Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
Yes — and the demand side just got its strongest field evidence yet, while the bear case got its first hard number. GPU rental prices are rising for the second time in three months, SemiAnalysis found even broken clouds can sell capacity, and Nvidia answered with the largest buyback authorization in corporate history. Against that, the break-even arithmetic now runs to $3.7 trillion of annual AI revenue by 2032 — the debate has moved from "will capex be cut" to "can revenue catch up before the financing buckles."
The demand signals keep running hot — even broken clouds are selling
The scarcity evidence is no longer inference from earnings calls; it's now been measured in the field. SemiAnalysis's ClusterMAX 3.0 — 77 providers tested, 200+ customer interviews — found that "demand for NVIDIA Corp. GPUs has become so intense that even cloud providers with badly functioning infrastructure can still find customers willing to pay healthy margins." Researcher Sam Harshe put it plainly: "If you've got GPUs," buyers will pay, even if a provider's Kubernetes is "completely broken" and its storage "barely works."
"He described 'ridiculous backwardation' — customers paying steep premiums for compute available within weeks rather than waiting months." — Yahoo Finance on ClusterMAX 3.0
The pricing tape confirms it. Nebius's second hike in three months lands Oct 1 — H200 up ~20% to $5.40/hr, B300 up ~21% — with the increase spreading into memory tiers (+41% on some SKUs), and JPMorgan and BNP Paribas both upgraded the stock on the pricing power (shattered.io). Crucially for the depreciation debate, companies are signing four-year contracts for H100 capacity — 2022-vintage chips still commanding earning power (Finding: gpu-deployment-gap-overbuild-debate). Vendor corroboration stacks on top: Oracle delivered 300,000+ GPUs last quarter, nearly triple the prior, and says demand still grows faster than supply.
What to watch: whether Nebius's Oct 1 hike sticks and spreads — a third consecutive quarterly price increase would be the cleanest falsification of the glut thesis available.
The break-even bar got its full arithmetic — and it's brutal
The bulls' own numbers now quantify the hole. FT Alphaville's "What is the AI capex breakeven rate?" stacks three independent estimates: Goldman's ~$300B annual AI revenue needed just to break even (against ~$70B of cloud revenue above the pre-AI trend, annualized in Q2 2026); Van Nieuwerburgh's $3.7T/yr by 2032 — 9.2% of US GDP — to earn a 10% return on a >$10T build-out; and Bernstein-Cummings' $2.4–3.8T range, requiring current incremental AI revenues to grow "between 13-45 times in just six years" (FT Alphaville).
"AI revenues very soon need to grow a LOT MORE for the size of these investment plans to make economic sense." — FT Alphaville
This is the shift from the last digest: the demand tripwires remain unfired, but the arithmetic of what "intact" ultimately requires is now assembled end-to-end, and it's roughly 4x the current run-rate just to break even, ~10x for historical ROIC (Finding: big-tech-capex-reaches-710b-in-2026). The financing-layer watch from last cycle — rate risk, hyperscaler CDS — is where this math eventually gets tested.
What to watch: whether any of the three break-even estimates gets revised down on cloud backlog conversion — that's the reconciliation signal between the pricing tape and the arithmetic.
Nvidia answered the circular-financing critique with the largest buyback in history
On September 28, Nvidia's board authorized a $150 billion increase to the repurchase program, raising the remaining total to $235 billion — "the largest share repurchase authorization increase in history," to be executed through fiscal 2028. The stock jumped ~3% (NVIDIA Newsroom). Huang tied it directly to the capex engine: "Our cash generation gives us the capacity to invest in the technologies that advance this transformation and return capital to shareholders."
The balance sheet behind it makes the signal credible: TTM revenue $302.97B (+105.9% YoY), 74.7% gross margin, $21.40B quarterly free cash flow, and just 0.2x debt/EBITDA (NVDA market view, Sept 29). Management is pricing its own stock as cheap while guiding FY2028 revenue growth of ~70% — the strongest insider counterweight to the circular-financing critique, since the cash funding the flywheel's customers is Nvidia's own, not borrowed (Finding: nvidia-record-q1-results-and-trillion-dollar-order-book+1). The flywheel itself kept widening the same day: Samsung committed $1B to Helix Digital Infrastructure, a KKR- and Nvidia-backed firm, bringing Helix's secured capital past $11B.
What to watch: the 10-Qs — whether the buyback pivot means marginally less capital going into strategic stakes that manufacture demand.
The China channel is moving from signal to pipeline — without formal approval
Last cycle's RTX Pro 5500 signal firmed into volumes and a timetable. Asia Times reports ByteDance alone is considering about one million of the workstation chips, with Nvidia planning to start shipping to China in late December "at roughly 500,000 units per quarter," and Tencent added to the list of companies asked to file purchase plans (Asia Times). But as of Sept 28, "no final approval, purchase order, or shipment has been publicly confirmed" (shattered.io).
The SKU is engineered to clear both governments — a GB202 workstation card that can run open-source inference but not training, falling "just below the export control red line while remaining powerful enough for real applications." An MIIT-affiliated columnist names the stake: "For Nvidia, every chip that enters the Chinese market helps extend the life of the CUDA ecosystem" (Finding: nvda-china-market-access-and-trump-xi-summit). At ~500K units/quarter this is a marginal revenue add, not a needle-mover — the value is strategic, keeping CUDA installed while Huawei concedes its own AI-compute production "could not yet meet domestic demand."
What to watch: formal MIIT approval or a confirmed order — the one-shot watch from last cycle still hasn't fired, and a walk-back would close the channel for good.
What surprised us
- The first hard write-down on the board is at the vendor, not a customer. Nvidia's $0.4B charge for excess H200 inventory is the only real number in the overbuild debate — and it's for the China-restricted SKU, which is a policy story, not a demand story. Burry is still short, but his depreciation argument now runs into four-year H100 contracts (Finding: gpu-deployment-gap-overbuild-debate
).
- The bull case concedes the deployment gap in softer form. Only 19 of 77 providers earned a Medallion rating; Crusoe ran clusters SemiAnalysis called "exceptionally unreliable." Usable capacity is scarcer than purchased capacity — which is exactly where a glut would hide first.
- The buyback and the bear case landed the same day. Burry betting against Nvidia Sept 28 while the board authorized history's largest repurchase is the debate in miniature — and prediction markets still put a downturn by Dec 31 at only ~10% (Yahoo Finance).
- The demand leg and the arithmetic are both right, for now. Prices rising and contracts lengthening can coexist with a $3.7T revenue requirement — the question is only which one runs out of runway first.