UAPI Grok Cloud Colab Burst Lane Review 2026-06-22
Grok review of cloud.colab.gpu.burst as transient burst capacity, not UAPI control plane.
Promotion Metadata
- Source feedback id:
119 - Source feedback ids:
119 - Promotion request id:
50 - Feedback category:
beneficial - Feedback source:
[REDACTED:high-entropy] - Feedback created at:
2026-06-22T18:27:40.181131+00:00 - Target slug:
[REDACTED:high-entropy] - Review state:
review - Reviewer note: AI lane auto-promoted redacted feedback into internal catalog context.
Source Feedback Body
Feedback 119
- Product:
uapi - Category:
beneficial - Related slug:
[REDACTED:high-entropy] - Source:
[REDACTED:high-entropy] - Created at:
2026-06-22T18:27:40.181131+00:00
Grok Cloud Colab Burst Lane Review (2026-06-22)
ALIVE: true
Lane: Grok Worker A / UAPI cloud/Colab burst
Sources: [REDACTED:high-entropy].md, [REDACTED:high-entropy].md, [REDACTED:high-entropy].md, [REDACTED:high-entropy].md, [REDACTED:high-entropy].md + uapi-grok-hardware-autorouter-cloud-gpu.md; public: https://research.google.com/colaboratory/faq.html, https://colab.research.google.com/signup, https://docs.runpod.io/serverless/overview, https://www.thundercompute.[REDACTED:high-entropy].
Role: Burst/Specialist, Not Control Plane
UAPI owns control plane (PlacementDecision, routing, policy, adapters, run records, artifact refs). Explorer SoT. Conductor handles lanes (master control 706; cloud runbook 707).
cloud.colab.gpu.burst = candidate transient external overflow when local lanes (nuc., cortex.) saturated. Specialist only: not primary, not interactive/low-latency, not control/state. Separate from NUC; own auth/config/cost required.
Auth / Cost / Session Risks
Colab: ephemeral VMs deleted on idle; free max ~12h (varies, not guaranteed). Paid tiers better but still max lifetime + soft limits. https://research.google.com/colaboratory/faq.html
Idle timeouts prioritize interactive notebooks; batch prone to termination. Free disallows some remote/distributed patterns.
Cost (2026): Pro $9.99/mo (100 CU), Pro+ $49.99/mo; paygo $9.99/100 CU. T4 ~1.5-2 CU/hr, A100 ~13-15 CU/hr. Units for runtime+setup; sessions end on exhaust. Unpredictable + cooldown risk. https://colab.research.google.com/signup https://www.thundercompute.[REDACTED:high-entropy]
Auth: Google + billing. No durable UAPI tokens; re-init per lease. Separate config only. GPU type variable (T4 common); Drive quotas; no SLA.
Artifact-First Job Shape
UAPI dispatches via artifact refs + contracts. Burst never gets live control or sessions.
Job = input refs + spec (route=cloud.colab.gpu.burst, params) + timeout + budget + output sink. Broker leases; runtime pulls artifacts, runs, writes results/logs/metrics/manifest to sink. Storage handoff only (no pipe/SSH). Enables audit, retry, lineage. Matches adjacent serverless job patterns (submit/poll/result/scale-zero). https://docs.runpod.io/serverless/overview
UAPI retains PlacementDecision.
Evidence Required Before Enable
- Isolated paid Colab config (no shared secrets).
- External hard cost cap + accounting (track/terminate outside provider).
- Artifact store roundtrip test (pull/push small payloads + results).
- Synthetic lifecycle: lease, pull, run non-model, timeout enforce, write outputs, observe.
- Health probe (avail, GPU realized, store latency).
- Placement fallback rules (local first; burst on overflow+budget).
- Codex contracts + PlacementDecision baseline. Operator gate.
- Sanitized evidence promoted; no activation pre other lanes evidence.
Assumptions / Blockers / Next
Assumptions: paid tier; reliable artifact store; synthetic validation precedes real. Blockers: no keys/config; no broker/contract yet; Colab notebook bias + ephemerality (FAQ); contracts unbaselined; SN850X separate. Next: Codex promote this+prior; companion contract notes; handoff+cap evidence before gate.
Report: ALIVE: true | Artifact: [REDACTED:high-entropy].md | Burst as specialist only. Local, read-only, cited public patterns. Next: contract notes + gates. End.