Why Exit Clauses Matter in GPU Contracts
GPU hardware contracts typically run 12-36 months with monthly commitments ranging from $50,000 for a small H100 cluster to $2M+ for a B200 supercluster. Once signed, switching providers is expensive: data egress fees, checkpoint migration, network reconfiguration, and retraining time can cost 15-40% of the contract value. AI teams that do not negotiate termination clauses upfront often remain locked into underperforming or overpriced infrastructure when their needs change.
The 2025-2026 GPU market has shifted from seller-favorable (3x list price, no termination rights) to buyer-favorable (15-30% discounts, negotiable terms). Providers now compete on contract flexibility. AI teams that fail to negotiate exit terms leave money on the table and operational risk unaddressed.
Early Termination Fees and Penalties
Standard early termination fees for GPU reserved contracts range from 50-100% of remaining commitment value with a 6-month minimum. A 24-month H100 contract terminated at month 12 would owe 50% of the remaining 12 months, or 6 months of fees ($180,000 on a 32-GPU H100 cluster at $2/hr/gpu). Some providers offer tiered termination that decreases over the contract term, e.g., 100% of remaining in year one, 75% in year two, 25% after 18 months.
Negotiation targets: push for a 3-month minimum (not 6), a declining percentage starting at 50% (not 100%), and termination without penalty if the provider fails performance SLAs two months consecutively. Some providers will agree to a soft landing clause where termination fees cap at 20% of paid-to-date value rather than remaining commitment. Get this in the MSA, not the SOW.
Data Egress and Migration Windows
Data egress fees from GPU providers range from $0.00/GB (Lambda, CoreWeave on dedicated connections) to $0.12/GB (mainstream clouds). For a 50TB training dataset and checkpoint store, egress at $0.12/GB adds $6,000 to migration cost. More critically, the migration window how long you have to transfer data after termination notice is often 7-14 days, which may be insufficient for multi-petabyte datasets.
Negotiate a 60-day migration window at minimum, with the provider retaining the obligation to keep the GPUs powered and network active during that period. Add a clause requiring the provider to provide a direct connect or bulk transfer mechanism (Snowball-style physical shipping) if the dataset exceeds 100TB. Also negotiate egress at provider cost (typically $0.02-0.05/GB) or waived entirely for checkpoint data.
| Line Item | Typical Default | Negotiated Target |
|---|---|---|
| Termination Notice Period | 14 days | 60-90 days |
| Early Termination Fee | 100% remaining | 50% declining to 20% |
| Minimum Commitment | 6 months | 3 months or waiver |
| Data Egress Fee | $0.08-0.12/GB | At cost or waived |
| Migration Window (post termination) | 7 days | 60-90 days |
| SLA Breach Termination | Not specified | After 2 consecutive months |
| Force Majeure Refund | No refund | Pro-rata if >30 days |
Performance SLA Breach and Recourse
GPU provider SLAs typically guarantee 99.5-99.99% availability, but the definition matters. Many count a node as available if it can ping even if the NVLink fabric is degraded or GPU clock frequencies are below spec. A meaningful SLA links credits to training throughput impact, not just uptime. For example, if cumulative NCCL all-reduce bandwidth drops below 90% of rated spec for more than 2% of measured intervals, the provider owes 10% of the month's fees.
Specific SLAs to negotiate include node availability (99.5% minimum for the GPU fleet, not individual nodes), network bandwidth guarantee (measured at the GPU level via NCCL benchmark, not switch port counters), and maintenance windows (maximum 4 hours per month, scheduled with 7 days notice, no degradation during training). If the provider fails any SLA metric for two consecutive months, you should have the right to terminate without penalty.
Force Majeure and Geopolitical Clauses
The standard force majeure clause in GPU provider contracts excuses performance when events outside the provider's control occur. The risk for AI teams is that a provider in a geopolitically unstable region (Taiwan for hardware sourcing, parts of Eastern Europe for data centers) could invoke force majeure for extended periods while you continue paying. The 2024-2025 Taiwan earthquake disruptions triggered force majeure claims by multiple providers whose HBM supply chains ran through TSMC.
Negotiate three modifications to the standard force majeure clause. One, a time limit after which you can terminate with full refund of prepaid but unused fees (30-60 days is standard). Two, an exclusion for events the provider could have mitigated by maintaining inventory in multiple regions. Three, a force majeure suspension of payment obligations not just delivery obligations so you stop paying when the GPUs are unavailable.
Equipment Return and Decommissioning
For dedicated hardware deployments (colocation, bare metal leases), the equipment return clause determines what happens at termination. Standard terms require the customer to return servers in their original configuration within 30 days, with damaged or missing components charged at replacement cost plus 15% restocking fee. GPU damage in particular is a frequent dispute point when providers claim thermal stress or voltage irregularity caused wear beyond normal use.
Negotiate a pre-installation baseline inspection report with photographs and GPU benchmark results. The return clause should reference that baseline, not undefined normal wear. Also negotiate a minimum 90-day return window to allow for data sanitization (NIST 800-88 purge of GPU memory and SSDs), and cap replacement charges at the depreciated value shown on the provider's own balance sheet, not the MSRP of new hardware.