AI DeploySelf-hosted LLMGet it deployed

Negotiate a GPU Hosting Contract for Your LLM

Prepare workload evidence and negotiate GPU capacity, purchase duration, service levels, charges, support, renewal, and exit terms for an LLM deployment.

For your LLM, negotiate an llm hosting contract that converts workload evidence into written commitments for GPU capacity, purchase duration, availability, extra charges, support, renewal, and exit. Treat public product pages as background; place every negotiated term in the applicable quote or order form, SLA, and final agreement. Do not accept phrases such as “GPU access” or “support included” until the operative document defines the metric, period, exceptions, owner, and enforceable remedy.

Build the evidence your LLM negotiation needs

Start with a record the provider can verify: model and configuration, GPU type and memory, operating windows, observed utilization, failed launches, storage and egress, support incidents, and expected growth. Label observed values separately from forecasts; if no history exists, call the estimate a forecast and state its assumptions.

MODEL_AND_CONFIGURATION = YOUR_VALUE
GPU_TYPE_AND_MEMORY = YOUR_VALUE
OPERATING_WINDOWS = YOUR_VALUE
OBSERVED_UTILIZATION = YOUR_VALUE
FAILED_LAUNCHES = YOUR_VALUE
STORAGE_AND_EGRESS = YOUR_VALUE
SUPPORT_CASES = YOUR_VALUE
GROWTH_ASSUMPTION = YOUR_VALUE

Attach the source records to the worksheet. Ask the provider to fill gaps in writing rather than replacing missing evidence with an unsupported promise.

Decide what your llm hosting agreement must guarantee

Translate the evidence into a service schedule that states:

  • the exact GPU model, count, memory, and whether resources are dedicated, partitioned, or shared;
  • the Region, Availability Zone, capacity start, capacity end, and commitment duration;
  • the provider’s responsibilities for compute, network access, storage, and data handling;
  • notice and communication duties for maintenance, faults, capacity loss, and interruption;
  • support intake, escalation, and the evidence required to invoke any remedy.

For each item, name the controlling document and precedence order. Present unresolved fields as proposed contract language, not verbal requests. If the quote and SLA conflict, do not sign until the final agreement resolves the conflict.

Make the GPU specification an exact entitlement

For a physical or bare-metal addendum, NVIDIA’s GeForce RTX 4090 specifications list 24 GB of GDDR6X memory, 450 W total graphics power, an 850 W required system power, and a 304 mm by 137 mm three-slot card for the GeForce RTX 4090 Founders Edition. Those figures describe that board; they do not define a rented GPU entitlement.

Require the hosting schedule to state:

  • the provider’s exact SKU, memory, and GPU count;
  • whether the advertised memory or compute is dedicated, partitioned, or shared;
  • replacement rules for faulty or unavailable hardware;
  • the acceptance test, measurement method, and remedy for failed delivery.

Do not accept only a GPU family or marketing label when your workload depends on a precise accelerator configuration.

Match the purchase mechanism to commitment and interruption risk

AWS’s billing and purchasing options describe On-Demand Instances as billed by the second; Savings Plans as a consistent usage commitment measured in USD per hour for a 1- or 3-year term; Reserved Instances as a commitment to a specific instance type and Region for a 1- or 3-year term; and Capacity Blocks as a way to reserve a cluster of GPU instances.

For interruptible capacity, AWS says a Spot Instance interruption notice is a best-effort warning issued two minutes before EC2 stops or terminates the instance, and it recommends checking every 5 seconds. Its Spot Instance interruption notices document this IMDSv2 path:

http://169.254.169.254/latest/meta-data/spot/instance-action

Require the contract to define checkpointing, draining, recovery, retained storage, and charges during interruption. Google Compute Engine’s Spot VM documentation says Compute Engine can preempt Spot VMs at any time, provides a best-effort shutdown period of up to 30 seconds, and excludes Spot VMs from the Compute Engine SLA. Treat those terms as provider-specific rather than portable guarantees.

For every commitment, state its start and end, cancellation rights, renewal mechanism, unused-capacity treatment, and consequences after expiry. Do not infer these terms from the product name.

