---
type: "post"
id: "where-compute-ends-up-country-by-country-71a4d2"
slug: "where-compute-ends-up-country-by-country-71a4d2"
title: "Where compute ends up, country by country."
description: "Epoch AI has published an interactive database estimating where AI chips have arrived since 2022, using customs records and domestic production data."
retrieval_nugget: "Epoch AI has published an interactive database estimating where AI chips have arrived since 2022, using customs records and domestic production data. Its estimates put the United States far ahead in cumulative compute capacity, but the figures are modeled allocations of chip sales—not a direct count of deployed data-center compute."
published_at: "2026-11-11"
updated_at: "2026-11-11"
record_date: "2026-11-11"
date_kind: "published_at"
topics: ["compute"]
entities: ["Epoch AI"]
related_patterns: []
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://epoch.ai/data/ai-chip-imports"
source_urls: ["https://epoch.ai/data/ai-chip-imports"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:where-compute-ends-up-country-by-country-71a4d2"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/where-compute-ends-up-country-by-country-71a4d2/","markdown":"https://newruntime.com/posts/where-compute-ends-up-country-by-country-71a4d2.md","json":"https://newruntime.com/posts/where-compute-ends-up-country-by-country-71a4d2.json"}
markdown_url: "https://newruntime.com/posts/where-compute-ends-up-country-by-country-71a4d2.md"
---

# Where compute ends up, country by country.

## Retrieval answer

Epoch AI has published an interactive database estimating where AI chips have arrived since 2022, using customs records and domestic production data. Its estimates put the United States far ahead in cumulative compute capacity, but the figures are modeled allocations of chip sales—not a direct count of deployed data-center compute.

## Summary
Epoch AI has published an interactive database estimating where AI chips have arrived since 2022, using customs records and domestic production data. Its estimates put the United States far ahead in cumulative compute capacity, but the figures are modeled allocations of chip sales—not a direct count of deployed data-center compute.

## What happened
Epoch AI’s “AI chip imports” explorer, updated Nov. 3, 2026, estimates the geographic distribution of AI chips using customs data and domestic production information. It offers maps, charts, tables and downloadable datasets, with views for compute capacity in H100-equivalents (H100e), trade value, unit counts and chip power. Users can filter by destination country, chip designer or chip type, and view values as absolute totals or percentages. Epoch says the data is free to use, distribute and reproduce with attribution under a Creative Commons Attribution license.

The database covers quarterly arrivals since 2022. Epoch collects customs statistics from more than a dozen national sources, focusing on four Harmonized System codes covering computer systems, servers, other computer units and parts such as GPU boards. For destinations without direct customs data, it infers estimates from exporters’ records. The method subtracts a non-AI import baseline anchored to 2021 trade, nets out re-exports so chips are counted once in their final destination, and converts trade value to chip-only value using measured ratios. Epoch then allocates quarterly totals from its separate AI chip sales data across countries in proportion to the resulting trade value. It assumes a two-month lag between a chip’s sale and its arrival.

In Epoch’s cumulative totals, the United States leads with $288 billion in trade value, 16.1 million units and 19.8 million H100e, with an H100e range of 14.2 million to 25 million. China is listed at $50.5 billion, 4.39 million units and 2.64 million H100e, with a range of 1.9 million to 4.19 million. The European Union follows at $23.2 billion and 1.39 million H100e, with a range of 1.03 million to 2.57 million. The page also lists estimates for countries including Malaysia, India, Japan, Australia and the United Kingdom.

Epoch distinguishes measurements from allocations. Trade value comes directly from customs records; country-level units and H100e inherit global sales totals, with customs data determining how those totals are distributed. Ranges for units and H100e span the 5th to 95th percentiles across alternative modeling assumptions. The dataset counts arrivals into a destination’s customs territory, which occur after sale but before data-center deployment. Chips without known thermal design power are excluded from chip-power totals, so those totals are lower bounds. Countries with fewer than 1,000 H100e under the selected view are grouped or left off the map because their estimates are too uncertain.

## What the archive adds
New Runtime’s earlier coverage has examined Epoch’s [Inference Performance Index](https://newruntime.com/posts/epoch-ai-launched-the-inference-performance-index-scoring-1bb9b1/) (Nov. 10, 2026), which compared modeled concurrent inference capacity across chips rather than relying on raw FLOP/s. This is a new dataset and a new view of AI infrastructure: instead of comparing chip performance, it estimates the countries where chips arrive.

## Trend signal
This supports the infrastructure-measurement thread in New Runtime’s earlier coverage of the [Inference Performance Index](https://newruntime.com/posts/epoch-ai-launched-the-inference-performance-index-scoring-1bb9b1/): useful infrastructure comparisons depend on a defined operational measure, not a headline hardware specification. The dataset also makes its allocation assumptions and uncertainty ranges visible. It does not directly measure deployed compute, so the distinction between estimated arrivals and usable capacity remains important.

## Why now
Epoch’s page was updated Nov. 3, 2026, and its estimates include arrivals through the available quarterly data. Its methodology says the country-level series starts in 2022 because that is when Epoch’s AI chip sales data begins; charts default to 2023 onward because that is when shipments clearly separate from the non-AI baseline.

## Why it matters
For teams choosing where to train or serve models, a country-level estimate can help frame questions about infrastructure concentration, regional supply chains and potential access constraints. The headline totals are striking, but they should not be used as a proxy for compute a team can actually rent: arrivals do not equal deployment, and deployment does not establish that capacity is commercially available, connected to a particular cloud, or suited to a workload.

The database is more useful as a map of estimated infrastructure flows than as a leaderboard of current usable capacity. Its downloadable quarterly data lets analysts examine how estimates change over time, while the ranges and documentation make assumptions inspectable. Builders can use it to identify where further diligence is warranted, then verify actual capacity, location, chip type, power and service availability with operators before making deployment decisions.

## Limits and uncertainty
Epoch describes the results as estimates, not direct observations of installed chips or operating compute. Customs records do not identify chip models, so some designer and chip-type breakdowns are allocated using estimated per-model prices and common chip mixes; Epoch says these splits mostly reflect the global mix rather than destination-specific evidence. The country totals also rely on an assumed two-month sale-to-arrival lag, a baseline derived from 2021 trade, and global sales totals that customs data only distributes geographically. The ranges capture variation across modeling assumptions, but do not eliminate uncertainty in source records or establish how much compute is deployed or available to buyers.

## Reader takeaway
Epoch’s explorer gives builders a documented, downloadable estimate of where AI chips have arrived—not a census of live compute. The United States leads the cumulative H100e estimate by a wide margin, but the modeled ranges and the gap between customs arrival and deployment matter as much as the headline. Use it to ask better infrastructure questions, then verify capacity at the operator level.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Compute](https://newruntime.com/topics/compute.md): Explore the compute topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)