08 / CONTRACT ECONOMICS
Understanding a GPU service agreement
An uptime percentage becomes useful when you know what it measures, which failures count, and what remedy follows. Apply different terms to the same incident.
START WITH THE DENOMINATOR
A working API can hide unavailable GPUs.
A provider’s control platform may keep responding while individual machines are down. A platform-only definition can report 100% uptime through those failures. A definition based on delivered GPU-hours includes them.
This fictional contract uses illustrative credit tiers. The exercise explains their arithmetic; actual terms and remedies require review of the agreement.
ILLUSTRATIVE CONTRACT SCENARIO
What counts as unavailable?
1,024 GPUs · $2.35/GPU-hour
730-hour billing month
GPU-HOURS IN THE MEASUREMENT WINDOW
747,520
1,024 GPUs × 730 hours
Platform-only measurement
Individual GPU failures are excluded from this definition.
Availability definition
For example, 500 GPUs down for 10 hours = 5,000 unavailable GPU-hours.
Assumed cost of lost work
—
5,000 GPU-hours × $2.35 × 3
Service credit under these terms
—
0% × monthly fees
The loss multiplier is an assumption, not measured business damage. Credits are calculated from monthly fees, so they can exceed this modeled loss and may only offset future bills.
Contract & incident assumptionsRates, measurement window & exclusions
Uptime promise
Measured over
the reference deal — numerical exposures use these values unless a row is explicitly labeled illustrative
Illustrative model: 730-hour month, the reference deal's live rate and capacity, credits computed as a tier percentage of monthly fees, and node failures spread evenly across the window. Tier schedules are a representative composite. The downstream-cost multiplier is a sensitivity input, not an empirical estimate.
THREE TERMS TO EXAMINE
Review the contract clauses14 clauses · review & export
Open a clause to compare the draft with a proposed change. Mark your priorities, copy redlines, or export your review.
Illustrative negotiation aid—not legal advice. The agreement is a fictional composite. Terms vary by provider, jurisdiction, and deal; use counsel for an actual agreement. Dollar figures are scenario exposure, are not probability-adjusted, and should not be added together.
the draft says"commercially reasonable efforts" toward a target date
you ask for0.5% of the prepay per day late; walk with a full refund past 45 days
why it mattersThe draft states a target date without an explicit delivery credit or delay-based exit right.
the redline
Provider shall use commercially reasonable efforts to make make the Committed Capacity available by the target date. If Ready-for-Service is delayed beyond the target date, Provider will credit Customer 0.5% of the Initial Payment per day of delay; if the delay exceeds forty-five (45) days, Customer may terminate this Order Form with a full refund of the Initial Payment.
effect of the proposed changer1 — The proposed text ties a missed delivery date to a credit and an eventual exit right. Both the trigger and remedy need to be defined.
the model1.5 × monthly — a 45-day slip: committed spend with no service and no remedy
the draft saysReady for Service on Provider's own validation
you ask forburn-in + acceptance tests to your spec; billing starts at your sign-off
why it mattersThe draft lets the provider determine readiness and start billing using its own validation.
the redline
Capacity is Ready for Service upon Provider's completion of its standard validation procedures. completion of burn-in and acceptance testing — sustained multi-day load, interconnect at specified bandwidth, HBM and ECC error rates within specified thresholds — and Customer's written sign-off.
effect of the proposed changer2 — The proposed acceptance test connects billing to agreed performance criteria and customer sign-off.
the model1 × monthly — first month billed on a cluster that fails your spec
the draft sayscommitted volume may not be reduced, ever
you ask for±15% at months 9 and 18, at the Committed Rate
why it mattersThe draft fixes committed volume for the term. A later fall in demand does not reduce the payment obligation.
the redline
Committed volume may not be reduced during the Committed Term. Committed volume may be adjusted upward or downward by up to fifteen percent (15%) at months nine (9) and eighteen (18), at the Committed Rate (true-up / true-down).
effect of the proposed changer3 — The proposed checkpoints permit a limited reduction in committed volume. Their timing and size determine how much demand risk remains.
the model0.15 × monthly × max(0, months − 9) — demand falls 15% at the first checkpoint; draft makes you pay it anyway for the remaining term
the draft sayspower-cost increases pass through on 30 days' notice, uncapped
you ask forcapped at 5% per contract year, documentation auditable
why it mattersThe pass-through provision can increase the effective rate even when the quoted base rate is fixed.
the redline
The Committed Rate is fixed for the Committed Term, except that Provider may pass through documented increases in power and facility input costs on thirty (30) days' notice, not to exceed five percent (5%) in any contract year, with supporting documentation subject to Customer audit.
effect of the proposed changer4 — The proposed cap limits annual increases and adds access to supporting documentation. It still allows some price movement.
the model0.05 × 12 × monthly — one year of drift at the 5% cap you're asking for, as a floor on uncapped exposure
the draft saysInitial Payment non-refundable unless you terminate for cause
you ask forrefundable pro-rata on any termination
why it mattersRefund rights determine how much prepaid cash can be recovered when the agreement ends.
the redline
The Initial Payment is non-refundable except upon Customer's termination for cause refundable pro-rata upon any termination of this Order Form.
effect of the proposed changer5 — The proposed wording returns unused prepaid amounts on termination. Its effect depends on the refund calculation and the other exit provisions.
the modelprepay$ — cash forfeited under the draft on any termination that isn't your cause
the draft saysnot in the draft — no MFN anywhere in the paper
you ask forno comparable customer pays less; audit annually
why it mattersThe draft contains no protection against a lower price being offered to a comparable customer.
the redline
[what's not here — this agreement contains no most-favored-nation (MFN) protection.]
