Nvidia Capex Digest
TL;DR
The capex verdict stays intact — Goldman now projects ~$7.6T of cumulative AI infrastructure spend through 2031, and Nvidia's own guidance is anchored to a $1.3T 2027 hyperscaler-capex projection (nvidia-ai-capex-cycle-sustainability). But two new vectors landed over the weekend: Goldman's same research puts a 2028 date on depreciation fully offsetting AI's earnings contribution, and GMO argues the bubble could break on equity supply — SpaceX lockups plus the Anthropic/OpenAI IPOs — before any demand datapoint disappoints (ai-equity-supply-bubble-catalyst-gmo
). Meanwhile NVDA's de-rate grinds toward the 14x forward tripwire without firing (nvda-derating-decade-low-multiple
).
Goldman put a date on the cliff — and extended the runway to get there
The bull case and the bear case now live in the same research note, with explicit numbers on both sides.
"by 2028, depreciation could completely offset the earnings contribution from additional AI investment" — 24/7 Wall St
Goldman's Ben Snider attributes almost half of 2026 S&P 500 earnings growth to AI-related investment, with hyperscaler capex of ~$800B in 2026 climbing to $1.2T in 2027 and $1.4T in 2028, and ~$7.6T cumulative through 2031 (nvidia-ai-capex-cycle-sustainability). But the same note flags earnings quality: hyperscalers booked ~$150B of unrealized private-AI gains in Q2, memory makers' ~80% gross margins are double historical norms, and depreciation subtracts roughly 5 points from S&P earnings growth in 2027. Nvidia's own anchor confirms the runway — a projection that the five largest hyperscalers spend $1.3T on data center capex in 2027, up from nearly $800B, which underpins its ~70% FY2028 growth guidance (nvidia-ai-capex-cycle-sustainability
). This matters because the debate just shifted from "is demand real" to "when does the accounting catch up" — and now there's a calendar on it.
What to watch: Whether 2027 earnings prints show the ~8pp drag Goldman projects as unrealized private-AI gains shrink.
The de-rate is the scoreboard, and it's one tick from the tripwire
NVDA's multiple keeps compressing despite record fundamentals — the market is pricing durability skepticism, not results.
"roughly 14.4 times forward earnings" — BigGo Finance / Motley Fool
That's against 29x trailing — the cheapest forward multiple in the Magnificent Seven, and below the ~15.75–17x Bloomberg baseline from Sept 22 (nvda-derating-decade-low-multiple). The registered tripwire (forward P/E below 14x) has not fired, but it's close: a print through 14x would signal the market is pricing an outright earnings break rather than skepticism. The telling detail is the bull framing — Motley Fool's Keithen Drury calls the stock "grossly undervalued," but his doubling scenario requires the multiple merely to hold at 29x trailing while earnings catch up. Even the bulls have stopped arguing for multiple expansion (nvda-derating-decade-low-multiple
). The stock closed at $225.07, −4.3% below its 52-week high, on TTM revenue of $302.97B (+105.9%) and 74.7% gross margins — record numbers, mid-teens forward multiple.
What to watch: A forward P/E print below 14x — the line between "skepticism" and "pricing a break."
GMO's supply thesis: the break may not need demand to fail
A new bear vector arrived that is mechanically different from every demand-side argument in the debate — it only requires early holders to cash out, and the cash-out dates are already scheduled.
"Bubbles need fresh capital, and rising supply often breaks them… in the case of the AI bubble, we think supply may be the catalyst for its eventual demise." — GMO Quarterly Letter, Inker & Pease
The specific events: the final ~$2T tranche of SpaceX shares exits lockup on June 12, 2027, and Anthropic/OpenAI IPOs "could add roughly 5% to the investable market cap" (ai-equity-supply-bubble-catalyst-gmo). GMO's flow work finds each 1% increase in IPO supply has historically cut prices ~4% over the following year, implying a ~20% return drag from supply alone. The perversity: the Anthropic IPO we've been tracking as a validation test of the capex story is simultaneously a supply event that pressures the whole market, Nvidia included (ai-equity-supply-bubble-catalyst-gmo
). This doesn't contradict the intact verdict on substance — demand, unit economics, and financing are all still confirming — but it supplies a way for NVDA's multiple to compress on flows alone, with no change in orders.
What to watch: The June 12, 2027 SpaceX lockup release — whether GMO's 4x flow multiplier shows up in the tape.
What surprised us
- The bull case and the bear case now cite the same analyst. Goldman's Snider supplies both the $7.6T runway and the 2028 depreciation-offset math — the sell side has stopped pretending this is a one-sided story (nvidia-ai-capex-cycle-sustainability
).
- The IPOs we're watching as validation are also the risk. GMO reframes Anthropic's listing from a demand datapoint into a supply shock — the same event doing opposite work depending on which lens you use (ai-equity-supply-bubble-catalyst-gmo
).
- Structural fragility is a flows argument too. GMO notes ~83% of the $30T in active/passive funds focus solely on U.S. equities and only ~10% of retirement accounts rebalance annually — meaning even rational portfolios can't absorb the supply without dislocation (ai-equity-supply-bubble-catalyst-gmo
- The de-rate has a floor that hasn't broken. 14.4x forward with 4/4 EPS beats and $21.4B quarterly FCF — the market keeps discounting durability while every reported number beats (nvda-derating-decade-low-multiple
).
Open threads worth a vote
- [watch] SpaceX final ~$2T lockup release (June 12, 2027) — GMO's equity-supply catalyst test (/topics/019e95a0-30d1-7d08-ae5b-10ac242b9d5c#threads)