Committed Term Rates
The price of guaranteed capacity by tenor, for the chips lenders finance. Contracted rates, built under our method from deals in public filings, shown against posted term rates.
Contracted Rates by Vintage
Notes on the chart
Every rate is restated to a three-year term, so contracts of different lengths share one scale. The restatement assumes rents fall 10% a year, so a dot can differ from its table cell. Money costs 10% a year, the same rate that puts prepaid contracts on a pay-as-you-go basis. Rates are GPU-only and pay-as-you-go, as in the table. Trend lines are fitted for 8-GPU servers, weighted toward larger and more recent contracts. Vera Rubin and the rack-scale systems have too few contracts for a trend.
| Silicon · tenor | 2023 H2 | 2024 H1 | 2024 H2 | 2025 H1 | 2025 H2 | 2026 H1 | 2026 H2 |
|---|---|---|---|---|---|---|---|
| H100 | |||||||
| 2Y | · | 2.87 | · | · | · | 1.40 | · |
| 3Y | 2.79 | 2.51b | · | · | · | · | · |
| H200 | |||||||
| 6M | · | · | 2.28 | 2.14 | · | · | · |
| 1Y | · | · | · | 2.14 | · | 1.67 | · |
| 2Y | · | · | 1.92 | 2.00 | · | · | · |
| 3Y | · | · | · | 2.20 | 1.74 | · | · |
| B200 | |||||||
| 1Y | · | · | · | · | 2.68 | · | · |
| 2Y | · | · | · | 3.75 | · | 3.12 | · |
| 3Y | · | · | · | · | · | 3.07 | · |
| B300 | |||||||
| 2Y | · | · | · | · | 3.11b | 4.12 | 4.19 |
| 3Y | · | · | · | · | 3.39 | 3.72b | 4.66 |
| 4Y | · | · | · | · | · | · | 4.18b |
| 5Y | · | · | · | · | · | 3.30 | 3.83 |
| Vera Rubin | |||||||
| 3Y | · | · | · | · | · | · | 7.15 |
| 6Y | · | · | · | · | · | · | 5.28 |
| System · tenor | 2025 H2 | 2026 H1 | 2026 H2 |
|---|---|---|---|
| GB200 NVL72 | |||
| 6M | · | 12.2c | · |
| 1Y | · | 13.8c | · |
| 3Y | · | 13.0 | · |
| GB300 NVL72 | |||
| 6M firm · cancellable | · | 50.2 | · |
| 5Y+ | · | · | 15.5f |
| 5Y+ · hyperscaler anchor | 10.1c | · | · |
| 5Y+ · with Vera Rubin · hyperscaler anchor | 11.7 | · | · |
Notes on the rack contracts
Rack-scale contracts state value, term and megawatts, and rarely a GPU count, so they are priced per megawatt of IT load per year. Prices are all-in: GPUs, network, power and site. A megawatt price does not convert to a GPU-hour price without a GPU count, so only contracts with a known count appear in the chart above. Hover for the contracts. Hyperscaler anchor rows are whole-campus, multi-year offtakes, often prepaid, and sit on their own line. c: holds a deal filed per GPU-hour with a GPU count, converted at 440 GPUs per IT MW (measured range 360 to 510, so ±15%). f: the filed value is a lower bound. Cancellable: the tenor is the firm term to the earliest exit; either party may end the contract on notice after that. With Vera Rubin: the contract covers GB300 and Vera Rubin and does not split them. Listed, not in the cells: SharonAI / NVIDIA ($11.3M, GB300, revenue share with credit support, not a plain rental); Firmus / NVIDIA ($12.7M, Grace Blackwell and Vera Rubin, press).
B300 deal by deal, with sellers, buyers, trends and filing links: Signed B300 contract rates, March to August 2026 →
The Term Ladder
| Silicon | On-demand | 3M | 6M | 1Y | 2Y | 3Y | 5Y |
|---|---|---|---|---|---|---|---|
B300 | 8.39 | — | — | 6.61 | 6.50 | — | — |
| 6.59 | 6.39 | 5.96 | 6.33 | — | — | — | |
| 4.50 | 4.32 | 4.15 | 4.11 | 3.68 | 3.10 | — | |
| 3.52 | 3.36 | 3.18 | 3.11 | 2.64 | 1.83 | — |
How the ladder is built
Every cell is the chip's on-demand rate times the chip's own tenor factor. On-demand is the median across neocloud sellers in every region, each seller counted once. A tenor factor compares each seller's term ask with its own on-demand ask in the same region. The factor is the median of those ratios across sellers. A chip never borrows another chip's factor. A longer tenor can price above a shorter one when sellers post it that way.
Hyperscaler reservation schedules stay out of the factors.
