Economics

Compute Cost Index

Cloud rate cards joined to Epoch AI ML-hardware throughput specs at SKU-region-pricing-tier grain (dense BF16 TFLOP/s; no 2:4 structured sparsity factor), normalised cell-by-cell to $/petaFLOP-day across the on-demand, spot, reserved-1yr, and reserved-3yr tiers. Refresh cadence is split: AWS, Azure, and OCI update daily; GCP, CoreWeave, and Lambda have no fetchable public pricing API and carry a manually refreshed, Playwright-verified snapshot — every row's own verification date renders in the table and serializes in exports. A parallel on-prem TCO buildup runs at 3-year and 5-year horizons; the gap to the cheapest cloud cell is the build-vs-buy boundary. Facility-level capacity estimates live at FLOP Engine.Hourly cloud pricing across AWS, Azure, GCP, Oracle, CoreWeave, and Lambda normalised to a single denominator ($/petaFLOP-day) that weights each provider's hourly rate by the GPU's actual throughput rather than its sticker price. AWS, Azure, and Oracle prices refresh daily; GCP, CoreWeave, and Lambda are verified manually, and every price on the page shows the date it was last checked. On-prem TCO at 3-year and 5-year horizons runs alongside; the gap between the two is the build-vs-buy boundary. The same denominator feeds frontier training cost estimates downstream.