nuclear AI datacenter targets 2027.
Good morning, it’s Friday, July 31.
Hyperscaler earnings are putting AI CapEx skeptics on the back foot, with Azure hitting $100 billion in annualized revenue at 43% growth, its fastest pace in four years. Samsung’s chip unit posted a roughly 250-fold earnings surge in Q2 2026, TSMC’s 1.4nm fab is running ahead of schedule, and CXMT’s $590 billion Shanghai debut is drawing Senate scrutiny over its Apple supply chain ties.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Hyperscaler Earnings Silence AI CapEx Skeptics, Lift Chip Stocks
Microsoft’s fiscal third quarter landed with enough force to move markets broadly. Azure crossed $100 billion in annualized revenue at 43% growth, the fastest pace in four years, and the company’s stock surged roughly 17% for its best single-day performance in nearly 18 years. Microsoft added close to $450 billion in market capitalization in a single session, surpassing Nvidia’s one-day record. Goldman Sachs, per Bloomberg coverage, framed the print as proof that Microsoft is showing Wall Street how the AI bet pays off.
Amazon matched the moment from its own direction. AWS posted 37% revenue growth, its fastest quarterly cloud print in 18 quarters, with margins climbing alongside it. Amazon then raised its total investment plan to $220 billion, removing any ambiguity about its infrastructure appetite. The two reports together rattled through the semiconductor supply chain: SK Hynix jumped 25% and Samsung rose more than 20% on the combined earnings optimism, with Nvidia and Micron also gaining on the broader chip rally.
Behind the demand:
TD Cowen analyst Derrick Wood described Microsoft’s quarter as a “Goldilocks” result, citing accelerating Azure growth and improving AI unit economics.
The 30-year U.S. Treasury yield hit a 19-year peak on the same session Microsoft shares rallied, complicating the backdrop for rate-sensitive equities.
QQQ, the Nasdaq-100 ETF, rallied alongside the Microsoft beat as broader tech sentiment shifted on the Azure numbers.
Austin: Azure can make decent margins and a quick return renting out GPUs, and even better margins running AI workloads on those GPUs to power enterprises and Microsoft’s own software. And this AI, by the way, increasingly runs on custom XPUs and CPUs:
Satya Nadella: “We also continue to modernize our fleet with our first-party innovation, alongside the latest from NVIDIA and AMD… our first-party CPUs and accelerators. Our Maia 200 AI accelerator — which offers over 30% improved tokens per dollar, compared to the latest silicon in our fleet — is now live in our Iowa (👋🏻) and Arizona datacenters. Our Cobalt server CPU is deployed in nearly half of our DC regions…”
Austin: AWS takes it further. Not only does AWS run first party workloads on first party silicon, but customers are renting this custom silicon directly or even asking to buy it outright!
Andy Jassy: “On the question about selling Trainium, we’re quite excited about what’s happening in our chips business… The fact that we have a multiyear multi-gigawatt commitments from the 2 largest AI labs in Anthropic and OpenAI and more and more companies, as I mentioned in my opening comments, using Trainium is exciting and promising. And so there are a lot of customers who are very excited about using it in the form that we’re providing right now. We do have an increasing number of customers who are interested in us providing the Trainium chips to them separate from even from our cloud and we’re actively having those conversations and exploring, and I expect there’s a real chance we’ll do that in the future.
Samsung Chip Profits Soar 250-Fold as Mobile Unit Posts Loss
Samsung Electronics posted Q2 2026 revenue of KRW 171.5 trillion with operating profit of KRW 89.5 trillion, a roughly 250-fold surge in semiconductor earnings driven almost entirely by AI-server demand for HBM and enterprise NAND. The mobile division couldn’t ride that wave: it slipped into a loss as elevated memory component prices compressed margins it couldn’t pass on to consumers. Samsung simultaneously announced plans to begin construction of a second fab in Taylor, Texas, this year and said it would commit 60% to 70% of memory output to long-term supply agreements with AI customers, a structural shift that tightens available spot supply further.
Kioxia reported a 45-fold surge in quarterly net profit on AI-driven NAND demand, yet still missed analyst estimates, and Bloomberg noted the shortfall clouded optimism about how much further flash memory prices can run. Kioxia’s outlook miss came alongside a 3-for-1 stock split announcement, a combination the market received coolly. Samsung’s warning that memory supply tightness will persist until at least 2028 frames both results: the AI-server end of the market is genuinely capacity-constrained, while consumer-facing segments sit in a cost squeeze of the same dynamic’s making.
On an adjacent front:
Qualcomm reported fiscal Q3 2026 revenue of $9.947 billion with weaker earnings, citing memory price inflation on smartphone demand, and plans to raise prices across most of its product portfolio starting in September.
