Nvidia Capex Digest
TL;DR
The capex story's substance got stronger this stretch — record TSMC revenue, rising GPU rental prices, no wobble in Nvidia's own numbers — but the market's verdict split in two: the bond market now charges an explicit "funding tax" on AI debt, and Nvidia's equity has de-rated to a decade-low forward multiple despite sitting near all-time highs on price (nvidia-ai-capex-cycle-sustainability). A new bear vector also crystallized: not financing capacity, but customer concentration at the top of the stack — Steve Eisman's claim that the ecosystem depends on two labs, one of which he thinks is in trouble (ai-customer-concentration-openai-anthropic
). Last cycle's question was "who finances the next trillion dollars" — this cycle, both markets started answering with a price.
The bond market started charging for the buildout — a literal "funding tax"
Credit markets stopped treating big tech AI debt as free money this week.
"Bond markets are officially penalizing scale… institutional investors demand concessions to absorb the 'supply deluge.'" — Nicholas Mugalli, per Benzinga via TradingView
The data behind the quote: Alphabet's credit spreads widened from 56 to 71bps as issue sizes scaled from $1.25B to $3.5B, and Meta and Nvidia's $4B offerings priced 15-19bps wider — while hyperscaler debt grew 92% year-over-year versus 4% for the rest of the investment-grade market (nvidia-infrastructure-leasing-circular-financing). This extends last cycle's circular-financing ledger ($366B of commitments, $279B of supply commitments): the marginal dollar of AI capex now has a rising market-clearing price, which is exactly the mechanism through which an intact demand story could slow itself. Note it's still not distress — cloud fundamentals accelerated, with 52% cloud/AI revenue growth in 2Q26 — but the financing debate now has a ticker.
What to watch: Whether the scale premium widens further and whether the hyperscaler CDS basket (already above 100bps, an 8-year high) approaches the 150bps tripwire.
The equity market's verdict: price near highs, multiple at a decade low
Nvidia's de-rate is the most under-discussed datapoint of the stretch.
"The stock has de-rated pretty significantly, which suggests a healthy dose of skepticism that the company's current earnings power is sustainable… it tells you the market is expecting less than what the consensus is currently estimating." — Eli Horton, TCW senior PM, per Quartz via Yahoo Finance
NVDA trades under 17x forward earnings — half its 2025 multiple, down from above 25x as recently as May — while revenue and net income are each expected to grow roughly 90%+ this fiscal year, and the stock sits just 2.7% off its 52-week high (nvda-derating-decade-low-multiple). Critically, this isn't a semiconductor-wide phenomenon: the SOX is up ~76% YTD with Intel and AMD more than tripling, while Nvidia is up ~22% — near the bottom of its own index's leaderboard. The de-rate maps precisely onto the three open debates: circular financing, memory-cost margin compression, and custom silicon. Huang's counter is that Nvidia is "the world's first and only growth value stock" and "incredibly misunderstood."
What to watch: The forward multiple itself — a slide below 14x would imply the market is pricing an outright earnings break, not just skepticism; a recovery above 20x would signal the skepticism is clearing.
Eisman moved the bear case from the bond market to the income statement
The new critique isn't about how the buildout is financed — it's about who ultimately pays for it.
"70% of AI revenue of the hyperscalers is just from OpenAI and anthropic… the entire ecosystem is dependent upon two companies, and one of them I think is in trouble, which is OpenAI." — Steve Eisman, September 21 podcast, per 24/7 Wall St.
The chain as traced by 24/7 Wall St.: Nvidia has invested nearly $50B in the Frontier AI Labs; Microsoft's restructured OpenAI deal gave it a ~27% stake in exchange for OpenAI contracting $250B in incremental Azure services; OpenAI reportedly burns ~$12.3B a quarter (ai-customer-concentration-openai-anthropic). If OpenAI can't generate the cash to consume what it promised, Microsoft's capex intensity starts looking like credit exposure to a loss-making lab rather than leverage to a boom. The caveat: Eisman offered no source for the 70% figure — but the concentration question is sharper than anything a 10-K discloses, and the Anthropic IPO is now the first audited public look at one of the two load-bearing companies.
What to watch: OpenAI's next funding round — investors approached at as much as a $1.5T valuation, and whether it prices or collapses is the loudest possible datapoint for the whole chain.
The physical layer keeps printing records — and the depreciation bears are losing the rental market
While the financing layer frayed, the demand and supply chain hit its strongest readings yet.
"Demand for GPU time is surging to record levels." — John McPeake, Rosenblatt, per MarketWatch via Morningstar
The single best test of Burry's $176B depreciation thesis — that GPUs are economically dead in 2-3 years — is the rental market for old chips, and old chips are getting more expensive: Nebius is raising pay-as-you-go GPU prices in October, with increases covering four-year-old H100s, and CoreWeave is contracting new capacity at higher prices (gpu-depreciation-debate-hyperscaler-earnings). Upstream confirms it: TSMC's August revenue hit a record NT$514.8B, up 53.3% year-over-year — its fourth consecutive record month — with 2nm reportedly booked out through 2026, and ASE raised 2026 capex to $10.5B on advanced packaging (nvidia-ai-capex-cycle-sustainability
). The caveat: rising rentals validate the asset economics, not the renters' balance sheets — CoreWeave's debt-to-market-cap sits at 0.9.
What to watch: Any hyperscaler shortening GPU useful lives or taking an impairment — that, not commentary, is the decisive test of the Burry thesis.
What surprised us
- The de-rate was absorbed by earnings, not price. Nvidia is near a decade-low multiple and near all-time-high price simultaneously — the market's skepticism has been paid for entirely by earnings growth, which means the multiple is a pure bet on durability, not on the past (nvda-derating-decade-low-multiple
).
- The funding gap got a number: $5.5T. Mugalli's math — ~$1.0T of organic cash flow plus ~$0.4T of equity covering "barely a quarter of the bill," leaving $2.1T to bond markets — turns a rhetorical critique into arithmetic with a rising price attached (nvidia-infrastructure-leasing-circular-financing
- Four-year-old H100s getting price increases is the anti-Burry datapoint. Obsolescence curves don't run backwards; if the oldest mainstream AI chip is re-pricing upward, the load-bearing wall of the $176B earnings-quality attack is losing empirical support (gpu-depreciation-debate-hyperscaler-earnings
).
- The bear case relocated mid-cycle. Last cycle it was financing capacity; this cycle Eisman pointed at end-demand concentration — the revenue side of the buyers may be more concentrated than any filing discloses (ai-customer-concentration-openai-anthropic
).
Open threads worth a vote
- [watch] NVDA forward P/E falls below 14x (de-rate extends to pricing an earnings break) — the equity market's live verdict, with a recovery-above-20x trigger in the other direction
- [watch] OpenAI's next funding round prices (or collapses) at ~$1.5T valuation talk — the single sharpest test of the end-demand chain
- [watch] Nvidia supply & capacity commitments exceed $350B in next 10-Q — carried over from last cycle; the $279B stack is still the hardest number in the financing debate