Reports

Super Intelligence reports

Published weekly. Newest first. Each report includes a summary and the full PDF.

Live launch and ready rate counts: Capacity IndexHow we test

Published reports31 reports

External Tester Review, November 2, 2026

November 2, 2026

An outside tester paid through x402 and ran a live H100 job. We checked every step: payment, machine order, boot, health checks and shutdown. Read the results.

Method

How we test

Controlled host failure drills, with the timings published.

On Oct. 14, 2026 we conducted a controlled failover drill on an internal Kilawatt test account. A GPU workload wrote a checkpoint roughly once per minute. We then terminated its host machine at the supplier mid run and measured the time to detect the failure, the time to provision a replacement on a different supplier, and the time to resume from the last checkpoint.

  1. 1Launch: start a short GPU workload on one supplier from an internal test account.
  2. 2Checkpoint: the workload writes a numbered checkpoint about every 60 seconds.
  3. 3Inject the fault: after checkpoint 4, we terminate the host machine at the supplier.
  4. 4Detect and replace: Kilawatt detects the failure, requests a replacement on a different supplier, and the replacement is provisioned.
  5. 5Resume: the workload reads back its latest checkpoint and continues.
  6. 6Record: every timestamp comes from the platform's own records, and we publish them.
RunPathTime to detectTime to provisionTime to resumeTotal recoveryResult
Run 1RunPod to Hyperstack238 s239 s175 s (manual)10.9 minFailure detected, replacement provisioned, checkpoint intact. Resume was performed manually because of a defect in our test script, since corrected.
Run 2RunPod to Vast.ai121 s119 s12 s (automatic)4.2 minFailure detected, replacement provisioned, resumed from checkpoint 4 automatically.
Excluded runVast.ai, no fault injectedn/an/an/an/aStopped before the drill. Machine boot took 14.6 minutes, consuming most of the 20-minute job window. Not a test result.

What the results show

Kilawatt detected both host failures without intervention and provisioned the workload again on a different supplier. The latest checkpoint was intact on the replacement in both runs. In run 2 the workload resumed from checkpoint 4 automatically and went on to write checkpoints 5, 6 and 7.

What the results do not show

  • Sample size: two runs do not establish a reliability rate.
  • Recovery time: recovery is measured in minutes, not milliseconds, because a replacement must be provisioned and verified.
  • Detection latency: 121 to 238 seconds is longer than our target. Reducing it is our first priority.
  • Workload: a simple step counting job on an internal test account, not a training run or a customer workload.
  • Independence: prepared by Kilawatt, not independent reporting.