Define the unit of capacity
Begin with the exact hardware configuration rather than a broad product family. Record GPU model, memory specification where verified, GPU count and whether resources are dedicated or shared. Also note whether the price covers a complete instance or one accelerator.
Suppose an eight-GPU instance is quoted at $24 per instance-hour. The arithmetic equivalent is $3 per GPU-hour. That normalization does not mean the supplier sells a single GPU for $3, or that all bundled CPU, memory and networking resources have been separately priced. Preserve both the original unit and the converted unit.
Compare configurations before ranking prices
A concrete H100 comparison shows why the product label alone is insufficient. Hyperstack’s public table lists H100 SXM 80GB at $3.20 per GPU-hour, with maximum allocations of 24 physical CPUs and 240GB RAM per GPU. Verda lists a one-GPU H100 SXM5 80GB instance at $3.25, with 30 CPUs and 120GB RAM. These inputs were checked on 13 September 2026.
The five-cent hourly difference does not establish a cost-per-job winner. CPU definitions, actual allocations, storage, region, networking and workload performance still need alignment. In particular, a maximum allocation is not a promise that every purchasable configuration includes it.
We therefore do not combine these products into a market median or draw a line between them as if they were different terms of one contract. A supplier comparison and a duration curve answer different questions.
Preserve the commercial terms
Keep the start date, duration, minimum spend, payment timing, renewal conditions and cancellation rights next to the rate. Capacity assurance and interruption conditions should be explicit. Unknown terms should remain unknown.
An upfront commitment and a monthly bill may show similar normalized hourly costs while imposing different cash-flow requirements. A comparison can show both, but it should avoid treating them as economically identical without an explicit financing assumption.
Observation time · Source · Provider · Region · Hardware · GPU count · Tenancy · Start date · Duration · Original price unit · Normalized rate · Payment terms · Capacity conditions · Data classification.
Separate observation from interpretation
A provider page is evidence of what was published when it was observed. It is not evidence that a transaction occurred at that price. Keep source links, timestamps and quote validity information so that another reader can check the record’s meaning.
Comparative statistics need a defined sample. A median across four selected offers is the median of those offers, not a global market benchmark. If some providers have no comparable contract, report their absence rather than substituting an estimate into the observed sample.
Publish the limits alongside the result
Explain taxes, storage, network transfer and support exclusions. State the exchange-rate source if currencies are converted. Show how monthly prices become hourly prices and whether the denominator is calendar hours, an assumed 30-day month or productive usage hours.
The Pricing Lab uses a dated public-price snapshot and a normalized 720-hour month for planning totals. Region and future-start availability remain unspecified. The displayed total is a calculation, not a supplier-issued invoice. Recheck the source before using the figures in a purchasing decision.
The strongest comparison is one whose exclusions are easy to find. Readers should be able to identify what is comparable, what differs and what still needs confirmation before making a procurement decision.