GPU cluster rental: how to rent an H100 or B200 cluster

Reserved clusters vs on-demand

A GPU cluster is a set of 8-GPU servers joined by an RDMA fabric (InfiniBand or RoCE), with shared storage and a scheduler. You can get one two ways.

• On-demand clusters. Self-serve, billed by the hour or second, no term. Runpod Instant Clusters go up to 64 GPUs (8 nodes) on demand over InfiniBand or RoCE v2 at 1,600 to 3,200 Gbps per node (Runpod). Together AI sells on-demand GPU clusters at $3.99 per H100 SXM GPU-hour with no term (Together AI). Good for a few days of work or a test run.
• Reserved clusters. You commit to a GPU count for a term in return for guaranteed capacity and usually a lower rate. Our view: this is the default for runs longer than a few weeks, because capacity is guaranteed for the whole run.

Reserved is not always cheaper per hour. Together AI's reserved H100 SXM rates on 2026-11-09 were $3.69 for 3 to 30 days, $3.45 for 31 to 90 days and $3.19 for 91 to 180 days, all below its $3.99 on-demand rate (Together AI). Lambda's 1-Click Cluster rates for the same chip were $6.16 (16 GPUs), $5.85 (64) and $5.54 (256), above its $4.59 8x on-demand instance (Lambda, schedule effective 2026-11-16). The cluster price buys InfiniBand between nodes and a managed scheduler, which a stand-alone instance does not have.

For B200 the pattern is the same. Together AI listed $7.99, $7.79 and $6.79 across the three reserved tiers against $8.19 on demand. Lambda listed $9.86, $9.36 and $8.87 for 16, 64 and 256+ GPU clusters. Compare chips first in H100 vs H200 vs B200, and current on-demand rates on our pricing page.

Published minimum terms

Most cluster pricing is by quote. These are the minimums providers publish, as seen on 2026-11-09.

Discounts that are published without a GPU-specific rate: CoreWeave says committed usage can save up to 60% off on demand. Latitude.sh says monthly prepaid servers are 50% cheaper than hourly and yearly 65% cheaper. Nebius says commitment discounts exist for multi-month large-cluster reservations but publishes no figures (CoreWeave, Latitude.sh, Nebius). Hyperstack publishes reservation prices "starting from" $2.72 for H100 SXM, $2.79 for H200 and $5.10 for B200, against on-demand rates of $3.20, $3.99 and $6.00, without stating the term (Hyperstack).

What a cluster quote should include

This is our checklist, not an industry standard. If a quote is missing a line, ask for it in writing.

1. Hardware. GPU SKU and form factor (H100 SXM, not "H100"), GPUs per node, node count, CPU and RAM per node, local NVMe.
2. Fabric. InfiniBand or RoCE, speed per node, topology (rail-optimized, non-blocking or oversubscribed), and whether every node sits on the same fabric.
3. Storage. File system type, capacity, measured throughput, and object storage for checkpoints.
4. Software. Managed Slurm or Kubernetes, who maintains drivers and images, and what health checks run.
5. Location. Data center and region, plus who owns the facility.
6. Dates. Delivery date, acceptance period, and when billing starts.
7. Price. Per GPU-hour and total contract value, what is included (storage, egress, support), and the payment schedule. Note whether any part is prepaid.
8. Failures. How fast failed nodes are replaced, whether spare nodes are held back, and the credits you get when targets are missed.
9. End of term. Renewal price, notice period, and whether you can extend or shrink.

Before the acceptance period ends, run nccl-tests across all nodes and a short run of your real training job. The GPU cloud checklist covers interconnect, storage and reliability checks in more detail.

Timelines

Few providers publish lead times, so we report only what they state.

• Runpod Instant Clusters and Together AI on-demand clusters are self-serve, with hourly or per-second billing and no term (Runpod, Together AI).
• Prime Intellect says dedicated cluster offers are "ready in 24 hours" and that cluster requests need no commitment (Prime Intellect).
• Blackwell capacity at several providers is "reserve now" or contact sales. Voltage Park lists B200, GB200, B300 and GB300 for reservation. Denvr's page says Blackwell is coming soon. CUDO quotes everything on request (Voltage Park, Denvr, CUDO).

