The Hardware Behind AI Has Several Centers

There is a habit of talking about “the AI supply chain” as if it were a single pipeline running through a single country. The trade data says otherwise. If you take the two product systems that sit closest to AI compute — integrated circuits and the data-processing machines and parts that house them — their combined exports grew from $544.0 billion in 2000 to $1.55 trillion in 2024, a 2.9-fold increase in nominal value. But that single number hides two very different maps.

Split the combined group into its halves and the divergence is immediate. Integrated circuits were $221.8 billion of exports in 2000 and $967.6 billion in 2024. Data-processing machines and parts were $322.2 billion in 2000 and $586.0 billion in 2024. Chips started as the smaller of the two systems and ended as the larger: in 2000 integrated circuits were 40.8% of the combined total and machines were 59.2%; by 2024 chips had risen to 62.3% and machines had fallen to 37.7%. The reading path for the rest of this piece is that contrast — chips versus machines — because the countries that dominate each are not the same.

A note on what this grouping is and is not. “AI-related” here is an editorial grouping of HS92 product codes, not a category that BACI or the customs nomenclature defines. The codes are broad: they capture all processors, controllers, memory and computer assemblies, and they cannot isolate GPUs or AI accelerators from ordinary chips and PCs. Nominal value reflects both price and quantity. Read the levels as the shape of a trade system, not as a count of AI chips.

Line up the 2024 exporter shares for each product group and you get three distinct leaderboards. For integrated circuits, the largest single exporter is “Other Asia, nes” at 23.9% of group world exports, followed by China at 17.6% and the Republic of Korea at 17.2%. For data-processing machines and parts, China stands alone at 40.5%, with “Other Asia, nes” at 13.8% and Vietnam at 10.2%. For the smaller optics group, China again leads at 28.7%, but here the United States (12.1%) and Japan (7.5%) are the next two — a different cast entirely.

The combined AI-compute group blends these into a top three of China 26.2%, “Other Asia, nes” 20.1% and Korea 12.3%, together 58.6% of the group. But the blend is misleading if you stop there: China’s lead in the combined group comes mostly from machines, while the chip leaderboard is headed by an aggregate, not by China.

That aggregate needs to be named carefully. “Other Asia, nes” is BACI country code S19 (numeric 490), an aggregate that folds in Taiwan. Taiwan is not a distinct reporting entity in this dataset, so the chip exports that the industry attributes to Taiwanese foundries appear in these rankings under “Other Asia, nes”, never under a separate Taiwan row. When this chart shows “Other Asia, nes” as the top integrated-circuit exporter, that is the convention at work, and it is labelled exactly as the source labels it.

Concentration is measured here with a Herfindahl-Hirschman Index computed on fractional exporter shares, so it runs from 0 (perfectly spread across exporters) toward 1 (a single exporter). Integrated circuits have been a concentrated business throughout: the exporter HHI was 0.103 in 2000 and 0.137 in 2024. Even so, no single country dominates chips — the top exporter (“Other Asia, nes”) holds under a quarter of the market, and 221 distinct exporters record positive integrated-circuit exports in 2024. Chip concentration is real but it is a concentration among a handful of leaders, not a monopoly.

Data-processing machines tell a concentrating story. Their exporter HHI rose from 0.071 in 2000 to 0.202 in 2024 as assembly consolidated. China's share of machine exports climbed from 8.8% in 2000 to 40.5% in 2024. The two systems therefore move in opposite directions on the map even as they grow together in value: chips spread their leadership across Other Asia, China and Korea, while machines pulled toward a single hub.

Optics, the smallest group, sits at an HHI of 0.118 in 2024 — less concentrated than either chips or machines, and led by China but with the United States and Japan close enough behind to keep it a genuinely multi-centre market. The small-multiples view lets you read each group’s trajectory on its own axis rather than forcing them onto one scale.

The 2024 snapshot is the end of a long reallocation. In 2000, integrated-circuit exports were already led by “Other Asia, nes” at 7.4%, but the United States (21.3%) and Japan (14.2%) were far larger chip exporters than they are today. The chip leaderboard has shifted toward East Asia, but its top of “Other Asia, nes” has been a fixture for the whole period — the Taiwan-centred foundry model shows up as a durable lead by the S19 aggregate, not as a recent disruption.

Machines moved more dramatically. China's share of data-processing machine and parts exports was 8.8% in 2000; by 2024 it had risen to 40.5%, and Vietnam emerged as the third-largest machine exporter at 10.2%. The combined AI-compute group grew 2.9-fold over the period, but chips grew 4.4-fold against machines’ 1.8-fold — which is why the chip half overtook the machine half in value despite machines being the early leader.

None of these movements should be read as a clean mine-to-model supply chain. These are gross export flows of broad product codes. A chip fabricated in one economy, packaged in a second and soldered onto a board in a third will appear in several of these flows at several stages; the codes do not net out re-exports or trace a physical route. The value is in the contrast between systems, which is robust to those caveats, not in any single shipment.

Switch from who exports to who buys, and the concentration shows up again from the demand side. This matrix takes a set of major integrated-circuit importers in 2024 and shows where each sources its chips, using bilateral flows from the raw trade file. These are source-country shares — the economy a shipment is recorded as coming from — not shipping routes, and they describe a regional cluster that trades chips intensively with itself.

China is the largest single buyer of integrated circuits, importing $199.8 billion in 2024, with its top recorded source being Rep. of Korea at 31.8%. Hong Kong, a major entrepot, imported $219.3 billion with China as its leading source at 33.5%. The United States imported $39.0 billion, leaning on Other Asia, nes (28.4%). Mexico, an assembly hub for the North American market, imported $24.3 billion with Malaysia its top source at 22.4%.

The supplier columns repeat across rows: the same East Asian economies — the “Other Asia, nes” aggregate, China, Korea, Malaysia, Singapore -- appear as top sources for buyer after buyer. That is the structural point of the whole article. Chips are not concentrated in one exporter, but the set of exporters that matters is small and shared, so a disruption to any one of them would be felt across many importers at once. Because these are source-country attributions, they cannot prove which physical fab made a given chip, and a shipment routed through an entrepot is credited to that entrepot, not to the original producer.

Because the whole grouping is an editorial choice, the honest move is to show the codes. The integrated-circuit group is four HS92 lines: 854211 (processors and controllers), 854219, 854220 and 854280. The largest of these by 2024 export value is the processors-and-controllers line, 854211, whose top exporter is the “Other Asia, nes” aggregate — the single code that most clearly carries the Taiwan foundry story.

The data-processing group is five lines: 847120 (whole computers, an older code), 847192 (input/output units), 847193 (storage units), 847199 (other ADP units) and 847330 (parts and accessories of those machines). The optics group — 854470, 900110, 900190 and 901380 — adds only $24.0 billion in 2024 and is deliberately excluded from the headline AI-compute total, which is why the combined figure is the four chip codes plus the five machine codes and nothing else.

These descriptions are broad on purpose, and they cannot separate an AI accelerator from a microcontroller, or a server motherboard from a consumer one. Read the cards as a map of which customs lines the story rests on, so that any reader can re-run the same grouping and get the same $1.55 trillion combined total. The point of publishing the codes is reproducibility, not precision about end use.