Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
Yes — and the demand leg got marked up again: Goldman now sees $1.2T of hyperscaler capex in 2027, GPU rentals are backwardated, and OpenAI's own chip benchmarks still can't displace Nvidia from the order book. The action moved to the financing layer: Goldman's bond desk is underweighting hyperscaler credit, the AI bond wave is being blamed for 2007-era Treasury yields, and Burry's write-off thesis got its first market rebuttal in the form of four-year H100 contracts. The stress is in delivery and financing, not demand.
Demand got re-priced up — and the rental market is now the strongest proof
The demand case hardened from guidance into real-time pricing. Bloomberg (Sept 25) reports spending by the five largest US hyperscalers is "set to increase by more than half next year to $1.2 trillion," with Goldman's Ryan Hammond putting the break-even test on record: hyperscalers "will need to generate roughly $300 billion in annual AI revenue in the next few years simply to break even on their investments" (Finding: big-tech-capex-reaches-710b-in-2026).
The cleaner signal came from SemiAnalysis's ClusterMAX 3.0 (77 providers tested, 200+ customers interviewed). Researcher Sam Harshe described "ridiculous backwardation," with customers paying "steep premiums for compute available within weeks rather than waiting several months" — and buying "even if a provider's Kubernetes... is 'completely broken' and its storage 'barely works'" (Yahoo/Benzinga). The detail that matters most for the depreciation debate: "companies are signing four-year contracts for H100 capacity, evidence that older Nvidia GPUs can retain significant earning power while compute remains scarce."
That's a demand read no guidance document can fake — buyers are locking multi-year terms on previous-generation silicon. Fortune added the macro milestone: data center spending now exceeds US residential housing investment (Fortune, Sept 20).
What to watch: whether any neocloud cuts prices this quarter — the first cut, not a hike, remains the cleanest demand falsifier.
The financing layer is now repricing ahead of demand
The credit market delivered its first explicit sell-side call on AI debt supply. Lindsay Rosner, head of multi-sector fixed income at Goldman Sachs Asset Management: "We believe there will be a lot of hyperscaler issuance... For that sector at large we are underweight knowing more issuance will come" (TradingView/GuruFocus). Brookings research in the same report estimates US AI infrastructure could require "roughly $10.3 trillion of investment through 2032," with risk migrating "away from Big Tech balance sheets and into less-transparent financing structures."
The neocloud equities show the market discriminating on financing structure, not GPU demand. CoreWeave carries $51.61B of debt at 13.6x EBITDA with Q2 free cash flow of -$5.74B, and trades 38.8% below its high; Nebius has worse leverage at 39.0x debt/EBITDA but its equity is +8.6% on the month near its high (Finding: off-balance-sheet-ai-infrastructure-financing-models+1). Same demand environment, opposite verdicts — the market is pricing the balance sheet, not the GPUs.
What to watch: hyperscaler bond issuance volumes — if supply keeps outpacing cash generation, funding cost becomes the new valuation debate, exactly as Goldman warns.
The capex story is now moving its own discount rate
The loop from last cycle's digest tightened: the 10-year Treasury printed 5.18% on Sept 24, the highest since 2007, with TechTimes attributing the move substantially to "a $570 billion AI bond" wave on top of the Fed's Sept 16 hike to 3.75%–4.00% (TechTimes, Sept 25; Finding: first-rate-hike-since-2023). NVDA is absorbing it anyway — $225.07, 13% above its 200-day average, P/E 28.5 — but every leg higher on the long end compresses growth multiples and raises the cost of the debt-funded slice of capex. The FOMC's next move (late Oct or mid-Dec) is now the single macro lever on this story.
What to watch: a second 2026 Fed hike — it's the one variable that hits every leg of the buildout simultaneously.
Custom silicon got its first measured win — at the margin, not the replacement
OpenAI published benchmarks for Jalapeño, its first custom inference chip, and the appendix names Nvidia's current flagships as comparators: GB200 and GB300. The claims: "Jalapeño delivered 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems" (OpenAI). Critically, the benchmarks ran on InferenceX, SemiAnalysis's public yardstick — not a house metric.
But read the hedge that protects Nvidia's order book: "We will continue to widely deploy accelerators from NVIDIA and other partners for both training and inference workloads," with Jalapeño deployment only beginning "by the end of the year" (Finding: custom-silicon-and-datacenter-hardware-competition+2). Meanwhile the Rubin ramp is producing commercial systems now — Supermicro announced it is shipping Vera Rubin NVL72 racks (72 Rubin GPUs, 20.7 TB HBM4 per rack). This is substitution at the margin, not displacement, and it's inference-only.
What to watch: whether Jalapeño actually enters production deployment by end of 2026 — the benchmarks are claims until silicon ships.
The overbuild debate got a champion and an immediate rebuttal
Michael Burry warned Big Tech's $3T AI infrastructure boom "could end in major write-offs, echoing 2000's capital cycle" (Seeking Alpha, Sept 25). The market's rebuttal arrived in the same dataset: SemiAnalysis's four-year H100 contracts, with the framing "So long as demand outstrips supply, old GPUs keep earning" (Finding: gpu-deployment-gap-overbuild-debate). Polymarket prices an AI-industry downturn by Dec 31 at about 10%.
The decisive fact: no hyperscaler or neocloud has announced a GPU write-down, deployment-percentage disclosure, or depreciation-schedule reversal. Until one fires, this is a narrative debate, not a fundamental event.
What to watch: the first GPU inventory write-down or depreciation-schedule disclosure from any major player — the tripwire that would convert this debate into fundamentals.
What surprised us
- Even broken clouds are selling. SemiAnalysis found buyers paying premiums for compute on providers whose Kubernetes is "completely broken" — scarcity has overwhelmed quality, which is about as clean a demand signal as markets produce.
- The bond market moved first. Before any equity crack, Goldman's fixed-income desk underweighted hyperscaler credit on expected issuance — the financing layer is repricing the boom before the demand layer shows any stress.
- The Rubin production-cut rumor sits awkwardly against reality. An unverified Chinese-language report claims Nvidia cut 2026 Rubin targets 25% on HBM4 delays — while Supermicro is already shipping Rubin racks. One of these is wrong, and the verification thread is open.
- Data centers now out-spend American housing. The macro crossover happened quietly, but it's the kind of statistic that reframes the political economy of the buildout.
Open threads worth a vote
- Verify the reported 25% cut to Nvidia's 2026 Rubin production target on HBM4 delays — if real, it constrains the FY2027–28 revenue ramp; if false, it's noise.