On September 11, OpenAI did something that sounds like a business going backwards: it stopped taking money. The company suspended new sign-ups and upgrades for ChatGPT Pro 20X, its top consumer plan at $200 a month, saying demand for its new flagship model, GPT-6 Astra, had outrun the compute it could supply. Existing subscribers keep their access, and the cheaper tiers and the pay-as-you-go API stay open. But the most expensive plan on the menu is, for now, sold out, and the reason is revealing.
Astra, released on September 3 and billed by OpenAI as its most powerful model, can "use a computer" the way a person does: filling in forms, browsing, writing and running code, grinding through workflows that stretch across days. That single capability quietly rewrote the company's cost math. In the old chatbot era, demand was roughly the number of users times how often they asked a question. In the agent era, one person can set several agents running at once, each churning for hours, so token consumption no longer tracks headcount. By one estimate cited internally, the number of active agents can grow 20 percent while the tokens they burn jump 300 percent.
That is why the top tier buckled first. According to calculations by the research firm SemiAnalysis, OpenAI starts losing money on the Pro 20X plan once a subscriber uses just 5.7 percent of the quota it includes. The people who pay $200 a month are precisely the ones who will use it: developers running long agent chains, engineers leaving code assistants going overnight, researchers dispatching hundreds of sub-tasks a day. Thibault Sottiaux, OpenAI's head of core products, admitted the tier "puts the most strain on our systems," and ruefully conceded he had underestimated Astra after once assuring users the company had "plenty of compute."
OpenAI has been here before. In late 2023, a post-launch surge forced it to pause new ChatGPT Plus sign-ups. What is different now is scale. The company's available computing power grew roughly 9.5 times between 2023 and 2025, and it still was not enough. So it is spending to catch up on a scale that is hard to picture: the $500 billion Stargate buildout, at least 10 gigawatts of Nvidia systems, a 6-gigawatt deal with AMD, and 10 gigawatts of custom chips through Broadcom, on the way to a stated $750 billion in infrastructure by 2035.
It is not alone in treating raw capacity as the prize. In the same stretch, Google committed about $15 billion to data centres in Finland, complete with a 22-year deal to buy nuclear power, its first such arrangement outside the United States. Qualcomm said Amazon could buy up to $60 billion of its AI data-centre chips, a bid to break beyond phones and become an infrastructure supplier. The race, increasingly, is for electricity, cooling, and silicon rather than for the next clever model.
There is a tell in how OpenAI handled the squeeze. On the very day it froze Pro 20X, it rolled out a slate of enterprise products: ChatGPT for Financial Services, a data-agent plugin, a public beta of its Agents API, and an API for its voice model. With a finite pool of compute and an IPO reportedly possible early next year, steering scarce capacity toward business customers who pay predictably, and nudging bursty individual demand onto the metered API, is simply good arithmetic. The lesson of the week is that in the agent era, the constraint on artificial intelligence is turning out to be stubbornly physical.