TL;DR
The capex story is intact — arguably stronger than last time — but the funding question is now the live one. VanEck's count puts 2026 hyperscaler capex at $844B, up 81%, with no cut anywhere in consensus, while the bear case shifted from "demand is slowing" to "the balance sheets doing the spending are increasingly leveraged to one assumption." Meanwhile the China option goes to summit week: Beijing, not Washington, is now the binding constraint on H200 access, and Thursday's Trump–Xi dinner is the largest unpriced swing factor.
The capex numbers accelerated through the selloff
The selloff didn't price a demand break — it priced a growth-rate inflection on spending that's still going up. VanEck's Nick Frasse put it directly: "The market didn't sell off because the build-out slowed down. It sold off because the build-out sped up." — VanEck thematic blog. The evidence: semiconductor earnings grew 131% in Q2 2026 while SMH fell 23% over the same stretch — the fourth-deepest of seven drawdowns since 2011 (VanEck).
The guidance stack is intact and mostly rising: Amazon ~$220B for 2026, Alphabet raised twice to $195–205B, Meta $130–145B, Oracle guiding FY2027 to $90–95B after a +235% YoY quarter (VanEck). VanEck's bottom line answers this cycle's brief head-on: "The growth rate is normalizing. The spend is not." — VanEck thematic blog. The cancellation tripwire from last cycle's digest — a capex cut or a cancelled data-center order — still has not fired.
What to watch: whether any company-level guidance revision in the coming earnings round is a real cut rather than a reclassification (Microsoft's $190B→$175B was the latter).
The bear case moved from demand to the balance sheet
The new bear arguments aren't about orders — they're about who's funding them and how concentrated the spend has become. Korea's NICE Credit Rating calculates that 82% of US AI infrastructure spending sits in five companies, with Amazon needing an additional $94.1B beyond internally generated funds to hit its FY2027–28 targets (NICE via BigGo). Apollo notes hyperscaler credit "rests on one consensus assumption," and the FT's bubble map marks Oracle and CoreWeave "bright red" (NICE via BigGo).
VanEck's own free-cash-flow math concedes the point: aggregate hyperscaler FCF troughs at −$50B in 2026 and −$67B in 2027 before recovering to +$489B by 2029 (VanEck). That's the honest shape of the bull case now: two years of negative cash flow, underwritten by debt markets that must keep believing. The McKinsey deployment stat is the other edge — 88% of organizations use AI somewhere, but only 6% have crossed a meaningful EBIT-impact threshold — runway if you're a bull, a ceiling on enterprise willingness to pay frontier prices if you're not (VanEck).
What to watch: hyperscaler debt issuance and any wobble in credit spreads — the funding flywheel, not order flow, is where the first crack would show.
Nvidia is now signing the cheques itself — beyond the hyperscalers
The demand base is widening past the Big-4, and Nvidia's own capital is doing the widening. In Australia, Nvidia and eight partners (IREN, NEXTDC, AirTrunk, Firmus, Sharon AI, ResetData, Megaport, CDC) announced up to 2GW of AI capacity by 2027 — more than double the nation's ~1.6GW of existing total data centre load — with Nvidia supplying the DSX platform and no hyperscaler on the partner list (Tech Insider). The structural read: "chipmakers, not just hyperscalers, are the ones signing the cheques for land and power" — Tech Insider.
The driver is sovereign AI — banks, miners, universities and government agencies wanting capacity kept inside Australian borders, with local Nvidia Cloud Partners unable to keep pace (Nvidia newsroom). Caveats matter: the 2GW is a working ceiling, not contracted capacity, and grid queues in NSW and Victoria are the binding constraint (Tech Insider). But it's concrete demand breadth from sovereign buyers and neoclouds — exactly the "new legs" pattern from last cycle, now with Nvidia as kingmaker rather than mere supplier.
What to watch: whether the DSX co-investment template replicates in other sovereignty-driven markets — each replication diversifies Nvidia away from Big-4 order concentration.
The China option goes to summit week
The H200 file has flipped: Washington licensed exports (at a 25% US fee, since Dec 8, 2025, Blackwell and Rubin excluded) but "the real slowdown has come from Beijing" — TipRanks. Chinese officials have limited big deliveries and pushed local firms toward homegrown suppliers (TipRanks).
The prize if Thursday's Trump–Xi dinner — with Huang, Altman, Cook and Qualcomm's Amon attending — moves Xi: KeyBanc's John Vinh estimates "Chinese tech giants would happily buy about 1.5 million H200 chips if given the green light," against a roughly $50B Chinese chip market (TipRanks). The asymmetry is unchanged from last cycle: roughly zero China revenue in the base case, ~$30B of order-book upside if Beijing relents, and a domestic-substitution headwind if the summit fails. Politically, Huang's sponsorship holds — Bessent repeated that the president is "completely aligned with Jensen Huang" (CNBC) — but the operative gate is now in Beijing, not the White House.
What to watch: the sidelines AI-executives meeting at Thursday's dinner — any H200 delivery signal is a direct TAM event in either direction.
What surprised us
- The selloff was a rate-of-change trade, not a demand trade. Earnings up 131% while SMH fell 23% is the cleanest statement yet that the market is pricing the deceleration of the acceleration — a much less scary bear case than the headlines implied (VanEck).
- Microsoft's "cut" wasn't one. The $190B→$175B move was a lease-treatment reclassification — the kind of headline that would have tripped the cancellation tripwire if read carelessly (VanEck).
- US corporate AI investment has surpassed housing for the first time (NICE via BigGo). That's a macro fact with bubble-echo implications no one in the debate has fully metabolized.
- Chip stocks rallied after the Fed hike. With the 10-yr at 5.01%, NVDA, MU and AMD bounced in the days after Sept 16 — discount-rate pressure is real but not currently binding on the trade (TradingKey).
Open threads worth a vote
- Does the cancellation tripwire stay untriggered through the next hyperscaler earnings round? — still the only clean falsifier, and Microsoft's reclassification shows why headline discipline matters.
- Does the Sept 24 summit unblock the H200-China channel, or harden it into a structural TAM loss? — KeyBanc's 1.5M-chip estimate is the number to hold.