There’s something strange happening right now in the tech industry. To some, it’s a bubble. To others, it’s just a market trend. To a certain class of wonkish statisticians, it’s a type of Jevon’s paradox, which we learned a lot about through the past few years of market activity.
One abiding truism in tech is that Nvidia is killing it: the company’s GPUs are the best, and everybody wants the best, so everybody is buying.
However, a scale-out of vendor services classically means that supply should meet or outstrip demand.
That’s not what the folks at Nebius are seeing as they auction off dormant data center compute.
Access and Ownership
Are Nvidia GPUs themselves rare? That’s kind of beside the point, for reasons related to post-cloud-era business strategy.
In other words, you don’t have to own Nvidia chips, Blackwell or otherwise, to use them. You just buy compute from a vendor. Like Nebius. This allows access without the burden of owning something.
Keep in mind that right now, Nvidia has just opened up five hundred billion in loans for GPU owners to treat them as conventional assets, like real estate.
Anyway, what Nebius is seeing is robust demand for their auctioned-off capacity, running on Blackwell chips, at relatively high bid prices.
Now, ironically, even as Nvidia GPU capacity is popular, data centers are not. It’s even possible that the pushback, in terms of municipal policy, protests by local residents, and other headwinds are contributing to the high price of vendor services. But data centers are being built. So the question remains: why are people outbidding each other at Nebius auctions?
I got the news from my favorite podcast, Nathaniel Whittemore’s AI Daily Brief, where a posted blurb announces that Nebius “cleared its Blackwell compute auctions at 15% above its previous record price for Hopper.”
The simplest answer would probably be that supply doesn’t always tamp down demand, not if something is popular enough.
Jevon’s Paradox at Work
If you’re not familiar with Jevon’s paradox, business people have been making a lot of use of it in the AI era.
Jevon’s paradox basically states that as the supply of something increases, demand will not slack if the consumer base simply uses more and more of the resource.
So in the agentic age, if buyers are using higher levels of Nvidia compute to create greater numbers of AI agents, that’s a plausible reason why price will stay stubbornly high, and scarcity will persist.
It’s simply an enormous step, from glorified chatbots that are isolated in a browser page, to agents that can go out and do things on their own. We’re seeing the cybersecurity ramifications with Mythos and new GPT models, and I guess we’re seeing it play out in the market as tenacious demand pressure.
The Old Stuff – and the New Efficient Stuff
AIDB analysis of the situation also uses another term – token austerity.
What I think this means is that engineers have been creating LLMs and systems that use less compute to do more tasks, models that function more with fewer tokens or parameters. But even that isn’t quelling the thirst for compute.
Also, the seemingly boundless appetite on the part of compute customers isn’t limited to services running on Blackwell architectures. The old gear, it turns out, is good, too.
Other news from AIDB is around CoreWeave, another data center compute provider, which reportedly signed contracts for A100-generation equipment extending into 2029, suggesting AI demand currently spans multiple Nvidia generations rather than moving exclusively to the newest chips.
“On August 11, 2026, CoreWeave disclosed that it had signed a contract for cloud capacity using the NVIDIA A100, a GPU that debuted in 2020, running through 2029,” writes Y Kobayashi at XenoSpectrum. “This serves as a concrete example showing that even as cutting-edge GPUs continue to be refreshed, multi-year demand persists for older generations as well.”
H100 and H200 Mvidia GPUs are similarly still in demand, and those closest to cutting-edge hardware are looking ahead to the expansion of Nvidia’s new Vera Rubin build. So that desire for Nvidia-based compute is comprehensive, not a flash in the pan.
Imagine, if you will, an auctioneer, yelling for bids, and not even getting into swing before escalating: ten, twenty, thirty, forty, etc. Now imagine a whole host of such auctioneers, doing their work in a vast hall, where IT constitutes 40% of all business being done from sea to shining sea.
It’s really kind of mind-blowing. This is the backdrop against which we operate conferences and lecture events at MIT, and elsewhere, to keep thinking about how the systems using all of this compute affect our societies. Because AI, it seems, is here to stay.
