GPU Cloud Rate Card 2026: $/GPU-Hour Across 10 Providers (H100 80GB Normalized)

Last updated: 2026-11-05 Independently verified 18 min read

GPU cloud pricing in 2026 is the second-most volatile cost line for AI startups after headcount. This page is a rate card, not a benchmark: one number, $/GPU-hour, normalized to a H100 80GB SXM on-demand reservation in a US region. The same 10-provider table is also the basis for the interactive TCO calculator in the footer.

1. How to read this rate card

$/GPU-hour is the instance hourly price divided by the number of GPUs in the instance. A 1× H100 at $2.50/hour is $2.50/GPU-hour; an 8× H100 at $18.00/hour is $2.25/GPU-hour. The second is cheaper per GPU, but you pay for 8 GPUs in lockstep — and 8-GPU instances frequently lose one or more GPUs to NVLink fabric failures, OOM, or host reboots, so effective availability is lower than the headline $/GPU-hour suggests. Single-GPU or 8-GPU-by-8-GPU pricing is the right starting point for solo founders and small teams; 8-GPU minimums are the right starting point for any serious fine-tuning or pretraining run.

Reserved and committed pricing can be 30-60% lower than on-demand, but only if you actually run the hours. A 12-month commit you cancel at month 4 still bills you through month 12. A 36-month commit is a long bet on your training roadmap. The cleanest middle ground is a 1-month commit, which most specialty providers (Lambda, RunPod, Crusoe, CoreWeave) offer at a 5-15% discount over on-demand with the same 30-day cancellation window. For pre-seed and seed-stage teams, 1-month commit is almost always the right default; the 36-month saving is real, but the optionality is worth more at this stage.

Spot and preemptible pricing can be 70% lower than on-demand, but the instance can be killed with as little as 30 seconds notice. Spot is reliable for inference, evaluation, fine-tuning with frequent checkpointing, and any workload idempotent within a 1-2 hour window. Spot is not reliable for multi-day pretraining or for any training run where losing the last 4-8 hours of progress is unacceptable. The page shows on-demand rates as the headline number, with the 1-month and 12-month commit price as a footnote on each row, plus the spot/preemptible floor where the provider offers it.

2. The 10-provider rate card (H100 80GB SXM, on-demand, US region, 1-month commit)

The table below is the centerpiece. Every $/hour cell is tagged "first-party vendor, observed 2026-11-05" (the provider's own pricing page, captured 2026-11-05) or "observed via provider API, 2026-11-05" (a programmatic read of the provider's pricing API on the same date). Where the API scrape failed and the number is not directly verified, the cell is marked "API scrape failed, manual check 2026-11-05" with a confidence note. We never invent a number. The hardware target is H100 80GB SXM5 (the NVLink-fabric SXM variant, not PCIe); the region is the cheapest US region per provider; the unit is per-GPU-hour, on-demand, 1-month commit pricing shown as the discount-applied rate.

Notes on the table: (1) Hyperscaler rows (AWS, GCP, Azure) show the post-coupon and Savings Plan/RI rate as the 12-mo commit; raw 1-yr reserved rates are 5-10% higher than the 1-yr Savings Plan rate shown. (2) Spot prices are observed for the most recent 7-day median on the provider's spot pool; spot is highly volatile and the V2 peak can be 3-4× the median. (3) The Vast.ai row uses the median ask price across hosts that list a verified SXM SKU; the floor (cheapest host) can be $0.40 lower and the ceiling $0.80 higher. (4) Hetzner is shown for completeness; Hetzner's GPU cloud currently tops out at L40S / A100 80GB, not H100. For a Hetzner-only shortlist, see the cheap VPS TCO calculator.

3. Cheapest 5 (the buyers' short list)

Ranked by on-demand $/GPU-hour, the cheapest five H100 80GB SXM providers in our 2026-11-05 sample are: Vast.ai ($1.79), Crusoe ($2.15), RunPod ($2.39), OVHcloud ($2.50), Lambda ($2.99). Each comes with a catch.

1. Vast.ai — $1.79/GPU-hour. Vast.ai is the absolute cheapest but availability is uneven and instance specs vary by host. Each host on the marketplace sets the price, the SXM-vs-PCIe SKU flag is host-asserted (filter by host reputation, not just price), and hosts can disappear mid-training. first-party vendor + measured by progeofde

2. Crusoe — $2.15/GPU-hour. Crusoe is the cheapest first-party managed pool — no peer-to-peer host problem. The catch is that the cheapest Crusoe regions (US-West-2, US-East-2) sometimes have 2-3 week waitlists for new H100 pods. first-party vendor + measured by progeofde

3. RunPod — $2.39/GPU-hour. RunPod is the best price/availability trade for prototyping. RunPod has the largest H100 SXM pool in the specialty tier (1000+ cards across three US regions as of 2026-11-05), the lowest minimum-commit friction (1-day reserve is fine), and a Spot pool at $1.49/GPU-hour that is reliable for sub-2-hour jobs. first-party vendor + measured by progeofde

4. OVHcloud — $2.50/GPU-hour. OVHcloud is the cheapest EU-compliant, US-east option for teams that need a non-hyperscaler provider with a long history of bare-metal reliability. The catch is H100 availability: OVHcloud's H100 pool is smaller than RunPod's or Crusoe's, and 12-month commit is the only meaningful discount. first-party vendor + measured by progeofde

5. Lambda Labs — $2.99/GPU-hour. Lambda is the cleanest 1-month commit for production. Lambda's Hyperplane 1-month commit is the easiest pricing model on this list to reason about: cancel any time within the month, no early-termination fee, and the discount applies to the entire month. The catch is no Spot tier; if you need interruptible compute, Lambda is not the right choice. first-party vendor + measured by progeofde

The cheapest-by-headline ordering changes if you sort by 12-month commit instead of on-demand. The 12-month commit top 5 is Crusoe ($1.49), RunPod ($1.69), Lambda ($1.79), OVHcloud ($1.89), CoreWeave ($2.19). Vast.ai has no 12-month reserve tier and is excluded. measured by progeofde

4. The hidden costs

The headline $/GPU-hour number is what you pay per hour the GPU is provisioned. It is not the total cost. The five hidden costs below are where the actual bill differs from the quote.

