GPU ECONOMICS & INFRASTRUCTUREENGLISH · GLOBAL EDITION
TensorCurve

Understanding the cost of compute.

Forward markets

A forward-start GPU quote is not a price forecast

Separate an offer available today from a prediction about what future capacity might cost.

Source-backed editorial draft. Public inputs and calculations are labeled; availability is unverified.

Two prices that answer different questions

A forward-start quote answers: what terms is a supplier offering now for capacity that begins later? A forecast answers: what price do we expect to observe in the future? The start date may be identical, but the evidence behind the two numbers is different.

Suppose a supplier gives a quote today for eight GPUs starting in three months. That quote may have a validity window, capacity conditions, payment requirements and a cancellation policy. A model’s estimate for the same future date has none of those contractual properties unless a supplier separately makes an offer.

Use labels that preserve the distinction

Label What it represents
Published rate A price publicly displayed by a provider, subject to its terms.
Provider quote A dated offer for specified capacity and conditions.
Model estimate A calculated value, including interpolation of missing observations.
Price forecast An expectation about a future market observation.
Synthetic example Invented data used to explain a concept or test an interface.

A published rate is not necessarily an executable quote. Availability and eligibility may still need confirmation. Similarly, a line connecting observations on a chart should not imply that intermediate contract dates can be purchased.

Future capacity is a product-specific question

AWS Capacity Blocks documentation provides an example of reserving GPU-based capacity for a future date. It also describes product-specific restrictions. This supports the distinction between a real future-capacity product and a forecast; it does not establish that every GPU provider supports the same dates, durations or cancellation terms.

When collecting data, retain the provider’s wording and the time at which it was observed. Record whether the quote is indicative, subject to availability or a firm offer. If the source does not say, leave that classification unresolved rather than upgrading an indicative figure to a guaranteed price.

Interpret the forward-start difference carefully

A higher quote for a later start could reflect a different product configuration, a change in supplier terms or the value of securing capacity for that interval. Price differences alone do not identify the cause. Before interpreting a premium, compare the same GPU count, region, tenancy, duration and payment basis.

Nor should a quote be treated as an unbiased prediction of future on-demand prices. A rental contract includes access and obligations. A forecast describes an expected observation. Those objects can carry different costs and risks.

How TensorCurve handles the boundary

The Pricing Lab now separates a public discount schedule from an actual forward-start offer. Its base row contains rates calculated from published inputs. Future-start rows show “Not published in this dataset.” There is no assumed premium, scarcity multiplier or forecast hidden behind those empty cells.

The empty rows describe our evidence, not the full commercial capabilities of the supplier. A sales team might provide a quote that is absent from the public table. That quote would need its own source, observation time, configuration, validity and permission to publish before joining the dataset.

Neither an on-demand price nor a discount schedule proves that capacity will be available when a project is ready. Keep the commercial availability check separate from the arithmetic. A missing quote is more useful than a smooth curve whose future prices have been invented.