Sam Altman Warns of 'Unsustainable Silliness' in the AI Compute Buildout
The AI industry's biggest builder just told everyone to stop building.
Sam Altman, the CEO of OpenAI and the man who turned large language models into a global infrastructure race, called the AI compute buildout "unsustainable silliness" in a recent interview. His target: the neocloud providers that are announcing massive datacenter capacity without the revenue or customers to back it up.
"Too many others have a 'cost is no object' mindset," Altman said. "We're just going to build out crazy amounts of compute even at higher prices for it."
The warning is remarkable not just for its bluntness, but for who is delivering it. Altman's OpenAI is the single largest consumer of AI compute on the planet, spending billions on GPU clusters from Microsoft and others. His company's very existence depends on the infrastructure buildout he is now calling reckless.
And the irony runs deeper. Just months ago, Anthropic CEO Dario Amodei accused Altman of "YOLO-ing" money into the compute buildout, spending without regard for return. Now Altman is the one waving the caution flag, warning that the industry's enthusiasm has outpaced its economics.
The Overcapacity Problem
Altman's specific concern is the neocloud sector. These are companies that build AI computing infrastructure and lease GPU capacity to startups and enterprises. They have proliferated in the wake of the AI boom, raising billions to build datacenters that they hope AI companies will fill.
The problem, Altman says, is that too many of them are building without customers.
"Unsustainable silliness" is how he described the dynamic. Providers announce massive capacity, investors pour money in, and the actual demand has yet to materialize. The result is a market where supply is growing faster than the applications that would consume it.
OpenAI's own expansion is different, Altman assured. It is profitable. It is backed by real demand from millions of users paying for ChatGPT, API access, and enterprise deployments. But he acknowledged that even OpenAI's position is not immune to the broader dynamics.
The Efficiency Paradox
Here is where Altman's warning gets really interesting for anyone watching the AI industry.
OpenAI is working aggressively to cut the cost of running its models. Every efficiency gain, every optimization, every architectural improvement makes inference cheaper. That is good for OpenAI's customers. But it also reduces the amount of compute needed to deliver the same value.
Altman admitted that if OpenAI succeeds in making its models dramatically cheaper and more efficient, today's expensive datacenter buildouts could turn into bad bets. The infrastructure that looked necessary at today's cost structure could become superfluous at tomorrow's.
This is the efficiency paradox of AI infrastructure: the better the models get, the less compute they need to deliver the same results. A datacenter built today for a model that costs $100 per query might be obsolete when the same model costs $1.
The Risk of a Downturn
Altman also acknowledged that a broad economic downturn could strain OpenAI's ability to pay for capacity it has already committed to. He described the risk as manageable, but the fact that he raised it at all is notable.
OpenAI has signed long-term, multi-billion-dollar commitments with Microsoft and other infrastructure providers. These contracts assume continued growth in AI demand. If the economy slows, if AI adoption plateaus, or if competitors erode OpenAI's market share, those commitments could become a significant burden.
The company has already signaled some strategic shifts. In August 2026, OpenAI cut contracts with Cursor and terminated some Musk-related deals, suggesting a tightening of focus. The "unsustainable silliness" comment fits a pattern of increasing caution from the company that started the AI boom.
What Comes Next
Altman did not rule out OpenAI eventually selling compute to third parties, though he said the company is not doing so yet. That is a significant change from the current model, where OpenAI consumes compute rather than providing it. If OpenAI ever becomes a compute provider itself, it would compete directly with the neocloud providers it is now warning about.
The broader question is whether Altman's warning reflects a genuine shift in OpenAI's strategy or a tactical message to the market. If the company is signaling that it expects compute costs to fall, it may be telling investors and competitors that the era of unlimited spending is ending.
For the neocloud providers building datacenters on the assumption that AI demand will grow forever, Altman's words are a warning shot. The party may not last as long as everyone hoped.
And for the rest of the industry, the message is clear: the man who built the machine is now questioning whether anyone can afford to run it.