On August 11, 2026, CoreWeave disclosed that it had signed a contract for cloud capacity using the NVIDIA A100—a GPU that debuted in 2020—running through 2029. This serves as a concrete example showing that even as cutting-edge GPUs continue to be refreshed, multi-year demand persists for older generations as well. However, it has not been confirmed that the same GPU manufactured in 2020 will operate for nine years, nor that this contract guarantees profitability. The specific reasons behind this individual contract have not been disclosed. To assess the residual value of older-generation GPUs, it's necessary to separate chip generation and use case from data center infrastructure, power, and contract terms.
What the A100 Contract Actually Confirms
CEO Michael Intrator stated during the earnings call that CoreWeave recently signed an A100 contract at an "attractive price," with a term extending to 2029. The A100 is an Ampere-generation GPU that NVIDIA announced in May 2020. The fact that it has been contracted through 2029 means that at least some customers have plans to pay for A100-based compute resources even nine years after its launch.
At the same time, the scope of what was disclosed is narrow. The customer's name and the number of GPUs were not made public. The total contract value, the start date, and the hourly unit price all remain unknown. The "attractive price" assessment also comes from CoreWeave's management, and there is no external material available to calculate the profitability of the deal on a per-contract basis. It should be understood not as CoreWeave selling A100 units directly to a customer, but as a contract for cloud capacity that incorporates A100 hardware.
CFO Nitin Agrawal explained that the capacity approaching renewal currently represents only a limited portion of the company's total equipment holdings. He also noted that the average selling price for older-generation GPUs is at or above the level seen about a year ago. While this single deal serves as evidence that buyers for older-generation GPUs have not disappeared, it does not necessarily mean that CoreWeave's entire older-generation fleet can be re-contracted under the same terms.
Not a Simple Nine Years from 2020 to 2029
The product announcement date, the date the GPU was purchased, the date it went live in a data center, and the end date of a customer contract all operate on separate timelines. NVIDIA announced the A100 in May 2020, but it remains unknown when CoreWeave acquired the GPUs allocated to this particular contract or when they were brought online. Therefore, there is no basis for asserting that "a GPU manufactured in 2020 will generate revenue for nine years through 2029."
Accounting useful life is also not counted from the product announcement date. CoreWeave's 2025 Form 10-K sets the estimated useful life of technology equipment, including GPUs, at six years. Since depreciation expense is recorded against fixed assets once they are placed in service, whether a contract running through 2029 exceeds the six-year depreciation period cannot be determined without knowing the equipment's activation date.
That said, there is continuity with explanations CoreWeave has repeated in the past. In its March 2026 investor materials, the company stated that its average contract term is roughly five years, that recent large-scale contracts run five to six years, and that the accounting useful life of GPUs is approximately six years. Additionally, the average price of A100 units reportedly rose during 2025. If equipment can be resold after its initial contract ends, it adds incremental revenue on top of the original investment return—a factor supporting the economics of re-contracting.
However, this doesn't mean that operating costs disappear once book value reaches zero. Power and data center rent still incur costs. The burden of maintenance, networking, and software operations also remains. To gauge the residual value of older-generation GPUs, what matters more than the length of the contract is how much the unit price and utilization rate after re-contracting exceed these ongoing costs.
Why Buyers Remain for Older-Generation GPUs
While the A100 falls short of the Blackwell generation in performance for cutting-edge, large-scale training, it hasn't lost its use cases. According to NVIDIA's specifications, the A100 80GB supports both training and inference, and can also be used for data analytics and HPC. A single unit can be partitioned into up to seven independent MIG instances. For inference and fine-tuning tasks that don't require peak performance, the A100—with its established software ecosystem and sufficient memory—remains a viable option. It can also be applied to scientific computing and to providing smaller allocations to multiple users.
Infrastructure considerations also play a major role. NVIDIA's DGX A100 houses eight A100 units in a 6U chassis with a maximum power draw of 6.5 kilowatts. In contrast, the DGX GB200 NVL72 is a rack-scale configuration that connects 72 GPUs as a single unit, consuming approximately 120 kilowatts per rack and equipped with liquid-cooling manifolds. Beyond the difference in performance generation, the two systems rest on entirely different assumptions regarding power supply, cooling, and rack design.
This comparison doesn't indicate power efficiency for achieving equivalent processing performance. What it does show is that replacing a rack of A100s with newer GPUs isn't simply a matter of swapping cards. In facilities where power supply and cooling infrastructure would need to be rebuilt, it can be faster to generate revenue by redirecting an already-operational, validated A100 cluster to a different use case.
At the end of CoreWeave's Q2 2026, active power stood at 1.5 gigawatts, while contracted power reached approximately 3.7 gigawatts. In a phase where the construction of deliverable capacity for customers hasn't kept pace with demand, existing compute resources themselves become scarce. The price of older-generation GPUs is supported not just by semiconductor performance, but by the value of already-operational power and infrastructure being immediately available.
Upside from Re-Contracting Versus CoreWeave's Capital Burden
The A100 contract is a positive signal for CoreWeave's business model. The company has explained that it structures equipment to achieve a certain return on investment within the term of the initial contract, and that reselling capacity after the contract ends generates incremental revenue. Revenue for Q2 2026 came in at $2,575 million, up 112% year-over-year, with backlog reaching $104.2 billion. Notably, this backlog figure does not even include the more than $25 billion in new commitments added at the start of Q3.
At the same time, evaluating the upside from re-contracting requires weighing it against the capital costs CoreWeave carries.
| Q2 2026 Metric | Amount |
|---|---|
| Capital expenditure | $9.4 billion |
| Depreciation and amortization | $1,393 million |
| Net interest paid | $640 million |
| Net loss | $626 million |
CoreWeave raised its full-year 2026 capital expenditure guidance to a range of $35 billion to $39 billion. As of the end of June, property and equipment stood at $46,736 million, and total liabilities had swelled to $72,046 million. Even if incremental revenue can be earned from older-generation GPUs, what will ultimately determine the company's overall return on investment is whether new capacity can be delivered to customers on schedule, whether borrowing costs can be kept in check, and whether major customers fulfill their long-term contracts.
Customer concentration also remains an issue. In its Form 10-Q, CoreWeave disclosed that its largest customer accounted for 71% of revenue in Q2 2025, and it anticipates that concentration among major customers will continue going forward. While the A100 re-contract clearly demonstrates depth in the market, the underlying structure—dependence on a small number of large-scale contracts—remains unchanged.
The A100 contract extending through 2029 makes clear that a residual market exists for older-generation GPUs. However, it's not possible to extrapolate the lifespan or profitability of the entire fleet from a single contract. What we'll want to confirm in the next earnings report is the scale of older-generation capacity approaching renewal, the unit price and duration of re-contracted deals, and the profit margins on services including managed inference. Once these are disclosed, it will become possible to judge, with actual numbers, whether the A100 has shifted from being merely "an old but usable GPU" to an asset that extends the timeline for investment recovery.