The key race: Nvidia's own Q2 print shows the memory shortage hardening into a multi-year constraint, not a one-quarter squeeze
Nvidia's August 26 fiscal second-quarter report was the sharpest single data point yet on how far memory scarcity has spread through the AI-hardware stack. Revenue hit $96.2 billion, up 106 percent year over year and above the $92.2 billion consensus, with data-center revenue at $89.0 billion (up 117 percent) and CFO Colette Kress guiding fiscal 2028 revenue growth to roughly 70 percent, well above the roughly 45 percent analysts had modeled - an outlook she called supply-constrained because demand is running ahead of what Nvidia can actually deliver. The clearest evidence of that constraint sat in a single disclosure: supply commitments, mostly for HBM memory, more than doubled in one quarter to $279 billion from $119 billion three months earlier, against a hyperscaler cloud backlog Nvidia says already exceeds $2 trillion. Non-GAAP gross margin came in at 75.0 percent but is guided to ease toward 74 percent in the third quarter, with memory cost the explicit reason.
The memory makers' own results show the shortage sharpening rather than easing. Samsung's Q3 HBM revenue is on track to surpass SK Hynix's for the first time, with Bernstein estimating roughly $12 billion in sales as Samsung's HBM4 revenue more than triples sequentially and HBM4 climbs above 60 percent of its total HBM mix; SK Hynix, which still holds an estimated 50 to 55 percent of the overall HBM market, actually saw HBM4 shipments decline about 27 percent from April even as it guides to a stronger ramp in the back half of the year. HBM4 pricing for Nvidia's own GPUs is now running $31 to $32 per gigabyte, nearly double HBM3e's $17 to $18, and DRAM and HBM capacity across Samsung, SK Hynix, and Micron remains sold out through all of 2027 - the same conclusion this overview reached a month ago, now with a full quarter of new results behind it.
Who's ahead: TSMC keeps pulling its US buildout forward, Intel converts one customer into a $20 billion bet, and Nvidia's dealmaking widens past pure compute
TSMC kept accelerating rather than merely executing its overseas plan: equipment installation at Arizona's Fab 21 Phase 2 begins in the third quarter of 2026, with N3 production targeted for late 2026 and N2 output from 2028, several quarters ahead of the facility's original schedule, part of the additional $100 billion Arizona commitment (at least four more 2nm-class fabs, taking the total US pledge to $165 billion) CEO C.C. Wei laid out on TSMC's Q2 call. Intel converted its single named 14A external customer, Tesla, into a much larger balance-sheet bet: an August 10 stock offering, originally sized at $15 billion, was upsized to $20 billion at $95 a share as AI-driven demand accelerated, funding capacity for a foundry business whose returns will not arrive before 2028 and that still rests on Tesla's Terafab commitment (in Grimes County, Texas, alongside SpaceX as the site's designated high-volume operator) as its only confirmed 14A logic customer.
Nvidia's own dealmaking widened past pure compute contracts this cycle. AWS and Nvidia announced August 26 an additional 2 million Blackwell Ultra, Rubin, and Rubin Ultra GPUs for 2027-2028 on top of AWS's prior 1-million-GPU commitment, and Nvidia took a strategic stake in Blackstone-backed Lancium alongside a partnership turning its 15-plus-gigawatt site pipeline into a deployment platform for Nvidia's DSX reference designs. Reporting in the days after Nvidia's earnings call, still unconfirmed by the company as of this update, put Nvidia near a roughly $12.9 billion acquisition of AI model-hosting platform Hugging Face - what would be its largest acquisition ever - alongside separate talks to invest in AI search company Perplexity at a valuation above $30 billion, extending a pattern in which Nvidia increasingly decides which AI-infrastructure and AI-software companies get built, not just which ones get its chips.
What decides it: the Kyber delay still has no real answer, and China's homegrown lithography is real progress that still falls well short of the leading edge
The sector's single biggest unresolved execution risk is unchanged from a month ago: SemiAnalysis's report that Nvidia's Kyber rack for Rubin Ultra is delayed more than twelve months to 2028, traced to a 78-layer PCB midplane that cannot yet be manufactured at reliable yields, with a stopgap two-rack alternative reportedly scrapped after cloud-customer pushback. Nvidia's only public response remains a short statement that its roadmap is intact, without confirming or denying a slip, and no new information surfaced this period to resolve the question either way - which matters because Kyber was meant to be the scale-up vehicle for 2027's highest-margin hardware and fiscal 2028 guidance already assumes demand Nvidia may not be able to fully supply on schedule.