3.4 Provider represents that the effective rate hereunder is no less favorable than that offered to any customer of comparable volume and term, and Customer may audit compliance annually (most-favored-nation).
effect of the proposed changer6 — A comparable-customer clause needs a precise comparison group and a usable audit process. It may be difficult to verify without access to pricing records.
the modeloptionality — deliberately unnumbered
the draft says"Available" = the platform API answers
you ask foravailability measured per Committed GPU-hour, node-level
why it mattersA platform-only definition can omit individual machine failures while the provider API remains reachable. The outage example compares the two measurements.
the redline
4.1 "Monthly Uptime Percentage" means the percentage of minutes in a calendar month during which the Service is Available. The Service is "Available" when Customer is able to access the Provider platform API. a Committed GPU-hour is delivered by a node that is operational and performing to specification; a GPU-hour is Unavailable if its node is failed, degraded below specified interconnect bandwidth or error thresholds, or withdrawn from service. Monthly Uptime is computed over Committed GPU-hours.
effect of the proposed changer7 — The proposed wording changes the measured unit to delivered GPU-hours. The outage example shows how that choice changes which incidents can trigger credits.
the model(0.10+0.10+0.25+0.10) × monthly — four breach months at the published schedule; platform-level = $0
the draft saysmaintenance uncapped; node failures excluded by name
you ask formaintenance ≤8h/month on 7 days' notice; the node exclusion struck
why it mattersThe draft excludes individual node and component failures even when those failures reduce usable capacity.
the redline
4.2 Excluded from the calculation: scheduled maintenance, not to exceed eight (8) hours per month on seven (7) days' notice; events beyond Provider's reasonable control; failures of individual nodes, GPUs, or components where the platform remains accessible; and suspensions under Clause 6.
effect of the proposed changer8 — The proposed wording limits scheduled maintenance and removes the named exclusion for individual-node failures.
the model(16/730) × monthly × 12 — uncapped maintenance at 24h/mo vs the 8h ask, over a year
the draft saysnot in the draft
you ask forfailed nodes replaced within 6 business hours from a 2% hot-spare pool
why it mattersThe draft does not specify a replacement deadline or a pool of spares reserved for this cluster.
the redline
4.5 Provider will replace any failed node within six (6) business hours from a hot-spare pool maintained at no less than two percent (2%) of cluster size (time-to-replace).
effect of the proposed changer10 — A replacement deadline and reserved spares address restoration of capacity. Review the clock, coverage hours, and remedy for a missed deadline.
the model0.25 × monthly — a quarter-month of capacity protected across the term
the draft saysnot in the draft
you ask for3 breaches in 6 months → terminate, no cure, pro-rata refund
why it mattersCredits for individual months do not themselves create an exit right for repeated failures.
the redline
4.6 If Monthly Uptime breaches the Committed Level in any three (3) months within a rolling six (6) month window (chronic breach), Customer may terminate for cause without cure period, with pro-rata refund of the Initial Payment.
effect of the proposed changer11 — The proposed repeated-breach trigger adds an exit right for a pattern of failures. It depends on which incidents the availability definition counts.
the modeloptionality — deliberately unnumbered
the draft saysevery breach gets 30 days' notice and cure
you ask for15-day cure for availability; chronic breach skips cure
why it mattersThe time allowed to remedy a breach affects when other contractual remedies become available.
the redline
Either party may terminate for material breach not cured within thirty (30) days of written notice, except that chronic breach under Clause 4.6 requires no cure period and availability breaches carry a fifteen (15) day cure.
effect of the proposed changer14 — The proposed text shortens the availability cure period and treats repeated breaches separately.
the model1 × monthly — one month of exposure before any remedy matures
the draft saysliability capped at 3 months' fees; credits the sole and exclusive remedy
you ask for12-month cap; confidentiality, data protection, gross negligence carved out
why it mattersThe cap, sole-remedy wording, credit restrictions, and claim deadline can each limit recovery.
the redline
PROVIDER'S AGGREGATE LIABILITY UNDER THIS AGREEMENT IS CAPPED AT FEES PAID IN THE THREE (3) MONTHS TWELVE (12) MONTHS PRECEDING THE CLAIM, AND SERVICE CREDITS ARE CUSTOMER'S SOLE AND EXCLUSIVE REMEDY FOR ANY FAILURE OF AVAILABILITY. THE CAP DOES NOT APPLY TO BREACHES OF CONFIDENTIALITY OR DATA PROTECTION, OR TO PROVIDER'S GROSS NEGLIGENCE OR WILLFUL MISCONDUCT.
