A 6.5 kW rack still fits in a data hall that a 140 kW Blackwell one would cook.
CoreWeave has booked Nvidia A100 through 2029. CEO Mike Intrator confirmed the contract on the company’s second-quarter earnings call, and the arithmetic is the story: the A100 shipped in May 2020, so a customer is paying to run that silicon nine years after launch, two full architectures past the point where it should have been scrap.
Jensen Huang went out of his way to amplify it. “The mighty A100 fleet are mission-capable from 2020 through 2029,” he posted, arguing that CUDA is what keeps the hardware alive: “Versatility makes it fungible. Fungibility drives utilization and extends durability, making NVIDIA compute a productive asset: rentable, durable and financeable.”
The real reason is power, not CUDA #
Huang’s framing is a software story. The more convincing explanation is electrical. An air-cooled DGX A100 pulls around 6.5 kW flat out, which slots neatly into legacy halls built for roughly 20 kW per rack. Nvidia’s current GB200 and GB300 NVL72 racks want 120–140 kW and direct-to-chip liquid cooling: call it six times the power, plus a cooling loop those buildings simply do not have.
So the A100 is not competing with Blackwell for a slot. It is sitting on floor space Blackwell physically cannot occupy, in buildings that are already built, already energised and already paid for. Given how long grid connections and new data centre construction take right now, that otherwise-dead capacity is worth something.
Memory helps too. The 80 GB HBM2E variant looks pedestrian next to HBM3E, but HBM is sold out across the industry into 2027, and 80 GB is 80 GB. TechPowerUp put it neatly: in a shortage, even old HBM is acceptable when there is enough of it per card.
Nobody is discounting #
Intrator’s other number is the one that should worry anyone waiting for cheap second-hand accelerators. He said pricing on prior-generation parts is at or above where it sat years ago, and told CNBC that a batch of H100s coming off an expired contract was rebooked immediately at 95 percent of the original rate. Old GPUs are not being liquidated. They are being re-let at full freight.
That drops straight into the running argument about depreciation. Hyperscalers assume five- or six-year useful lives for AI silicon; Michael Burry has accused them of understating depreciation by $176 billion between 2026 and 2028 by stretching those assumptions against a cadence that now ships a new Nvidia architecture roughly every year. A 2029 contract on 2020 hardware is a point for the long-life camp, though one customer signing one deal is evidence rather than proof, and CoreWeave has an obvious interest in the answer.
For everyone else the implication is dull and unwelcome. While 2020 silicon still earns rent, none of it reaches the secondary market, and the AI build-out keeps absorbing the same HBM, packaging and substrate capacity that consumer graphics cards need. The wait for a cheap 80 GB card on eBay just got several years longer. Reported by Tom’s Hardware.