Austin: No surprise Samsung crushing it with HBM and NAND flash given that prices for both have gone to the moon. But new capacity is coming online across in the industry in 2027 and 2028. Yes, demand for memory and storage will outstrip supply for some time, but that demand-supply gap will shrink, and at some point, prices will stabilize. They’ll even come back down. Don’t forget that…
Vik: Memory and storage will continue to have pricing power for some more time. We don’t think the end is around the corner regardless of how this week has gone in the markets. Mobile business units of both Qualcomm and Samsung cannot turn profits in this environment; its the age of the AI datacenter, like it or not. Samsung has something to offset the barbell, Qualcomm really doesn’t.
TSMC Races Intel on Packaging While 1.4nm Fab Beats Schedule
TSMC is developing an EMIB-like substrate interconnect technology in partnership with Kinsus to counter the risk of customers drifting toward Intel’s advanced packaging offerings, according to TrendForce. The approach uses a high-density bridge embedded in an organic substrate, broadly comparable to Intel’s Embedded Multi-die Interconnect Bridge, and is intended to offer an alternative for AI chip customers who find CoWoS supply constrained. CoWoS capacity is expected to remain tight through 2026, per Wccftech, which adds urgency to TSMC’s push toward a complementary packaging path. The timing matters: Unimicron told Digitimes that Intel’s own next-generation EMIB-T won’t reach mass production until 2027, leaving a window TSMC is now moving to occupy.
On the process side, TSMC’s 1.4nm fab in Hsinchu is running ahead of schedule, with the first building now expected to complete before April 2027, according to TrendForce. Mass production is targeted for mid-2028, a timeline Wccftech says has been tightened partly by TSMC’s use of AI tools to manage construction risk and reduce exposure to high-temperature hazards on site. The two tracks, packaging and process, are advancing in parallel as TSMC works to hold its position across the full chip production stack.
On an adjacent front:
Samsung is reportedly in a $200 billion foundry deal with Broadcom framed as a direct challenge to TSMC’s AI manufacturing dominance.
Digitimes argues Samsung’s HBM4 base-die production gives it a manufacturing angle in advanced memory that neither TSMC nor Intel can readily replicate.
Austin: TSMC is increasingly taking Intel Foundry and advanced packaging seriously. Making an EMIB alternative is a shrewd decision. Call it an alternative to CoWoS capacity tightness… or just call it an alternative to Intel. Regardless, TSMC is smart to pursue this. More on EMIB vs CoWoS in this Chipstrat article.
Vik: Two things: (1) call it TSMC EMIB or whatever, but where is the capacity to push into mass production? (2) TSMC is starting EMIB-like packaging now? Intel will eat their packaging lunch before that.
CXMT’s $590 Billion Debut Draws Senate Fire Over Apple Supply Chain
Chang Xin Memory Technologies hit the Shanghai exchange running, its market cap crossing $590 billion after a $4.4 billion IPO that MSCI fast-tracked into its indexes within days. The company now holds 7.67% of global DRAM supply, enough to rattle shares of SK Hynix and Micron. Senators Chuck Schumer and Tom Cotton sent Apple a formal warning not to source chips from CXMT or YMTC, citing national security grounds. Bloomberg reported the letter; Global Times shot back that it distorts market competition and raises costs for US firms.
Behind the IPO sits a decade of patient state-backed capital: CXMT burned through roughly $5 billion before turning profitable, funded largely through Hefei’s municipal investment model. On the tooling side, SMEE is expanding immersion DUV production out of Shanghai, though Asia Times notes its technology still trails ASML by four generations. That gap matters less if volume fills the space that tightening Western export controls create.
Ripples from the debut:
CNBC reported Apple is weighing CXMT and YMTC chips as memory costs climb, making the Senate letter a direct intervention into a live procurement decision.
Chinese chip stocks broadly sold off as investors rotated out of tech following the CXMT listing.
CXMT’s roadmap targets further node shrinks through 2028, according to Digitimes research.
Austin: Everyone uses memory… not just Apple. Will letters be sent to Qualcomm? To AMD and Intel? What about startups? I’m sure there are sourcing folks at all the chipmakers who are asking (behind closed doors) what the monetary and political costs of entertaining CXMT memory are.
Vik: Everyone feared that CXMT will flood the memory market and cause prices to plummet. Regulations like these prevent it from happening. Rest easy— the big 3 still have a stranglehold on memory.
AI Power Crunch Spurs Nuclear Pacts, Fuel Cells, and Off-Grid Bets
Crusoe Energy and Aalo Atomics are building what they describe as the world’s first nuclear-powered AI data center, sited at Idaho National Laboratory with a demonstration target of 2027. The project pairs Aalo’s microreactor design with Crusoe’s GPU cluster, bypassing grid constraints entirely. Crusoe is simultaneously pressing ahead in Texas, where it has filed permits for two additional data centers at its Goodnight Campus in Armstrong County, a site already carrying a $29 billion-plus price tag. Texas grid rules now require co-located AI campuses to operate off-grid during curtailment events, a regulatory nudge that makes dedicated generation more attractive to every developer in the state.
