Billing account: paid, active. Two projects tested, both blocked.
WHAT FAILS
Every Cloud Run GPU deployment is refused at admission:
ERROR: metadata.annotations[run.googleapis.com/maxScale]:
You do not have quota for using GPUs without zonal redundancy
Reproduced in europe-west4, europe-west1, us-central1 and us-east4.
Reproduced with BOTH --gpu-zonal-redundancy and --no-gpu-zonal-redundancy.
Reproduced with BOTH “gcloud run deploy” and “gcloud run services replace”.
PROJECT 1 - (PII Removed by Staff)
Created 2026-07-19. The documented automatic grant of 3 L4 GPUs
(zonal redundancy off) on first deployment never fired.
Cloud Quotas API:
NvidiaL4GpuAllocNoZonalRedundancyPerProjectRegion → details {} (0)
NvidiaL4GpuAllocPerProjectRegion → details {} (0)
both: quotaIncreaseEligibility = NOT_ENOUGH_USAGE_HISTORY
I submitted quotaPreferences/qwen36-l4-ewest4 on 2026-09-01T18:11:44Z.
Auto-denied: grantedValue “0”, stateDetail “We cannot grant the preferred
quota ‘1’ … at this moment.”
PROJECT 2 - (PII Removed by Staff) (created today, same billing account)
A brand-new project, exactly the case the docs describe for the automatic
grant. Enabled run.googleapis.com, deployed the public
us-docker.pkg.dev/cloudrun/container/hello image with
–gpu 1 --gpu-type nvidia-l4 --no-gpu-zonal-redundancy.
Same refusal. Quota API on the new project reports the same
NOT_ENOUGH_USAGE_HISTORY and 0 for both L4 quotas.
CpuAllocPerProjectRegion in europe-west4 is 20000, so the documented
“subject to availability depending on your CPU and memory capacity”
caveat does not apply here.
This shows the gate is at the billing-account level, not the project level:
creating a fresh project does not trigger the automatic grant.
WHY I CANNOT SELF-SERVE
- Console: cannot request quotas with no prior usage.
- Cloud Quotas API: auto-denied, as above.
- Support case: impossible. cloudsupport.googleapis. reports
CreateCaseRequestsPerDayPerProject = 0, and Cloud Quotas marks that
limit NOT_SUPPORTED, so it cannot be raised either.
WHAT I NEED
NvidiaL4GpuAllocNoZonalRedundancyPerProjectRegion = 1
Region europe-west4 (us-central1 equally acceptable)
Project (PII Removed by Staff)
WORKLOAD
One serverless OpenAI-compatible LLM inference endpoint: llama.cpp
llama-server serving Qwen3.6-35B-A3B (GGUF UD-IQ3_S, 12.74 GiB) on a
single NVIDIA L4. Service config: min-instances=0, max-instances=1,
concurrency=1. At most one L4 is ever active, none while idle.
The container image and the model are already built and staged in the
project. The quota is the only blocker.
Asking for the minimum: 1 GPU, zonal redundancy off. Happy to provide
project details privately.