Every number needs a receipt.
Our comparison is designed around reproducibility: preserve the original rate, attach a source, then normalize only the fields needed to make offers comparable.
Collection
Production collectors will pull public pricing APIs where providers expose them and use published pricing pages for the remainder. Each record stores its source URL, observation time, source currency, billing unit, and raw configuration name.
Normalization
Hourly node prices are divided by the physical GPU count to produce a per-GPU hourly rate. Monthly commitments remain labeled as reserved; spot rates never compete silently with on-demand rates.
per_gpu_hour = source_hourly_total / physical_gpu_countAvailability
Providers expose capacity differently. Production records will retain the freshest reported state and mark it available, limited, unavailable, or unknown. Stale availability will be downgraded rather than presented as live inventory.
Known limits
List prices can exclude tax, disks, data transfer, support, minimum terms, or negotiated discounts. Two nodes with the same GPU can differ significantly in CPU, memory, interconnect, storage, and delivered performance.