The raw materials of AI

What a unit of compute, a unit of memory and a unit of energy actually cost — split into CAPEX and OPEX, tracked from public sources.

There is no single price for "AI". There is a price for a TFLOP, for a gigabyte of memory, for a GPU-hour and for a megawatt-hour — and those four move for different reasons. This dashboard tracks each of them separately, from sources anyone can audit.

GPU-hour, H100

Awaiting the first run with the ComputePrices key configured.

Compute, BF16

Awaiting the first run with the ComputePrices key configured.

Henry Hub gas

Awaiting the first run with the EIA key configured.

US industrial power

Awaiting the first run with the EIA key configured.

Market — what it costs today

Daily prices. A GPU-hour and a gas contract are the two AI inputs with a real, observable, moving market price.

GPU rental price

Median on-demand quote across cloud marketplaces, USD per GPU-hour

Awaiting the first run with the ComputePrices key configured.

ComputePrices.com

Cost of compute

USD per PFLOP-hour — the rental price divided by the accelerator's dense BF16 throughput, which is what makes a consumer card and a datacentre module comparable

Awaiting the first run with the ComputePrices key configured.

ComputePrices.com × Epoch AI

Natural gas, Henry Hub spot

USD per million BTU, daily — the marginal fuel that sets the marginal power price in most US markets

Awaiting the first run with the EIA key configured.

U.S. Energy Information Administration

History — the long curve

The secular decline in the capital cost of compute, and the twenty-year record of the energy that runs it.

Capital cost of compute at launch

USD per TFLOP/s of dense BF16 throughput, at each accelerator's release price

No data yet — the scheduled pipeline has not produced this series.

Epoch AI — Machine Learning Hardware (CC BY)

Capital cost of compute at launch, FP32

USD per TFLOP/s of FP32 single precision, at each accelerator's release price — the older, non-tensor measure, which is the only one the pre-2017 hardware reports and so the only one that reaches back that far

No data yet — the scheduled pipeline has not produced this series.

Never compared against the BF16 chart above: the two measure different silicon paths, and a ratio between them is meaningless. · Epoch AI — Machine Learning Hardware (CC BY)

Capital cost of memory at launch

USD per GB of on-package memory, at each accelerator's release price

No data yet — the scheduled pipeline has not produced this series.

Epoch AI — Machine Learning Hardware (CC BY)

Energy efficiency at launch

Dense BF16 TFLOP/s per watt of TDP — the hinge between the CAPEX you buy and the OPEX you then pay

No data yet — the scheduled pipeline has not produced this series.

Epoch AI — Machine Learning Hardware (CC BY)

US retail electricity price

US cents per kWh, monthly average by end-use sector and state

Awaiting the first run with the EIA key configured.

U.S. Energy Information Administration

Natural gas, monthly since 1997

USD per million BTU, monthly average of the Henry Hub spot price

Awaiting the first run with the EIA key configured.

U.S. Energy Information Administration

Semiconductor producer prices

US producer price index for semiconductor and related device manufacturing — the closest free proxy for what the supply chain charges

Awaiting the first run with the FRED key configured.

FRED / U.S. Bureau of Labor Statistics

How to read this

Dense throughput, never sparsity
Every TFLOP figure is dense BF16/FP16. Vendors often quote a doubled number that assumes 2:4 structured sparsity; mixing the two conventions would silently halve the cost per TFLOP of newer parts.
A launch price is not a market price
Datacentre accelerators have no public list price — they are sold on contract. Those figures are the best public estimates, and they describe the moment of launch, not what the same silicon costs today.
Rental is the only observable price
A GPU-hour is transacted continuously and quoted publicly, which makes it the one genuine market price for AI compute. It bundles capital, power, cooling and margin — that is a feature for a cost dashboard, and a caveat for a hardware comparison.
The history accumulates from here
Snapshot sources publish today and nothing else, and no one backfills them for free. Each scheduled run commits one dated observation, so the archive is built forward rather than bought.
Energy prices are averages, not your bill
Retail sector averages are published; the power purchase agreements that actually supply large data centres are private. Virginia and Texas are broken out because they carry the largest data-centre loads in the US.

Sources

SourceUsed forLicence
Epoch AI — Data on Machine Learning HardwareAccelerator release prices, throughput, memory and TDPCC BY 4.0
ComputePrices.comCloud GPU rental quotesPublic API
U.S. Energy Information AdministrationHenry Hub gas price, retail electricity pricesPublic domain
FRED — Federal Reserve Bank of St. LouisProducer price indices, redundant energy seriesAttribution requested

© 2026 Daniele Martini. All rights reserved.

We use cookies to analyze site traffic and enhance your experience. By clicking "Accept", you consent to our use of cookies.