generated: '2026-08-17'
method: searched
source: https://gpu-instances.shadow.tech/docs/faq/
limit_count: 0
note: >-
  Shadow publishes no request-rate limits for the Shadow GPU API and no rate-limit response headers.
  What it does publish is a quota model — the OpenStack per-project resource quota, plus a commercial
  instance limit — which constrains what an API caller can create rather than how often it can call.
  Recorded as limit_count 0 because no numeric request-rate limit, window, burst, or exhaustion status
  code is published anywhere on the developer surface. Searched the docs FAQ, the quota section, the
  limitations page and the pricing FAQ; no numbers are given for any quota either.
rate_limits: []
response_headers:
  published: []
  note: >-
    No X-RateLimit-*, RateLimit-* or Retry-After documentation exists, and no unauthenticated
    operational endpoint was available to observe headers on. The only anonymous endpoint on the
    platform is the Keystone version document (GET https://auth.<region>.os.shadow.tech/v3), which
    returned no rate-limit headers when probed on 2026-08-17.
exhaustion_status_code: not-documented
quotas:
- scope: per-project
  kind: resource-quota
  resources: [vCPU, RAM, storage, volumes, instances]
  limit: not-published
  enforcement: >-
    Creation fails on exhaustion. Documented as a first-line cause of both instance launch failure
    ("Insufficient quota limits for CPU, RAM, or storage") and volume creation failure ("Exceeding your
    storage quota").
  visibility: 'Current quota usage is viewable in the dashboard; the docs give no CLI/API command and
    publish no default values.'
  increase_process: '"Request quota increases through proper channels if needed" — no self-service path
    and no documented turnaround.'
  source: https://gpu-instances.shadow.tech/docs/faq/
- scope: per-account
  kind: instance-limit
  limit: not-published
  enforcement: commercial
  increase_process: >-
    Published verbatim as: "The limit can be revised after several regular billing cycles. Contact our
    Sales team for quick validation and to avoid service interruption." The instance ceiling is
    therefore time-and-sales gated, not a technical limit an integrator can plan against.
  source: https://gpu-instances.shadow.tech/en/
- scope: per-volume
  kind: storage-qos
  limit: 'Write IOPS 60,000 (max 120,000); Read IOPS 60,000 (max 120,000); Write throughput 250 MB/s;
    Read throughput 500 MB/s'
  applies_to: [Wood, Silver, 'Gold (coming soon)']
  enforcement: Cinder QoS
  caveat: >-
    Shadow publishes the SAME QoS figures for all three storage tiers, so the published numbers do not
    distinguish the tiers they are attached to. Quoted as published.
  source: https://gpu-instances.shadow.tech/docs/advanced/storage-classes/
preemption:
  applies_to: Spot instances
  behaviour: 'Not guaranteed — preemptible based on availability'
  notice_period: not-documented
  note: >-
    Spot preemption is the most consequential runtime limit on this platform and Shadow documents no
    notice period, no preemption signal, and no API event an agent or autoscaler could subscribe to.
  source: https://gpu-instances.shadow.tech/en/
cross_links:
  conventions: conventions/shadow-conventions.yml
  plans: plans/shadow-plans-pricing.yml