The daily record also carries direct posted term asks. We use them as a check. Each shows the gap to the ladder value: B300 1Y 7.21 (+9%) · B300 2Y 6.71 (+3%) · B200 3M 7.64 (+19%) · B200 6M 6.09 (+2%) · B200 1Y 6.02 (−5%) · H200 3M 4.15 (−4%) · H200 6M 3.64 (−12%) · H200 1Y 3.40 (−17%) · H200 2Y 3.58 (−3%) · H100 3M 2.73 (−19%) · H100 6M 2.54 (−20%) · H100 1Y 2.44 (−22%) · H100 2Y 2.73 (+4%) · H100 3Y 2.24 (+22%).
Posted vs Contracted Rates
| Silicon · layer | On-demand | 3M | 6M | 1Y | 2Y | 3Y | 5Y |
|---|---|---|---|---|---|---|---|
| B300 Posted | 8.39 | — | — | 6.61 | 6.50 | — | — |
| Contracted | — | — | — | — | 4.10b | 4.10b | 3.74 |
| Gap | — | — | — | — | −37% | — | — |
| B200 Posted | 6.59 | 6.39 | 5.96 | 6.33 | — | — | — |
| Contracted | — | — | — | — | 3.29 | 3.07 | — |
| Gap | — | — | — | — | — | — | — |
| H200 Posted | 4.50 | 4.32 | 4.15 | 4.11 | 3.68 | 3.10 | — |
| Contracted | — | — | — | 1.67 | — | — | — |
| Gap | — | — | — | −59% | — | — | — |
| H100 Posted | 3.52 | 3.36 | 3.18 | 3.11 | 2.64 | 1.83 | — |
| Contracted | — | — | — | — | 1.40 | — | — |
| Gap | — | — | — | — | −47% | — | — |
How to read the contracted rates
Each contracted rate is our construction from filed contracts at its tenor. A filing states a rate, or a unit count, a term and a value. We divide the value by GPUs times term hours and bill every hour at a flat price. We leave out deals priced under a government subsidy.
Each cell holds the filed deals at its tenor from the last twelve months. A deal counts fully when signed, and its weight falls in a straight line to zero at twelve months, so it never drops out on a set day. Weight also rises with the fourth root of the GPU count. A cell shows only while its deals carry the weight of at least one deal about nine months old. The vintage table in Section 01 shows every deal.
B300 2Y: 4.10 weighted, range 3.11 to 4.19. Bundled inside: 3.11 estimated GPU-only (filed 3.15, includes managed services, 20% prepaid). Prepaid inside: 4.12 pay-as-you-go (filed 4.01, 21% prepaid); 4.19 pay-as-you-go (filed 4.08, 20% prepaid). B300 3Y: 4.10 weighted, range 3.39 to 4.76. Bundled inside: 3.40 estimated GPU-only (filed 3.59, includes storage, CPU); 4.13 estimated GPU-only (filed 4.29, includes storage). Prepaid inside: 3.61 pay-as-you-go (filed 3.45, 30% prepaid); 3.81 pay-as-you-go (filed 3.70, 17% prepaid). B300 5Y: 3.74 weighted, range 3.30 to 4.16. B200 2Y: 3.29 weighted, range 2.69 to 3.74. Prepaid inside: 3.74 pay-as-you-go (filed 3.65, 23% prepaid). B200 3Y: 3.07. Prepaid inside: 3.07 pay-as-you-go (filed 2.94, 25% prepaid). H200 1Y: 1.67 weighted, range 1.59 to 1.75. Prepaid inside: 1.75 pay-as-you-go (filed 1.65, 100% prepaid). H100 2Y: 1.40.
B300 4Y: 4.18. Bundled inside: 4.18 estimated GPU-only (filed 4.13, includes CPU servers, storage, 25% prepaid).
Renewal options, listed and never pooled: B300 year-4 option: 2.13 · B300 year-5 option: 1.75.
Filed, listed and kept out of the cells: H100 3Y 2.66 (superseded by the 2,048-GPU contract) · H100 3Y 2.32 (term sheet, superseded by the executed contract) · B200 3Y 3.09 (terms under renegotiation) · B200 3Y 3.04 (superseded by a later contract) · B200 3Y 4.26 (chip not disclosed).
Every rate is a hosted GPU-hour. Data center, power and networking stay in, as GPU clouds sell them. A deal that bundles only those enters at its filed rate.
A bundled deal also prices storage, CPU servers or managed services, with no split. It enters at a GPU-only estimate. For three B300 deals we price the CPU servers and storage out one by one. Every other bundled deal enters at its filed rate less 5%, the size of those three deductions. The filed value appears on hover.
Paying upfront buys a lower hourly rate. A contract with a filed prepayment enters at its pay-as-you-go rate, at a 10% cost of money. The upfront payment is credited evenly over the term unless the filing gives a schedule. The filed rate appears on hover.
SharonAI states its B300 rate inclusive of taxes. We treat the $3.30 as pre-tax, as the B300 note does.
Disclosure selection is a limit. The sellers that disclose unit counts are small listed companies.
Per-chip pages: H100 · H200 · B200 · A100 · all chips →
Compute Credit Index Research, GPU Committed Term Rates, https://ccir.io/term, retrieved 2026-10-15.