China's parallel lithography track produced a genuine milestone without closing the gap that actually matters. Shanghai Yuliangsheng, a startup with ties to Huawei and SiCarrier, has moved its domestically built immersion DUV scanner from trials into low-volume production, targeting five units delivered in 2026 to SMIC, Hua Hong, and CXMT and roughly 20 units by 2027; the tool is native to 28-nanometer features, with 7-nanometer or 5-nanometer output reachable only through costlier multi-patterning, and some critical components are still imported from Japan. Goldman Sachs now projects China's overall chip self-sufficiency gap could narrow to just 34 percent by 2035, but the firm and the tool's own specifications agree on the binding constraint: advanced-node lithography, the layer AI accelerators actually need, remains years from domestic reach even as China's mass-produced 28-nanometer-class capability becomes real.
The money and the rules: Washington opens a second front against ASML's China exposure, and Nikon and Canon give it a real competitive test
The US escalated its pressure on ASML's remaining China business onto a new, more direct track: the Trump administration is now preparing to press the Dutch government to force a near-total ban on ASML's sale and servicing of its deep-ultraviolet lithography tools in China, going beyond the EUV and advanced-DUV restrictions already in place, according to reporting around August 20 that also described Dutch officials as irritated by the move. That direct approach runs alongside, rather than instead of, the legislative track: the MATCH Act, which would push allied nations to match US restrictions on advanced chipmaking-tool sales, remains bundled with the AI OVERWATCH Act and Chip Security Act in the Senate's NDAA manager's amendment, but had not secured a floor vote as of this update, and no formal Bureau of Industry and Security enforcement action has yet followed Commerce Secretary Howard Lutnick's earlier claim that an EUV system or component may have reached China, a claim ASML continues to deny.
ASML also picked up its first real commercial pressure in over a decade, from two Japanese rivals rather than from Washington. Nikon is aggressively undercutting ASML's ArF immersion DUV pricing and has signaled a cross-compatible next-generation platform targeted for fiscal year 2028, a move that mainly compresses pricing and encourages dual-sourcing in the 7-to-28-nanometer node range where Nikon still has real precision credibility, without reaching EUV or the leading edge ASML alone still commands. Canon is pursuing a different opening with nanoimprint lithography, marketed at 5-nanometer-equivalent resolution and lower cost, though adoption remains nascent and largely confined to memory and specialty logic. Neither threatens ASML's roughly 94 percent share of the lithography-equipment market outright, but together they mark the most credible commercial challenge the company has faced since Sony's smaller camera-lens-adjacent pushes years ago.
What to watch through 2027
Watch Nvidia's fiscal third-quarter report, expected in late November, for whether the roughly $108 billion revenue guide holds and whether gross margin actually bottoms near 74 percent, and watch for any harder confirmation, or a real denial, on the Kyber 2028 delay. Watch TSMC's mid-October Q3 report for whether 2nm capacity reaches its 100,000-wafer-per-month year-end target and whether Arizona Fab 21 Phase 2's accelerated equipment-installation timeline holds. Watch ASML's Q3 report on October 14, guided to 11 to 12 billion euros in net sales, and watch whether the direct US-Dutch pressure on ASML's remaining DUV business in China produces an actual new export rule, separate from whatever happens to the MATCH Act inside the NDAA.
Watch whether Nvidia's reported Hugging Face acquisition and Perplexity investment talks convert into signed, confirmed deals. Watch Samsung's and SK Hynix's full second-half results for whether HBM4 revenue and shipments behave as guided, including whether Samsung's quarter-over-quarter HBM4 tripling actually lets it overtake SK Hynix in HBM revenue. Watch Shanghai Yuliangsheng hold its five-unit 2026 delivery target to SMIC, Hua Hong, and CXMT, the clearest near-term test of China's lithography-independence pace, and watch whether Intel's PDK 0.9 design kit ships on schedule later in 2026 and produces a second named 14A customer alongside Tesla. Watch whether Nikon's FY2028 ArF platform draws a first named foundry customer, the real test of whether its pricing challenge to ASML becomes more than a discount offer.
Sources
- CNBC - Nvidia earnings takeaways: Huang forecasts 70% fiscal 2028 revenue growth
- The Motley Fool - Nvidia just locked in a $279 billion bet on memory
- TrendForce - Samsung, SK hynix's HBM4 push puts pricing in the spotlight for 2H earnings
- Tom's Hardware - China begins mass production of homegrown immersion DUV lithography machines
- NL Times - US preparing to force Netherlands to ban ASML from selling to China
- CNBC - Intel upsizes stock offering to $20 billion at $95 per share as AI demand accelerates
