AI Chips and Semiconductors: Who Makes Them, the Supply Chain, and the Chokepoints

Executive summary

The AI chip market is one of the most concentrated in modern technology, with three chokepoints stacked on top of each other. First, at the accelerator level, Nvidia dominates: it held roughly 86% of the data-centre AI GPU market in 2025 by segment-revenue estimates, and Epoch AI attributes more than 60% of the world's installed AI computing capacity to Nvidia chips, with Google's TPUs and Amazon's Trainium making up much of the rest. Second, at the manufacturing level, almost every leading-edge AI chip — Nvidia, AMD, Google, Apple — is physically fabricated by Taiwan's TSMC, which held about 70% of the pure-play foundry market in 2025 (TrendForce) and more than 90% of the advanced-node capacity required for cutting-edge AI processors. TSMC in turn depends on a single supplier, ASML of the Netherlands, for the extreme-ultraviolet (EUV) lithography machines that print those chips. Third, at the memory level, high-bandwidth memory (HBM) is a bottleneck controlled by three firms, led by SK Hynix at about 62% of Q2 2025 bit shipments. Layered over all of this is geopolitics: US export controls introduced in October 2022 and tightened repeatedly since restrict the sale of advanced AI chips to China. The stakes are large and growing — global semiconductor sales reached a record $791.7 billion in 2025, up 25.6%, with AI the primary driver (SIA). Understanding AI chips means understanding this vertical stack of near-monopolies, from design to fabrication to memory, and the export-control regime wrapped around it. Market-share figures reflect the specific metric cited in each case.

Key findings

What is an AI chip? (GPU vs TPU vs ASIC)

An AI chip is a processor specialised for the massively parallel maths — mainly matrix multiplication — that powers training and running neural networks, and the dominant type today is the GPU (graphics processing unit), led by Nvidia. Unlike a general-purpose CPU, which executes a few complex tasks quickly in sequence, an AI accelerator packs thousands of simpler cores that run many calculations at once, which is exactly what deep learning requires.

There are three broad families. GPUs (Nvidia's H100/H200/Blackwell, AMD's Instinct MI300/MI325) are flexible and dominate the merchant market. TPUs (Tensor Processing Units) are Google's custom AI chips, built specifically for its own workloads. ASICs (application-specific integrated circuits) are custom silicon designed for one job — Amazon's Trainium and Inferentia, and Google's TPU, are ASICs; they trade flexibility for efficiency and cost at scale.

The table below summarises the main AI-chip types, what they do, and a real-world example of each.

Who makes AI chips? (the companies and their market shares)

AI chips are designed by a small group of companies led overwhelmingly by Nvidia, which held roughly 86% of the data-centre AI accelerator market in 2025 by segment-revenue estimates. AMD is the main merchant challenger with its Instinct line but holds a low-single-digit to mid-single-digit share of data-centre AI GPUs. The other large volumes come not from merchant vendors but from cloud providers designing their own chips: Google (TPU), Amazon (Trainium/Inferentia), Microsoft (Maia) and Meta (MTIA).

The distinction between designing and manufacturing matters. Nvidia, AMD, Google, Apple, Amazon and Microsoft are 'fabless' — they design chips but own no fabs. The physical manufacturing is outsourced, almost entirely to TSMC. So 'who makes AI chips' has two answers: Nvidia designs the accelerators the market runs on, while TSMC builds them.

On installed capacity rather than sales, Epoch AI attributes more than 60% of total AI computing capacity to Nvidia chips, with Google and Amazon making up much of the remainder — a reminder that the merchant-share and installed-compute pictures differ because hyperscalers deploy large fleets of their own silicon.

Who manufactures AI chips? (the TSMC and ASML chokepoint)

Nearly every leading-edge AI chip in the world is physically manufactured by a single company: TSMC (Taiwan Semiconductor Manufacturing Company), which held about 70% of the pure-play foundry market in 2025. TSMC produces the advanced 3nm and 5nm chips for Nvidia, AMD, Apple and Google, and controls more than 90% of the advanced-node capacity that cutting-edge AI chips require. This makes its Taiwan fabs the most critical — and most geopolitically exposed — point in the AI supply chain.

TSMC itself depends on a chokepoint above it. The extreme-ultraviolet (EUV) lithography machines needed to print the smallest features are made by exactly one company on Earth: ASML of the Netherlands. No ASML, no leading-edge chips — which is why EUV export restrictions are a lever in US-China tech policy. Below TSMC sit other single-source tools from Applied Materials, Lam Research and Tokyo Electron.

