Photo: Steve Jurvetson / Wikimedia Commons / CC BY 2.0 (https://creativecommons.org/licenses/by/2.0)
The OpenAI CEO says an AI infrastructure glut is coming, and it could reshape the economics of the entire tech sector
Sam Altman thinks the AI industry might be building too many data centers. The OpenAI CEO warned in a recent podcast interview that a compute oversupply could emerge within the next two to three years, potentially upending the massive infrastructure bets being placed by the world’s largest tech companies.
The case for too much compute #
Altman laid out several scenarios that could tip the balance from scarcity to surplus. The big ones: AI models becoming dramatically more efficient, the cost of intelligence per unit dropping sharply, abundant cheap energy coming online, and breakthroughs that let models run locally instead of in massive data centers.
In his words, a compute glut “will likely happen for sure” within two to three years, or possibly on a longer timeline of five to six years. That’s a wide range, but the confidence in the direction is notable.
One training run at OpenAI recently consumed as much compute as the company’s entire previous footprint, according to Altman. That sounds like demand is exploding. But the flip side is that if models learn to do the same work with a fraction of the resources, all those shiny new data centers become very expensive paperweights.
The current reality is still one of scarcity #
To be clear, nobody is swimming in excess compute capacity right now. Microsoft CEO Satya Nadella has pointed to power availability and data center construction timelines as the binding constraints on AI infrastructure today.
This creates an awkward timing problem. The industry is simultaneously dealing with not enough compute today and potentially too much compute tomorrow. Companies are making capital allocation decisions now, committing tens of billions of dollars to infrastructure that won’t come online for two to four years, based on demand projections that assume current growth rates hold.
There’s already early evidence of the shift. As of mid-2026, Meta announced plans to sell off excess AI compute capacity, signaling that at least one major player had already reached a point of internal surplus.
What this means for investors #
The compute oversupply thesis has real implications for a massive swathe of the market. Consider the chain of companies benefiting from the AI infrastructure buildout: chip designers like Nvidia, cloud providers like Amazon, Microsoft, and Google, data center operators and REITs, power companies positioning themselves as AI energy suppliers, and construction and cooling equipment manufacturers.
If Altman is right and a glut materializes on the shorter end of his timeline, utilization rates at data centers could fall significantly, compressing margins for cloud providers and potentially causing brutal repricing for companies further up the supply chain. Energy economics add another variable. Altman flagged abundant cheap energy as one of the conditions that could accelerate an oversupply. Cheaper energy lowers operating costs, but it also lowers barriers to entry, which means more competition and more potential overcapacity.
For investors in AI-adjacent crypto infrastructure, like decentralized compute networks and GPU tokenization protocols, the implications are particularly sharp. These projects are built on the thesis that compute is scarce and getting scarcer. An oversupply would undermine that narrative at its foundation. Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our