Nvidia Capex Digest
TL;DR
The capex story survived its first real stress test this week: a selloff sparked by the "pace the frontier" essay was fully retraced in two days, and every hard datapoint — Nvidia's Q3 guide of ~$108B, hyperscaler 2027 capex raised to ~$1.12T — says demand is still accelerating (nvidia-ai-capex-cycle-sustainability). But two new fronts opened that last cycle's digest only gestured at: the Fed hiked and named AI capex as an inflation driver, and Michael Burry's depreciation attack moved the bear case from "the spending is crazy" to "the profits funding it are overstated" (gpu-depreciation-debate-hyperscaler-earnings
).
The pace panic came and went — in two trading days
The "slow down" argument finally hit the tape, and the market answered it within 48 hours by buying everything back.
"Markets are trying to figure out what the pace [of AI development] is, and there will be winners and losers depending on what that pace is." — Thrivent's David Royal, per CBS News
Monday's repricing was real: Nvidia slid 3.4%, Micron 5.3%, and the iShares Semiconductor ETF fell 5.6% — its worst day since July 1, with Intel down 7% (CBS News). The rebuttal was faster than the selloff: Huang pushed back, Trump called the concerns a hoax, and Piper Sandler's channel checks described demand as "full-steam ahead... well ahead of available supply" (Invezz). Gabelli's Hendi Susanto landed the sharpest counter-argument: if frontier labs slow down, "second and third tier players will seize the opening to catch up" — inference, enterprise AI and physical AI don't depend on frontier-training scale anyway (Invezz).
The reason this matters: the pacing debate has now been priced once, and the clearing price was "no change to orders." That's a stronger signal than if the essay had been ignored.
What to watch: Whether Amodei, Musk and Altman escalate from essays to anything concrete — a single disclosed order delay would re-open the trade.
The Fed joined the bear case — by name
Last cycle flagged the macro squeeze as the real threat; this week it arrived on schedule, and with an added twist: the FOMC is now explicitly blaming the boom for inflation.
"Domestic spending has been resilient, productivity growth strong, and capital investment is robust." — Chair Kevin Warsh, per Reuters
The September 16 hike — 25bp to 3.75–4.00%, unanimous, the first since 2023 — came with a dot plot showing 16 of 18 policymakers projecting at least one more hike by end-2026, and PCE return to 2% pushed back to 2029 (nvidia-ai-capex-cycle-sustainability). Reuters listed "capital spending from the artificial intelligence boom" alongside tariffs and the energy shock among the pressures that "kept price pressures intense enough that the Fed felt it needed to raise." With the 10-year at 5.0% — a 19-year high per the WSJ — the marginal financing cost of the ~$1.1T 2027 buildout is rising, and the central bank is now a named participant in the capex debate (/markets/macro/2026/09/17).
What to watch: The December FOMC — the dot plot makes a follow-up hike the base case, and each one extends the squeeze on long-duration semiconductor multiples.
The demand datapoints refuse to crack
Underneath the macro noise, every guide went up, not down. Jefferies now pegs combined capex for Microsoft, Amazon, Alphabet, Meta and Oracle at ~$1.12T in 2027 versus ~$818B in 2026 — a $302B single-year increase — with Amazon raised to ~$220B and Oracle lifted to $90–95B for FY2027 (InfotechLead). Nvidia's own quarter was its fourth consecutive beat: $96.22B revenue, Data Center at $89.0B (+117% YoY), and Q3 guided to ~$108B with Vera Rubin "ramping into full production" at CoreWeave, Google Cloud, Azure and OCI (Stocks Down Under). Morgan Stanley sees capex growth decelerating in 2028 — not before (nvidia-ai-capex-cycle-sustainability).
What to watch: Nvidia's Q3 FY27 report on November 18 — Vera Rubin is guided to ~20% of data-center revenue, making it the first hard read on the next product cycle.
The bear case migrated to the buyers' accounting
The most substantive new argument of the week doesn't touch Nvidia's revenue at all — it attacks whether the customers can afford to keep paying it at reported profitability.
"[Burry estimates] ~$176 billion of understated depreciation and overstated profit across the industry between 2026 and 2028." — Value The Markets
The claim: hyperscalers depreciate GPU servers over five to six years when the economic life is two to three, overstating Oracle's earnings by close to 27% and Meta's by ~21% by 2028 — and Burry has backed it with shorts on Nvidia and Palantir (michaeljburry.substack.com). The rebuttal comes from the rental market: CoreWeave's CEO says expiring H100 contracts rebooked at ~95% of original price, and Huang points to H100 rental rates up 22% in a month to $3.28/hour, plus A100s (2017 vintage) contracted through 2029 (Value The Markets). As Value The Markets notes, "a chip can fall substantially in value and still generate plenty of rental income, so neither measure alone settles" the question.
Why it matters: depreciation schedules are the earnings-quality tripwire. If any major hyperscaler shortens useful-life assumptions or takes GPU impairments, reported AI earnings compress industry-wide without a single order being cancelled — and higher rates plus faster-recognized depreciation means weaker buyer free cash flow and more reliance on off-balance-sheet financing channels.
What to watch: Every hyperscaler earnings report from here — a useful-life change is the trigger that fires this thread.
What surprised us
- The Fed's language, not its rate. The hike was expected; Reuters listing AI capital spending as an inflation driver was not. The buildout is no longer just being squeezed by rates — it's being cited as a cause of them, which is a feedback loop (Reuters).
- The Anthropic IPO is becoming the flywheel's biggest turn yet. Nvidia in talks to anchor up to $10B of a potential $100B raise at ~$2T valuation — the largest IPO in history, targeted before the November midterms — extends the chipmaker-finances-its-own-demand pattern while the DOJ Groq probe sits open (Reuters). Anthropic's annualized revenue, for context, went from ~$9B at end-2025 to over $65B by July.
- The accounting split is inside the data. Meta extended GPU useful lives in 2025 while Amazon shortened them the same year — the "industry standard schedule" Burry attacks is already diverging in practice (SEC 10-Ks).
- KKR doubled private-credit AI deal volume to $80B in H1 2026 — the alternative-financing channel keeps widening exactly as rates make it more expensive, which cuts both ways (nvidia-ai-capex-cycle-sustainability
).