OpenAI Has a Compute Problem. The Whole Industry Is About to Find Out.

The most important thing Greg Brockman said last week had nothing to do with benchmarks.

"We're heading to a compute-powered economy," OpenAI's president told me in an interview published Thursday. "There is not going to be enough compute in the world to meet the demand."

That is the story. Not GPT-5.5's scores on Terminal-Bench 2.0, not the debate over whether scaling laws are still alive, not the usual round of model-vs-model comparisons. The story is that the most consequential AI company in the world is running out of the hardware it needs to run its own products.

The evidence is not subtle. OpenAI CFO Sarah Friar said in a separate interview that the company is passing on opportunities because it lacks compute. "If you do not have it, you do not have revenue," she said. "That is one thing I know for sure." OpenAI discontinued its Sora video application and reallocated the underlying hardware to higher-priority work. Token pricing for GPT-5.5 runs double that of its predecessor. The company recently closed a $122 billion funding round, a significant portion of which is earmarked for securing future compute capacity.

These are not the signs of a company executing a plan. They are the signs of a company under resource pressure.

Inside OpenAI, the constraint is already reshaping how work gets done. More than 85 percent of the company's employees now use Codex every week, across every function, including finance, communications, and marketing. The compliance team used it to review 71,637 pages of K-1 tax forms, cutting two weeks off the prior year's timeline. Brockman described the shift in stark terms: the adoption of agentic coding tools inside the company went from "night and day before December versus after December."

The outside world is noticing. Jensen Huang called GPT-5.5 a "huge achievement" and "a great example of the arrival of agentic AI that does actual work, not just answers questions." NVIDIA deployed it across their own engineering workflows. This is the hardware company's CEO acknowledging that the software has reached a threshold.

What makes the moment significant is the structural nature of the constraint. Brockman was careful to frame this not as a temporary shortage but as a fundamental mismatch between demand and physical infrastructure. "We are talking 10 million GPUs, 50 million GPUs," he said. "Maybe you could squint and get a little bit more than that. But getting to billions of GPUs, getting to one GPU per person, that is not in anyone's cards." The 8 billion GPUs-per-person figure that some use as a total addressable market framing is, in his telling, not a ceiling to reach but a number that will never be reached.

The implication is not that AI will stop improving. It is that access to improvement will be rationed. The companies and labs that hold GPU allocations will determine what gets built and what gets killed, which products ship and which get quietly discontinued, who can train at the frontier and who is pushed to the application layer. This is an economics story wearing a technology costume.

It is also not a story about OpenAI alone. Anthropic recently tightened usage caps on Claude during peak hours. Google has committed $40 billion to Anthropic in part to secure compute access. The entire industry is making multi-year commitments to chipmakers, building custom silicon, and rationing internal tools. The constraint is real and it is shared.

The counterargument is that GPU production is scaling and new fabs are coming online, that the shortage is a moment in time rather than a permanent condition. That may be true. TSMC is expanding. Blackwell supply is increasing. But the timelines for building new data centers and commissioning new chip lines run in years, not quarters, and the demand curve shows no sign of flattening.

What changes next is not open to much debate. The labs with compute will set the agenda for the rest of the industry. Custom silicon development, already underway at OpenAI, Google, Amazon, and Meta, becomes a survival priority rather than an optimization project. Smaller labs and academic researchers face a narrowing window of access to the training infrastructure that defines the frontier.

Brockman said OpenAI's chip team is "punching above their weight compared to other programs out there". Whether that turns out to be true, and whether it matters more than the relationship with Nvidia, is the next question. But the broader point is already settled: the bottleneck is physical, it is here, and it is not moving.