NVIDIA’s Jensen Huang Restates the GPU Durability Thesis — and the Depreciation Debate It Can’t Yet Settle

An X post, a rising H100 rental print, and the one accounting question on which the entire AI capital cycle turns.

What Happened

On Monday, September 7 — in an X post, not an earnings call, a keynote, or a formal interview — Jensen Huang replied to compute-price tracker Ornn Exchange, whose index showed the roughly three-year-old H100 GPU renting for more this month than last. Huang’s exact words, as corroborated by two secondary reports (the tweet timestamp could not be independently verified; X was unfetchable at publication): “NVIDIA compute is fungible, durable and highly rentable. It is a productive, revenue-generating asset.” That is the verbatim post. Characterizations you may have read elsewhere — such as “every GPU keeps generating revenue long after deployment” — are paraphrase, not his words.

The post is a one-line reprise of Huang’s signed August 11 NVIDIA essay, “AI Factory Compute Is Becoming an Investable Asset Class,” which used almost the same formula at greater length. It names no opponent: this is an implicit rebuttal to depreciation skeptics, not a direct response to Michael Burry, Jim Chanos, or any named analyst. On the rental rate that prompted it: Benzinga’s rendering of Ornn’s post put the H100 at roughly $3.28 per hour, up approximately 22%; Ornn’s own index page showed the rate settling nearer $3.17 on September 7. Both figures are attributed reads of Ornn’s data, not a settled or NVIDIA-confirmed rate — treat them accordingly.

The supporting evidence for GPU durability that circulates alongside this story — that NVIDIA CFO Colette Kress noted six-year-old A100s still running at full utilization in Q4 FY26; that CoreWeave has signed an A100 rental contract running to 2029; that Huang’s August essay described the A100 fleet as mission-capable from 2020 to 2029 — is real, but it comes from those separate sources. It is not in this post and should not be folded into the quote. Huang is also, plainly, an interested party: he benefits directly from the market believing GPUs are durable assets, which is the minimum discipline any reader should apply before accepting his framing.

The Structural Read

Strip away the X-post format and the question Huang is answering is this: are the GPUs underneath the entire AI capital boom durable assets that keep earning, or fast-depreciating hardware whose long book lives are flattering reported profits? That question is the hinge on which every supply-chain financing move, every hyperscaler capex forecast, and every AI-infrastructure fund thesis currently turns. We covered the capital-deployment side of this in our Wistron supply-chain raise piece and in this week’s synthesis; the accounting assumption underneath all of it is exactly what Huang is publicly defending.

The bull case is coherent. If compute is fungible (any H100 can run any workload), durable (it keeps earning for years, not quarters), and highly rentable (spot demand exceeds supply even three years post-launch), then the hundreds of billions flowing into GPUs and data centers look like an investable asset class — not a pile of soon-to-be-obsolete silicon. The market evidence available today supports this: Ornn’s index shows H100 rents rising, not falling; Kress has separately noted six-year-old A100s at full utilization; CoreWeave’s 2029 contract is a real cash-flow instrument, not a forecast. NVIDIA has institutionalized the thesis by anchoring a $500B+ financing structure with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — Wall Street’s largest capital allocators are pricing in durability with real money.

The bear case is equally coherent, and it operates on a different time horizon. Burry’s $176 billion understated-depreciation argument (November 2025, attributed) says the accounting is already wrong: booking five-to-six-year useful lives on hardware that will be economically obsolete in two to three years means today’s reported profits include future write-downs that haven’t been recognized yet. Ben Emons adds a supply-side mechanism: a flood of cheap domestic Chinese AI chips could weaken the collateral value of GPUs that are currently backing those $500B+ financing structures. Neither argument says old GPUs aren’t earning today — they are. The bears’ claim is that the book life assumption is wrong, and that the gap between accounting and economic reality will surface as Blackwell and Rubin flood capacity and cheaper supply arrives.

Three Implications