Buyer's guide · compiled 2026-09-10

The GPU contract gotcha list

Reserved GPU contracts are where the deliverability problem becomes a legal problem. Below: the clauses that burn buyers, compiled from published negotiation guides, provider rating methodologies, and practitioner writeups — each with its source. The pattern in all of them: the contract prices a number of GPUs, and the disputes are always about everything the number doesn't say.

Capacity guarantee vs uptime SLA conflation

Providers advertise 'Capacity guarantee: Yes' without percentages, measurement windows, or remedies. A capacity guarantee governs access to your reserved allocation; an uptime SLA governs whether it works. Demand exact percentages, per-node vs cluster-averaged measurement windows, and specific credits for misses — don't infer a number. [source]

Idle billing and hidden utilization floors

Reserved contracts bill flat hourly rates whether the cluster runs or sits idle; some contracts price the discount against a guaranteed minimum utilization, converting a capacity option into a usage obligation. Flex reservations (holding fee + usage rate) reduce idle penalty. Breakeven: a 45% discount only beats on-demand above ~55% utilization. [source]

Uncapped overage rates referencing floating market prices

'Standard on-demand rates' in an overage clause is a pointer to whatever the provider charges that day, not a number. Negotiate a multiplier ceiling (e.g., 1.5x reserved rate max) and a pre-agreed 10-20% burst allowance instead of open-ended exposure during shortage spikes. [source]

Egress fees and undefined exit mechanics

Hyperscalers impose high data-transfer fees that function as switching costs (most neoclouds do not charge egress). Negotiate capped or waived egress at term end, defined data export formats, and dated migration assistance — e.g., NVIDIA's 30-day transition window as a concrete, dated obligation rather than goodwill language. [source]

Capacity reclaim / preemption clauses on discounted compute

Some providers sell discounted idle compute but retain the right to reclaim it with seven days' notice if a higher-paying customer appears — fine for fault-tolerant batch work, disastrous mid-training-run if you didn't read the clause. [source]

Hardware substitution rights breaking ASC 842 lease treatment

Multi-tenant products often retain rights to swap underlying hardware, which can fail the identified-asset test and disqualify the contract from ASC 842 lease accounting — changing balance-sheet treatment. Contracts under 12 months typically avoid capitalization; finance teams can use accounting treatment as leverage against 3-year locks. [source]

Late delivery with no escape clause

Late GPU deliveries are widespread in the GPU cloud industry; MSAs must specify exact delivery dates with escape clauses (or penalties) for delays, since the customer's own downstream commitments start on the promised date. [source]

Prepayment funding the provider's own buildout

Contracts over one year typically require ~20% of total contract value prepaid, and neocloud suppliers use these prepayments to finance their own infrastructure deals — meaning buyers carry counterparty/completion risk on capacity that may not exist yet. [source]

No-cancellation policies and penalty asymmetry

AWS Capacity Blocks have a strict no-cancellation policy; Google allows cancellation only before provisioning; most neocloud contracts impose significant financial penalties. Secondary-market resale of unused reservations is an emerging escape valve. Also watch auto-renew on multi-year commitments without explicit sign-off. Negotiable levers: step-down clauses at 12/24-month checkpoints, hour banking, and SKU-generation (not specific SKU) commitments allowing H100->H200 swaps. [source]

SLA-compliant but unusable clusters (link-flap loophole)

NIC flaps lasting one microsecond every minute can cause NCCL stalls of minutes to half-hours, yet some providers claim SLA compliance because nodes technically stayed 'up.' SLAs must define health in terms of workload-relevant metrics (flap rates, fabric error thresholds), not node ping. [source]

SLA credit mechanics: future credits vs real-time deductions

Credit mechanisms vary — some providers issue future bill credits (locking you in to collect your remedy), others deduct downtime from the current bill in real time; ClusterMAX 2.0 flags real-time deductions and explicit breach penalties as what buyers should demand. [source]

Discount tiers vs term-length lock-in math

March 2026 benchmarks: 6-month terms run 20-30% below on-demand, 1-year 40-50%, 3-year 55-72%; providers resist sub-6-month terms on premium GPUs. Multi-vendor bidding cuts 15-20%, demand consolidation 15-30%, and quarter-end timing improves leverage — but the discount only pays if utilization exceeds (1 - discount%). [source]

How the tracked providers stack up

The same clauses, provider by provider, from each seller's own published terms (full detail + sources on every provider page). n/p = not published — the seller states it nowhere we could find. Absence is a finding: an unpublished SLA is the link-flap loophole with better manners.

As published 2026-09-14 · truncated for scanning — the provider pages carry full text and source links.

The through-line

Nearly every gotcha above is an unspecified deliverable: uptime measured on nodes while the fabric flaps, capacity "guaranteed" without a number, hardware swapped under an identified-asset test, overage priced against a floating reference. This is exactly what a delivery specification exists to fix — ASSAY-1 defines the record a contract can cite so "8×H100" stops being a negotiation and starts being a specification.

And the number every clause above orbits — what deals like yours actually close at — is exactly what list prices can't tell you. The street-price pilot aggregates redacted real prints, contribute-to-see.