Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
The demand base finally got its hard disclosure: Anthropic's S-1 shows revenue accelerating violently — $9B run-rate in December, above $65B by July — but against $518B of mostly fixed compute obligations. Nvidia answered the bear case with the largest buyback authorization in history, while Burry moved his bubble clock up to 2027 and the financing flywheel reached its most circular turn yet: Broadcom lending Anthropic up to $42B, and Nvidia reportedly weighing a $10B anchor stake in Anthropic's own IPO.
The Anthropic S-1 turned the demand-base question into arithmetic — and the arithmetic cuts both ways
The first audited disclosure from a pillar of the demand base confirms acceleration, not the plateau the bears flagged. Per the prospectus-based reporting: 2025 revenue ~$4.6B, annualized run-rate ~$30B by April 2026, and above $65B by July 2026 (anthropic-ipo-compute-obligations). Against that sits the obligation stack:
"Anthropic expects to funnel $518 billion on cloud, computing, and infrastructure obligations in the coming years." — Anthropic's $2 trillion IPO comes with a $518 billion bill (Yahoo Finance/TheStreet)
The Motley Fool's read adds the teeth: $161.2B of Broadcom-linked equipment leases "that Anthropic has limited ability to cancel," with minimum-spend requirements on hyperscaler commitments even if capacity goes unused — roughly 80% of ~$414B effectively fixed (anthropic-ipo-compute-obligations). So the deceleration thesis has no confirmed-data support yet, but the same document quantifies the asymmetry: fixed obligations against usage-based revenue (82.6% of sales) and an $8B operating loss.
What to watch: the IPO itself, expected in November after the midterms — the cleanest single market test of whether public investors will underwrite that structure.
Nvidia's balance sheet cast the biggest vote yet — a record buyback on top of an intact P&L
The cash-engine leg of the story isn't just intact; management just staked the most capital in corporate history on it. Nvidia's board authorized a $150B increase to its buyback — the largest single authorization ever — bringing the remaining program to $235B through fiscal 2028, eclipsing Apple's $110B record (nvidia-record-q1-results-and-trillion-dollar-order-book+1).
"Our cash generation gives us the capacity to invest in the technologies that advance this transformation and return capital to shareholders. This authorization reflects our confidence in the long-term opportunity ahead." — Nvidia press release, Sept 28
The pace is funded, not levered: $39B repurchased in H1 FY2027 alone against $74.4B of operating cash flow, with debt/EBITDA at 0.2x. The stock closed the week at $233.95 with a $5.65T market cap and RSI(14) at 83 — deep overbought, with no cushion for a guidance wobble at Q3 earnings (nvidia-record-q1-results-and-trillion-dollar-order-book+1). Reuters' counter-framing is the honest one: the buyback is also "a big bet on its own stock as intensifying AI-chip competition weighs on its performance relative to peers" — management buying its own demand story at ~29.6x trailing earnings while Burry holds mid-$100s puts.
What to watch: whether the buyback pace holds through the November earnings print — it's the balance-sheet vote on the demand base, updated quarterly.
Burry moved his clock up and attacked the depreciation defense with 1968
The bear case is now a dated, priced, falsifiable position rather than a narrative. Burry said he's "moving timelines up" from a 2028 base case, covered his outright shorts, and replaced them with puts on Nvidia (mid-$100s strike, September 2027 expiry), Palantir, Micron, Nebius, Oracle and the Nasdaq 100 — the index puts struck in the $24,000s against ~$30,300, an implied ~20% drawdown (gpu-deployment-gap-overbuild-debate).
"If the spending stops or slows, it all comes apart." — Business Insider on Burry, Sept 29
His second strike hit Nvidia's Sept 27 investor slide on retained GPU value: his Oct 2 post "Don't Believe Your Lyin' Eyes" quoted Adam Smith's The Money Game on the 1960s computer-leasing mania — the same "practically infinite" demand rhetoric, the same levered leasing structure, 58 years ago (gpu-deployment-gap-overbuild-debate). The demand-side counter strengthened in the same week: AWS hiked EC2 Capacity Block prices ~15% across A100–B300, the first hyperscaler rental gauge, joining Nebius's two hikes — while the neocloud price-cut tripwire still hasn't fired.
What to watch: any hyperscaler capex cut, GPU write-down, or deployment-gap disclosure — until one lands, both books stay open.
The flywheel reached inside the demand base: vendor lending, an IPO anchor, and an unpriced SPV
The financing layer extended in a genuinely new direction this cycle — the chip supply chain is now lending to the labs that buy the chips. Reuters' filing coverage reports Broadcom has agreed to lend Anthropic up to $42 billion, while the Apollo/Blackstone $35B tranche for Anthropic's chip installations priced at 8.5%, about 275bps over Treasuries (off-balance-sheet-ai-infrastructure-financing-models+1). And the most circular turn of all:
Nvidia is reportedly considering up to a $10B anchor investment in the IPO — the chipmaker's balance sheet directly propping the valuation of its own demand base. — off-balance-sheet-ai-infrastructure-financing-models
+1
Amazon's $8B Grace Blackwell sale-leaseback into an SPV remains unpriced — terms, spread and residual-value assumptions undisclosed — against ~$220B of planned 2026 capex, with Amazon aiming "to strengthen its balance sheet" (off-balance-sheet-ai-infrastructure-financing-models+1). That pricing is the datum that would set the market's clearing yield on installed-GPU collateral. Scarcity pricing (the AWS hike) and off-balance-sheet ownership moves arriving the same week is the tell of a maturing compute-finance ecosystem — and the residual-risk question from last cycle's digest is now fully inside the demand base, not just around it.
What to watch: the first disclosed Amazon SPV term sheet and any $500B-platform guarantee level — both would finally put a price on who eats GPU residual risk.
What surprised us
- The S-1 vindicated the bulls on trajectory and the bears on structure — in the same document. Revenue went $9B → $30B → $65B+ in seven months, yet ~80% of the obligation stack is effectively fixed against usage-based revenue. Both halves of the prior digest's debate got their evidence from one filing.
- Nvidia anchoring Anthropic's IPO would be circularity in its purest form — the supplier underwriting the valuation of a customer that committed $414B to buy its class of hardware. A decline would be equally informative about how Nvidia itself values demand durability.
- Burry's 1968 parallel landed harder than expected because the analogy isn't just sentiment — it's the leasing-financing structure, which is exactly what the Broadcom $42B loan and the Apollo/Blackstone tranche are building.
- The buyback is a third front in the debate nobody asked for: management returning $46.1B to shareholders in six months while the credit market refuses to underwrite the assets behind the cash flows.
Open threads worth a vote
- [watch] Nvidia Confirms or Declines $10B Anchor Investment in Anthropic IPO — the cleanest single signal on whether Nvidia's balance sheet will directly back its own demand base's valuation.