A quote needs more than an hourly rate
Imagine two offers for the same GPU configuration. Both say “three months” and both show a dollar-per-hour rate. One starts today; the other starts in December. These offers solve different scheduling problems. The first provides capacity for immediate work. The second is relevant to a project whose data, model or team will only be ready later.
Without an explicit start date, a comparison can mix currently usable capacity with future availability. Without a duration, it can mix a flexible rental with an obligation that continues after the workload ends. Neither error is fixed by converting every number into dollars per GPU-hour.
Separate observation, start and duration
A price record should include three time fields. The observation time tells you when the offer was seen. The start date tells you when access begins. The duration tells you the contracted period of access or payment. An expiry time, if the supplier provides one, tells you how long the quote remains valid.
| Illustrative record | Start | Duration | Observation |
|---|---|---|---|
| Immediate rental | 13 Sep 2026 | 3 months | 13 Sep 2026 |
| Future-start rental | 13 Dec 2026 | 3 months | 13 Sep 2026 |
Both records were observed on the same day. Their start dates differ. A later quote observed in December would be another record, even if its start date and duration matched the second row.
Read a published commitment curve
A term curve changes contract duration while holding the product constant. A forward-start curve instead changes when access begins. The homepage uses Verda’s H100 SXM5 80GB instance with 30 CPUs and 120GB RAM. Its published on-demand base is $3.25 per hour for one GPU.
The supplier lists discounts of 2%, 3%, 4% and 8% for commitments of one, three, six and twelve months. Applying those discounts gives $3.1850, $3.1525, $3.1200 and $2.9900 per GPU-hour. TensorCurve calculates these values; the source publishes the base and discount schedule. The chart uses equal spacing for duration categories, not a continuous calendar axis.
Source checked 13 September 2026: Verda official pricing. Calculations use the published on-demand rate and reserved discounts; they are not separately quoted transaction prices.
This is one supplier’s tariff-derived commitment curve. It does not measure the global GPU market. The pricing table does not establish a region-specific offer or a guaranteed start date. Those fields remain unverified. A price for a twelve-month commitment does not tell us what a twelve-month contract beginning next December would cost.
Do not confuse a discount with capacity assurance
Payment commitment and capacity reservation also need separate fields. A lower effective rate does not, by itself, establish that a particular machine will be ready for a particular workload on a particular day.
For a concrete product example, AWS describes Capacity Blocks for ML as a way to reserve accelerated instances for a future date. Separately, AWS Savings Plans exchange a usage commitment for a different pricing arrangement. These documents describe different commercial mechanisms; their terms should be checked individually rather than grouped under one generic “reserved GPU” label.
What to record before comparing
Keep the original quote beside its normalized rate. Record the GPU model and count, region, tenancy, start date, duration, payment schedule and cancellation conditions. If a field is missing, mark it unknown instead of filling it with an industry assumption.
Then ask a practical question: can both offers serve the same workload over the same period under comparable terms? If the answer is no, keep them in separate comparison groups. A smaller set of genuinely comparable records is more informative than a large table that conceals contractual differences.