OpenAI said demand for its new flagship model, GPT-6 Astra, has become so strong that the company may need to temporarily halt new ChatGPT Pro subscriptions if current usage levels continue.
Tibo, OpenAI’s head of core product and platform and a contributor to Codex, wrote in a post on X that Astra demand was "truly unprecedented." He said OpenAI was using every method available to support the load, but added that even after seeing several periods of very steep growth, he had never seen demand at this scale. Tibo said the company’s immediate priority was to preserve service quality for existing users. If Astra demand remains where it is now, OpenAI may have to pause new Pro subscriptions.
Capacity strain appeared in Astra’s first week
The supply-demand imbalance was visible from the first day of launch. After introducing GPT-6 Astra on Sept. 3, OpenAI did not immediately release it to all paid subscribers. The company instead used a phased rollout, initially giving access to some enterprise and security-related customers before gradually extending availability to Pro, Business, Plus, and Codex.
Because many paying users still could not access Astra after the launch, OpenAI Chief Executive Officer Sam Altman apologized at one point for the disorder around the rollout and said the team was working as quickly as possible to expand model access capacity.
Tibo also said at the time that OpenAI would issue one "banked reset" for each day a paid-plan user had to wait for Astra access, allowing reset opportunities to be stored and used later. OpenAI’s official explanation showed that eligible Plus, Pro, and Business users on Sept. 3 and Sept. 4 received banked resets, a sign that Astra’s early days were marked by meaningful capacity limits.
From delayed access to possible limits on new demand
At first, the issue was when existing paying users would get access to Astra. Tibo’s latest comments show OpenAI is now considering steps aimed at limiting new demand as well.
If Astra usage keeps rising quickly, the company could stop taking new ChatGPT Pro subscriptions and reserve limited inference capacity for current paying users. That would mark a shift from managing delayed access for existing subscribers to controlling incoming demand at the subscription level.
Infrastructure, not interest, is the key constraint
The situation highlights a separate bottleneck in the race to deploy frontier AI models. Once model capability improves, the limiting factor for broad product adoption may not be the model itself, but whether GPUs, data centers, networks, and inference infrastructure can keep up.
ABMedia wrote that OpenAI’s compute procurement has already reached what it described as a nation-scale infrastructure effort. The Stargate project, launched in 2025, targets $500 billion in spending over four years and 10GW of AI data center capacity. In the same year, OpenAI and Oracle pushed forward with as much as 4.5GW of new capacity, and Stargate capacity under development at one point exceeded 5GW and more than 2 million chips. By September 2025, OpenAI said it had planned nearly 7GW of infrastructure investment worth more than $400 billion.
Even so, Astra still ran into capacity pressure after launch. ABMedia said that suggests inference demand for frontier models may be rising faster than the pace of data center expansion. The report also said OpenAI has continued to look overseas for additional compute, including a multi-year agreement with Nvidia-backed Firmus using capacity from two data centers in Malaysia.
What Astra is signaling
If the pattern continues, where each new increment of compute is quickly absorbed by model demand, Astra would point to more than just enthusiasm around one release. It would also suggest the AI infrastructure buildout is still far from over. ABMedia said the market now needs to watch not only whether major technology companies are willing to build data centers, but whether AI-driven revenue and demand can continue to absorb investment on this scale.
For now, Astra appears to be sending a positive demand signal for OpenAI: the bottleneck is still insufficient supply, not insufficient demand.