Look up an H100 on our GPU board and you’ll see the same silicon quoted at wildly different prices. That spread is a reliability market. Once you can read it, you can save 50-70% on workloads that don’t need the guarantees you’d be paying for.
The three tiers
Secure / on-demand is classic datacenter capacity: the provider owns the hardware, runs it in a real facility (often tier 3/4), staffs support, and offers an SLA or something like one. RunPod’s Secure Cloud and Lambda’s on-demand instances live here. You pay the most per hour, and the machine will still be there tomorrow.
Community is the same rental interface, but the hardware belongs to third-party hosts vetted by the platform. RunPod’s Community Cloud is the canonical example. Prices drop 20-50%. Hosts are usually reliable, but power, networking and physical security are whatever that host has, and machines can disappear from inventory.
Marketplace is the open order book: Vast.ai lists thousands of machines, from proper datacenters down to a gaming rig in someone’s garage, each with its own price, reliability score and verification status. This is where the true price floor lives (our trends page tracks it daily), and where you shoulder the most variance.
What goes wrong on cheap tiers
- Interruption. Spot/interruptible rentals can be reclaimed when someone outbids you. Fine for checkpointed training or batch inference; miserable for a customer-facing API.
- Host disappearance. A community host can go offline: hardware fault, power cut, or the host simply left the platform. Your persistent volume may be stuck on that machine.
- Slow networking. Consumer-hosted GPUs often sit behind residential uplinks. Downloading a 140 GB model onto a machine with 100 Mbps ingress burns three hours of rent before you run a single token.
- No compliance story. If you need SOC 2, HIPAA, or even just “our data never leaves a real datacenter,” community and marketplace tiers are disqualified on day one.
Choosing in practice
Match the tier to the blast radius of an interruption:
| Workload | Tier |
|---|---|
| Production inference API | Secure / on-demand |
| Fine-tuning with checkpoints | Community or interruptible marketplace |
| Batch offline inference | Marketplace, cheapest verified host |
| Experiments, evals, one-offs | Marketplace, whatever’s cheapest |
| Anything with compliance requirements | Secure, full stop |
Two practical habits: on marketplace platforms, filter to verified datacenter hosts for anything that runs longer than a day; the premium over unverified hosts is small and buys real reliability. And always check ingress bandwidth before renting a machine to serve large models; a cheap GPU with slow networking is often the most expensive option once you count setup time.
The floor prices on our per-GPU pages show the tier next to every quote, so you’re never comparing a marketplace ask against an SLA-backed on-demand price without knowing it.