How much power is enough for artificial intelligence? No one knows, not even OpenAI CEO Sam Altman or Microsoft CEO Satya Nadella.
This has put software-based companies such as OpenAI and Microsoft in a bind. Much of the tech world has focused on computers as a major obstacle to the development of artificial intelligence. And while tech companies scramble to secure power, those efforts have delayed GPU purchases to the point that Microsoft has apparently ordered too many chips for the amount of power that has shrunk.
“The supply and demand cycles in this particular case you can’t really predict,” Nadella told the BG2 podcast. “The biggest problem we’re facing now is not overkill in computing, but it’s a power and it’s a kind of ability to get [data center] constructions are done fairly quickly near power.’
“If you can’t do that, you might have a bunch of chips in inventory that I can’t connect. In fact, that’s my problem today. It’s not a matter of chip supply, it’s the fact that I don’t have hot shells to connect,” Nadella added, referring to the commercial real estate term for tenant-ready buildings.
In a way, we’re seeing what happens when companies accustomed to dealing with silicon and code, two technologies that scale and grow quickly compared to massive power plants, have to step up their efforts in the energy world.
For more than a decade, US electricity demand has been flat. But over the past five years, demand from data centers has begun to grow, outpacing utilities’ plans for new capacity. This has led data center developers to add power in so-called behind-the-meter arrangements, where electricity is fed directly into the data center, bypassing the grid.
Altman, who also participated in the podcast, believes that problems could arise: “If a very cheap form of energy is released on a mass scale soon, then a lot of people will be extremely burned with the existing contracts that they have signed up for.”
Techcrunch event
San Francisco
|
13-15 October 2026
“If we can continue this incredible reduction in cost per unit of intelligence — let’s say it’s been an average of 40 times for a given level per year — you know, that’s like a very scary exponent in terms of infrastructure development,” he said.
Altman has invested in nuclear energy, including fission startup Oklo and fusion startup Helion, along with Exowatt, a solar startup that collects the sun’s heat and stores it for later use.
However, none of these are ready for widespread deployment today, and fossil-based technologies such as gas-fired power plants take years to build. Additionally, orders placed today for new gas turbines likely won’t be fulfilled until later this decade.
That’s partly why tech companies are adding solar at a rapid clip, driven by the technology’s cheap cost, emission-free power and ability to rapidly deploy the technology.
Subconscious factors may also be at play. Solar PV is in many ways a parallel technology to semiconductors, and one that has been discarded and commercialized. Both photovoltaics and semiconductors are built on silicon substrates, and both production lines begin as modular components that can be packaged together and connected in parallel arrays that make the complete component more powerful than any individual unit.
Because of the modularity and speed of solar development, the pace of construction is much closer to that of a data center.
But both still take time to build, and demand can change much faster than either a data center or a solar project can be completed. Altman admitted that if AI becomes more efficient or if demand doesn’t grow as he expects, some companies may be saddled with idle power plants.
But from his other comments, he doesn’t seem to think that’s likely. Instead, he seems to strongly believe Jevons paradoxwhich says that more efficient use of a resource will lead to greater use, increasing aggregate demand.
“If the price of computation per similar unit of intelligence or whatever—however you want to think of it—dropped by 100 tomorrow, you’d see usage increase by much more than 100, and there would be a lot of things that people would want to do with that computation that just don’t make any economic sense at the current cost,” Altman said.
