Nvidia Capex Digest
TL;DR
The verdict stays intact, but this cycle the pressure moved from the demand side to the financing side: the 10-year Treasury hit a 24-year high of 5.33% — just 16bp below the 5.5% "AI capex crack" tripwire — while Michael Burry put ~$1.1B of notional behind a 2027 break (nvidia-ai-capex-cycle-sustainability). Meanwhile the demand leg got its strongest second-order confirmation yet: Micron printed $54.23B of quarterly revenue with HBM sold out deep into next year (micron-ai-memory-pricing-power
). The IPO window that was supposed to validate all of it just slammed shut.
The demand leg keeps hardening — this time from the memory side
The buildout is no longer asking for faith; it's showing up in suppliers' order books. Micron's fiscal Q4, reported Sept 30, printed $54.23B of revenue against $11.32B a year earlier — nearly 5x — with non-GAAP gross margin above 86% and data center sales up 11x (micron-ai-memory-pricing-power):
"> $100 billion in total cumulative long-term 'take-or-pay' contract value locked in across top hyperscale cloud providers and AI GPU customers" — Invezz
That matters because HBM tightness only binds if GPU shipments keep rising — hyperscalers locking take-or-pay floors on memory is second-order confirmation the accelerator buildout is still accelerating (micron-ai-memory-pricing-power). Nvidia's own numbers agree: TTM revenue of $302.97B (+105.9% YoY), 74.7% gross margin, four straight EPS beats, and the $150B buyback expansion to $235B (nvidia-ai-capex-cycle-sustainability
). The symmetric risk is worth naming: if hyperscalers ever cut GPU orders, Micron's tightness becomes inventory build and the pricing-floor narrative breaks (micron-ai-memory-pricing-power
).
What to watch: Micron's Q1 FY2027 guide of $61.5B ±$1.5B — the next checkpoint on whether memory tightness persists.
The price of money is the new front — and the tripwire is 16bp away
The stress in this cycle arrived through the bond market, not the order book. The 10-year Treasury hit 5.33–5.342% on Oct 1 — the highest since 2002 — with the 30-year at 5.67% and yields "up for a 7th consecutive month for the first time since 2011" (nvidia-ai-capex-cycle-sustainability). That leaves only ~16bp to the 5.5% Man Group "AI capex crack" level this digest has been tracking since last cycle — and the driver is energy, not the Fed: Deutsche Bank's Jim Reid notes Brent crude is "up +42.0% from its lows at the end of June," which drove "a major global bond selloff" (nvidia-ai-capex-cycle-sustainability
).
Burry's move is the same thesis in put form. He closed his NVDA/MU/NBIS/PLTR/CRWV shorts and replaced them with ~$1.1B of put notional — including NVDA September-2027 puts struck in the mid-$100s against a $229 close — after pulling his AI-bubble timeline from 2028 to 2027 (nvidia-ai-capex-cycle-sustainability). The Ares line he quoted is the sharpest statement yet of the cancellation-clause risk:
"It would take only a season in which AI revenue disappoints the capital expenditure underwriting it... The legal documents contemplate that decision." — Startup Fortune
What to watch: The 10-year crossing 5.5% — the registered tripwire, last read 5.33% and closing fast.
The IPO window shut — the equity exit channel that validated AI capex is closed into 2027
Last cycle's digest flagged Anthropic's listing as the hardest public-market test of AI capex economics. That test just got harder to run. OpenAI postponed its IPO to 2027 and is targeting a ≥$30B private round at ~$1.4T; an internal FT-sourced presentation expects cumulative negative free cash flow of $278B through end-2030 (ai-ipo-window-equity-exit-channel). Around it, the postponements stacked: Oura pulled its $2.2B IPO the day before listing, SB Energy delayed a $5B IPO after a $3.2B net loss on under $139M of H1 revenue, and Nscale's roadshow is expected to slip (ai-ipo-window-equity-exit-channel
).
Anthropic still intends to go — but moved from October to November, after the midterms (anthropic-ipo-public-market-test). Its reconciled numbers soften the headline: the $42B net loss includes a ~$34B non-cash accounting charge; the real operating loss was $8.06B on ~$4.6B of revenue (anthropic-ipo-public-market-test
). But the filing also surfaced the cycle's most concrete cancellation risk: agreements to pay SpaceX up to $84.5B for Nvidia GPU capacity — double what SpaceX's own S-1 disclosed — terminable by either party on 90 days' notice (anthropic-ipo-public-market-test
). That 90-day clause is exactly what the Ares/Burry argument points at.
What to watch: Anthropic's November pricing — the last, largest public validation available before the window reopens in 2027.
What surprised us
- The stock barely flinched at any of this. NVDA closed at $228.38, 2.9% below its 52-week high, at 28.91x trailing earnings — a 24-year-high 10-year and a $1.1B bear put position left no mark (nvidia-ai-capex-cycle-sustainability
).
- Micron beat enormously and the stock fell anyway. After a 200%+ YTD run, Invezz reads the sell-off as "positioning, not fundamentals" — even a 5x revenue print can't clear the bar (micron-ai-memory-pricing-power
).
- SoftBank is the chain's weakest visible link. OpenAI's largest investor, committed up to $30B in tranches, just issued $11.15B of BB+ rated bonds to fund the third tranche — a leveraged holder of the leveraged buyer (ai-ipo-window-equity-exit-channel
).
- The rates driver is oil, not the Fed. Markets price only ~39% odds of an October hike; the yield surge is energy inflation (nvidia-ai-capex-cycle-sustainability
Open threads worth a vote
- [watch] 10-yr Treasury crosses 5.5% — Man Group's "AI capex crack" level (last read 5.33%, ~16bp away) (/topics/019e95a0-30d1-7d08-ae5b-10ac242b9d5c#threads)
- [watch] Nscale IPO: roadshow expected to slip, but still on as of Oct 1 (/topics/019e95a0-30d1-7d08-ae5b-10ac242b9d5c#threads)