Nvidia Capex Digest
TL;DR
The capex story didn't just survive — it ratcheted up: Huang publicly projected 2027 chip sales roughly doubling, Nvidia's own estimate for top-5 hyperscaler 2027 capex rose to $1.3T, and TSMC printed a record +53.3% YoY August (nvidia-ai-capex-cycle-sustainability). But the debate has fully migrated from "will they spend?" to "can the buyers finance it?" — hyperscaler default protection sits at an 8-year high while 42% of fund managers now name AI capex the likeliest origin of a systemic credit event (nvidia-infrastructure-leasing-circular-financing
).
The demand ratchet moved up again — this time from Nvidia's own mouth
The bull case stopped being defensive this week and went on offense, with the company itself revising the future upward in public.
"2027 chip sales are projected to roughly double compared to 2026... The unconstrained is significant." — Jensen Huang, per 24/7 Wall St.
The hard numbers back the rhetoric. Nvidia pegs top-5 hyperscaler 2027 capex at $1.3 trillion, up from $800 billion this year — above the ~$1.1T consensus the last digest tracked (Globe and Mail). TSMC's August revenue hit a record NT$514.81B, +53.3% YoY, nowhere near the <20% softening tripwire (Quartz), and Meta raised the floor of its capex range to $130–145B (24/7 Wall St.). Supply, not demand, is the binding constraint: Huang says Nvidia has supply for only ~70% of demand, and supply obligations swelled to $279.0B largely on Vera Rubin memory procurement (24/7 Wall St.).
Why it matters: last cycle the market re-priced the "pace the frontier" essay and concluded nothing had changed in orders. This week the demand side got stronger, with FY2028 consensus EPS climbing from $12.67 to $15.57 in 90 days on 39 upward revisions and zero cuts (24/7 Wall St.).
What to watch: Whether Huang's doubling call shows up in customer order books at the November 18 Q3 report — or whether "supply-constrained" becomes the excuse that caps it.
The credit market is pricing the buyers, not the vendor
The bear case now lives in credit spreads, and it has a clean market-priced signal for the first time.
"The widening in hyperscaler CDS is not driven by dealer hedging of new issuance." — Apollo's Torsten Slok, per TradingView/Benzinga
Slok's data: 5-year CDS on an Amazon/Google/Microsoft/Oracle basket climbed above 100 basis points — the highest in his eight-year series — while bank CDS sit flat near 40bps. The issuance curve explains why: the five hyperscalers sold ~$121B of IG bonds in 2025 versus a ~$28B/year 2020–24 average, and 2026 issuance passed $180B by midsummer (TradingView/Benzinga). Buyer strain is visible in the P&L: Oracle burned $5.4B of free cash flow in a single quarter against $28.5B of capex, and Meta's Q2 FCF collapsed to $784M from $8.55B a year earlier (TradingView/Benzinga).
This is the key structural point: the $500B Wall Street financing alliance with Apollo, BlackRock, Blackstone and others was sold as de-risking the buildout by moving it off customer balance sheets — but the credit market is pricing the system anyway, and the platform's definitive agreements remain unexecuted MOUs (nvidia-infrastructure-leasing-circular-financing). With the 10-year at 5.01% and a follow-up Fed hike signaled, the carry on exactly these debt-funded structures keeps rising (/markets/macro/2026/09/18).
What to watch: The CDS basket itself — a move through 150bps would be the loudest stress datapoint yet on the buyers' solvency.
Anxiety and conviction are rising together
The most interesting datapoint isn't the fear — it's that the fear isn't changing anyone's behavior.
"AI hyperscaler capex continues to be seen as the most likely source of a systemic credit event (per 42%)." — BofA's Michael Hartnett, per TradingView/Benzinga
In BofA's September survey (190 managers, $512B AUM), 42% named AI infrastructure spending the top systemic credit risk — up from 38% — and a record net 33% say companies are overinvesting. Yet 79% expect no spender to announce a cut this year, up from 71% (TradingView/Benzinga). Meanwhile the Sept 14 slowdown-call selloff (Nebius −5.5%, CoreWeave −6.8%) found "no evidence of cancelled chip orders or paused buildouts" (Simply Wall St.).
Why it matters: professional allocators have essentially declared this the most dangerous trade in markets while increasing their expectation that it continues. That combination — high stated risk, high realized momentum — is exactly the setup where the trigger, when it comes, comes from a financing event rather than a demand miss.
What to watch: The October Anthropic IPO — up to $100B at ~$2T valuation with Nvidia reportedly weighing a $10B anchor stake — as the hardest public-market test of AI capex economics (Reuters).
What surprised us
- Nvidia is now the most aggressive forecaster of its customers' spending. The $1.3T top-5 hyperscaler capex estimate is above sell-side consensus — the vendor out-bidding the bulls on its buyers' behalf is a new dynamic, and it raises the stakes if 2027 orders disappoint (Globe and Mail).
- Credit markets are separating Nvidia from its customers. Hyperscaler CDS at an 8-year high while bank CDS sit flat means the market is pricing buyer leverage, not a general credit squeeze — Slok explicitly rules out dealer hedging as the driver (TradingView/Benzinga).
- The $500B financing platform is still just MOUs. Two months after the August 10 announcement, definitive agreements remain unexecuted — the de-risking story is a promise, not yet plumbing (nvidia-infrastructure-leasing-circular-financing
).
- SpaceX is quietly a top-tier AI capex story. Its AI segment revenue jumped 247% YoY to $2.56B, with $15.83B of $18.37B quarterly capex going to AI compute — and Musk committing to "build exclusively on NVIDIA" (24/7 Wall St.).