AI progress

The state of artificial intelligence progress

Compute, models, chips, data centers and benchmarks: the full trajectory of AI, quantified and continuously updated.

Largest training run

5 x 10^26 FLOP

Grok 4 · xAI

Tracked data-center power

11.86 GW

across 75 sites

Tracked data-center compute

12.5M H100e

H100 equivalents

Top capability index

161.7

GPT-5.6 Sol (max) · ECI

Datasets

AI models

AI models3,568

records

Training compute, parameters, dataset size, hardware and accessibility for the machine learning systems that define the field, from the 1950s to today.

Compute capacity, electrical power, capital cost and build-out timeline for the biggest AI training and inference sites in the world.

Every known large GPU cluster and AI supercomputer, with its chip count, compute throughput, power draw and the date it came online.

GPUs, TPUs and custom accelerators with their throughput at each precision, memory bandwidth, power draw, process node and price.

The chip types tracked through the AI supply chain, with their compute in H100-equivalents, memory, power and approximate unit cost.

Revenue, valuation, funding, headcount and compute spend for the labs and platforms at the centre of the industry.

Every evaluation tracked, from saturated classics to the frontier tests models still fail — plus the Epoch Capabilities Index that ties them together.

Open

Views across the data

Compute map

Every located data center and GPU cluster on one map, sized by the compute it holds.

Open

Rankings

The largest data centers, the biggest training runs, the fastest chips — each with its current answer.

Open

Organizations

Every lab with a tracked model, and everything it has built.

Open

Countries

Where models are trained, and where the compute physically sits.

Open

Comparisons

Frontier models head to head, specification by specification and benchmark by benchmark.

Open

How this is measured

Almost nothing here is disclosed by the labs. Training compute is reconstructed from hardware counts and training time; data-center capacity is inferred from construction permits, satellite imagery and power contracts; chip supply is modelled from earnings and shipment data. So every estimate carries its uncertainty with it — a confidence tag on an attribution, a 5th-to-95th percentile band on a supply figure, a note on how a compute number was derived. Those qualifiers are kept on the page rather than rounded away.

The underlying research is Epoch AI's, published under a Creative Commons Attribution license. What is added here is structure: one durable URL per model, per data center, per cluster, per chip, per benchmark — server-rendered, cross-linked and citable, instead of locked inside an interactive explorer.

Read the full methodology →

SourceEpoch AI, 'AI Models'. Published online at epoch.ai. Retrieved 2026-07-29 from https://epoch.ai/data/ai-models. Licensed under CC BY 4.0.

Last synced 2026-07-29