How we stopped making customers prove their GPU went down
Most GPU cloud providers put the burden of proof on you. If a machine degrades or dies mid-job, you have to notice it, you have to prove it, and you have to file a claim — usually within a narrow window, or the credit is forfeited.
We built the opposite. On Kilawatt Cloud, the platform watches every active job itself. When something goes wrong, you're credited automatically, with a timestamped record, no claim required.
What runs on every active job now
A health sweep checks every running instance on a short interval — provider heartbeat, GPU temperature where the provider exposes it. If a machine goes dark, the outage window gets backdated to the last confirmed-healthy moment, not to whenever the system happened to notice.
On recovery — or immediately if the machine is confirmed dead — your wallet is credited automatically: full rate for downtime, half rate for degraded performance. It shows up as a line item on the job's receipt. Severe failures trigger an automatic attempt to move the job to another provider, using the same failover logic already proven in production.
Not a simulation — a real failure, mid-build
A live job on Vast.ai genuinely failed during image build and never reached a running state. We didn't force this fault to test the system; it just happened, and here's what got recorded and done about it, automatically:
Metric Recorded value
Job 1ef01462-30d9-464d-8f9d-1cc0bc9cd27c · Vast contract 51996473 · RTX A4000
Billed $0.06
Down window detected 02:37:41.759Z → 02:40:19.222Z (157.5 seconds)
Automatic credit issued $0.01 · wallet balance 276¢ → 277¢
Receipt line sla_credit_usd: 0.01 · net_billed_usd: 0.05
Failover attempted RunPod, Hyperstack, Lambda — all logged in the append-only ledger
Where the industry standard requires you to notice an outage and file a claim — often within a 24-hour window or the credit is forfeited — this system detects and credits the affected amount automatically, in the time it takes the job itself to fail and recover: 157.5 seconds, start to finish.
A public, permanent accountability record
Alongside this system, we launched a public status page at kilawattcloud.dev/status: live component health for payments, provisioning, the API, and GPU health monitoring, plus a permanent, dated history of past incidents and how each was actually resolved. This is the same standard companies like AWS and Cloudflare hold themselves to, applied to a company our size.
Recent incident history:
• Sep 15 — Billing race condition → Resolved
• Sep 16 — Payment window / OFAC cache → Resolved
• Sep 21 — RunPod GPU substitution → Resolved
• Sep 22 — SLA health monitoring → In progress (this launch)
What's next
Downtime detection and automatic credit are live and proven in production. The next phase, already underway, extends this same protection to deeper in-machine hardware fault detection. We'll publish that the same way we published this: with real jobs and real receipts, not before it's proven.
Why this matters if you're evaluating GPU providers
A reliability guarantee is only as good as how it behaves when nobody's watching. This system removes the step where you have to notice, prove, and request your own credit, and it keeps a permanent, public record of exactly how the platform has performed over time. That's the standard we're holding every piece of this infrastructure to: prove it in production, publish it, and stand behind the receipt.
Kilawatt Cloud — zero-quota GPU orchestration across H100, H200, and A100 capacity, with automatic failover and 15-minute metered billing. kilawattcloud.dev
Top comments (0)