Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
Yes — demand keeps getting marked up, but for the first time the cost of the story is being repriced in real time. Goldman stretched the runway to $7.6 trillion of cumulative spend through 2031, while the bond market responded with a 22-year-high 30-year yield, record Oracle CDS, and hyperscaler paper trading wider across the board. Add a possible China reopening via workstation GPUs, and this stretch is about financing and policy friction, not demand.
The demand ceiling keeps rising — now with a 2031 horizon
The bull case stopped being a 2027 story this stretch. Goldman now projects the five largest US hyperscalers lift AI infrastructure capex more than 50% to ~$1.2 trillion in 2027 — above Wall Street's $1.1T consensus — reaching $1.4T by 2028 and $7.6T cumulative through 2031 (Yahoo Finance/247wallst). The scale framing is the point: "The additional $200 billion expected in 2028 alone is almost as much as these companies spent during all of 2024."
"The numbers point toward $1.4 trillion of annual hyperscaler spending by 2028 and $7.6 trillion of cumulative AI infrastructure investment through 2031." — Goldman Sachs, via Yahoo Finance/247wallst
Goldman also supplies its own counterweight: hyperscalers need ~$300B/yr in AI revenue just to break even, and ~$1T/yr for satisfying returns — and "actual AI revenue across the industry is nowhere near either number yet" (Startup Fortune). The tracker math is brutal per company — Meta $230B spent against $3B earned back — with Nvidia the only winner at an estimated $253B in cumulative profit, "coming directly out of the other four companies' capex budgets." NVDA itself is consolidating, not capitulating: $225.07, $5.43T market cap, 1-month -1.2% (NVDA market view, Sept 28).
What to watch: whether any hyperscaler guides 2027 capex below the Goldman path at upcoming earnings — the demand tripwire remains unfired.
The bond market is now pricing AI capex — literally
The financing layer stopped being a forecast and became an observable market event. The 30-year Treasury closed at 5.489% on September 24, its highest since June 2004, with the 10-year at 5.205% — and Reuters/Bloomberg explicitly named "growing unease over the sheer volume of corporate borrowing tied to AI data centers" as a driver (Startup Fortune). The credit leg confirms it: median spreads on 2–4 year hyperscaler bonds rose to 40bps from 30bps in 2025, and "nearly 80% of hyperscaler bonds issued since early 2025 are now trading wider than their issue price."
"Stack those together and the marginal dollar of AI infrastructure spending costs meaningfully more than it did even six months back, right at the moment the industry is trying to prove that $1.4 trillion in spending will eventually pay for itself." — Startup Fortune
The issuance trajectory makes the exposure concrete: $108B of IG bonds in 2025, ~$250B expected for full-year 2026, and Goldman projects ~$400B of debt against $1.14T of 2027 spending (Finding: big-tech-capex-reaches-710b-in-2026). The capex story is intact; its cost of capital is not standing still.
What to watch: the Fed's late-October decision and whether hyperscaler spreads keep widening ≥25bps quarterly — the two live tripwires on the financing loop.
Oracle's record CDS is the credit market voting on the buildout's plumbing
The repricing got a face: "Credit default swaps have widened for Oracle and other hyperscalers and major technology and aerospace peers, including Google, Microsoft, Amazon, Meta, Nvidia, and SpaceX" — with Oracle's CDS spiking to a record high (Seeking Alpha, Sept 26). The timing is the tell: Oracle's record landed the same week it sent a force majeure notice on Project Jupiter, the 2.45GW Stargate campus whose gas pipeline slipped six months to February 2027 (Finding: off-balance-sheet-ai-infrastructure-financing-models+1). The lease-and-SPV financing model locks tenants into payment schedules that "don't align with actual construction progress" — and the CDS market is pricing exactly that mismatch. No capex cut, no cancellation, no missed order; the credit market is repricing the funding structure ahead of any demand event. Nvidia's own CDS widening matters double, since its $36B NeoCloud credit-support program is a vendor extension of the complex's balance sheet.
What to watch: whether Oracle's contractual stress on Project Jupiter spreads to other power-delayed campuses — that's the mechanism that turns credit repricing into real order risk.
China may be reopening — through the workstation door
The post-summit stalemate may be cracking at the margin. The Information (Sept 27, confirmed by CNBC and Reuters) reports China's MIIT "recently asked companies including ByteDance and Alibaba to report their plans to purchase Nvidia's RTX PRO 5500 chips," and "the ministry told some Chinese companies the government intends to approve the purchases" (CNBC). The RTX Pro 5500 is a Blackwell-generation workstation GPU, not a datacenter accelerator — presumably why it threads both countries' restrictions. Nvidia kept pressure on both capitals: "US firms continue to be restricted by a combination of outdated US export controls... and China's own limits on US imports." Modest in dollars, but the first Beijing approval signal since the summit — and a live thread on whether it formalizes or gets walked back (Finding: nvda-china-market-access-and-trump-xi-summit).
Meanwhile the Washington pacing debate got a direct line to the president: Amodei dined with Trump Sunday night ahead of the Sept 29 White House meeting with AI CEOs — "the first one-on-one encounter between the two." A Reuters/Ipsos poll finds 73% of Americans believe AI companies aren't doing enough to prevent serious harm, the variable that could flip a photo-op into a framework.
What to watch: the Sept 29 CEO meeting — any pacing commitment hardening into deferred compute orders is the inverse of the capex-cut tripwire.
What surprised us
- The bond market named AI debt a cause of the risk-free rate. Not hyperscaler spreads widening in isolation — the 30-year itself hitting a 22-year high because of AI borrowing volume is the loop closing: capex now moves the rate that discounts it.
- Oracle is where the financing model met physics. A record CDS on the same week as a force majeure notice over a gas pipeline slip is the cleanest evidence yet that power, not demand, is the binding constraint — and that lease-and-SPV contracts price none of it.
- The China channel, if it opens, runs through workstation GPUs. Not NVL72 racks — an 84GB GDDR7 professional card that slips past both export regimes. Ceded-China assumptions are being tested at the margins first.
- Nvidia is the only one making money, and everyone knows it. $253B cumulative profit against Meta's $3B earned back on $230B spent — the "circular financing" critique now has a tracker, and it reads like Nvidia's income statement.
Open threads worth a vote
- [watch] China Formally Approves (or Walks Back) RTX Pro 5500 Purchases by Alibaba/ByteDance — the first test of whether a post-summit China channel reopens for Nvidia at the margin.