H100 GPU Cluster Rental Pricing and Contract Terms Explained

The H100 rental market quotes prices that differ threefold for what looks like the same chip, and training teams sign on the hourly rate without noticing what the tiers underneath actually sell. A cluster rental is not eight GPU rentals stapled together — the interconnect, the scheduling, and the contract terms decide whether a multi-week campaign survives — and the procedure for renting one well is different from renting instances. This page prices the tiers honestly and walks the rental procedure.

Prerequisites: Know the Tiers Behind the Prices

H100 rental prices span roughly three tiers with reasons: marketplace and spot capacity near 1.7 to 2 dollars per GPU-hour (median marketplace pricing, subsidized or shared-pool economics with availability variance), dedicated cloud capacity from roughly 2 to 3 dollars (provider-operated, consistent trust), and contract terms around 2.35 to 2.50 on annual commitments — with the SemiAnalysis price index documenting annual contract rates rising about forty percent in a year as supply tightened. The threefold spread is not noise: it prices different trust structures, availability guarantees, and interconnect quality, and a training campaign that needs eight GPUs talking reliably pays for the tier where they do.

The hardware context explains the subsidized floor: an H100 costs on the order of twenty-five thousand dollars to buy, so a two-dollar hour that runs around the clock barely clears the hardware's financing — community discussion of exactly this arithmetic is why below-market rates deserve the questions the next section formalizes. Within-tier comparison beats across-tier shopping: rate tables that average fifty-seven providers mix the tiers into a meaningless middle, while the honest comparison is same-tier on cost per effective compute, the normalization the rating systems built for exactly this.

The Rental Steps: From Requirements to Signature

Rent in five steps: define the campaign shape (GPU count, duration, interconnect minimum — the fabric requirement that separates cluster rentals from GPU rentals); shortlist by tier (marketplace for fault-tolerant bursts, dedicated for production campaigns, contract for multi-month programs); demand the six terms (fabric throughput guarantee, capacity availability commitment, scaling and burst terms, support response, data handling including egress cost, exit and migration); test before the campaign (the acceptance phase); and structure payment to milestones where the provider allows it. The signature mistake is renting clusters like instances — the interconnect and scheduling obligations live in the contract, not the spec sheet.

The dedicated and contract tiers are where the six terms land on receptive paper — providers operating single-tenant capacity price the guarantees rather than fighting them, the structure of dedicated H100 environments such as the ones OneSource Cloud delivers in U.S. data centers, where fabric behavior is contracted and monthly pricing is flat with zero egress fees. Marketplace tiers can still win for fault-tolerant burst work — the terms list simply shrinks to what a restartable job needs.

Verification: Acceptance Tests Before the Campaign

Run three acceptance tests before committing the campaign: a fabric test (sustained all-reduce at the contracted throughput — cluster throughput is only real under the collective that will carry the training), a scheduling test (gang placement of the campaign's job shape, startup and restart behavior), and a node-health soak (hours of thermal and error-rate stability across all nodes — the marginal node that fails in week two was detectable in hour two). The verified state: test results in writing, tied to the contract terms they exercise, so any later shortfall is a remedy conversation rather than a renegotiation.

Treat the tests as part of the rental, not a favor — providers organized for cluster campaigns schedule them into onboarding, and the results document becomes the campaign's baseline: the first week's performance conversation references it, and the renewal negotiation starts from it. For the contract-tier route with delivered fabric guarantees, the acceptance phase is documented alongside the OneSource Cloud Managed AI Infrastructure model — the tests stay yours, the terms already exist.

FAQ

Why do H100 hourly prices vary threefold between providers?

Because they price different things: marketplace rates near two dollars reflect subsidized or shared-pool capacity with availability variance and lighter trust, dedicated rates carry provider-operated consistency, and contract rates buy availability commitments — compare within tiers and on cost per effective compute (the rating methodology that normalizes for real throughput), never headline to headline.

What should a multi-node H100 rental guarantee that single-GPU rentals ignore?

The fabric and its schedule: sustained interconnect throughput under all-reduce (the traffic your training actually generates), gang scheduling for the job shape, and node-health stability across every node — the terms single-GPU rentals never exercise and cluster campaigns cannot survive without; on dedicated capacity like OneSource Cloud enclaves these arrive as contracted guarantees with flat monthly pricing rather than best-effort clauses.