Legacy GPUs Are Turning Into Long-Lived Profit Machines
The growing profitability of older GPUs in large-scale AI infrastructure is a shift where legacy hardware economics, power-constrained data centers, and long-term Nvidia GPU contracts combine to make past-generation chips more valuable over time than many new deployments. CoreWeave is the clearest evidence that AI compute profitability is now as much about asset duration and infrastructure fit as it is about raw performance. Its second-quarter revenue climbed to 2.58 billion as demand for AI computing power keeps outpacing what the company can build, and that growth rests on fleets of Nvidia GPUs that industry observers once assumed would be obsolete in a few years. In other words, the AI boom is rewarding operators who sweat their silicon for longer, not those who chase every new architecture at any cost.
CoreWeave’s Revenue Surge and the New AI Demand Curve
CoreWeave’s revenue more than doubled in the latest quarter to 2.58 billion as customers race to rent GPU infrastructure for training and inference workloads. That explosion has pushed the company to the front of the GPU infrastructure scaling story: it rents data center capacity built around Nvidia chips to AI builders who cannot get enough compute on their own timelines. Part of the recent momentum comes from a partnership to provide AI cloud infrastructure to defense and intelligence agencies, a segment competitors have struggled to penetrate. At the same time, CoreWeave is expanding its footprint with new data centers in Asia-Pacific, signaling that demand is no longer confined to a handful of big tech labs. As its CEO put it, AI cloud demand is spreading into software, industry, financial markets, enterprise workflows, and national security. The demand curve is broadening exactly as older GPUs are proving they can stay productive far longer than finance teams once modeled.
Nine-Year Nvidia GPU Contracts and Legacy Hardware Economics
The most provocative part of CoreWeave’s model is not its growth, but its willingness to bet on legacy hardware economics. The company has signed a contract for Nvidia A100 GPUs that runs into 2029, nine years after that architecture debuted. That means chips deployed in 2020 now have a contracted revenue tail that extends beyond the six‑year depreciation schedules big cloud providers once defended. CoreWeave’s CEO has said that a batch of H100s coming off an expired contract was rebooked at 95% of the original price, while Nvidia’s finance chief notes that A100s sold six years ago still run at full utilization. These data points show a new reality: GPU value is not collapsing on two‑ or three‑year horizons. Instead, AI infrastructure scaling is being optimized around keeping prior-generation SKUs profitable for as long as workloads and power envelopes allow.
Power-Constrained Data Centers Make Old GPUs the Rational Choice
Why would anyone commit to nine-year-old GPUs in a market obsessed with the latest Nvidia architecture? Because many data centers cannot feed or cool the newest racks. An air‑cooled DGX A100 system draws about 6.5 kW at full load and fits within legacy halls designed for roughly 20 kW per rack. By contrast, current GB200 and GB300 NVL72 racks draw 120–140 kW and need direct‑to‑chip liquid cooling. Those power and cooling requirements are roughly six times what older facilities can deliver, so Blackwell cannot simply move into the same rooms where Ampere is installed without rebuilding core infrastructure. The energized, air‑cooled capacity holding A100s has no higher‑value alternative use, and renting nine‑year‑old GPUs is better than leaving that capacity dark. In practice, that makes older GPUs economically attractive for inference workloads: they match the building, not just the model.
A $100B Market Built on Cheaper Compute and Long Tails
CoreWeave’s contracted power grew to 3.7 GW in the second quarter and had reached 4.2 GW while only 1.5 GW was online. That gap between active and committed capacity shows how far ahead customers are booking GPU infrastructure. Its revenue backlog has already risen to 104 billion, even before counting more than 25 billion in fresh commitments. Analysts argue that as GPU computing costs fall, a long tail of organizations will start training and fine‑tuning proprietary models, expanding CoreWeave’s total addressable market and potentially lifting annual revenue above 100 billion over time. Managed inference has already hit 100 million in annual recurring revenue shortly after launch. The lesson is blunt: AI compute profitability is shifting from short, aggressive upgrade cycles to long‑running, high‑utilization fleets. In this world, the winners will be operators who treat GPUs like long‑lived infrastructure assets, not disposable gadgets.
