Wall Street treats GPUs like disposable electronics. CoreWeave (CRWV -2.34%) is running them like long-duration infrastructure.

What's the deal with this stock?

CoreWeave is borrowing tens of billions of dollars to buy Nvidia (NVDA +0.03%) chips and rent them to AI software builders. The bear case against the stock is simple: Graphics chips age quickly. If Nvidia launches faster processors every 12 months, three-year-old chips should become useless, trapping CoreWeave in a cycle of taking on new debt just to replace dying hardware.

That theory has a flaw. The real-world contract data shows older chips are not dying.

Fully depreciated on paper. Fully rented in reality. Image source: The Motley Fool.

Wall Street thinks GPUs decay like iPhones. The contracts say otherwise.

On its Q2 2026 earnings call, CoreWeave disclosed a multi-year renewal on a cluster of Nvidia A100 GPUs extending into 2029. Nvidia introduced the A100 in 2020. That means an architecture introduced in 2020 can still generate contracted cash flow nearly a decade later.

Management explicitly laid out the mechanism: As earlier-generation fleets finish their initial contracts, they deliver returns in subsequent years because much of the underlying capital burden has already been paid down. Wall Street models assumed older GPUs would be retired or deeply discounted. Instead, they are staying online at attractive prices.

Older chips do not need to win benchmarks to make money

Scarcity explains part of this demand. Power connections take 18 to 24 months to build, so software teams rent whatever active compute is available today. But the longer-term mechanism is workload cascading.

The newest chips naturally get the jobs where speed matters most. Older chips can move down the stack to fine-tuning, inference, coding workloads, and smaller models where customers care more about cost than benchmark leadership.

As the market matures, compute demand splits by cost and task:

• Frontier pre-training: Uses top-tier processors like Blackwell because raw speed dictates training time.

• Fine-tuning and domain adaptation: Moves to previous-generation chips like the H100, where customers can trade some speed for a lower cost.

• Inference and everyday applications: Operates on older architectures like the A100.

An older GPU does not need to beat a new chip on raw speed. It only needs to be cheap and reliable enough for tasks that do not justify top-tier pricing.

The second lease is where the real economics live

The economics of a GPU cluster change after its initial contract. During the first lease (typically three to five years), customer payments recover a substantial portion of the original hardware cost and service the debt used to buy it. By the time that lease expires, the unrecovered capital burden sitting on those servers drops significantly.

Second and third leases operate on a different financial equation. Power, facility rent, and maintenance still cost money. But the heavy debt service and capital recovery are largely finished. Every dollar of revenue from a renewal generates higher incremental returns on already-recovered capital.

That creates an asset-management flywheel: The initial contract carries most of the capital burden. The extension monetizes an asset whose original investment has already been substantially recovered.

Lenders are now betting on what happens after year three

Debt markets are beginning to price in this residual value. Historically, lenders matched GPU debt maturities directly to the length of the underlying customer contract. CoreWeave's $2.6 billion DDTL 5.5 debt facility broke that pattern. The debt carries a five-year maturity, but the customer contracts backing it average only three years.

That gap matters. Lenders are accepting exposure beyond the initial customer lock-in, which suggests they are increasingly comfortable underwriting some residual economic value after the first contract ends.

Nvidia is building a broader framework around this concept. It partnered with a consortium of six major asset managers and banks (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR) to mobilize over $500 billion of private capital, explicitly aiming to turn AI compute into a recognized investable asset class. Reporting indicates Nvidia may even provide limited residual-value support on certain debt structures. When equipment suppliers and private credit markets back residual value, GPU compute stops looking like disposable tech hardware and starts looking more like financed infrastructure.

Verdict: The asset machine is working, but leverage is the price of admission

CoreWeave is making a clear trade-off. It is taking on heavy debt and significant near-term interest drag to build a massive pool of infrastructure. That strategy fails if older GPUs lose their rental value quickly.

The A100 extension shows that older GPUs can remain economically useful far longer than the bear case assumes. CoreWeave's real advantage is not just getting new Nvidia chips first. It is extracting cash from second and subsequent contracts on hardware the market assumed would be obsolete.

Watch what happens as the first H100 contracts mature and how cheaply CoreWeave can continue financing project-level infrastructure. If older fleets keep renewing at attractive economics while SPV financing remains cheap, the residual-value thesis gets stronger.