CoreWeave's 2029 A100 deal puts a number on GPU residual value

On its August 11th earnings call, CoreWeave told investors it had signed a contract for A100 capacity extending into 2029, at what it described only as "an attractive price" ^[1]. Two days later Jensen Huang amplified the point on X, writing that "The mighty A100 fleet are mission-capable from 2020 through 2029" and crediting CUDA with keeping Ampere, Hopper and Blackwell systems upgradeable across their operating lives ^[2][3].

Nvidia introduced the A100 on May 14th, 2020 as the first data-center GPU built on Ampere ^[4]. A contract running into 2029 puts that silicon to work as much as nine years after launch, across several generations of faster parts ^[5]. For anyone modelling accelerator fleets on a three-to-five-year strand, that is a material difference.

The financing mechanics are where it bites. Chief Financial Officer Nitin Agrawal told investors that a typical five-year customer contract requires heavy upfront capital spending, funded through debt, customer prepayments and corporate capital, and that CoreWeave expects the initial contract to repay the asset-level debt attached to a cluster ^[6][7]. Anything after that, whether renewal or resale, arrives without the original financing burden ^[8]. On CoreWeave's own framing, a nine-year mission-capable life against a five-year initial contract leaves roughly four years of earning capacity on already-deleveraged iron ^[3].

That matters because the rest of the quarter is heavy. CoreWeave reported $2.575 billion in second-quarter revenue with a $626 million net loss, $640 million in net interest expense, and $9.4 billion in capital expenditures ^{[9][10][11][12]}. Interest alone ran near a quarter of revenue ^[1], and capex was roughly 3.7 times quarterly revenue ^[2]. Extending revenue lives is one of the few levers that improves the return on that spend without new customers.

The secondary-market signal is the more interesting datum. CoreWeave said its earlier Ampere and Hopper fleets remain largely sold out and that prices for older generations are holding near or above levels seen a year earlier ^[13], with customers running prior-generation infrastructure across a wider set of workloads while the newest clusters take the most demanding jobs ^[14]. Nvidia's documentation supports the plumbing: its CUDA architecture matrix lists Ampere toolkit and driver support as ongoing, the driver API is backward compatible so newer drivers run applications compiled against older toolkits, and the framework support matrix still lists A100 alongside Hopper and Blackwell ^[15][16][17].

Two caveats travel with this. Software compatibility does not close the performance gap between a 2020 A100 and current hardware; it only keeps the older part addressable for jobs that do not need the fastest training or inference ^[18]. And the 2029 contract is one data point, not proof that A100 clusters generally stay utilized or economically competitive that long, with workload requirements, electricity costs, customer pricing and newer-generation performance deciding which survive ^[19][20]. Nvidia is not a neutral witness either: it bought $2 billion of CoreWeave Class A shares at $87.20 on January 26th and agreed to help CoreWeave secure land, power and buildings using its own balance sheet ^[21][22].

Watch for whether renewal pricing on prior-generation capacity shows up as a disclosed line rather than a characterisation, and whether any operator publishes the useful-life assumption it actually books.