Nvidia CEO Jensen Huang says rising AI chip rental prices prove that older hardware keeps earning, after rates for a three-year-old chip climbed 22% in a month.
The chip is the H100, the training processor behind the first wave of generative AI. It now rents for $3.28 an hour. Huang calls that durability. The longer record complicates the claim.
Older AI Chip Rental Prices Jump 22% in a Month
Index data circulated on X puts the hourly rate at $3.28, a gain of roughly 22% over the past month. Huang shared the chart and described Nvidia compute as fungible, durable, and revenue-generating.
The rebound cuts against standard accounting. Hyperscalers write down graphics processing units (GPUs) over roughly five to six years. Michael Burry argues the real useful life runs shorter, and he expanded his Nvidia short in late August.
Why One Month Does Not Settle the Depreciation Debate
History cuts both ways here. On-demand capacity for the same chip once cost $7 to $8 an hour at large cloud providers, and early rates ran higher still. Against that base, $3.28 remains a steep markdown.
Supply also explains part of the move. Newer Blackwell systems stay reserved for the biggest buyers, so older clusters keep filling inference workloads. Tight power and memory keep the whole market expensive.
Then there is who pays the rent. CoreWeave, a cloud firm that buys Nvidia chips and rents them out, carried $35 billion of debt at the end of June. Nvidia, meanwhile, agreed to rent back unused capacity from such partners. Critics count that among the off-balance-sheet AI deals that flatter demand.
Nvidia posted record second-quarter revenue of $96.2 billion in August, and the stock rose more than 4% after the report. Rents for aging silicon give Huang a fresh talking point. A second month of gains would turn that point into a trend.