Bloom Energy’s stock jumped 25.3% after a Q2 earnings beat driven by fuel-cell contracts with AI data center operators, and the company raised full-year guidance. Schneider Electric raised its data-center demand outlook for the second time this year, citing accelerating orders from hyperscalers. Delta Electronics lifted its 2026 capital expenditure to NT$70 billion and expects AI-related sales to exceed 25% of total revenue, with HVDC shipments slated to begin in Q4 2026 and scale through 2027.
Also, a battery startup (Antora Energy) raised $550 million to supply grid-scale storage to AI data centers, per Bloomberg.
Vik: This is great! Nuclear behind-the-meter power sources are the holy grail for onsite power generation in datacenters. High capacity factor, clean energy, but they do need a ton of water. I have written about this in the past — AI datacenters drink more water than you think (free post).
Sector Watch
Power
Monolithic Power Systems reported Q2 2026 results with adjusted EPS up 54% year-over-year and Q3 outlook topping $1.14B, driven by AI data center power demand. (GlobeNewswire)
Allegro MicroSystems reported Q1 FY2027 revenue of $259M, up 27% year-over-year, marking a recovery in its magnetic sensor and power IC business. (Allegro MicroSystems)
Murata Manufacturing raised its full-year outlook while warning AI infrastructure spending will eventually plateau, signaling caution from a key passive-component supplier. (Bloomberg Tech)
Compute
AMD signs a 15-year data center lease with Core Scientific valued at up to $14B, securing dedicated AI compute capacity at scale. (Barchart.com)
Xsight Labs raises $300M to expand its switch silicon portfolio targeting edge, aggregation, and telco AI infrastructure. (Light Reading)
Together AI raises $800M at an $8.3B valuation as enterprises accelerate adoption of open-model AI cloud alternatives. (startupfortune.com)
Memory
Renesas reported Q2 2026 financial results with non-GAAP profit of 132.7 billion yen and issued updated consolidated forecasts, with DDR5 MRDIMM chipset demand cited as a growth driver. (Business Wire)
Seagate roadmap targets 50 TB HAMR hard drives in 2027, extending its areal density lead for high-capacity data center storage. (techpowerup.com)
Montage Technology debuts a CXL 3.2 MXC prototype, advancing China’s domestic memory expansion interface silicon ecosystem. (intlbm)
Foundry
UMC moves ahead with fab expansions in Tainan and Singapore, adding mature-node capacity to serve diversified end markets. (Taiwan News)
DB HiTek foundry expansion hits a setback as unresolved water supply constraints delay the planned capacity increase. (thelec.net)
Rapidus accelerates its advanced chip supply push as AI-driven component price inflation intensifies pressure on Japan’s domestic semiconductor roadmap. (JStories)
Packaging
Samsung Electro-Mechanics advances its Mi-RAE growth initiatives into commercialization phase, targeting next-generation substrate and component markets. (thelec.net)
Samsung Electro-Mechanics and SoulBrain reportedly expand their partnership to develop glass substrate materials for AI chip packaging. (digitimes)
Sony begins phased restart of its Kyushu image sensor plant following the 7.1-magnitude earthquake, with full production timeline still being assessed. (Bloomberg Tech)
Optics & Networking
Corning targets a $10B addressable opportunity in AI data center co-packaged optics and silicon photonics as it converts fiber leadership into rack-level wins. (SDxCentral)
GigaDevice expands its optical communication IC portfolio, positioning Chinese chipmakers in the fast-growing transceiver component market. (dataweek.co.za)
LG Electronics deepens its Nvidia partnership to pursue physical AI projects spanning manufacturing and robotics applications. (The Elec)
Edge
Silicon Motion and MediaTek will co-present an FMS 2026 keynote on intelligent storage solutions for next-generation AI-ready automotive platforms. (Silicon Motion)
AMD launches Ryzen Embedded AI X100 SoC series and a full physical AI product stack including modules and dev kits targeting edge and robotics deployments. (ServeTheHome)
Allegro MicroSystems posted Q1 FY2027 revenue of $259 million, up 27% year-over-year. (Allegro MicroSystems)
Data Centers
Nexus Data Centers is in advanced talks to secure $15B in financing for a Google-backed Anthropic data center in Texas. (WSJ)
DeepSeek is developing a massive AI data center in Inner Mongolia, extending China’s sovereign AI infrastructure buildout. (Bloomberg.com)
Wistron will build a 10 MW AI computing center in Tainan, adding ODM-operated AI infrastructure capacity in Taiwan. (Taiwan News)
Policy & Trade
Anthropic CEO Dario Amodei publicly rejects a proposed ban on open-weight AI models while calling for tighter chip export controls and mandatory safety testing. (ESG News)
Korea plans a $14B sovereign wealth fund allocation to AI investments following the market rout, signaling state-level commitment to the technology sector. (Bloomberg.com)