All runbooks
runbookuapiredacted view

UAPI Grok Cloud Provider Burst Comparison 2026-06-22

Grok comparison of Colab-style notebook burst, RunPod-style serverless/on-demand, and lab-hosted GPU route.

Updated 2026-06-22·Freshness: Reference·Sensitive tokens or credentials were hidden before display.

UAPI Grok Cloud Provider Burst Comparison 2026-06-22

Grok comparison of Colab-style notebook burst, RunPod-style serverless/on-demand, and lab-hosted GPU route.

Promotion Metadata

  • Source feedback id: 131
  • Source feedback ids: 131
  • Promotion request id: 62
  • Feedback category: beneficial
  • Feedback source: [REDACTED:high-entropy]
  • Feedback created at: 2026-06-22T18:32:58.796700+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 131

  • Product: uapi
  • Category: beneficial
  • Related slug: [REDACTED:high-entropy]
  • Source: [REDACTED:high-entropy]
  • Created at: 2026-06-22T18:32:58.796700+00:00

Grok Cloud Provider Burst Comparison (2026-06-22)

ALIVE: true
Lane: Grok Worker D / UAPI burst provider comparison
Sources: agent-work-queue-v2-unblocked.md; local (grok-cloud-colab-burst-lane-review.md, grok-cloud-gpu-broker-contract-notes.md + priors); public: https://research.google.com/colaboratory/faq.html, https://colab.research.google.com/signup, https://docs.runpod.io/serverless/overview, https://docs.runpod.io/serverless/pricing, https://www.runpod.io/pricing.

Scope

UAPI broker for transient burst (e.g. cloud.colab.gpu.burst). UAPI owns PlacementDecision + leases + artifact handoff (pre-lease sink input; post sink outputs). Adapters execute only. Local NUC/lab first. Compare Colab notebook burst, RunPod serverless/on-demand, lab-hosted GPU for adapter fit, cost, lifecycle, risk.

Colab-style Notebook Burst

  • Ephemeral Jupyter VMs (deleted idle + max lifetime). Free ~12h max (varies); Pro+ ~24h w/ CUs. https://research.google.com/colaboratory/faq.html
  • GPUs: T4 common, L4, A100 (varies). Prioritizes interactive notebooks; batch/SSH/distributed often terminated.
  • Auth: Google + billing/Drive. Re-init per lease; isolated config mandatory.
  • Cost: Pro $9.99/mo (100 CU) or PAYG $9.99/100CU. T4 ~1.5-2 CU/h, A100 ~5-15 CU/h (runtime+setup). Soft exhaust terminate. https://colab.research.google.com/signup
  • UAPI fit: Weak (notebook bias, Drive quotas, ephemerality conflicts lease+artifact model). Low priority overflow only.

RunPod-style Serverless / On-demand

  • Serverless (Docker handler endpoints, scale-to-zero or active workers) or dedicated Pods.
  • GPUs: RTX 4090 (~$0.69/h), A100 (~$1.4/h), H100 (~$2.9/h) on-demand; per-sec serverless (e.g. A100 ~$0.00076/s). FlashBoot reduces cold starts. https://www.runpod.io/pricing
  • Lifecycle: Config idle/exec timeouts (scale-to-zero good for burst); pre-warm volumes; no idle cost on Flex.
  • Auth: API keys + network volumes. Supports isolated per-lane endpoints.
  • UAPI fit: Strong (handler aligns artifact pull/run/push; scale-to-zero; configurable timeouts map broker lease/contract).

Lab-hosted GPU Route

  • Internal dedicated (NUC: Arc 140T iGPU + NPU on Core Ultra 9 285H ~64GiB shared; VPS/Cortex relay; Proxmox passthrough possible). https:// (local inventory only)
  • Lifecycle: Operator-reserved/persistent. No cloud idle-delete; re-bootstrap post-use. Contention with Conductor.
  • Auth: Internal/lab-mediated (no external cloud billing). Broker lease still applies.
  • Cost: Amortized hardware+power (zero marginal $/hr; hard capacity cap).
  • UAPI fit: Highest (locality to lab sinks/storage SN8100; preferred before external). e.g. lab.cortex.* or lab.gpu.burst.

Key Trade-offs

  • Ephemerality: Colab (high, unpredictable) > RunPod (controlled) > Lab (lowest).
  • Cost predictability: Lab (fixed) > RunPod (per-sec variable) > Colab (CU burn + soft limits).
  • GPU variety/perf: RunPod (H100+) > Colab (A100) > Lab (140T/NPU specific).
  • Broker integration: RunPod/Lab strong (headless job+artifact); Colab requires wrappers, high risk.
  • All need UAPI hard timeout/budget/artifact caps per broker notes. Placement prefers local/lab then burst.

Assumptions: Paid cloud tiers; reliable sinks; small pre-warm manifests; local first policy. Blockers: No isolated configs; Codex PlacementDecision/lease contracts unbaselined; no synthetic evidence; separate SN850X/NUC tasks. Next: Promote to Codex; per-provider adapter specs + gates; operator isolated-tier validation post-contract.

Report: ALIVE: true | Artifact: [REDACTED:high-entropy].md | Three-way burst provider compare (Colab/RunPod/lab) for UAPI. Cited public+local. <3800 chars. Local only. Handoff: matrix. COMPLETE. End.