Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
The capex story is intact on every demand gauge we track — but the challenge has moved to a new layer. Nebius's second price hike landed on schedule and JPMorgan is now underwriting rising compute pricing, while NVDA holds 2.9% below its high on a 5.26% 10-year that just matched its 2007 peak. The genuinely new threat isn't silicon benchmarks: DeepSeek and Huawei open-sourced a CUDA alternative, attacking the software moat Nvidia itself says is the thing to attack.
The demand falsifier stays unfired — Wall Street has started underwriting scarcity
The pricing tape didn't just hold this cycle; it acquired sell-side sponsorship. The Nebius hike landed Oct 1 as scheduled — the second increase in three months — at roughly 17% for H100 and 21% for B300 (Yahoo Finance), and JPMorgan upgraded CoreWeave to Overweight citing "stronger compute pricing and CoreWeave's use of shorter-term contracts at premium rates."
"higher prices are adding 5 to 10 percentage points to contribution margins on new contracts compared with earlier agreements" — Yahoo Finance
That margin line is the sharpest new datapoint: rising prices aren't just surviving, they're expanding profitability on renewals. CoreWeave's contracted power grew from 3.1 GW at end-2025 to 4.2 GW as of Aug 11, and Nebius converted the same scarcity into a $236.2M adjusted EBITDA profit on Q2 revenue up 514%. The counterweights are real — Goldman expects the largest US hyperscalers to spend $800B in 2026 against $1.1T consensus in 2027, while warning that "the rate of both capex growth and upside surprises diminish[es] relative to recent quarters" (FT Alphaville). Prices rising twice in three months across four GPU generations is a market rationing scarcity, not digesting a glut.
What to watch: whether a third consecutive Nebius increase lands — the cleanest glut falsification available remains a price cut that never fires.
The moat's software layer just took its first measured hit
The competitive challenge to Nvidia has so far been measured in silicon benchmarks; this cycle it moved to the software layer — the part Nvidia's own filing says matters most. Reuters reports DeepSeek is adapting and open-sourcing parts of its stack for Huawei's Ascend processors, "making it easier to run and optimize AI workloads without Nvidia hardware," centered on TileLang, "an open-source programming language DeepSeek says offers a 'simpler programming model' than CUDA" (Benzinga, Sept 30). The pair also built a 128-chip Ascend 950 supernode.
Nvidia flagged this exact risk in its own 10-Q (quarter ended July 26, 2026), warning that its export-driven absence from China has helped rivals build "larger developer and customer ecosystems to challenge us worldwide" (SEC EDGAR) — against a CUDA base of more than 7.5 million developers. Nobody thinks workloads leave CUDA quickly (CSIS's Gregory Allen testified the shift "would likely take years"), but the threat isn't displacement — it's a slow erosion of the switching costs that underpin Nvidia's pricing power, now happening in the open.
What to watch: adoption traction for TileLang outside DeepSeek — the first third-party lab shipping Ascend-optimized workloads would mark the crack widening.
Rates matched the 2007 peak — and the stock is absorbing it anyway
The macro bear case hardened another notch: as of September 29, the 10-year Treasury stands at 5.26% — at the very 2007-peak level last cycle's digest measured distance from, and the highest since 2007 (macro view, Oct 1, 2026). Every basis point on the long end compresses the multiple the market pays for AI cash flows and raises the cost of the debt-funded layer of hyperscaler capex.
"NVDA trades at $228.38, a $5.49T market cap, 28.91x earnings with a PEG of 0.479" — NVDA view, Oct 1, 2026
The stock is holding near its 52-week high (2.9% below, 90.3rd percentile of the range) while the long end makes new cycle highs — a 2.2-beta vehicle with no multiple cushion if rates keep rising. The fundamentals doing the lifting: TTM revenue of $302.97B up 105.9%, 74.7% gross margins, four-for-four earnings beats, and $21.40B of quarterly free cash flow against $38.86B total debt (NVDA view, Oct 1, 2026) — the balance sheet that funds the $235B buyback and the credit support behind the neocloud financing stack. Analysts' mean target of $327.70 sits ~43% above spot.
What to watch: the next FOMC decision, and whether the 10-year breaks through 5.26% into territory unseen since pre-2007.
What surprised us
- The attack came through software, not silicon. Everyone was watching Ascend benchmark numbers; instead DeepSeek open-sourced a programming language. The moat's most defensible layer per Nvidia's own filings is the one now under direct, public assault.
- JPMorgan is underwriting scarcity as a margin story. "5 to 10 percentage points to contribution margins" on new contracts means rising GPU rental prices are flowing to neocloud profitability, not just headlines — the sell-side has moved from observing the pricing power to recommending it.
- Goldman's caution is about the rate of surprise, not the level. Expecting capex to beat consensus while flagging diminishing upside surprises is a subtle downshift — the buildout is still growing, but the easy beats may be behind it.
- Polymarket still prices bubble-burst odds by Dec 31 at ~9% (Benzinga) — the prediction market and the pricing tape agree the near-term glut scenario is a tail, not a base case.
Open threads worth a vote
- First GPU-Collateral Insurance Agreement (Nvidia–Insurer Talks, FT) — still open from last cycle; fires on a signed deal, named pricing, or a confirmed abandonment of the talks.