[peterschmidt85]

Nebius GPU offers currently report unknown availability, so users cannot see quota or capacity restrictions and dstack may attempt configurations that cannot be provisioned.

Add opt-in checks through capacity_advisor: true to show these statuses and guide provisioning and fabric selection. The option defaults to false because checking adds latency and requires caching quota and capacity data.

Match advice by region, platform, preset, spot mode and fabric. Catalog fabric compatibility and configured allowlists remain enforced in both modes. Missing, stale or failed advice stays unknown and eligible.

Notes

• Fetching quota/capacity from Nebius adds roughly 1–3 seconds. Cache both for 10 minutes, matching AWS's quota-cache TTL. Pricing stays offline; the existing 3-minute offer cache can extend reuse of a fetched snapshot to ~13 minutes.
• Preserve automatic reservation use. Reserved capacity keeps otherwise blocked on-demand offers eligible as unknown, since tenant-wide advice does not check project reservation limits. Spot offers use only preemptible advice.
• Support an optional shared cache scoped by tenant and credential IDs. Factory wiring: https://github.com/dstackai/dstack-factory/pull/22. Project quota/balance checks remain separate.
• Keep the existing 10-second timeout per page and reuse tenant discovery for project selection and advice.

Getting offers

Local backend-collection medians on November 5, with pricing already loaded; 3 cold/new-backend runs and 9 refreshes. Cold ranges are in parentheses. AWS: 11 regions / 2,049 offers. Nebius: 9 regions / 169 offers. HTTP overhead is excluded.

Scenario	AWS	Nebius, default	Nebius, capacity_advisor: true
New backend, empty provider caches	1.77 s (1.72–1.92 s)	0.97 s (0.59–2.31 s)	1.50 s (1.35–1.59 s)
Offer refresh, provider data already cached	37.7 ms	5.1 ms	3.8 ms
New backend using shared provider caches	36 ms	757 ms	789 ms

The shared-cache row uses Factory's cache-injection pattern. Nebius still performs tenant/project discovery for each new backend. AWS checks quotas and zones, without querying live free capacity for ordinary offers.

[peterschmidt85]

Made capacity advisor opt-in via capacity_advisor: true in the Nebius backend configuration. It defaults to false: existing configurations make no capacity-advice requests and keep native availability unknown.

Enabling it shows quota/availability statuses and applies advice to provisioning and fabric selection, with a 10-minute advice cache. The PR body now includes an AWS latency comparison with the option disabled and enabled.