Runpod

• Pure usage-based — no monthly fee, per-minute Pods and per-second Serverless.
• H100 SXM $3.29/hr · A100 80GB $1.49/hr · L40 $0.99/hr · budget GPUs from $0.27/hr.
• Secure Cloud (tier-3+ DCs, SLA) for production; Community Cloud for cheap dev.
• Runpod Flash ships a Python file to a serverless GPU — no Docker required.
• Real multi-GPU: clusters up to 64 GPUs with InfiniBand, no enterprise sales call.

The real question: what does a GPU-hour actually cost?

Almost every GPU-cloud comparison gets derailed by branding. The number that matters to an AI team is brutally simple: how many dollars does one hour of an H100 cost, and can I actually get one today? On the hyperscalers the honest answer in 2026 is "expensive and usually reserved." An AWS p5 instance — eight H100s — lists near $98/hr on-demand, which works out to roughly $12.25 per H100-hour before you factor in the capacity reservation you almost certainly need to get one at all. Runpod's pitch is that it collapses that to one rentable H100 SXM at $3.29/hr, on demand, with per-minute billing and no commitment. That is the whole story, and it is why Runpod earns a place on most AI teams' shortlist.

The second-order point is billing granularity. A reserved hyperscaler instance bills whether you use it or not. Runpod Pods bill per minute and Serverless bills per second of actual request processing — which is the only honest way to price bursty inference. If your traffic is spiky, scale-to-zero Serverless means you stop paying the moment the queue empties.

Put the two together and the economics get interesting. A team fine-tuning a 7B model overnight on a single A100 80GB pays roughly $1.49/hr — call it about $12 for an eight-hour run — and then shuts the Pod down. The same eight hours on a reserved hyperscaler instance is billed against a commitment you signed weeks earlier, whether the GPU was busy or idle. For research and experimentation, where you spin compute up and down dozens of times a week, that difference compounds into the single largest line item you control. The discipline Runpod rewards is simple: provision when you need it, stop when you don't, and let per-minute and per-second billing do the rest.

There is also a supply story behind the price. H100s have been scarce on the hyperscalers for two years, which is exactly why getting one on AWS often means a capacity reservation and a wait. Runpod's distributed model — a mix of Secure Cloud datacenters and a Community Cloud of peer providers — means flagship GPUs including B200, H200, and H100 are generally bookable in minutes. For a team that needs to start training today, availability is as much a feature as price.

Runpod pricing in 2026 — the full GPU-hour table

There is no platform fee layered on top — the GPU-hour and storage rate is the bill. Promotional credits for new accounts are awarded at Runpod's discretion; verify the current rate at signup, since flagship GPU pricing moves as supply changes.

Runpod vs Lambda Labs, Vast.ai, and AWS

The GPU-cloud market splits into three archetypes, and Runpod deliberately sits between them. The comparison that matters is the effective per-H100-hour cost paired with whether you can actually get the hardware.

The takeaway: Vast.ai will sometimes beat Runpod on raw price, but reliability is a coin-flip; Lambda matches Runpod on-demand but has no serverless tier and tighter capacity; AWS is the most expensive and the hardest to provision. Runpod's edge is the combination — marketplace-adjacent pricing, hyperscaler-grade availability, and a genuine serverless option none of the others ship.

Runpod at a glance — the spec sheet

What you actually get

How to get an H100 running on Runpod in five steps

1. Sign up free through the partner link

   No credit card commitment beyond a small balance to start metering. There is no subscription, so you only ever pay for compute time used.

2. Pick Secure Cloud or Community Cloud

   Production work → Secure Cloud (SLA, tier-3+ DCs). Experiments and non-critical batch → Community Cloud for the lower rate.

3. Choose a GPU and a template

   Select an H100/A100/L40 and a one-click template (PyTorch, vLLM, ComfyUI) or your own Docker image. The Pod spins up in seconds.

4. Work over SSH or Jupyter

   Attach a persistent network volume so your data and checkpoints survive a restart. Per-minute billing runs only while the Pod is on — stop it when you're done.

5. Promote to Serverless for production

   When you ship, deploy the model as a Serverless endpoint (or use Flash for a Python-only path). It scales to zero between requests so idle time costs nothing.

Who should use Runpod — and who shouldn't

Where Runpod earns its keep

Four workloads cover the bulk of what teams actually run on it. LLM fine-tuning on a budget is the obvious one — rent an A100 for $1.49/hr instead of $3+/hr elsewhere, fine-tune Llama or Mistral in a few hours, and shut it down before the next billing minute ticks over. Production inference at scale is the Serverless story: deploy a vLLM or TGI endpoint that scales from zero to N workers and only bills while it's serving requests, which is the right shape for any product with uneven traffic. AI agents that need persistent GPU state live on Pods, where a persistent network volume keeps context and checkpoints across restarts so a long-running, multi-step pipeline doesn't lose its place. And image and video generation services spin up A1111, ComfyUI, or a video-model template and serve generations to users at marketplace prices — a category where GPU cost directly sets your margin.

The common thread is that none of these workloads wants a yearly reservation. They want flagship hardware on tap, billed by the minute or second, with the freedom to switch GPU class as the model or the traffic changes. That is precisely the gap Runpod fills between the cheap-but-flaky marketplaces and the reliable-but-expensive hyperscalers.