There is a point in the AI race when demand stops being measured only in users and starts being measured in compute.
OpenAI appears to have reached that point. The company has temporarily paused new sign-ups and upgrades for its $200-per-month ChatGPT Pro plan, citing unprecedented demand for GPT-6 Astra and the strain that the highest-usage tier places on its systems. Existing Pro subscribers are unaffected, while Plus, Go and API access remain available.
The decision is notable because OpenAI is not restricting access to Astra because of a lack of interest. It is doing the opposite: demand has become large enough to affect capacity planning.
Why Astra Is So Demanding
GPT-6 Astra, released on September 3, represents a different kind of workload from earlier generations of ChatGPT.
The model is designed not only to generate text, but to operate computers, navigate websites, update CRM systems, fill out forms, analyze data, create websites, test software and perform complex multi-step tasks. OpenAI reports a 72.6% score on OSWorld 2.0, compared with 65.7% for GPT-5.6 Sol, while also reporting a 57.9% score on Terminal-Bench 4.0.
That matters because an AI that performs work consumes resources differently from an AI that simply answers a question.
A request to summarize a paragraph may require a relatively short interaction. An agent working inside a browser or software environment can involve multiple actions, observations, reasoning cycles and tool calls before the task is complete.
The more capable the agent becomes, the more valuable each interaction can be — and the more computationally expensive it can become.
The $200 Plan Became the Pressure Point
OpenAI’s product leader Thibault Sottiaux said demand for Astra was “unprecedented” and warned that the company might need to pause new Pro subscriptions if the growth continued.
The company ultimately chose to restrict the highest-usage tier rather than reduce access across the entire product.
It is a familiar problem in OpenAI’s history.
In November 2023, the company temporarily paused new ChatGPT Plus sign-ups after demand following its first developer conference exceeded available capacity. But the scale of the AI infrastructure supporting OpenAI has changed dramatically since then.
The difference today is that the bottleneck is emerging around a much more capable generation of AI.
AI Infrastructure Is Becoming Part of the Product
The episode reveals something easy to overlook when looking only at model benchmarks.
The quality of an AI model is increasingly inseparable from the infrastructure capable of serving it.
OpenAI has been expanding its computing capacity through partnerships and large infrastructure commitments. Fortune reports that the company expects significant additional capacity from Nvidia, AMD and Broadcom systems beginning in the second half of 2026, while its broader Stargate initiative represents a planned $500 billion, four-year AI infrastructure buildout.
Yet even that expansion does not eliminate the problem.
As models become better at reasoning and acting, users have more reasons to use them — and sophisticated agentic tasks can consume considerably more compute than traditional chatbot interactions.
That creates a feedback loop:
Better models → more useful tasks → more usage → more compute demand.
The Bigger Story
The temporary pause in Pro subscriptions is therefore more than a pricing or availability decision.
It is a small demonstration of a much larger transition taking place across the AI industry.
For years, the central question was whether companies could build models powerful enough to attract users.
Now another question is becoming equally important:
Can they build enough infrastructure to serve what those models make possible?
GPT-6 Astra is pushing AI further toward software that can act rather than simply respond. OpenAI’s decision to temporarily stop selling its most demanding subscription tier shows the physical economics behind that transition.
The AI race is no longer only about who builds the smartest model.
It is also about who can build — and afford — the infrastructure required to let millions of people use it at scale.
