GPU and AI compute markets compared, verified and priced

Decentralized GPU markets price like any marketplace: the cheapest listings are the cheapest for a reason. The ranking above weighs what the cheap listings hide - verification of work, egress costs, and interruption behavior - as heavily as the rate itself, because those three decide whether a $0.28 RTX 4090-hour or a $2 H100-hour becomes a finished job or a support ticket.

The verification question nobody asks

On a hyperscaler, the datacenter contractually guarantees the GPU. On a marketplace, a stranger's rig claims to be an H100. The mechanisms that separate serious networks from naive ones [2][3]:

• TEE / attestation - the node proves what hardware it runs via a hardware root of trust. Strongest answer, currently rare.
• Checkpoint sampling - the network re-runs slices of your job on other nodes and compares results. Good middle ground for inference.
• Reputation with skin in the game - host bonds get slashed for failed verifications. Common on the bigger networks.
• Nothing - a dashboard and trust. More common than anyone admits; ask the question before the job [4].

Match verification to workload: sampled checkpoints tolerate marketplace economics for inference; long unattended training runs do not.

Live from our database - caveats and sources in each linked review.

Egress: the fee that decides real cost

GPU-hour prices are quoted without the bill that arrives at the end: getting your data out. Trained checkpoints, datasets, render output - moving 500 GB off a cheaper platform at $0.05-0.10/GB can cost more than the compute saved. Three questions before the first job: what egress costs, whether there is a cap, and whether storage bills separately. The storage hub covers the same economics for data; for compute the pattern repeats.

Interruption is a pricing feature, not a bug

Cheap GPU-hours are cheap because someone with a higher bid can take them. The rate prices preemption in - your job design has to match [2]:

• Checkpoint every few minutes - treat any job as if the machine disappears at a random point.
• Keep state external - resume from object storage, not the node's disk. Rented disks vanish with the lease.
• Diversify across nodes for long jobs - one node's interruption should cost minutes, not days.
• Bid the interruption, not the rate. The honest comparison at equal workload is finished-job cost: rate times expected runtime plus the expected cost of preemptions. A rate twice as high with zero preemption often wins on long jobs.

The gpu-prices page tracks the live spread. The current feed snapshot shows the spread between camps: H100 minimums near $2.56/hr, A100 near $0.84/hr, and commodity 3090s near the floor - the same GPU-hour costing an order of magnitude more one camp over.

How to test a market before committing

Replay, don't benchmark. Provider marketing quotes medians; your bill lives in the tail:

1. Replay a production day. Run yesterday's real request/job traces against two candidate markets and compare completion rates at p95, not the median [7].
2. Time-to-first-tensor. Cold-start provisioning on decentralized markets ranges from seconds to many minutes depending on image size and host queue - measure it on the cheapest tier you plan to use. The rtx-4090 hub shows the commodity end of that spread ($0.36/hr at the moment).
3. Kill-test the checkpoint. Kill the job at minute three, restore on another node, and time the resume. If your pipeline cannot survive this, the marketplace tier is not for that job.
4. Price the full round trip. Rate x hours + egress + storage + failed-job overhead. The finished-job cost is the only number that matters, and it regularly reorders the price table.

Where the networks differ

The ranking splits the compute population into three honest camps:

• Marketplaces (Vast-style) - real hardware diversity, host reputation, spot pricing that actually clears. Best for fit-tolerant workloads and small teams [1].
• On-chain lease markets (Akash) - utilization data is public [1], settlement on-chain, capacity real but thinner. Good for reproducible infra-as-code workloads.
• Token-incentivized networks (io.net, Render, Nosana) - supply grows on emissions, rates can price below cost, and long-run pricing rides the token loop. Fine for experimentation; budget for volatility.

Model fit is the first filter, before price: a 7B model at FP16 needs ~14 GB plus KV cache, 70B needs ~140 GB - which is why the LLM inference hub separates VRAM tiers before rates. The median H100 rate in our feed sits at $2.56/hr; the GPU workload tool does the arithmetic; the rent-gpu guide covers the practical setup.