Nvidia GPUs Are Becoming Wall Street Collateral. Nebius's 30% Rally Shows How.

A 2020-vintage A100 data-center chip, six years old and one architecture behind the current generation, is the basis for a multi-year rental commitment CoreWeave signed this month (WccfTech, Aug. 13 2026; CNBC, Aug. 12 2026; Nebius Q4/FY 2025 financials). The same week, investor Ben Pouladian laid out the mechanism on the Podcast Alpha / etn. interview: GPUs are now Wall Street's newest financeable asset class, and roughly $10 trillion in idle capital can be routed at them (Podcast Alpha, Aug. 17 2026). The $10 trillion figure is an analyst estimate, not a verified pool of capital. It is the size of the math, not the size of the check.

The shift is structural, not a Burry-versus-Nebius story. For most of the AI buildout, the dominant bear case has been that the chips depreciate faster than the cash flows they generate, so the lenders cannot underwrite them. CoreWeave's A100 commitment through 2029 is the visible proof. Nvidia CEO Jensen Huang told WccfTech that CUDA software upgrades keep the chip productive into 2029, and the rental commitment is the public demonstration Pouladian cites for his Boeing 747 analogy: a GPU fleet behaves, on a lender's balance sheet, like a long-duration, cash-flow-backed asset that holds its value as long as the underlying demand holds (WccfTech, Aug. 13 2026).

Nebius is the public-market receipt. The company reported a pipeline reportedly worth about $9 billion and contracts from Microsoft and Meta, and the stock jumped after earnings (Nebius Q4/FY 2025 financials; CNBC, Aug. 12 2026). That is the practical shape of the new math: a chip fleet financing structure, multi-year hyperscaler contracts, and a stock price that re-rates on the proof that the structure is real. CNBC framed the same data-center story the day before as a direct rebuttal of the prevailing AI trade bear case. The data centers have to "counter" that bear case for the financing logic to keep working (CNBC, Aug. 12 2026).

The bear cases are intact, and the re-rating does not retire them. AI demand has to hold up long enough for the long-duration contracts to cash-flow. Power-grid permitting, already the binding constraint in Northern Virginia and Phoenix, sets the ceiling on how many of these fleets can actually be plugged in. And the financing structure leans on a small number of GPU customers, so a single hyperscaler pulling a contract is a much larger event for Nebius and CoreWeave than for a diversified cloud. Those are the watch items, and they are the conditions under which the GPU-as-collateral frame works or breaks.

The next milestone is concrete. CoreWeave's A100 commitments run through 2029. Nebius's pipeline is multi-year. If a major GPU customer pulls a contract, or if permitting slips past a publicly dated interconnection date, the financing math re-prices fast.