Nvidia Capex Digest
TL;DR
The verdict reads intact for a sixth straight read, but the character of the debate shifted again: the bear case went forensic — Burry counting $3T of Big-5 commitments behind third-party debt, "the customer is now the lender" (nvidia-infrastructure-leasing-circular-financing) — while compute itself got a public, exchange-traded forward curve on Oct 5 (gpu-compute-futures-financialization
+1). Meanwhile the IPO exit channel closed further (Nscale joined the postponement list) and OpenAI formally replaced its listing with a $30B private round at $1.4T (ai-ipo-window-equity-exit-channel
). Demand confirms at the physical layer; the argument is now entirely about who's financing it and what the collateral is worth.
The bear case entered its forensic phase — and picked the right target
The argument moved from opinion to accounting autopsy. Michael Burry's Substack covering all five hyperscalers zeroes in on Oracle's $664B RPO (+$209B YoY) and the structure behind it:
"The customer is now the lender to Oracle." — Benzinga, Oct 5
He counts $3T of aggregate Big-5 purchase commitments, future leases, and guarantees backing third-party debt, and calls part of Oracle's future revenue growth a "financing gross-up, manufactured by the structure rather than by usage" (nvidia-infrastructure-leasing-circular-financing). The market has already voted on the most-levered leg: ORCL is down 50.72% over the last year while NVDA sits 0.5% below its high. The reason this matters is that the critique targets the exact seam the financing web depends on — credit support migrating from hardware to customer contracts, precisely the structure Reuters flagged when it noted GPUs won't underwrite like aircraft (nvidia-infrastructure-leasing-circular-financing
).
What to watch: whether Oracle's de-rating starts infecting neocloud and NVDA multiples — or whether the Amazon $8B Grace Blackwell SPV prices, which would extend the web rather than break it.
Microsoft's capex "cut" was an accounting migration, not a demand signal
The most-misread datapoint of the cycle resolves cleanly on inspection. Microsoft's calendar-2026 capex guide fell from ~$190B to ~$175B — but per Amy Hood, "Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged" (Yahoo/TIKR). The mechanism: data-center useful lives stretched from 15 to 25 years, and future data-center leases reclassified from finance leases (in capex) to operating leases (in operating cash flow). The cash didn't go anywhere — FY2026 capex rose ~80% to $115.95B, while FCF slipped to $66.99B and NTM Market-Cap/FCF reads 116.92x against a 42.44x mean (nvidia-infrastructure-leasing-circular-financing).
Demand underneath is untouched: commercial RPO at $678B, Azure guided to ~45% growth, and "customer demand continues to exceed available capacity" still the operative sentence (nvidia-ai-capex-cycle-sustainability). But note the irony: stretching useful lives to 25 years is exactly the practice Burry's depreciation critique targets — Microsoft just did it in the open, as an accounting optimization.
What to watch: the Oct 27 fiscal Q1 report — the first quarter carrying the new lease treatment. FCF above Q4's $19.6B suggests the squeeze is stabilizing; any GPU useful-life change would fire the Burry watch.
Compute got a price — and the forward curve is mildly bearish
The financialization milestone landed on schedule. Today CME and Silicon Data listed H100 and B200 Rental Index futures on NYMEX — the first exchange-traded forward curve for GPU compute, each contract a month's rent for one GPU (CME press release).
"Just as oil fueled the 20th century economy and evolved from spot trading into a global derivatives market, our futures contracts will now turn compute into a standardized, tradable commodity." — Pete Keavey, CME, via CME press release
The reference levels are now public record: B200 at $5.86/hour, H100 at $2.77/hour (FT). The telling detail: Silicon Data's B200 term curve sits below spot — the market is already pricing in easing once Rubin volume arrives (gpu-compute-futures-financialization+1). This cuts both ways for the capex story: a public curve gives lenders and insurers a hedge (potentially unlocking more debt), but it also makes residual-value assumptions testable by market price instead of dueling blog slides — the depreciation debate just got a referee.
What to watch: first-month volume and open interest, due for a check around Nov 5 — the test of whether institutional money treats compute as a real commodity, and whether ICE/Ornn or Nodal lists next.
The demand leg keeps confirming at the physical layer
While the argument rages upstairs, the orders haven't moved. TSMC hit another record as Apple, Nvidia, AMD, Qualcomm and MediaTek reportedly raised 2nm orders by 10–20%, pushing TSMC's end-2026 2nm capacity target to ~120,000 wafers/month and 2026 capex to $60–64B (invezz). Stifel's Sahej Singh framed the only debate left: "The debate, in our view, is not about demand but near-term margin durability" (nvidia-ai-capex-cycle-sustainability). The tape agrees: NVDA closed $233.95, RSI(14) at 83, TTM revenue $302.97B (+105.9% YoY) — a strongly extended tape, but not a cracking one (nvidia-ai-capex-cycle-sustainability
).
What to watch: TSMC Q3 earnings Oct 15, then Microsoft Oct 27, then NVDA Q3 FY2027 on Nov 18 — three checkpoints that test margin durability, lease accounting, and demand in that order.
The exit channel narrows as the private market formally substitutes
The IPO pullout wave now includes the neocloud leg. Morningstar confirms "SB Energy, NScale, and Holtec—all tied to the AI data center buildout—have also reportedly pushed back their IPOs in recent weeks" (Morningstar). Q3 saw just 32 US listings, the lowest since early 2024, and Renaissance Capital's diagnosis cut through the euphemism: postponed IPOs "may cite 'adverse market conditions' when the reality looks closer to normalized market conditions" (ai-ipo-window-equity-exit-channel).
The substitute is private money at rising prices: OpenAI is targeting at least $30B at a ~$1.4T pre-money valuation, up from $852B post-money in March, after pushing back its IPO (Reuters/Bloomberg). That's the opposite of a demand collapse — but every postponed IPO (Oura, Nscale, SB Energy, Holtec) is a priced-in-public no-vote on the valuations underpinning the buildout, and the longer price discovery stays private, the harder the eventual public read becomes (ai-ipo-window-equity-exit-channel).
What to watch: whether OpenAI's $30B round actually closes at the target valuation — the cleanest private-market check on end-demand financing.
What surprised us
- The depreciation debate became a literature argument. Burry answered Nvidia's Sept 27 retained-value slide with the 1968 Money Game — the 1960s computer-leasing bulls who said demand was "practically infinite" and put up "at least three percent cash" (Benzinga). Nvidia supplies the data; Burry supplies the analogy that the data itself is the bubble tell (gpu-depreciation-debate-hyperscaler-earnings
). Neither side has touched an order.
- The first public forward curve for compute opened bearish. The B200 term curve below spot means the market's opening position on Blackwell rental persistence is cautious — before the futures even traded (gpu-compute-futures-financialization
+1).
- Microsoft's "capex cut" was the cycle's biggest head-fake. The spending is unchanged; only the accounting container moved — and it moved in the direction Burry's useful-life critique predicts (nvidia-infrastructure-leasing-circular-financing
).
- S&P 500 net investment/GDP at 2.07% — higher than any time in four decades except the 2000-tech-bubble aftermath (Benzinga). That's the macro frame for the whole debate in one number.
Open threads worth a vote
- CME compute futures first-month volume/OI check (Nov 5) — did institutional money show up, and is the B200 term curve's implied easing deepening as Rubin ships?
- Microsoft fiscal Q1 2027 report (Oct 27): first quarter with new lease treatment — FCF vs $19.6B, capex vs the >$50B guide, and any GPU useful-life change.