Provisioning reliability: RunPod vs Vast
Renting a GPU is cheap and nominally instant. We rent across providers at volume and report how long it took until a node accepted SSH, whether it ever did, and how it failed. This is a high-level look at roughly 240 rentals on RunPod and Vast. The full breakdown is available on request (see below).
We found that the share of paid, running instances that never become usable ranged from 0 to 62 percent depending on where you rent.
| Tier | GPU | Instances | Failed to become usable | Median boot |
|---|---|---|---|---|
| RunPod Secure | A40 | 38 | 0% (0/38) | 16s |
| RunPod Community | RTX 3090 | 38 | 3% (1/38) | 18s |
| Vast verified | RTX 3090 | 40 | 25% (10/40) | 53s |
| Vast unverified | RTX 3090 | 40 | 62% (25/40) | 38s |
Provisioning reliability differs by more than an order of magnitude between providers, even on the same hardware type. The per-host reliability scores the providers publish are poor predictors of this.
The gap survives on identical hardware
We attempted to rent H100 SXM and A100 SXM on both providers, with a generous eight-minute window to reach SSH.
| GPU | Provider | Instances | Failed to become usable | Median boot |
|---|---|---|---|---|
| H100 SXM | RunPod | 19 | 5% (1/19) | 23s |
| H100 SXM | Vast verified | 20 | 25% (5/20) | 44s |
| A100 SXM | RunPod | 19 | 0% (0/19) | 20s |
| A100 SXM | Vast verified | 20 | 10% (2/20) | 42s |
On the H100, we found a five-fold difference in the rate of getting a working node. RunPod brought up nearly everything in about 20 seconds, while the Vast instances that came up took roughly twice as long, and the rest timed out of our upper limit.
Not all the failures are the same
On the cheap RTX 3090 hosts, every Vast failure was an instance that started, billed, and simply never became reachable in the window. None crashed. A large part of this might be image distribution, as RunPod serves a cached image, so the container is ready in seconds, while Vast pulls a fresh image onto each marketplace host, which is slow on consumer hardware. A renter who supplies a smaller or pre-cached image would narrow that part of the gap.
The premium hosts are different. Several of the Vast failures there were instances that started and then died mid-load, one within 18 seconds of launch, not slow boots that needed more time. So this is not purely an image-distribution artifact.
A high reliability score doesn't mean your node will boot
Vast publishes a per-host reliability score and steers renters toward high values. It does not separate the hosts that work from the ones that don't. Among the premium verified hosts we rented, the ones that failed to provision carried scores from 0.971 to 0.998, the same range as the ones that succeeded.