Old A100s, New Money: The Longevity Shock
A100 GPU profitability describes how Nvidia’s 2020 Ampere-based AI accelerators still generate high, contracted revenue years after deployment, thanks to continuous AI demand, legacy data center power limits, and cloud providers’ ability to monetize older hardware through inference workloads and managed services. CoreWeave has now signed a contract for Nvidia A100 GPUs that runs into 2029, meaning 2020 silicon will still be bringing in money nine years after launch. That is not a rounding error; it overturns the old assumption that GPUs become economic dead weight in a few years.
This is happening because AI itself has changed. CoreWeave’s CEO argues that AI is no longer confined to frontier model labs and is instead embedded in software, industrial systems, financial markets, enterprise workflows, and national security missions. He also describes training, inference, evaluation, and improvement as a single continuous loop where production systems feed back real-world data and drive constant model updates. In that world, compute is not a one-off capital spike. It becomes an ongoing requirement that grows with every deployed application and every cycle of improvement. The surprise is that this ongoing demand is not only soaking up new GPUs; it is keeping Nvidia aging chips fully booked.
Power Limits Make Legacy GPUs Economically Relevant
The key to understanding legacy AI hardware economics is not benchmark charts; it is power and facilities. An air-cooled Nvidia DGX A100 system draws about 6.5kW at maximum load and fits inside legacy data halls designed for roughly 20kW per rack. By contrast, Nvidia’s current GB200 and GB300 NVL72 racks pull around 120kW to 140kW and demand direct-to-chip liquid cooling, roughly six times what those older halls can feed and cool. That physical mismatch is doing more to preserve A100 profitability than any software miracle.
Those older halls are already wired, cooled, and paid for. They cannot host the latest liquid-cooled monsters without a rebuild, but they can run A100 fleets comfortably. The energized, air-cooled capacity holding those A100s has no higher-value use, so renting nine-year-old GPUs is better than leaving racks dark. From a GPU infrastructure longevity standpoint, this is rational: the bottleneck is power density, not raw FLOPs. As long as some workloads can live within the A100’s performance and energy envelope, those chips remain economically competitive for inference and mid-tier training.
Continuous AI Demand Turns Old Chips Into Cash Flow Machines
The economic twist is that AI workloads never stop. CoreWeave’s leadership describes a world where training, inference, evaluation, and improvement form a continuous loop and production models feed real-world data back into new experiments. Compute is no longer a one-time requirement at the start of a model’s life; it becomes an ongoing requirement that grows with every application and every improvement cycle. In that environment, demand spills down the stack. Frontier models may claim the latest silicon, but there is a long tail of inference, fine-tuning, and evaluation tasks that do not care about being one architecture behind.
This is why pricing for prior-generation SKUs is at or above where it was years ago. According to Nvidia’s CFO, A100s sold six years ago still run at full utilization. That is a quote-worthy reversal of the old view that GPUs have a one-to-three-year datacenter service life, as one Google architect once suggested. Instead of sliding into obsolescence, A100 fleets are booked out under multi-year contracts, with the latest deal stretching contracted revenue on 2020 silicon past the six-year depreciation schedules that hyperscalers previously defended. Continuous AI compute demand has created a durable, profitable secondary market where yesterday’s flagship is today’s inference workhorse.
CoreWeave’s Bet: Squeezing Value From Every Watt and Die
CoreWeave is the clearest example of a business model built around extracting value from aging hardware. Its A100 contract running into 2029 shows a deliberate choice to lock in revenue from older GPUs instead of ripping and replacing them. The company’s contracted power has already grown to 3.7 GW in one quarter and reached 4.2 GW as of a recent Monday call, while only 1.5 GW is online. Customer commitments already cover nearly triple the capacity the company can currently deliver. That imbalance pushes CoreWeave to treat every powered rack as a cash-generating asset, whether it holds Blackwell or Ampere.
The economics are blunt: that mismatch between what old halls can support and what new racks require keeps A100 fleets in service regardless of how gracefully the silicon ages. Renting nine-year-old GPUs is better than leaving energized capacity unused. At the same time, CoreWeave is trying to climb the stack with managed inference services and an “AI native platform” built around the continuous loop of training, inference, evaluation, and improvement. The legacy chips do the heavy lifting while the company sells higher-margin services on top, reshaping GPU infrastructure economics away from one-off capex and toward recurring, hardware-agnostic income streams.
The Real Lesson: Obsolescence Is Now a Financial Choice
The story of A100 GPU profitability is not about nostalgia for old hardware; it is about who controls the bottlenecks. Power delivery and cooling, not transistor counts, decide whether Nvidia aging chips stay online. As long as legacy facilities cannot host 120kW, liquid-cooled racks, air-cooled A100 systems drawing 6.5kW per node have a comfortable niche. In parallel, AI’s shift to a perpetual loop of deployment and improvement means there is always one more inference job that fits on last-generation hardware.
This flips the traditional view of GPU infrastructure longevity. Concerns that faster GPU cadences would unravel AI financing assumed that hardware obsolescence was inevitable on a one-to-three-year clock. Instead, we are seeing that obsolescence is now a financial decision. Operators can stretch useful life past six years if the power envelope, workload mix, and contracts line up. The conclusion is uncomfortable for anyone hooked on annual upgrade narratives: in AI infrastructure, the winning strategy may be less about chasing every new architecture and more about learning to profit from the silicon you already have.