The table below shows the concentration of the global pure-play foundry market in Q2 2025, the layer where AI chips are actually made. TSMC's share is not a rounding advantage — it is a structural near-monopoly on the capacity that matters.

What is HBM and why is it a bottleneck for AI chips?

HBM (high-bandwidth memory) is stacked DRAM mounted right next to the GPU to feed it data fast enough, and it is the second scarcest input in AI hardware after the accelerator itself. Modern AI accelerators are frequently 'memory-bound' — the compute cores sit idle waiting for data — so the amount and speed of HBM directly caps real-world performance. A Blackwell-class GPU carries stacks of HBM3E worth a large fraction of the chip's cost.

The HBM market is a tight oligopoly of three firms. In Q2 2025, SK Hynix led with about 62% of bit shipments (up from 55% a year earlier), Micron held about 21% and Samsung about 17%. SK Hynix is Nvidia's lead HBM supplier, and its qualification advantage on newer generations has let it capture the premium AI demand while Samsung lost ground.

Because HBM production is hard to expand quickly and is largely pre-sold to Nvidia and other accelerator makers, memory has repeatedly been the gating factor on how many AI chips can actually ship. The 2026 competition is shifting to HBM4.

What does the AI chip supply chain look like? (the layers and players)

The AI chip supply chain is a vertical stack of specialised layers, and each layer is dominated by one to three companies, which is what makes the whole chain fragile. From design tools at the top to assembly at the bottom, a disruption at almost any single layer can bottleneck the entire industry.

This concentration is not accidental — leading-edge semiconductors demand extreme capital and know-how, so each step consolidated around whoever reached the frontier first. The result is a chain where Nvidia (design), TSMC (fab), ASML (lithography) and SK Hynix (memory) each hold near-veto power over AI hardware output.

The table below maps the main layers of the supply chain to their key players.

What do US export controls on AI chips to China do?

US export controls, first imposed in October 2022 and tightened repeatedly since, restrict the sale of advanced AI chips and chipmaking equipment to China. The October 2022 rules banned exports of AI chips at or above the capability of Nvidia's A100, plus the tools to make them; updates in October 2023 and December 2024 closed workarounds and expanded the equipment list. CSIS's Gregory Allen described the aim as to 'choke off China's access to the future of AI.'

Nvidia responded by designing compliance-tuned chips for China — the H800, then the H20 — deliberately throttled to stay under the thresholds. In April 2025 the Commerce Department declared the H20 (which had powered DeepSeek's breakthrough model) newly non-compliant, illustrating how the thresholds move. Hundreds of thousands of H20s had already been sold, earning Nvidia an estimated $12-15 billion.

The regime is not static. In December 2025 the US moved to review licences for more capable chips — the Nvidia H200 and AMD MI325X — for export to China on a case-by-case basis, loosening but not dismantling the controls. The effect has been to fragment the market: China is pushed toward domestic alternatives (notably Huawei's Ascend line), while access to the leading edge remains gated in Washington.

Who owns the world's AI compute? (the Epoch AI picture)

Nvidia chips account for more than 60% of the world's total installed AI computing capacity, according to Epoch AI, with Google and Amazon — running their own TPUs and Trainium — making up much of the remainder. This installed-base measure is distinct from annual sales share and captures the cumulative stock of compute actually deployed.

The stock is growing at an extraordinary rate. Epoch AI finds total AI computing capacity across all chip designers has grown about 3.3x per year since 2022 — a doubling roughly every seven months — while the installed base of Nvidia compute specifically doubles about every ten months. Compute, in other words, is compounding far faster than the broader economy or even Moore's Law.

Ownership is also concentrating. Epoch AI estimates that five hyperscalers — Amazon, Google, Meta, Microsoft and Oracle — collectively held about 71% of the world's cumulative AI compute in Q4 2025, up from 63% in Q1 2024. So while Nvidia supplies most of the silicon, a handful of cloud giants own most of the machines built from it.

How much do AI chips cost and how big is the buildout?

AI chips are among the most valuable manufactured products on Earth, and demand for them turned semiconductors into a record market — global chip sales hit $791.7 billion in 2025, up 25.6% year over year (SIA), after $627.6 billion in 2024. Memory alone jumped 78.9% in 2024, and logic (which includes AI accelerators) was the single largest category — both surges driven by AI.

A single flagship data-centre GPU sells for tens of thousands of dollars, and a full AI server rack combining dozens of them can cost several million. That economics is why Nvidia's data-centre revenue and TSMC's advanced-node revenue both roughly doubled through the AI cycle, and why hyperscalers are committing hundreds of billions of dollars in capital expenditure to buy them.

