Key takeaways
- On 10 September 2026 OpenAI paused new Pro subscriptions, seven days after launching Astra, citing strain from demand its product lead described as unprecedented.
- Inference is now reported at roughly 55% of AI-optimised infrastructure spending, up from about a third in 2023. The constraint moved from training to serving live traffic.
- Reported lead times of 36 to 52 weeks for data centre GPUs, and 18 months for hyperscaler allocations, mean capacity cannot be added inside your planning horizon.
- Rationing does not arrive as an outage. It arrives as rate limits, slower responses, smaller context windows and restricted access to the newest model.
- Your fleet needs a defined behaviour for degraded capacity. If it does not have one, it has a default, and the default is failing tasks at the worst moment.
Does ChatGPT recommend your competitor instead of you?
Check how AI assistants describe your company, and whose name they give when someone asks for a recommendation in your category. Run an audit today, from $19.
On 10 September 2026, seven days after launching its newest model, OpenAI stopped accepting new subscribers to its highest tier. Not a price change, not an outage. It stopped selling.
For anyone running agents against a hosted model, that is the interesting kind of event, because it says the constraint is no longer money.
What actually happened, and why it matters operationally?
A provider concluded that the cheapest way to protect service for existing customers was to stop adding new ones. The tier it paused was the one it identified as placing the most strain on its systems; other tiers stayed open.
Read that as a signal about the shape of the market rather than about one company. Inference is now reported at roughly 55% of AI-optimised infrastructure spending, up from about a third in 2023. The industry’s bottleneck moved from training runs to serving live traffic, and serving live traffic is what your fleet does.
| Constraint | Reported figure | What it means for planning |
|---|---|---|
| Inference share of AI IaaS spend | ~55% | Serving, not training, is the contested resource |
| Data centre GPU lead time | 36 to 52 weeks | Capacity cannot arrive this quarter |
| Hyperscaler allocation lead time | ~18 months | Your provider's expansion is already committed |
| Demand against supply through 2027 | 1.4x to 1.6x | Scarcity is the baseline, not the exception |
How does rationing actually reach you?
Not as a status page incident. That is the part teams are unprepared for.
It arrives as a rate limit that is tighter than last month. As latency that drifts up at peak hours. As a context window quietly reduced on a cheaper tier. As the newest model being available to some accounts and not others. Every one of those is individually survivable, none is announced as rationing, and together they change what your fleet can promise.
An outage triggers your incident process. Degradation does not trigger anything, which is why it is more expensive.
Why retries make this worse
The instinct when a call fails is to retry it. Against a rate limit, a retry is the same request arriving again, consuming the same quota, at a moment when quota is the scarce thing. A fleet with aggressive retry policy converts a mild squeeze into a self-inflicted outage, and does it fastest at peak, which is exactly when the squeeze happens.
What should a fleet do when capacity tightens?
Decide in advance, and encode it as routing rather than as error handling.
| Task class | Behaviour under constraint | Why |
|---|---|---|
| Interactive, user is waiting | Smaller model, same task | A worse answer now beats a good answer nobody waited for |
| Batch, deadline in hours | Queue and drain off-peak | Moves load to where capacity is free |
| Irreversible action | Fail loudly, do not downgrade | A cheaper model deciding a payment is not a degradation, it is a different system |
| Speculative or exploratory | Shed first | Nobody notices, and it is usually the largest share of spend |
The last row is where most of the recoverable capacity lives. Fleets accumulate speculative work: parallel plan candidates, retries of things that will not succeed, background enrichment nobody reads. Under normal conditions that is a cost problem. Under a squeeze it is the thing you shed to keep the interactive path alive, and you can only shed it if it is tagged.
What can you actually control?
Less than a second provider, more than nothing.
Multi-provider routing helps against one vendor’s bad week and does not help against a market-wide squeeze, because reported lead times and the 1.4 to 1.6 supply gap apply across the market. What you control is the demand side: how much of your traffic genuinely needs the newest model, how much speculative work you generate, and whether your fleet knows the difference.
Three things worth doing before your provider makes the decision for you.
- Tag every task with a class. Interactive, batch, irreversible, speculative. Without the tag no policy is implementable.
- Test the degraded path. Set your own rate limit below the provider’s in staging and watch what the fleet does. Most teams discover the answer is “retries until something gives up”.
- Measure the speculative share. What percentage of last week’s tokens went to work whose output nobody consumed? That is your headroom, and it exists whether or not there is a squeeze.
The question for your next capacity review
Not “do we have a second provider”. Ask: if our rate limit were cut in half at nine tomorrow morning, what would the fleet do, and who would decide? If the answer is that it would retry until it recovered, the decision has already been made and nobody made it.
Frequently asked questions
Why did OpenAI pause Pro subscriptions?
The company said Pro placed the most strain on its systems and paused new sign-ups on 10 September 2026, a week after launching its Astra model. Other tiers including API, Go and Plus remained available. Its product lead for core products described demand for Astra as unprecedented.
What does capacity rationing look like for an API customer?
Rarely as an outage. It shows up as tighter rate limits, longer response times, reduced context windows, queueing at peak hours, and access to the newest models being restricted to some customers before others. Each is survivable individually and none is announced as rationing.
How should an agent fleet handle provider rate limits?
Decide in advance which tasks may be delayed, which may run on a smaller model, and which must fail loudly rather than silently degrade. Encode that as routing policy rather than as retry logic, because a retry against a rate limit is just the same request arriving again.
Can you avoid capacity risk by switching providers?
Only partially. Reported GPU lead times of 36 to 52 weeks and demand exceeding supply by an estimated 1.4 to 1.6 times through 2027 apply across the market. A second provider helps with a single vendor's bad week. It does not help with a market-wide squeeze.
Method and sources
- TechCrunch, reporting on OpenAI pausing Pro subscriptions on 10 September 2026 at 1:59 PM PDT, seven days after the 3 September launch of Astra, including quoted remarks from Thibault Sottiaux.
- 2026 infrastructure analyses reporting inference at roughly 55% of AI-optimised infrastructure-as-a-service spending, against about a third in 2023.
- 2026 supply chain analyses reporting data centre GPU lead times of 36 to 52 weeks, hyperscaler allocation lead times of around 18 months as of mid-2026, and demand exceeding supply by an estimated factor of 1.4 to 1.6 through 2027.
- 2026 analyses describing grid capacity overtaking GPU availability as the primary constraint on data centre expansion.
- No figures in this article are our own measurements. Figures on lead times and market share come from published analyses and are cited as reported.