Quick answer
What is the cheapest GPU cloud for AI right now?
The cheapest GPU cloud for AI depends on the GPU. As of 29 October 2026, the lowest self-serve on-demand prices we found were Massed Compute for the H100 ($2.73), H200, B200 and L40S, Thunder Compute for the A100 80GB ($1.09), and Salad for the RTX 4090 ($0.33) and RTX 5090. GPUwithCrypto prices each GPU at least 15% under a named on-demand reference, paid in crypto: H100 $2.32/hour.
Which cloud has the cheapest on-demand price for each GPU?
According to GPUwithCrypto's price check of 29 October 2026: the lowest self-serve on-demand price per GPU-hour, the next provider, and the median of all providers listing that GPU. GPUwithCrypto prices each GPU at least 15% under a named on-demand reference read the same day. We do not claim to be the cheapest for every GPU; spot and marketplace offers (Vast.ai) can be lower.
| GPU | Cheapest on-demand | Runner-up | Median (providers) | GPUwithCrypto | Crypto at the cheapest |
|---|---|---|---|---|---|
| H100 80GB | $2.73 (Massed Compute) | $2.89 (RunPod, PCIe) | $4.16 (16) | $2.32/hour | Yes: BTC |
| H200 141GB | $3.62 (Massed Compute) | $3.99 (Hyperbolic) | $5.29 (13) | $3.07/hour | Yes: BTC |
| B200 | $5.43 (Massed Compute, 8-GPU node ÷ 8) | $5.99 (Hyperbolic) | $8.50 (9) | $4.61/hour | Yes: BTC |
| A100 80GB | $1.09 (Thunder Compute) | $1.35 (Massed Compute) | $2.23 (13) | $0.92/hour | No: card or preloaded credit |
| L40S 48GB | $0.97 (Massed Compute) | $1.09 (RunPod) | $1.67 (11) | $0.82/hour | Yes: BTC |
| RTX 4090 | $0.33 (Salad) | $0.89 (RunPod) | $0.61 (2) | $0.28/hour | Yes: USDC, USDT, RENDER |
| RTX 5090 | $0.50 (Salad) | $0.84 (Spheron) | $0.84 (3) | $0.42/hour | Yes: USDC, USDT, RENDER |
Prices checked on
USD per GPU-hour, taxes excluded. Salad's High Priority prices rise on 1 November 2026: RTX 4090 to $0.44, RTX 5090 to $0.60. Voltage Park's "starting at $1.99" H100 is not counted: its pricing cards say "Contact for pricing". The full provider list is in the GPU rental price index.
Is a spot or marketplace GPU cheaper?
Yes, for every GPU. These offers can be taken back, change with each host, or run on community machines, so they are listed apart from on-demand prices. Each is labelled with what the provider calls it.
| GPU | Lowest spot / marketplace | What it is | Other low offers |
|---|---|---|---|
| H100 80GB | $0.79 (Nebius) | Preemptible "from" price, dynamic | Jarvislabs spot $1.54; RunPod Community Cloud $1.99 (PCIe); Akash bids from $2.04 |
| H200 141GB | $0.79 (Nebius) | Preemptible "from" price, dynamic | Jarvislabs spot $2.29; CoreWeave spot $2.62 |
| B200 | $0.99 (Nebius) | Preemptible "from" price, dynamic | CoreWeave spot $4.26; Verda spot $4.39 |
| A100 80GB | $0.89 (Jarvislabs) | Spot, interruptible | Akash bids from $1.07; Verda spot $1.11; RunPod Community Cloud $1.19 |
| L40S | $0.74 (Nebius) | Preemptible "from" price, dynamic | RunPod Community Cloud $0.79; Verda spot $0.94 |
| RTX 4090 | $0.16 (Salad) | Batch priority, "most interruptible" | RunPod Community Cloud $0.34; Akash bid $0.34 |
| RTX 5090 | $0.25 (Salad) | Batch priority, "most interruptible" | Akash bid $0.57 (none available); RunPod Community Cloud $0.69 |
Prices checked on
Multi-GPU prices divided by the GPU count. Vast.ai is quoted separately below: its price-feed licence asks for a data licence before use in a benchmark.