Our view: if a provider has the nodes racked and idle, delivery is fast. If they are installing for you, the date in the quote depends on their supply chain and power, so tie billing to acceptance, not to signature.

GPU brokers and marketplaces

Brokers, aggregators and marketplaces sit between you and the clouds that own the hardware. They help with price discovery and short-notice capacity. Here is how four of them describe themselves.

Shadeform

An aggregator across about 18 clouds. Shadeform says "you pay the same as going direct to the clouds in our network," so the listed price is the underlying cloud's price. Its public feed showed 304 instance types on 2026-11-09, of which 65 had current availability. The cheapest available H100 was a PCIe instance at $2.50 per GPU-hour. Reserved contracts go through a request form with a 12-month minimum term (Shadeform, reserved). You get one API, CLI and MCP server across clouds.

SF Compute

An order book for blocks of GPU time on 8-GPU nodes. Buyers set a node count, a delivery window and a maximum price per node-hour, and tenants can resell unused reserved time (SF Compute, docs). eWEEK reports the platform takes roughly 10% per transaction (eWEEK). Live prices need an account, so we could not record them.

Vast.ai

A marketplace where hosts set prices that change with supply and demand. On 2026-11-09 the lowest live offers were $2.27 for H100 SXM (median $2.87), $3.43 for H200 (median $4.34) and $5.50 for B200 (median $8.75) per GPU-hour, billed per second with no minimum hours (Vast.ai). Reserved terms of 1, 3 or 6 months are optional. Hardware, network and data center quality vary by host, so check each offer's details.

Prime Intellect

A compute marketplace that also runs its own clusters. Its compute page offers 1 to 256 GPUs by the hour with no commitment, and listed H100 at $3.50, H200 at $5.48 and B300 at $6.64 per GPU-hour on 2026-11-09 (Prime Intellect). SemiAnalysis gave it Bronze in ClusterMAX 3.0 and called it "the best marketplace tested" (SemiAnalysis). It raised a $130 million Series A in July 2026 (Prime Intellect).

Others

Mithril runs a bid-based spot market and "flexible reservations" that credit paused or returned time. Its published 7-day spot averages were $1.51 per GPU-hour for a 1x H100 and $2.625 for an 8x H100 instance (Mithril, flexible reservations). SemiAnalysis describes it as a marketplace wrapping three Nebius availability zones (SemiAnalysis). Hyperbolic lists supplier capacity at "starting from" rates refreshed weekly, for example B200 from $5.99 (Hyperbolic).

Our view on brokers: they are useful for comparing prices and finding a few nodes fast. For a long, large reservation, know which provider owns the hardware and who you call when a node fails. Ask the broker to name the underlying provider before you sign.

Negotiating a cluster (editorial)

These tips are our editorial view, not sourced facts.

• Get three written quotes for the same spec. Same GPU SKU, node count, fabric, storage, term and start date. Otherwise you are comparing different products.
• Trade term for price, not prepayment for price. A longer term is a fair reason for a lower rate. Paying months upfront moves the provider's financing risk onto you. If you prepay, ask for a bigger discount or an escrow.
• Ask for a burn-in period. A few days to run nccl-tests and your own job before billing starts, with the right to reject bad nodes.
• Put failure handling in the order form. Replacement time for a failed node, spare capacity, and credits.
• Ask for flexibility. An option to add nodes at the same rate, or to step down GPU count after the main training run ends.
• Watch list price changes. Crusoe announced on-demand increases effective 2026-12-01, with H100 going from $3.90 to $4.50 (Crusoe). In a rising market, a fixed rate for the term is worth more.
• Check the counterparty. Our neocloud market overview covers debt and customer concentration across the sector.

Marketplaces and aggregators

Providers in our directory that sell capacity from several hosts or clouds.

Unfamiliar term? See the GPU cloud glossary. Profiles: Shadeform, Vast.ai, SF Compute, Prime Intellect, Together AI, Lambda.