16 of 18 real GPU jobs required automatic failover mid-run. Zero were dropped.
That's not a number from a pitch deck. It's what we found when we audited our own infrastructure before saying a word about it publicly. Here's the full breakdown, and why it matters.
The paradigm shift, in one sentence
No single GPU provider, not RunPod, not Vast.ai, not Lambda, has enough of the right hardware, in the right place, at the right time, to carry serious AI workloads alone anymore. Most neoclouds, including CoreWeave, Lambda, and Nebius, are single-provider by design. When their capacity runs out or a job fails, there's no second provider to catch it. Multi-cloud stopped being a hedge and became the default for any team doing real AI work.
Kilawatt Cloud sits above the providers, RunPod, Vast.ai, Lambda, and Hyperstack, through a single API, and routes workloads to wherever real capacity exists, with automatic failover built in. Not zero failure, no infrastructure layer can promise that. No single point of failure.
The audits: what we actually checked
A multi-cloud broker is only as trustworthy as its weakest unverified claim. So we ran everything for real before publishing a word of it.
Live failover audit. Across a run of real completed jobs paid for through our x402 stablecoin rail on Base mainnet, 16 of 18 jobs (about 89%) required an automatic failover to a second provider mid-run. Zero jobs were dropped. On one day in that window, RunPod failed 100% of its attempts, all 11, and Lambda picked up the failover traffic for the first time in our logs. A separate database check confirmed two more real historical failover events on earlier dates, independent of that test run.
Provisioning speed audit. RunPod: 4.7 seconds to ready. Vast.ai: roughly 92 to 109 seconds depending on API vs. SSH-verified readiness. Lambda: about 260 seconds. Real gaps a routing layer has to account for, not smooth over.
Billing and provisioning safety audit. Before trusting our own metering, we went looking for ways it could fail and found seven real gaps: launches that weren't fully pre-authorized against balance, no per-key rate or concurrency limits, no customer-set spend caps, no automatic reaping of abandoned instances, a broken Vast.ai provisioning endpoint, a failover path that silently dropped instead of falling through, and gateway-launched instances that weren't tracked for termination. All seven were fixed and re-verified against real instance creation and destruction on RunPod and Vast.ai, not just reviewed in code.
Where the audits found real gaps, we published the before-and-after instead of quietly patching them. That's the standard: a number we haven't tested against live infrastructure doesn't go in front of a customer.
The ROI on our roadmap for enterprise
Everything above is live today. What we're building toward is where the ROI case for enterprise compounds. We're actively sourcing dedicated GPU capacity to broker larger B300 and B200 clusters for enterprise teams whose needs sit above what any single spot marketplace can reliably fill. Every brokered deal is structured with an upfront prepayment covering supplier cost plus our margin before we ever pay a provider, so an enterprise customer gets a fixed, predictable number instead of a variable bill that moves with provider pricing swings.
The math is straightforward once you sit with it. A stalled training run on a single-provider cloud costs real money sitting idle while a team scrambles for capacity elsewhere. A vendor relationship built one supplier at a time costs real engineering hours. Kilawatt collapses both into one relationship, one API, and one bill, backed by a routing layer that's already proven it catches failures before they become the customer's problem.
Where this is headed
As GPU demand keeps outpacing any single provider's supply, this shift only accelerates. The operators who win won't be the ones with the most hardware, they'll be the ones buyers can actually verify. We're building toward a real, multi-million-dollar infrastructure company, and the work above is exactly what that's built out of.
Top comments (0)