Skip to main content
Two different limits can refuse a deployment: your account’s quota, and the platform’s total serverless capacity. They fail differently and they need different responses.

​
Your quota

Defaults for a beta account: The third is the one people trip over: it is the sum of configured workersMax, not the number of workers currently running. Two deployments at workersMax: 3 each already total 6 and exceed the cap, even if both are scaled to zero and costing nothing. Exceeding any of them returns a 400 naming the limit:
The same checks run on update: raising workersMax on an existing deployment re-evaluates your totals. Resume does not re-check them — a paused deployment never gave its share back, so there is nothing to re-evaluate.
Delete deployments you are done with rather than leaving them paused. A paused deployment still occupies your deployment count and its workersMax still counts against your worker total.
To raise your limits, email [email protected] with what you are running and the scale you need.

​
Platform capacity

Serverless GPU capacity is finite and shared. When the platform-wide ceiling is reached, creating a deployment returns 409:
Raising workersMax on an existing deployment hits the same ceiling but reports it as a 400:
This is not about your account — you can be well inside your quota and still get it. What to do:
  • Ask for fewer workers. workersMax: 1 frequently fits when 3 does not, and scale-to-zero means a lower ceiling costs you nothing when idle.
  • Retry later. Capacity frees up as other deployments drain.
  • Do not retry in a tight loop. The answer will not change within seconds, and the queue API’s rate limits apply here too.

​
GPU availability

Capacity is also per-SKU. Each GPU row in the catalog carries an availability badge, refreshed hourly: A throttled worker on the workers panel is this problem showing up at runtime: the deployment is fine, there is simply nowhere to place the worker at that moment. Throttled workers are not billed. If a SKU you depend on sits at low or none repeatedly, deploy against a different one — the catalog spans several VRAM tiers, and a job that fits in 48 GB does not need to wait for 80.

​
Other limits worth knowing

Balance limits are covered separately in Billing — a credit hold you cannot cover refuses creation with a 402, which is a funding problem rather than a quota one.