Nvidia Capex Digest
TL;DR
The verdict reads intact for a seventh straight read, but the pressure has migrated from demand to the price of the buildout (nvidia-ai-capex-cycle-sustainability). The marginal AI borrower now pays ~11% — JPMorgan's Volta loan is a new high in all-in yield for the space — while demand confirms at a new layer: emerging-market sovereign-adjacent buyers signing binding Rubin orders (nvidia-infrastructure-leasing-circular-financing
). Meanwhile memory scarcity crossed from supplier economics into Nvidia's own price sheet, the first visible cost-tax on the buildout's unit economics.
The financing web is pricing its outer edge — and the biggest banks are still expanding
The story is no longer whether money exists for AI infrastructure, but what the marginal unit costs. JPMorgan is pitching ~11% yield on a $5B leveraged-loan sale for Volta Infrastructure — "one of the highest borrowing costs seen in the market for risky debt that's become popular for financing the AI boom," per Bloomberg a new all-in high for the space (Yahoo/Bloomberg). The structure is the tell: a $3.7B tranche buys ~36,000 Nvidia GPUs that themselves serve as collateral, and the deal only underwrites because "Anthropic has already signed a six-year agreement for compute capacity at the site" (nvidia-infrastructure-leasing-circular-financing).
The counterweight arrived the same day: KBW reads BofA's $250B infrastructure pledge, Morgan Stanley's $1.5T initiative, and Nvidia's $500B Wall Street platform as "clear signals of confidence from the largest lenders that this capex cycle has room to run beyond 2026," with C&I loan growth accelerating to a double-digit pace (EnergyNow/Reuters). Both facts are true at once — the edge pays 11% while the core expands commitments. That's a buildout funded at a rising cost of capital, not a refused one. Note also the buyer base is narrowing: the "strip" structure excludes CLOs, so this is genuine balance-sheet money (nvidia-infrastructure-leasing-circular-financing).
What to watch: the Volta loan itself — a pricing at ~11% institutionalizes double-digit AI debt costs; a pullback would be the sharpest financing-leg bear datapoint yet.
Demand is diffusing beyond the hyperscalers — and the aggregate number keeps climbing
The bull case's demand-diffusion leg got its cleanest confirmation yet. AM Intelligence (the AI-infrastructure platform of Greenko's promoters) signed binding orders for 20,000 Rubin GPUs for India and Malaysia, lifting committed frontier compute to ~29,000 GPUs and nearly 100 MW, with a plan to add 300 MW — "requiring over $20 billion in capital expenditure" — over 15 months (ETTelecom). Compute demand is no longer confined to four US hyperscalers and two labs (nvidia-ai-capex-cycle-sustainability).
The aggregate is moving the same direction: Goldman Sachs now projects hyperscaler capex may hit $1.2 trillion in 2027, up from the ~$725B the big four guided for 2026 (MarketWatch). The equity agrees — NVDA closed $238.90 at its 52-week high, $5.65T market cap, with four straight EPS beats (nvidia-ai-capex-cycle-sustainability).
What to watch: whether the demand-diffusion layer keeps signing binding orders or whether rising debt costs (Volta at 11%) starts choking off exactly these non-hyperscaler buyers.
Memory scarcity just showed up in Nvidia's own price sheet
The cost side of the buildout got its first clean datapoint. Nvidia is launching a 64GB DGX Spark at $4,999 — "That is $1,000 more than the original 128GB model's $3,999 launch price, despite offering half as much memory" (Yahoo/GuruFocus). The 128GB model now retails ~$6,950, roughly 75% above debut, and the driver is structural: "Micron has warned that memory shortages could worsen through 2027 and 2028, with no clear timeline for supply and demand to return to balance" (micron-ai-memory-pricing-power).
So far this is a transfer, not a brake — demand strength is letting Nvidia pass the cost through. But it means the memory layer is now capturing scarcity rent from the buildout, and if prices stay elevated, "customers may become more selective about how aggressively they expand AI infrastructure" (micron-ai-memory-pricing-power). The standing tripwires: Nvidia gross margin (74.7% as of the July quarter, with the bottoming watch at ≤72%) and whether HBM shortage forces a downgraded Rubin Ultra configuration.
What to watch: the Rubin Ultra memory spec at launch — a downgraded configuration would be the first demand-visible casualty of the shortage.
The bear case is a timing call, not a demand disproof
The loudest bear this week put a number and a date on the pop: Panmure Liberum's Joachim Klement sees an AI downturn taking the S&P 500 to 5,000 (~-35%) by end-2027, and "I think the entire bubble will pop either in 2027 or 2028" (MarketWatch). His mechanism is the "capex pickle": investors punish hyperscalers for raising capex, but with Goldman's $1.2T 2027 figure circulating, any capex shortfall would damage sentiment toward the beneficiaries just as badly. His advice: "cash and fetal."
Nothing in this identifies a cancelled order — it's a sentiment-timing mechanism, and it cuts against the forensic accounting critique from last cycle rather than reinforcing it. Meanwhile the macro pressure builds quietly: the 10-year sits at 5.28%, 22bp from the Man Group "AI capex crack" tripwire at 5.5% (nvidia-ai-capex-cycle-sustainability).
What to watch: the 10-year crossing 5.5% — the financing-cost tripwire that would bind all of this together.
What surprised us
- The strip structure is a quiet admission. JPMorgan excluded CLOs — the biggest leveraged-loan buyers — from the Volta deal, meaning the outer edge of the financing web is being funded by bespoke balance-sheet appetite, not machinery demand (nvidia-infrastructure-leasing-circular-financing
). That's a narrower buyer base than the headline $625B two-year borrowing total implies.
- A 64GB computer costing more than a 128GB one is the cleanest inflation datapoint of the cycle. Memory scarcity stopped being a Micron margin story and became a line item in Nvidia's price sheet — the first cost-push evidence that survived contact with demand (micron-ai-memory-pricing-power
).
- India and Malaysia are now demand anchors. The diffusion trade everyone talks about showed up as binding Q2-2027 Rubin deliveries from a Greenko-linked platform — not a memorandum, an order (nvidia-ai-capex-cycle-sustainability
).
Open threads worth a vote
- [watch] Volta $5B leveraged loan prices (or fails) at ~11% all-in yield — commitments were due within days of the Oct 5 report; a pricing confirms balance-sheet appetite for GPU-collateralized paper at double-digit yields, a pullback signals the leveraged market refusing the web's outer edge.