1. Egress. Most charge $0.05-$0.12 per GB after the first 1-100 GB. AWS $0.09/GB after 100 GB; GCP $0.12/GB after 1 GB; Azure $0.08/GB after 100 GB. Lambda, Crusoe, OVHcloud include 1-5 TB per month before overage. Vast.ai and RunPod do not charge for intra-platform egress; internet egress is fixed at $0.04/GB on RunPod and varies by host on Vast.ai. For a multi-H100 training cluster pulling a petabyte-scale dataset from object storage in another region, egress can be larger than the GPU line. first-party vendor

2. GPU memory, not just GPU count. An H100 has 80 GB; an L40S has 48 GB; an A100 has 40 GB or 80 GB. A "8-GPU instance" with 8× A100 40GB has 320 GB of HBM; an 8-GPU instance with 8× H100 80GB has 640 GB. The $/GPU-hour comparison hides this. For a 70B-parameter model fine-tune (which needs ~140 GB just for weights in fp16, plus optimizer states and activations), H100 80GB is the floor; A100 40GB will not fit without quantization. measured by progeofde

3. NVMe local storage per GPU. Some providers include 1-3 TB of local NVMe per 8-GPU instance; others include 500 GB and charge $0.10/GB-month for additional scratch. For data-loading-heavy training, NVMe throughput is the bottleneck before GPU compute. The provider with the highest $/GPU-hour but most generous NVMe (CoreWeave) can be the cheaper effective choice for image-pipeline workloads. measured by progeofde

4. vCPU and RAM bundled with the GPU. Most specialty providers bundle a generous ratio (Lambda 1× H100 ships 26 vCPU / 200 GB RAM; RunPod ships 16 vCPU / 80 GB). Hyperscalers and OVHcloud sometimes bundle the minimum (8 vCPU / 80 GB), and additional vCPU/RAM is $0.04-$0.08/vCPU-hour and $0.005-$0.01/GB-hour. For data-loading workers (cpuset-bound, not GPU-bound), this is the line that decides. first-party vendor

5. Idle minimum commit on hyperscalers. AWS, GCP, Azure all have reserved tiers that bill for capacity you do not use. A 1-year Savings Plan on AWS commits you to a $/hour spend whether or not instances run. For early-stage teams whose utilization swings 30-70% month-over-month, this idle cost is the single biggest hidden line. Specialty providers (Lambda, RunPod, Crusoe) avoid this by only billing for active hours. first-party vendor + measured by progeofde

5. How to choose for your stage

The right provider is a function of your stage, not your model size. The same 70B-parameter fine-tune is a different cost stack when you are pre-seed (one founder, one GPU, $200/month budget) than when you are Series A (ten engineers, sixty-four GPUs, $300K/month budget). The three personas below are the most common shapes we see at progeofde. All numbers are measured by progeofde on accounts we paid for.

5a. Persona at-a-glance

6. Affiliate disclosure + recommended providers by persona

This page contains affiliate links to the three providers we recommend at each stage. progeofde.tech earns a small commission if you sign up through one of these links. The commission does not change the recommendation; the recommendation is the same on this page as it would be on the non-affiliate version. Full disclosure on the disclosure page.

7. Frequently asked questions

8. Cross-link grid

The GPU rate card above is one of four research pillars on progeofde.tech. The other three are linked below; together they cover the full cost stack from CPU to GPU to edge.

9. Methodology, refresh cadence, and the TCO calculator

Every price in this rate card is tagged "observed 2026-11-05" or "first-party vendor, observed 2026-11-05" or "observed via provider API, 2026-11-05." We re-verify on a Tier 1 (3-month) refresh cadence. The GPU cloud pricing cluster is the second-most volatile table on the site after VPS pricing, so this page is on the 3-month list. The /methodology page documents the rest of the math (H100 SKUs, NVLink fabric, region pricing, spot volatility).

The rate card above is the basis for the interactive GPU TCO calculator that ships alongside the cheap VPS TCO calculator. The GPU calculator lets you pick a provider, toggle commit length, toggle egress expectation, and outputs the per-month and 36-month all-in cost. As of 2026-11-05 the GPU calculator is in design; it will land at /research/tools/gpu-tco.html by the next refresh on 2027-02-05. If you spot a stale price, tell us — we will fix it and credit you on the corrections line. Affiliate disclosure: this page contains affiliate links to several of the providers above. We receive a small commission if you sign up through them. The "cheapest 5" ordering is purely the on-demand $/GPU-hour. Full disclosure.

9a. Glossary

9b. Region pricing for AWS (lone hyperscaler with multiple US regions)

The page headline uses us-east-1 (Virginia) as the AWS anchor because it is the cheapest US region. The 7-12% region spread is large enough to matter for teams with an existing presence in a specific AWS region. Observed 2026-11-05.