Nvidia Capex Digest
TL;DR
The verdict reads intact for a fifth straight time, but this cycle the story moved from the tape to the argument: Nvidia entered the GPU residual-value debate publicly — and got fact-checked within days (gpu-depreciation-debate-hyperscaler-earnings). Meanwhile the yield curve split in two — the short end relieved by a weak jobs report, the long end at the highest levels since 2002 — and the public IPO window kept deteriorating at the edges even as private credit bridged the gap (nvidia-ai-capex-cycle-sustainability
; ai-ipo-window-equity-exit-channel
). The capex story is intact; the financing plumbing underneath it is getting more elaborate by the week.
The residual-value debate went public — and Nvidia's own citations cut against it
The ROI debate has a new protagonist: Nvidia itself. On Oct 1 the company published a blog arguing AI factories are productive, durable, and fungible — and a fact-check of the five outside sources it cited found it quoting "the most favorable datapoint in each contested area" (gpu-depreciation-debate-hyperscaler-earnings). The clearest case is the rental term curve:
"five-year term price as a share of one-month is A100 80.2%, H100 59.8%, H200 43.7%, B200 53.8%, B300 53.8% — which Ornn summarizes as '44 to 60 percent for Hopper and Blackwell.'" — FourWeekMBA
Nvidia quoted only the A100 — the outlier survivor, not the pattern. But here's the twist: the underlying data also weakens the bear case. Sprout reports Microsoft ran GPU generations for 6.9 to 8.8 years, with six-year-old A100s selling at 38–43% of launch price — well above the 25% residual lenders assume (gpu-depreciation-debate-hyperscaler-earnings). This matters because residual value is the collateral floor for the entire financing web — Seeking Alpha's framing: if Nvidia convinces lenders GPUs retain value beyond 3–4 year depreciation, "AI infrastructure financing could continue scaling." The blog, the Howden Re insurance talks, and the CME compute futures listing are all the same campaign from different directions.
What to watch: Whether any hyperscaler actually shortens GPU useful life or takes an impairment — Sprout's data says the opposite is happening in practice, and that remains the falsification test.
The yield curve split: relief at the short end, a 2002-era squeeze at the long end
The macro leg bifurcated. A weak September jobs print (+29K vs +84K expected) collapsed October hike odds to ~17% and dropped the two-year 10bp to 4.69% — Jefferies' Thomas Simons called it "the final nail in the coffin for an October hike" (nvidia-ai-capex-cycle-sustainability). NVDA touched an intraday record $237.88, closing at $233.95 with the Nasdaq-100 at a record.
But the long end — the actual financing cost of the ~$1.1T 2027 buildout — went the other way:
"5% isn't the straw that breaks the camel's back, but it's another heavy sack on an already tired hump. If you don't take some weight off, collapse is just a matter of time." — Max Gokhman, Franklin Templeton, via BigGo Finance
The 10-year tested 5.30% intraday, its fifth consecutive weekly climb and the highest levels since 2002 — leaving the Man Group 5.5% "AI capex crack" tripwire roughly 20–25bp away (nvidia-ai-capex-cycle-sustainability). The short end drives sentiment; the long end drives whether the buildout can be funded. Only one of those moved favorably.
What to watch: The 5.5% tripwire on the 10-year — the margin is now measured in basis points, not turns.
The public equity exit channel is closing; private credit is substituting
The IPO pullout wave is now "a catalog, not a cluster of anecdotes" (ai-ipo-window-equity-exit-channel). Oura pulled a $15B IPO the day before listing; OpenAI postponed to next year; Anthropic slipped from October to November; SB Energy — SoftBank- and Nvidia-backed — delayed a $5B raise because "potential investors baulked at the valuation." The channel that was supposed to recycle AI infrastructure exposure from private balance sheets to the public market keeps deteriorating at the edges.
What replaces it is private bridge capital. Nscale, the holdout still going ahead with its NYSE listing, secured $3.36B in convertible notes ahead of it — $2.36B led by Third Point, plus a $1B tranche from Nvidia due mid-November, converting at IPO — against a claimed $103B contract backlog (ai-ipo-window-equity-exit-channel). The Information's counterpoint is why the public market is being asked to take the leap: the backlog depends on campuses that don't exist yet. This is workable while private capital is willing — but it concentrates risk in a narrower lender set, on the same path as the SPVs, leasebacks, and converts mapped in the last digest (nvidia-infrastructure-leasing-circular-financing
).
What to watch: Nscale's actual IPO pricing — the cleanest live test of whether any public appetite for neocloud economics remains.
What surprised us
- Nvidia fact-checked itself into a corner. Its own cited source (Sprout) contains the strongest rebuttal yet to the Burry depreciation thesis — 8.4-year V100 lives, A100s never retired at AWS — while another cited source (Ornn) shows current-generation term decay that contradicts the blog's cherry-picked A100 quote (gpu-depreciation-debate-hyperscaler-earnings
). The data cuts both ways, and Nvidia only quoted one way.
- The financialization milestone arrived on schedule. CME's H100/B200 rental-index futures list on NYMEX Oct 5 — the "chips as an investable asset class" push is now an exchange-traded product, with B200 +27.6% YTD but a term curve already below spot (nvidia-ai-capex-cycle-sustainability
).
- The bear's best exhibit is now a financing condition, not a demand claim. ~$5.5T of cumulative AI capex through 2030, ~74.5% debt-financed — and the same analysis concedes the whole thing hinges on GPU collateral values (nvidia-ai-capex-cycle-sustainability
- Nvidia is now a bridge lender to its own ecosystem's IPOs. The $1B Nscale convert tranche is a vendor financing its own exit channel while public buyers balk (ai-ipo-window-equity-exit-channel
).