4.3 If Monthly Uptime falls below the Committed Level of %, Customer's sole and exclusive remedy is a Service Credit per the Schedule below, applied against future invoices only, provided Customer submits a written claim with supporting logs within thirty (30) days of the incident. Credits are calculated automatically from Provider telemetry, which Customer may audit on request.
effect of the proposed changer15 — This redline leaves the liability cap and sole-remedy language in place. Any wider cap, carve-out, or additional exit right requires separate consideration with counsel.
effect of the proposed changer9 — Automatic credits based on auditable telemetry reduce the claim-administration burden. The measurement and dispute process still need to be defined.
the model9 × monthly — recoverable delta on a catastrophic claim: 12-month ask minus 3-month draft cap
the draft saysyou may not assign; they may, to anyone
you ask formutual consent; exit if they're bought by your competitor
why it mattersThe draft gives the provider broader assignment rights than the customer.
the redline
Customer may not assign this Agreement without Provider's prior written consent. Provider may assign this Agreement to any affiliate or successor, including in connection with a merger, acquisition, or sale of assets. Neither party may assign this Agreement without the other's prior written consent, not to be unreasonably withheld. If Provider undergoes a change of control to a competitor of Customer, Customer may terminate on ninety (90) days' notice.
effect of the proposed changer16 — The proposed text makes assignment consent mutual and adds an exit right for a specified change of control.
the modeloptionality — deliberately unnumbered
the draft sayssupplier shortages and delays are force majeure
you ask forsupply-chain delay doesn't excuse Clause 1 delivery
why it mattersThe draft includes supplier shortages and delays among events that can excuse performance.
the redline
Neither party is liable for delay caused by events beyond its reasonable control, including acts of government, power or facility failures, and shortages or delays of suppliers or component manufacturers. Delays in Provider's supply chain do not excuse the delivery obligations of Clause 1.
effect of the proposed changer17 — The proposed wording prevents supplier delays from excusing the delivery obligation. Other force-majeure events remain subject to the clause.
the model1.5 × monthly — the same slip, pre-excused as force majeure
Interpretation & limitationsLimits & negotiation notes
- Decision supported
- Which contract issues deserve negotiation priority?
- Primary output
- Scenario exposure, proposed position, and redline
- Model boundary
- Composite agreement; exposure is not expected value or additive savings
The definition of "unavailable" determines which incidents count. Platform- or region-level definitions can exclude node, GPU, memory, and network-component failures while the API remains reachable. Meta's Llama 3 team reported 419 unexpected interruptions during 54 days of training on 16,384 H100s, with hardware failures occurring at the node level; Modal has also described recurring machine-level faults across a multi-cloud GPU pool. Open the definition of "available" row and compare platform-level with instance-level measurement in the outage experiment.
The measurement window changes the threshold result. A one-hour outage is approximately 99.86% availability in a 730-hour month, but approximately 99.989% over a year. Specify the window, data source, aggregation level, and claim mechanics rather than relying on the headline percentage alone.
Credit recovery depends on the measured failure. Use the outage experiment: a total platform outage may trigger a tier, while recurring node-level failures may not register under a platform-only definition. Credits are commonly applied to future fees rather than paid as cash, which can reduce their value when the relationship is ending.
Operational commitments complement service credits. Time-to-replace, dedicated spares, burn-in and acceptance testing, telemetry access, and termination rights address failure response and recurrence directly. Review them alongside the availability percentage and credit schedule.
sources
- Meta, "The Llama 3 Herd of Models" (2024) → arxiv.org/abs/2407.21783
- Modal, "Keeping 20,000 GPUs healthy" (Jan 2026) → modal.com/blog/gpu-health
- tier schedules: representative composite of published cloud compute SLAs, not any single contract.
- AWS, "Amazon Compute Service Level Agreement" → aws.amazon.com/compute/sla
- CoreWeave, "Terms of Service" — canonical platform legal terms → docs.coreweave.com/docs/policies/terms-of-service
- Lambda, "Terms of Service" — cloud terms (credits: no cash value, offset future fees only) → lambda.ai/legal/terms-of-service
- CoreWeave, Form 10-K, fiscal 2025 (filed Mar 2026) — take-or-pay committed contracts and initial-payment structure, SEC EDGAR → sec.gov/Archives/edgar/data/1769628/000176962826000104
- contract language: original composite drafted for this desk — structurally patterned on published GPU-cloud SLAs and reported committed-deal structures; no clause reproduces any real agreement.
- $ column: modeled figures are computed from the reference deal with the formula shown in each row; illustrative figures are stated scenarios; optionality is deliberately unnumbered.
- all parties, dollar figures, and dates in the agreement are fictional and illustrative.
illustrative composite · model version 2026.08 · methods, legal limitation & sources