When is spot or marketplace cheaper than GPUwithCrypto?
Always, on 29 October 2026: for all seven GPUs, at least one spot, preemptible or marketplace offer was below the GPUwithCrypto price. Vast.ai's public price feed went lower still, with an RTX 4090 from $0.13 per hour.
What you trade is certainty. Nebius says its highest preemptible rate is always one cent below on-demand, so the "from" price moves. Salad's Batch tier is its "most interruptible". A marketplace host sets its own terms. Instances are on-demand, not spot: nobody can take your GPU back while your balance covers it, and the instance runs until you stop it.
Spot suits jobs that checkpoint often and restart cleanly. A long fine-tuning run or a live inference server is usually worth an on-demand GPU. The spot GPU explainer covers the trade-off in detail.
Which of the cheapest GPU clouds accept crypto?
Most of them. Massed Compute, cheapest on-demand for the H100, H200, B200 and L40S, takes Bitcoin, with no ID step documented, but has no auto-recharge for crypto. Salad, cheapest for the RTX 4090 and RTX 5090, takes USDC, USDT and RENDER on Solana; crypto payments cannot be refunded.
Thunder Compute, cheapest for the A100 80GB, lists only a saved card or preloaded credit in its billing docs. RunPod, runner-up for the H100, L40S and RTX 4090, takes crypto after KYC. Hyperbolic's crypto support could not be confirmed.
GPUwithCrypto takes BTC, USDT, USDC, ETH, SOL, LTC with no ID, and prices each GPU at least 15% under a named on-demand reference. The crypto comparison of 33 GPU clouds has the full rules.
What should I check besides the hourly price?
Billing increments, minimums and data charges change the real cost of a short job. Facts from each provider's own pages, read on 29 October 2026; "—" means not stated.
| Provider | Billing increment | Minimum | Egress / storage | Payment / ID |
|---|---|---|---|---|
| GPUwithCrypto (ours) | Per minute | $25 top-up | Local NVMe included, wiped on delete | BTC, USDT, USDC, ETH, SOL, LTC; no ID |
| Massed Compute | Hourly rate, increment not stated | No long-term commitment | "No Bandwidth Charges"; local storage included | Card or BTC; no ID step documented |
| RunPod | Per second | $10 to start | Container disk $0.10/GB-month; idle volume $0.20/GB-month | Card, or crypto after KYC |
| Thunder Compute | Per minute | — | No egress fees; 100 GB disk per GPU, then $0.219/GB-month | Card or preloaded credit |
| Salad | Per second | $5 | — | Card or USDC, USDT, RENDER |
| Jarvislabs | Per minute | $10 | Free in-region egress; disks $0.10/GB-month | Card; bank transfer above $500 |
| Lambda | Per minute | None for instances | Local SSD included | Major credit cards only |
| Crusoe | Per hour or per second | No minimum commitment | No ingress or egress charges; persistent disk $0.08/GiB-month | Non-prepaid credit card |
| Nebius | — | $25 | Egress and ingress free | Card; bank transfer for companies |
| Spheron | Per minute after a 20-minute minimum runtime | $15 by card, $50 by crypto | Bandwidth and NVMe included | Card or USDC |
| Paperspace / DigitalOcean | Per second, 5-minute minimum | $25 authorization for crypto | 15,000 GiB transfer included per Droplet | Card, PayPal, stablecoins |
| OVHcloud | Per second, 1-minute minimum | — | Local NVMe/SSD included | Card, PayPal; ID sometimes |
| Google Cloud | Per second after a 1-minute minimum | Postpaid | Local SSD bundled | Card or bank; no crypto |
Prices checked on
Why are AWS, Google Cloud and Azure so expensive per GPU?
