The thesis that AI capital spending would eventually collide with a revenue ceiling has been running into a concrete problem this cycle: revenue is compounding faster than the ceiling. NVIDIA's fiscal Q2 FY27 result — $96.2B total revenue, $89.0B from data centres, 75% gross margin, and a Q3 FY27 guide of $108B — was published against a baseline that explicitly assumes zero data-centre compute revenue from China (https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027). That the accelerator monopolist is guiding to sequential growth while writing off its second-largest historical market is the kind of print that changes the arithmetic.
The revenue side has moved
The manuscript's Essay 04 anchor held AI-attributable revenue at roughly $50B against aggregate hyperscaler capex of ~$725B — a ratio of ~16×. The NVIDIA data-centre line alone annualises to $356B on the Q2 rate and $432B on the Q3 guide. Layer in Microsoft's Azure, which surpassed $100B annualised for the first time and grew 43% year-on-year with 30 million Microsoft 365 Copilot paid seats (https://www.microsoft.com/en-us/Investor/earnings/FY-2026-Q4/press-release-webcast), and the composite AI-attributable revenue base is now demonstrably above the $200B threshold the manuscript defined as falsification criterion #25. At the accelerator sub-metric alone, the capex-to-revenue ratio has compressed from ~16× to ~1.8-2.0×, which is inside the criterion #26 threshold of less than 3×. These are not near-misses; both thresholds have been breached.
Capex has not flinched in response. Microsoft's Q4 FY26 capex was $35.8B for the quarter and $115.9B for the full fiscal year — largely free-cash-flow funded, which means the balance sheet is not under strain. Combined with the NVIDIA–SK Group $500B+ and NAVER–Brookfield 200 MW sovereign-AI anchors tracked in July, the aggregate commitment continues to expand rather than contract.
The foundry layer confirms the signal
TSMC's monthly revenue series (https://investor.tsmc.com/english/monthly-revenue) has now run ahead of every prior base case in the corpus: July 2026 at NT$467,580M (+44.7% YoY), August 2026 at NT$514,806M (+53.3% YoY and +10.1% month-on-month), with YTD through August at NT$3,386,870M (+39.3% YoY). The prior anchor — TSMC exiting CY2026 at ~$150-160B USD equivalent — looks too conservative; the revised base case is ~$170-180B. The August month-on-month re-acceleration is the critical datum: sequential cooling has reversed, and the foundry-layer leading indicator continues to point away from any near-term demand compression.
ASML's position widens alongside it. The September 8 cluster of press releases (https://www.asml.com/en/news) confirmed a High-NA EUV collaboration with Intel Foundry to accelerate industry readiness and an expanded strategic collaboration with Samsung Electronics for next-generation manufacturing. Individual press-release URLs returned 404 this cycle, but the artefacts are verified via the newsroom index. ASML now has High-NA EUV customer relationships across TSMC, Intel Foundry, and Samsung — the lithography monopoly has widened its commercial base rather than compressed.
Friend-shoring accelerates up the stack
SK Hynix's groundbreaking at West Lafayette, Indiana on 2026-08-28 (https://news.skhynix.com/) is the first non-Korean HBM production footprint from the oligopoly leader. Capital detail and DoE CHIPS Act tie-in were inaccessible at the direct URL this cycle, but the structural implication is clear: HBM4 supply is beginning to diversify geographically, which matters for both the passive-flow inclusion path — SK Hynix's NASDAQ ADR listing in July compounds this — and for the broader friend-shoring thesis. Intel Foundry's ASML High-NA collaboration and Samsung's expanded ASML agreement on the same September 8 add a second and third leading-edge logic node to the US-Korea friend-shoring cluster.
NVIDIA integrates vertically
The $12.93B acquisition of Hugging Face (2026-09-03, https://nvidianews.nvidia.com/news) is the most strategically consequential single move this cycle. The merchant-GPU monopolist is now competing for the developer-tools and open-model-hub layer as well as the chip layer. Cash/stock split and antitrust posture are not disclosed at newsroom-index level — the DoJ, FTC, and EU Commission review is a material watch item — but the strategic logic is clear: substituting NVIDIA now requires displacing CUDA, NVLink Fusion, and the Hugging Face model hub simultaneously. The substitution barrier has widened. On 2026-09-10 (https://nvidianews.nvidia.com/news), d-Matrix's Raptor XPU adopted NVLink Fusion, Palantir announced sovereign-AI supply-chain integration, and Skild AI added physical-AI robotics to the ecosystem. Anthropic's Fable 5.1 and Mythos 5.1, released 2026-09-01 (https://www.anthropic.com/news) — the fourth frontier-model refresh in 90 days — sustains the training-and-inference demand signal that underpins the entire stack.
The China-zero guide deserves a separate note. NVIDIA publishing a $108B quarterly guide without any assumed China contribution is a revealed-preference disclosure: the H20/B20/B30 licence-approval regime is now operationally treated as fully binding at the revenue line. The $108B ex-China is what the Western hyperscaler clouds, sovereign anchors, and neocloud tier can absorb in a single quarter without any Chinese contribution. If BIS relaxes the regime in either direction, the guide moves.
What would change my mind
Two readings would break this view. First: TSMC September and October year-on-year growth falling below +25% for two consecutive months — that would signal the foundry-layer leading indicator has turned and the August re-acceleration was a one-month aberration. Second: NVIDIA's Q3 FY27 revenue missing the $108B guide by more than 5%, or gross margin printing below 70%. Either would indicate end-market absorption has hit a limit the China-zero baseline cannot paper over and that accelerator-side revenue compounding has stalled. A CoreWeave refinancing pricing more than 150 basis points wider than 2025 issuance would be a secondary signal on the debt-transmission channel.
What I'm watching next
The BIS docket for any H20/B20/B30 licence-approval cadence disclosure; the NVIDIA–Hugging Face antitrust review filing with DoJ, FTC, and EU Commission; SK Hynix Indiana capital detail and CHIPS Act 45X exposure — direct URLs returned 404 this cycle and it remains a priority re-source; TSMC September and October monthly revenue prints; and NVIDIA Q3 FY27 in October for the gross-margin and guide read-through. A dedicated rotation on advanced packaging — tracking BESI, ASM Pacific, Disco, and Amkor at higher resolution given CoWoS-S/L as the binding 2026-2028 sub-bottleneck — is the most warranted new slot.