Build a total-cost worksheet

Ask for an itemized quote and map each charge to the usage or event that triggers it. Fill this calculation from the quote or bill, not from an assumed rate:

HOURLY_RATE = YOUR_VALUE
BILLED_HOURS = YOUR_VALUE
STORAGE_RATE = YOUR_VALUE
STORED_AMOUNT = YOUR_VALUE
DATA_RATE = YOUR_VALUE
DATA_AMOUNT = YOUR_VALUE
TRANSFER_RATE = YOUR_VALUE
TRANSFER_AMOUNT = YOUR_VALUE
PROVIDER_QUOTED_AMOUNT = YOUR_VALUE
OTHER_QUOTED_LINE_ITEMS = YOUR_VALUE

COMPUTE = HOURLY_RATE × BILLED_HOURS
STORAGE = STORAGE_RATE × STORED_AMOUNT
NETWORK = DATA_RATE × DATA_AMOUNT + TRANSFER_RATE × TRANSFER_AMOUNT
TOTAL = COMPUTE + STORAGE + NETWORK + PROVIDER_QUOTED_AMOUNT + OTHER_QUOTED_LINE_ITEMS

AWS instance-state billing documentation says a stopped instance is not billed for instance usage, but EBS volumes and Elastic IP addresses can incur charges regardless of instance state. It also says each stop-to-running transition starts a new billing period with a one-minute minimum, and an instance is billed while in the stopping state when hibernated.

Request separate lines, where offered, for compute, commitments, storage, public addresses, snapshots, gateways, load balancers, data transfer, support, monitoring, taxes, and every startup or shutdown action. Mark each line as included, variable, one-time, or not offered. Recalculate the worksheet for an idle period, repeated stop-and-start cycles, an interruption, additional storage, and higher traffic; a low base rate alone does not show total cost.

Put availability and remedies in the SLA

Under the Amazon Compute SLA, checked on 2026-10-01, AWS separates Region-Level and Instance-Level commitments. The Region-Level commitment targets at least 99.99% monthly uptime when all running instances are spread across at least two AZs in the same Region, while the Instance-Level commitment is 99.5% for a single instance; at that level, “unavailable” means no external connectivity.

The same credit and remedy provisions say AWS will not charge for a Single EC2 Instance unavailable for more than six minutes in a clock hour, without a claim; service credits are a percentage of the monthly bill excluding one-time Reserved Instance upfront payments; a credit is issued only when it exceeds one US dollar; and the request must reach AWS Support by the end of the second billing cycle after the incident. Credits apply to future EC2 payments rather than a refund, and Region-Level and Instance-Level claims cannot be stacked. The exclusions include causes outside AWS’s control, customer actions or inaction, customer equipment or software, and suspension under the agreement; service credits are the sole and exclusive remedy.

Check that your architecture and chosen purchase option match the SLA conditions. State how downtime is measured, what evidence you must provide, who receives the claim, and whether any separately negotiated remedy supplements or replaces the SLA.

Write support, renewal, and exit clauses

Support, renewal, and exit terms should be operative clauses, not sales notes:

  • Support: define severity levels, intake channel, response and restoration targets, clock start, escalation path, and coverage outside business hours.
  • Renewal: state the notice date, price-adjustment process, automatic renewal or non-renewal, commitment expiry, early termination, and any credit or refund treatment.
  • Exit: define export of model artifacts and data, deletion deadlines, storage and egress after termination, and transition assistance.
  • Change control: require notice and consent for changes to GPU SKU, Region, Availability Zone, capacity, or service architecture.

Attach named contacts and response procedures. If a term remains “best effort,” “included,” or “available on request,” specify its objective boundary and the consequence of missing it.

Verify the agreement before signature

Match every schedule field across the quote, order form, SLA, support exhibit, and final agreement. Confirm that no placeholder remains, legal entities and dates are complete, totals use the same units, and document precedence is clear.

Common failure points

  • A GPU family is named, but the exact SKU, memory, count, and allocation method are missing.
  • The comparison uses only the compute rate and omits storage, addresses, transfer, or commitment charges.
  • A commitment lacks a clear start, end, cancellation path, renewal rule, or expiry outcome.
  • An SLA percentage is quoted without its architecture conditions, exclusions, claim process, or remedy.
  • Interruptible capacity has no checkpoint, recovery, retained-storage, or interruption-charge plan.
  • A verbal assurance is missing from the documents that the agreement says control.

The steps are compiled from the official documentation linked in the text and should be checked against the current version before use. Nothing was run or measured for this article. This is general information and not legal advice.

Sources