Because their GPU price includes the whole VM, and most of their large GPUs come as 8-GPU machines. On 29 October 2026 an on-demand H100 cost $6.88 per GPU-hour at AWS (us-east-1), $11.06 at Google Cloud (Iowa) and $12.29 at Azure (East US), against $2.73 at Massed Compute. The price index shows the H100 median falling from $4.16 to $3.49 without them.
Cheapest GPU cloud: questions
What is the cheapest cloud GPU for AI?
Per GPU-hour on-demand, the RTX 4090: $0.33 at Salad until 31 October 2026, then $0.44. GPUwithCrypto rents one for $0.28/hour. For 80 GB of VRAM, the cheapest on-demand was the A100 80GB at Thunder Compute, $1.09.
What is the cheapest H100 cloud right now?
On-demand, Massed Compute at $2.73 per GPU-hour on 29 October 2026, and GPUwithCrypto at $2.32/hour. Preemptible H100s started lower, from $0.79 at Nebius, but can be taken back.
Is spot always cheaper than on-demand?
On 29 October 2026, yes, for all seven GPUs we track. Spot and preemptible GPUs can be interrupted, so they suit jobs that save checkpoints.
Which cheap GPU cloud takes crypto with no ID?
Massed Compute (Bitcoin) and Salad (USDC, USDT, RENDER) describe no ID step and had the lowest on-demand prices for six of the seven GPUs. GPUwithCrypto takes BTC, USDT, USDC, ETH, SOL, LTC with no ID, at least 15% under a named on-demand reference for each GPU.
Is the cheapest GPU cloud reliable?
We did not measure uptime, so we do not rank on it. What we can compare is the type of offer: on-demand GPUs stay yours while you pay, while spot, preemptible and marketplace offers can end early.
All sources (read 29 October 2026)
Sources (22)
- Pricing (VM) vm.massedcompute.com/pricing — Massed Compute (read 2026-10-29)
- Crypto: automatic payments vm-docs.massedcompute.com/docs/billing/crypto/automatic-payments — Massed Compute (read 2026-10-29)
- Pricing thundercompute.com/pricing — Thunder Compute (read 2026-10-29)
- Billing thundercompute.com/docs/billing.md — Thunder Compute (read 2026-10-29)
- SaladCloud price changes, November 2026 blog.salad.com/saladcloud-price-changes-november-2026 — Salad (read 2026-10-29)
- Billing docs.salad.com/general/explanation/billing — Salad (read 2026-10-29)
- Pricing runpod.io/pricing — RunPod (read 2026-10-29)
- GPU marketplace hyperbolic.ai/marketplace — Hyperbolic (read 2026-10-29)
- Prices nebius.com/prices — Nebius (read 2026-10-29)
- Pricing jarvislabs.ai/pricing — Jarvislabs (read 2026-10-29)
- Pricing coreweave.com/pricing — CoreWeave (read 2026-10-29)
- Pricing verda.com/pricing — Verda (read 2026-10-29)
- GPU prices API console-api.akash.network/v1/gpu-prices — Akash Network (read 2026-10-29)
- Public GPU pricing feed — Vast.ai (read 2026-10-29)
- Pricing lambda.ai/pricing — Lambda (read 2026-10-29)
- Cloud pricing crusoe.ai/cloud/pricing — Crusoe (read 2026-10-29)
- Pricing spheron.network/pricing — Spheron (read 2026-10-29)
- GPU Droplets pricing digitalocean.com/pricing/gpu-droplets — DigitalOcean (read 2026-10-29)
- Public Cloud prices ovhcloud.com/en/public-cloud/prices — OVHcloud (read 2026-10-29)
- Accelerator-optimized VM pricing — Google Cloud (read 2026-10-29)
- EC2 on-demand price list, US East (N. Virginia) — AWS (read 2026-10-29)
- Azure Retail Prices API, East US — Microsoft (read 2026-10-29)