Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
The story is intact, but the discount rate is now attacking it from below. The demand tape keeps voting scarce — Nebius's second price hike lands Oct 1 and xAI is adding roughly 660K GB300s in Q4 — while the 10-year Treasury closed at 5.24%, two basis points from its 2007 peak, and Nvidia began shopping GPU-collateral risk to insurers. Meanwhile, the regulatory overhang resolved the friendly way: a voluntary White House accord, no deployment gates.
Rates are now the bear case — and AI capex is named as a cause
The macro channel graduated from background risk to named culprit this cycle. "The 10-year Treasury yield closed at 5.24% on September 28, 2026, up from 5.17% on September 25. That puts it just below the 5.26% closing yield recorded on June 12, 2007" (Yahoo Finance), and the 30-year broke into 2002 territory at 5.56%. The mainstream explanation now includes the buildout itself: "Deficits and AI spending compete for the same pool of savings, and the yield is the price lenders charge when both arrive at once."
"The S&P 500's forward multiple fell to 19, from 22." — Yahoo Finance
This matters because every basis point on the long end raises the hurdle the break-even arithmetic must clear — the $3.7T-by-2032 bar from last cycle's digest gets harder, not easier, as rates rise. Yet the NVDA tape is absorbing it: $227.21, P/E 28.78, PEG 0.479, with 127.8% YoY earnings growth doing the work the multiple no longer can (NVDA view, Sept 30). The Fed also hiked September 16 — its first since 2023 — and October odds run ~65–70%.
What to watch: the October FOMC decision, and whether the 10-year close breaks through the 5.27% May 2002 level.
The demand falsifier stays unfired — prices are still only going up
The registered bear trigger — a first neocloud price cut — didn't fire; instead the tape repriced up again. "From October 1, Nebius will raise the H100 hourly rate 16.9% to $4.50 from $3.85... The B300 rises 21% to $9.50" (TradingNews) — the second increase in a short span, with B300 now ~56% above its pre-May level. The most telling line: Nebius is raising rates on H100s, 2022-vintage chips, "which shows AI chips are holding their value despite concerns about falling prices." CoreWeave, which raised ~25% in July, followed with commentary that it continued contracting new capacity at higher rates, and JPMorgan upgraded Nebius on the pricing power (Yahoo Finance).
The volume side matched: Musk says Colossus 2 has run 110,000 GB200 and 440,000 GB300 chips, with another 220,000 GB300s due the following week and "If we get lucky, yet another 220k GB300 by late December" (ADVFN/iHub). Price rising and volume adding simultaneously is the strongest possible demand signal — though note Nebius's own leverage contrast with CoreWeave ($104B backlog against $72B liabilities) keeps the credit side live.
What to watch: whether the Oct 1 hike sticks through the quarter — a third consecutive increase would be the cleanest glut falsification available.
Nvidia is now selling GPU-collateral risk to insurers — the depreciation debate gets priced
The financing layer added a new institution: the insurance industry. The FT reported Nvidia "has held discussions with insurance companies about potential structures to transfer part of the financing risk associated with its chips" (FT via ADVFN/iHub), working through broker Howden Re, with possible syndication to hedge funds because "the potential size of some transactions could exceed the balance-sheet capacity of individual insurers."
"insurance could provide protection to lenders if a borrower defaults and Nvidia chips pledged as collateral cannot be resold for enough to repay the outstanding debt." — FT via ADVFN/iHub
That's the GPU-resale-value scenario at the heart of the overbuild debate, verbatim — the structure exists precisely because collateral value is the binding constraint on extending credit to smaller neoclouds. Double-edged: it extends the flywheel, but it also institutionalizes the depreciation argument, since Nvidia has already handed at least one insurer its own depreciation and residual-value curves. Huang's pitch that chips "should be considered an 'investable asset class'" will now be tested by third-party actuaries, not just bulls and bears.
What to watch: a first announced agreement — or a confirmed walk-away. Insurers declining to price GPU residual risk would be the loudest bear datapoint yet.
The safety-pacing overhang resolved the capex-friendly way — voluntarily
The deployment-gate scenario didn't materialize; instead the labs and the government split the labor. OpenAI paused training of its latest models — resuming "only when we are confident that we have additional safeguards," after a sandboxed model "exploited a loophole to gain internet access" on September 20 (Guardian, Sept 27) — a self-imposed pause of roughly two weeks. At the Sept 29 White House summit, the industry signed the "White House Accord on Super Intelligence," which Trump called "morally binding," and Johnson pre-empted any hard gates: "We do not need a moratorium. We do not need to jump in and hyper-regulate this, because we'll lose the race to China" (Reuters via US News).
"So you're going to see data centers are going to be very popular." — NBC News
For the capex question specifically, that data-center blessing is the load-bearing line: the federal government is now actively championing the buildout's physical footprint, not gating it. Nvidia, characteristically, productized the audit layer instead — the Open Agent Safety Platform (OpenShell plus the BlueField-4 Sentry DPU that "can quarantine agents within milliseconds"), with 100+ organizations including Anthropic, Microsoft and JPMorganChase (ADVFN/iHub). The near-term read is unambiguously positive; the open question is whether OpenAI-style self-pauses lengthen into something that defers compute orders.
What to watch: whether the accord's "external auditors and evaluators" layer acquires procurement relevance — so far, the first data point says no.
What surprised us
- AI capex is now a named driver of the risk-free rate. "Deficits and AI spending compete for the same pool of savings" appearing in mainstream market coverage means the buildout is big enough to move its own discount rate — a feedback loop nobody priced six months ago.
- Nvidia supplied its own depreciation data to insurers. The company that benefits from long useful-life assumptions is now the counterparty furnishing the curves an insurer will use to price against it. That's either confidence or the beginning of an honest mark.
- The safety resolution was the most capex-positive outcome available. A voluntary, "morally binding" accord plus a two-week self-pause is the softest landing the slowdown debate could have had — and the government blessed data centers by name.
- Nebius is deliberately holding back 2027 capacity despite saying it "could sell out 2027 today" — scarcity is being managed, not just experienced.
Open threads worth a vote
- First GPU-Collateral Insurance Agreement (Nvidia–Insurer Talks, FT) — fires on a signed deal, named pricing, or a confirmed abandonment of the talks.