Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
The capex story is intact on every demand gauge — Goldman lifted its trajectory to $1.2T for 2027 and $7.6T through 2031, Amazon raised EC2 ML reservation prices ~15%, and NVDA printed its first record high since May on a Morgan Stanley Top Pick reinstatement (big-tech-capex-reaches-710b-in-2026, nvidia-record-q1-results-and-trillion-dollar-order-book
+1). But the financing layer just crossed its biggest threshold: Amazon is negotiating to securitize ~$8B of installed Grace Blackwell chips, and Goldman's own asset-management arm has gone underweight hyperscaler debt on issuance supply. The bear case also went mainstream — Fidelity's macro desk and Morgan Stanley are now carrying it.
The demand tape keeps printing up — and now the biggest buyer is raising prices too
The capex trajectory is being revised in one direction only. Goldman's Ryan Hammond now projects the five largest hyperscalers at ~$800B in 2026 (up 94% from 2025), $1.2T in 2027, and $1.4T in 2028, with backlogs exceeding $1.5 trillion (big-tech-capex-reaches-710b-in-2026).
"Hyperscaler cloud revenues have accelerated sharply this year, annualizing roughly $70 billion above the pre-AI trend in Q2 2026, and announced revenue backlogs exceed $1.5 trillion." — Ryan Hammond, Goldman Sachs, via Investing.com
The pricing gauge followed: days after Nebius's October 1 hike took effect, Amazon announced a ~15% increase on EC2 Capacity Blocks for ML — the reservation product spanning Nvidia chips from the A100 to the B300 — with the price-cut tripwire this topic watches for still untriggered (big-tech-capex-reaches-710b-in-2026). When the largest buyer of AI compute raises reservation prices while committing to $200B of 2026 capex, that's scarcity being priced at the retail layer, not demand softening.
What to watch: hyperscaler Q3 earnings in late October — whether capex guidance follows Goldman's trajectory upward or the buyback-to-capex rotation (top-5 buybacks fell 64% year-over-year) accelerates.
The financing flywheel just reached hyperscaler scale — and Goldman's credit desk is stepping back
The off-balance-sheet financing model crossed its biggest threshold yet: Amazon is sounding out investors on moving ~$8B of Grace Blackwell chips already deployed across a dozen-plus US data centers into a special-purpose vehicle, then leasing them back (off-balance-sheet-ai-infrastructure-financing-models+1).
"If completed, this could ease AMZN's AI-capex burden without reducing NVDA demand, but shifts GPU [depreciation risk to investors]." — syndication comment, via Quartz
The need is visible in the cash-flow math: Amazon's trailing-twelve-month free cash flow has gone negative at roughly -$7.6B, Meta's quarterly FCF collapsed to $784M, and Amazon has issued the equivalent of nearly $100B in bonds this year (off-balance-sheet-ai-infrastructure-financing-models+1). The buyer side is balking — Goldman Sachs Asset Management's Lindsay Rosner: "We believe there will be a lot of hyperscaler issuance... For that sector at large we are underweight knowing more issuance will come." Nvidia's demand is untouched by all of this — Amazon keeps using every chip — but the question of who prices GPU residual risk five years out now moves to credit investors, against the 89–91 cents template Oracle's Jupiter loan already set.
What to watch: the Amazon SPV's pricing — whether GPU paper closes at investment-grade-adjacent spreads or airline-asset spreads tells you which world the credit market thinks it's in.
The overbuild critique went mainstream — and the demand base's own S-1 leaked
The "dead money" argument is no longer a blogger position. Fidelity's Director of Global Macro Jurrien Timmer said the AI trade "has been dead money for more than three months," citing token expenditures and H100/A100 lease rates "flat to down" (gpu-deployment-gap-overbuild-debate).
"More than half of the GPU servers sold between 2026 and 2028 might not have anywhere to be plugged in." — Morgan Stanley power-gap research, via Ed Zitron's "Dead Money"
The arithmetic behind the critique hardened too: Zitron pegs annual AI revenue across Google, Microsoft, Amazon, Oracle and SpaceX at ~$183B — 64% of it from Anthropic and OpenAI — against Goldman's $308B breakeven, and the leaked Anthropic S-1 shows the biggest demand pillar spent $12.65B in operating expenses to make $4.6B of revenue (gpu-deployment-gap-overbuild-debate). There's also a direct hit to the bull case's rental-rate evidence: SEC-filed five-year contracts show B300s locking in $3.30–4.30/hour, less than half the B200 spot index bulls cite. Spot and contract are telling different stories, and the contract tape is the one that amortizes capex.
What to watch: whether the rental-rate contradiction (Nebius/Amazon hikes vs. SEC-filed $3.30/hr B300 contracts) resolves — it's the single cleanest datapoint separating demand scarcity from contract-market distress.
The equity doesn't care — NVDA made a record high anyway
While the credit layer wobbled, NVDA rose to an all-time high above $237 on October 2 — its first record since May — after Morgan Stanley reinstated it as semiconductor Top Pick with a $300 target (nvidia-record-q1-results-and-trillion-dollar-order-book+1).
"[The land, power, data-center-shell, financing and component constraints] all play to their strengths." — Morgan Stanley, via TradingView/Stocktwits
Barclays' "napkin math" added roughly $30B of hyperscaler revenue upside to its 2026–27 estimates, and Nvidia completed its own final $10B investment into OpenAI the same week (nvidia-record-q1-results-and-trillion-dollar-order-book+1). The bull logic has adapted to the constraint story rather than denying it: when megawatts, not chips, are the bottleneck, the vendor best positioned across the stack wins. The tension with the sections above is stark — the equity is pricing Nvidia as the constraint-arbiter while the credit market prices the customers' financing as the risk.
What to watch: Nvidia's Q3 FY2027 print on November 17, guided to $108B revenue at ~74% gross margin — the next hard test of the order book.
What surprised us
- The same week produced both a record high and an underweight. Goldman's strategists project ~$400B of 2027 hyperscaler capex funded by debt while GSAM goes underweight hyperscaler credit — the house is long the equity story and short the financing of it, and both can be right for a while.
- The S-1 leak changed the debate's quality. Anthropic spending $2.75 to make a dollar isn't a vibe; it's an audited number on the demand pillar that's 64% of the bear case's revenue arithmetic. The overbuild debate now has filings in it.
- Contract rates contradict spot rates. The bulls' rental-hike tape (Nebius, Amazon) and the bears' SEC-filed $3.30/hr B300 five-year contracts can't both describe the same market — someone's evidence is structural and someone's is episodic, and it's not yet clear which.
- Morgan Stanley is on both sides of the tape. Its power research ("half of 2026–28 GPU servers may have nowhere to plug in") is the bears' best quote, while its desk reinstated NVDA as Top Pick — the constraint thesis cuts both ways, and the sell-side is trading it that way.
Open threads worth a vote
- [watch] Amazon $8B GPU Sale-Leaseback SPV Finalization (or Collapse) — a closing extends the GPU-collateral flywheel to hyperscaler scale; an abandonment signals credit investors won't price GPU residual risk even against an AA credit. Either outcome is the most informative datapoint of the next cycle.