The Q2 CY2026 hyperscaler earnings window has closed, and the verdict is unambiguous: the AI capex ceiling thesis is not finding support — it is being actively contradicted by the data, and this is the strongest single-cycle refutation in the research corpus. Meta's second-quarter result and full-year guide revision are the load-bearing new datum: the company raised the floor of its FY2026 capex range from $125B to $130B, narrowing the band to $130-145B (https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx), which signals committed physical slot-allocations and contracted hardware rather than a discretionary top-end stretch — a more decisive indicator of capex durability than an upper-end walk would be.
The arithmetic has run away
Three falsification criteria are now simultaneously in motion. Criterion #25 asked whether composite AI-attributable revenue would exceed $200B per annum. It has been breached. NVIDIA's data-centre segment generated $89B in Q2 FY27 and guided $108B for Q3 on a China-zero baseline — an annualised run-rate of $432B at the accelerator layer alone. Microsoft's Azure crossed $100B annualised at the Q4 FY26 print. Meta's Q2 revenue was $60.80B (+28% YoY), with Zuckerberg stating AI is "accelerating our core business today, powering our next generation of products, and opening the door to entirely new enterprise opportunities." Criterion #25: breached, direction contradicts.
Criterion #26 asked whether capex-to-revenue would compress below 3× at the composite hyperscaler-service level. At the accelerator sub-metric it already has: $89B of quarterly NVIDIA data-centre revenue against $700-800B of aggregate hyperscaler capex implies roughly 2×. The formal composite breach is gated by disclosure discipline — Microsoft, Alphabet, Amazon and Meta do not disaggregate AI-service revenue from cloud or advertising-with-AI-lift — but the directional read is unambiguous. Criterion #27, asking whether AI-service gross margin holds above 60% at two or more hyperscalers for three consecutive years, is the one leg that has not moved. NVIDIA reported 75% gross margin at the accelerator layer, one step upstream, but none of the four major hyperscalers discloses an AI-service segment margin. The silence is itself inferential: they could disclose if the numbers were flattering.
TSMC's trajectory supplies the demand-side corroboration: July revenue +44.7% YoY, August +53.3%, year-to-date +39.3%, with an August monthly sequential of +10.1%. These numbers are incompatible with any near-term capex-ceiling scenario.
Moat extension, not compression
NVIDIA's $12.93B acquisition of Hugging Face (https://nvidianews.nvidia.com/news), closed 3 September and paired with NVLink Fusion adoption in the d-Matrix Raptor XPU on 10 September, is the most strategically significant development outside the earnings prints. The merchant-GPU incumbent has extended vertically into the dominant open-model hub and developer-tools layer. Substituting NVIDIA now requires displacing an ecosystem — CUDA, NVLink Fusion, and the Hugging Face model registry — not merely qualifying a rival chip vendor. Paradoxically, a more durable NVIDIA moat strains the commoditisation limb of the ceiling thesis while simultaneously strengthening the arithmetic-collapse limb: if revenue and margin hold, the capex-to-revenue ratio compresses via revenue growth rather than a capex cut.
On the physical layer, PJM's 2027/28 Base Residual Auction cleared at the FERC price cap — the third consecutive cap-clear (https://www.utilitydive.com/search/?q=PJM+2027+2028+auction) — anchoring capacity-revenue durability for merchant-power names such as Vistra through the 2028 delivery year. GE Vernova's August launch of MV-UPS systems purpose-built for AI factories opens a new competitive front in rack-adjacent power conditioning against Vertiv; whether that translates into backlog displacement will show in Vertiv's Q3 earnings. At the foundry and advanced-packaging layer, SK Hynix's Indiana HBM fab groundbreaking on 28 August and ASML's High-NA collaboration with both Intel Foundry and Samsung on 9 September deepen the friend-shoring picture at a layer that is structurally 12-24 months from any viable substitution path.
Anthropic's governance operating line continues to compound in ways that will eventually appear in an S-1 cost structure. The appointment of Mariano-Florentino Cuéllar — former California Supreme Court justice and Carnegie Endowment president — as Chief Global Affairs Officer on 4 August (https://www.anthropic.com/news), coupled with the Bernanke Long-Term Benefit Trust appointment in July and more than $30M in disclosed governance-related expenditure inside sixty days, constitutes the third senior governance hire in thirty days. The Model Hardware Standard preview on 27 August and the Fable 5.1 and Mythos 5.1 releases on 1 September (https://www.anthropic.com/news) — the fourth frontier-tier refresh in 90 days — are the product-layer expression of the same compliance-cost cadence. This operating footprint will compress frontier-lab gross margins when S-1 disclosure eventually forces the number into the open.
What would change my mind
A downward revision of more than 10% to the combined CY2026 capex guides of Microsoft, Alphabet, Meta and Amazon at Q3 2026 earnings (October–November print) would be the cleanest falsification signal. Two consecutive TSMC monthly prints below +25% YoY — the September read is due mid-October — would indicate demand cooling before it surfaces in hyperscaler guidance. NVIDIA's fiscal Q3 FY27 gross margin falling below 70% would flag merchant-GPU pricing-power compression. An Anthropic S-1 revenue print below $30B annualised would break the $965B post-money valuation's implied arithmetic and push criterion #25 back toward the ceiling view. Any one of these alone would require reassessment; all four together would break the thesis.
What I'm watching next
Q3 2026 hyperscaler earnings — specifically whether Meta or any other hyperscaler walks its FY guide down rather than reaffirming. The Alphabet and Amazon Q2 CY2026 primary releases remain deferred from this cycle (abc.xyz/investor and ir.aboutamazon.com each failed to return the relevant press releases) and need sourcing before the next scan. CoreWeave's refinancing is in its fifth consecutive scan-gap; pricing more than 150 basis points wider than the 2025 Blackstone issuance would trip the private-credit transmission channel. On the physical side, GE Vernova's Q3 backlog data will reveal whether the MV-UPS launch is winning orders against Vertiv. Finally, the sovereign-AI-capex theme has accumulated enough primary observables — the Korean sovereign axis alone spans the SK Group $500B+ multi-year anchor, NAVER-Brookfield 200 MW, GEV-LS Electric VSC-HVDC JV, and GEV-Studsvik-Samsung 1.2 GW Sweden nuclear — to warrant a dedicated rotation scan, and I am placing it on the must-add list.