The buildout is capital-intensive at every layer: a leading-edge fab costs upwards of $20 billion, a single EUV machine tens of millions, and HBM capacity years to expand. These lead times are why supply has repeatedly lagged demand across the AI cycle.

What are the risks in the AI chip supply chain?

The defining risk of the AI chip supply chain is geographic and corporate concentration: the majority of leading-edge AI chips are fabricated in Taiwan by one company, TSMC, making the industry acutely exposed to any disruption in the Taiwan Strait. A conflict, blockade or natural disaster affecting Taiwan would hit the entire global AI industry at once, with no near-term substitute at the leading edge.

Single points of failure repeat up and down the stack: ASML is the sole maker of EUV lithography, SK Hynix supplies most premium HBM, and Nvidia's CUDA software ecosystem locks in developers even where rival hardware exists. Any one of these can gate output or pricing.

Layered on top is policy risk. Export controls, tariffs and subsidy races (the US CHIPS Act, the EU Chips Act, China's domestic push) are actively redrawing the map, and thresholds shift with each administration. The market's response — reshoring fabs to Arizona, Japan and Germany, and China building domestic capacity — reduces single-point risk over time but at higher cost and slower pace.

What is the outlook for AI chips and semiconductors?

The near-term outlook is continued AI-driven growth on top of an already concentrated structure: analysts and the SIA expect the semiconductor market to keep climbing past its 2025 record of $791.7 billion, with AI logic and HBM memory the fastest-growing segments. Nvidia's Blackwell and successor architectures, AMD's Instinct roadmap and expanding hyperscaler ASIC fleets will drive the next wave of demand.

Three shifts are worth watching. First, custom silicon: hyperscalers keep expanding TPU/Trainium/Maia fleets to reduce Nvidia dependence, gradually eroding the merchant-share picture even as Nvidia's absolute volumes rise. Second, geographic diversification: TSMC Arizona, Samsung Texas, Intel and new fabs in Japan and Europe slowly reduce Taiwan concentration. Third, the memory frontier moving from HBM3E to HBM4.

The structural takeaway is unlikely to change quickly: AI compute is a vertical stack of near-monopolies — Nvidia in accelerators, TSMC in fabrication, ASML in lithography, SK Hynix in memory — and the companies that control each chokepoint capture most of the value. All forward-looking figures are estimates and depend on demand, capacity build-out and policy.

Scoreboard (machine-readable data)

Every headline indicator with its value, period, source and confidence. Free to reuse under CC BY 4.0.

↓ CSV · ↓ JSON

Methodology & verification

Figures are drawn from primary and specialist sources and each is attributed inline with the specific metric it measures. Nvidia's data-centre AI accelerator share (~86%, 2025) reflects segment-revenue estimates and market trackers; its installed-compute share (>60%) comes from Epoch AI's AI chip datasets, which are a different measure (deployed capacity, not sales). Foundry shares are TrendForce pure-play foundry data for 2025/Q2 2025. Advanced-node share (>90%) is an industry estimate of leading-edge capacity, marked Medium confidence. HBM shares are Q2 2025 bit-shipment estimates. Semiconductor-sales totals are SIA figures for 2024 and 2025. Export-control dates and the Gregory Allen quote are from CSIS. Where sources report rounded values or ranges, the rounding is preserved; no figures were interpolated or invented, and any figure that could not be tied to a named source was omitted.

Data dictionary

Frequently asked questions

Glossary

AI accelerator

A processor specialised for the parallel maths of neural networks (training and inference); GPUs and custom ASICs such as TPUs are the main types. ↗

Foundry (pure-play)

A contract chip manufacturer that fabricates chips designed by other (fabless) companies; TSMC is the largest, with about 70% of the market in 2025. ↗

HBM (high-bandwidth memory)

Vertically stacked DRAM placed beside a GPU to deliver very high memory bandwidth; a key bottleneck in AI accelerators, led by SK Hynix. ↗

EUV lithography

Extreme-ultraviolet lithography, the technology used to print the smallest chip features; the machines are made only by ASML of the Netherlands. ↗

Sources

• Epoch AI — AI chip production & compute capacity
• Epoch AI — Data on AI Chip Owners
• SIA — Global semiconductor sales 2025 ($791.7B)
• TrendForce — Global foundry market share
• CSIS — US export controls on advanced AI chips
• Astute Group / Semiecosystem — HBM market share

Related reports & data

• The Biggest AI Data Centers
• AI Compute Index
• National AI Power Index