The published fee schedule.
Every number below is buyer-side, versioned, and served from the same frozen snapshot the billing engine refuses to boot without. What is not listed here is stated, not implied.
| GPU tier | $/day | $/hr | Price floor $/day |
|---|---|---|---|
| GB300 | $499.35 | ~$20.806 | $240.00 |
| B300 | $431.75 | ~$17.99 | $200.00 |
| B200 | $321.10 | ~$13.379 | $165.00 |
| H200 | $184.35 | ~$7.6812 | $100.00 |
| H100 | $132.90 | ~$5.5375 | $56.10 |
| A100 | $90.00 | $3.75 | $40.50 |
| L40S | $60.00 | $2.5 | $23.70 |
| RTX 4090 | $14.00 | ~$0.58333 | $10.20 |
| RTX 3090 | $9.00 | $0.375 | $9.00 |
| CONSUMER | $7.00 | ~$0.29167 | $6.00 |
The price floor is the rate below which no operator may sell that tier. ON_DEMAND settles per minute when a rental stops early, with a 1-day minimum (1-hour on hourly rentals).
INFERENCE workloads on datacenter tiers are billed at 20% below the on-demand rate.
Consumer tiers (CONSUMER, RTX_4090, RTX_3090) are already inference-priced; the discount never applies to them.
Applied to per-model token rates for inference, by the serving operator's standing tier. Standing is earned on track record, not on hardware: an operator reaches the middle tier on account age and served request volume, and the top tier is assigned by the platform for confidential-compute capacity. None of these tiers certifies which GPU an operator runs. Per-model base rates are served by the models catalog, not this schedule.
- OTHER
- Operator-declared custom rates with no canonical retail or floor; excluded from every published table.
- SPOT
- No published price. Register row V-0243 (open) records that the documented scheduling preference is not implemented; the preemption half is real. A SPOT line returns only when that row closes honour-or-withdraw.
- RESERVED
- No published price. Per the C4-b acceptance, a reserved contract is a paid claim on operator uptime: no capacity guarantee or fence is sold. The tier remains selectable at checkout without a published schedule price.
- x402
- x402 amounts are payment ceilings disclosed per-call in the 402 challenge; they are not prices and are not scheduled.
- operatorPayout
- Every number here is buyer-side. Operator crediting is a separate internal calculation; this schedule makes no payout claim.
- modelBaseRates
- Per-model token rates are served by /v1/models from ModelPricing; this schedule versions the multiplier regime only.
- agentLane
- Agent runtime pricing is dark and unpublished.
retailPerDay is canonical; retailPerHour = retailPerDay / 24, unrounded. Serving surfaces apply their own declared rounding and must say which.
This page displays $/hr at 5 significant figures. A value marked ~ is a display rounding: its exact quotient has more digits than 5 significant figures can carry, and the canonical number is always $/day. Raw, unrounded values: /v1/public/fee-schedule.
Delivery is covered by the SLA credit policy: the availability target, the automatic credit mechanism, and the measurement basis.