GPU clouds, compared on price

Nebius and CoreWeave are the only neoclouds rated Platinum in SemiAnalysis ClusterMAX 3.0 (Sep 2026), and both run managed Kubernetes plus Slurm on InfiniBand. Nebius publishes lower on-demand rates (H100 $4.50 vs CoreWeave $6.155 per GPU-hour on 2026-11-09) and holds SOC 2 Type II and ISO 27001. Caveat: both lean on debt; CoreWeave’s 10-Q shows $35.55B of debt and 36% of revenue from one customer, and Nebius issued about $10B of convertible notes in 2026.

Also consider CoreWeave

Denvr Dataworks lists the lowest fixed on-demand H100 price we found on 2026-11-09: $2.45 per GPU-hour on 8x SXM nodes with 3200G InfiniBand. Vast.ai’s lowest live marketplace offer was $2.27 for H100 SXM, and among ClusterMAX-medaled clouds Hyperstack is cheapest at $3.20. Caveat: Denvr is “Not recommended – Unavailable” in ClusterMAX 3.0, and lower “from” claims elsewhere (Latitude.sh $1.68, Voltage Park $1.99) were not shown as available offers.

Also consider Vast.ai, Hyperstack

Hyperstack has the lowest clear on-demand B200 price among ClusterMAX-medaled providers, $6.00 per GPU-hour on 2026-11-09, with a Bronze rating and SOC 2 Type II. For GB200 NVL72, CoreWeave is the only provider publishing an on-demand rate ($10.50 per GPU-hour); Nebius ($8.50), Lambda ($7.69 from 2026-11-16) and Together AI ($8.19) also publish B200 on demand. Caveat: SemiAnalysis noted Hyperstack relied on manual node remediation, and its GB200 and GB300 have no public price.

Also consider CoreWeave, Nebius, Lambda, Together AI

Prime Intellect is the only marketplace with a ClusterMAX 3.0 medal (Bronze), and SemiAnalysis called it the best marketplace tested; it lists H100 at $3.50 and H200 at $5.48 per GPU-hour and raised a $130M Series A in July 2026. Vast.ai is the alternative for breadth and price discovery, with live offers on about 20 GPU models from $2.27 for H100 SXM on 2026-11-09. Caveat: SemiAnalysis saw reliability issues in Prime Intellect’s burn-in, and Vast.ai states SOC 2 Type I only.

Also consider Vast.ai

TensorWave runs only AMD Instinct (MI300X to MI355X) and is rated Silver in ClusterMAX 3.0, where SemiAnalysis called its monitoring exemplary. For self-serve, DigitalOcean lists MI300X at $2.59 and MI325X at $3.80 per GPU-hour on demand, and Crusoe lists MI300X at $3.45 (2026-11-09). Caveat: TensorWave publishes no prices, and SemiAnalysis saw RCCL collective hangs in testing.

Also consider DigitalOcean, Crusoe

Runpod publishes its full rate card (H100 SXM $3.99, B200 $7.99 per GPU-hour on 2026-11-09), bills per second with no minimum, holds SOC 2 Type II and ISO 27001, and raised $100M in June 2026. Lambda (ClusterMAX 3.0 Silver) is the step up for multi-node work, with 1-Click Clusters from a 2-week minimum. Caveat: Runpod is only in ClusterMAX 3.0’s Participation tier, and Lambda’s public prices take effect 2026-11-16.

Also consider Lambda

Nebius is headquartered in Amsterdam with GPU regions in Finland, France, Spain and the UK, is rated Platinum in ClusterMAX 3.0 and holds SOC 2 Type II and ISO 27001. Verda keeps its whole fleet in Helsinki, states GDPR compliance, is rated Bronze and raised a $189M Series B in September 2026. Caveat: Verda’s on-demand prices are high (H100 $4.987 per GPU-hour on 2026-11-09), and SemiAnalysis notes construction delays at Nebius’s Béthune site in France.

Also consider Verda

Together AI runs serverless and dedicated model inference next to its GPU clusters, and publishes dedicated inference rates of $5.49 per H100 and $8.99 per B200 GPU-hour, plus on-demand, preemptible and 3–180 day reserved cluster prices. Runpod is the alternative for serverless GPU endpoints, with published per-second rates (H100 $5.39, B200 $10.79 per hour) and a reported 20B+ serverless requests served. Caveat: SemiAnalysis downgraded Together AI to Bronze in ClusterMAX 3.0 for reliability, and Runpod is in the Participation tier.

Also consider Runpod

Firmus is rated Silver in ClusterMAX 3.0, operates in Singapore and Australia with Malaysian sites under development, raised US$2B in August 2026 and names OpenAI as an anchor customer. GMI Cloud, also Silver, has APAC capacity including a Taiwan AI Factory. Caveat: Firmus publishes no GPU prices, and GMI publishes only “from” prices without saying whether they are on-demand.

Also consider GMI Cloud

Latitude.sh publishes hourly bare-metal GPU server prices (from $1.68 for a 1x H100 server and $8.00 per GPU on an 8x B300 server with 800 Gbps RoCE, as of 2026-11-09), holds SOC 2 Type II and is owned by ASX-listed Megaport. Hydra Host is the alternative for a wider bare-metal catalog, including B300 and GB300 reservations. Caveat: Latitude.sh is in ClusterMAX 3.0’s Participation tier, and Hydra Host is rated “Not recommended – Underperforming” with ambiguous published prices.

Also consider Hydra Host