A chip that is three years old is supposed to be worth a fraction of what it cost new. That is how depreciation schedules on Wall Street treat computing equipment, and for decades it matched reality. On Moonshots with Peter Diamandis, the panel worked through a number that points the other way: the hourly cost of renting NVIDIA's H100 processors rose 22% in a single month, to $3.28 per hour.
The figure comes from an index run by Ornn, a company that builds financial products around computing capacity. Before anyone discussed it, Diamandis read out the disclosures: Ornn is a Link Ventures portfolio company, it sits in the AI venture fund he runs with Dave Blundin, founder and general partner of Link Ventures, and Alex Wissner-Gross, a computer scientist and founder of Reified, advises the company. Later in the discussion, one of the panelists said he has a financial interest in Ornn and has written essays and made announcements on its behalf. Emad Mostaque said he has no financial interest in it.
What the index actually measures
An H100 is a graphics processor from NVIDIA's Hopper generation, the kind of chip used to train and run AI models. Most companies do not buy them outright; they rent time on them from cloud providers, priced by the GPU-hour. Ornn's index tracks that rental price, and the company says its measurements come from completed transactions rather than advertised offers, surveys or estimates.
That distinction matters for reading the 22% move. It describes what buyers and sellers actually paid to use Hopper chips over a month, not what the hardware itself would fetch in a resale, and not the prices a provider happens to be posting on a web page. Rental income and residual hardware value are related, but they are not the same measurement.
NVIDIA's chief executive, Jensen Huang, responded to the index on X by calling the chips "fungible, durable, and highly rentable, a productive revenue-generating asset."
"The worst thing you could ever buy is a chip"
Blundin said he had spent the morning fielding texts from large quantitative trading funds asking about Ornn. His explanation for the attention was a change in what a chip is as an asset. "Moore's law died. Chips are not commoditizing," he said. For his whole life, he argued, "the worst thing you could ever buy is a chip and put it in your closet because it depreciates faster than anything on the planet."
In his account that has now reversed: a GPU bought a year ago is worth more, and high-bandwidth memory — the fast memory stacked alongside AI chips, known as HBM — is up fivefold in value on his estimate. He expects the trend to hold at least until large new fabrication capacity comes online, and said that if demand for intelligence keeps finding new uses, "it may never go the other direction." Diamandis called it scarcity inside an abundance story and pointed to HBM supply and new fab projects as the solutions arriving in response.
One panelist, who also said he has a financial interest in Ornn, proposed the index itself as the test for the thesis. If the trend reverses, he said, it will show up there: the line starts going down instead of up. For this stage of AI, the same speaker argued, "the flops, the tokens and the outcomes, those are the commodities of this moment" — adding, without wanting it taken as financial advice, that they are "the oil of the singularity at this point in time."
Why an old Hopper can be worth more than it was
Mostaque, who said his company received some of the first Hoppers when he was running Stability AI, offered the operating explanation. A GPU, in his description, is "a means of transforming electricity into intelligence." What changes over time is not the chip but what you can run on it.
Think about the model that fitted on a Hopper when the chips first arrived three or four years ago, he said, and compare it with the level of intelligence that fits on the same hardware now: "It's way more. It's faster. It's cheaper. It's better." The chip converts the same electricity into far more useful output than it did on the day it shipped, which is why he did not find the rental price surprising. He tied that to demand arriving from several directions at once — visual and physical applications hitting the market alongside text.
There is counterpressure on the same mechanism. Newer architectures that cut the memory a model needs to serve users, such as the DeepSeek release the panel discussed elsewhere in the episode, change how much hardware a given workload requires. Mostaque's own example of demand came from the same company: he said DeepSeek had put out a note that day asking anyone building with 2,000 chips or more to get in touch, and that other large labs were making similar requests in their research papers — "call for chips, anyone."
The warning for companies that plan to do AI later
The practical consequence Blundin drew was about access rather than asset prices. One of the biggest mistakes corporate chief executives are making, he said, is assuming compute will be available when they want it, because it always has been. He said that anyone calling Amazon's AWS right now for NVL72 systems — NVIDIA's rack-scale machines — would be told the capacity is sold out years into the future.
His conclusion was that a company cannot treat AI as something to start later. It needs a data center and compute plan "really in the next couple of months," in his framing, or it gets frozen out. Diamandis folded the same point into an investment theme he expects to hear about at the Moonshot Summit: data centers, energy, fabs and anything else tied to compute facing what he called near-infinite demand.
What the month's move does not settle is an argument the panel has had before. In an earlier Moonshots discussion of NVIDIA's $500 billion financing plan, Salim Ismail raised the risk that a cheaper architecture strands equipment before the financing attached to it is repaid, while Mostaque pointed to paid-off A100s still earning on stable workloads. A rising rental index is a reported observation about one month in one market for one generation of chip. It tells you what Hopper time traded for; it does not tell you what the hardware will be worth, or that technological risk has left the calculation.