Skip to content
ContentLora

    Tip: press / anywhere to search.

    Analysis

    Why AI chip supply is so concentrated

    The leading AI chips depend on a short chain of specialists. Designers such as NVIDIA rely on foundries such as TSMC, buy memory from three suppliers, and use advanced packaging.[1][2] The finest layers are printed with EUV tools, a technology ASML describes as unique to it.[3] New fabs outside Taiwan and new chip designs are spreading supply, but slowly.

    Editor reviewedUpdated SemiconductorsArtificial intelligence
    Show:

    AI chip supply is concentrated because each step in making an accelerator depends on a handful of firms. This page sets out the evidence, what it suggests, and where the views differ.[1]

    The chain behind an AI chip

    NVIDIA designs its chips and uses foundries such as TSMC and Samsung to make them.[1] It buys memory from three named suppliers: SK Hynix, Micron and Samsung.[2] It packages its chips with CoWoS, a TSMC technology that places processors and high-bandwidth memory on a shared interposer.[4][5] NVIDIA says its supply chain is mainly concentrated in Asia.[6]

    The finest chip layers are printed with EUV lithography, which ASML describes as unique to ASML.[3] ASML recognized revenue on 48 EUV systems in 2025, worth €11.6 billion.[7] At TSMC, technologies at 7 nm and below made up 74% of 2025 wafer revenue.[8]

    Why it stays concentrated

    Demand has grown quickly. NVIDIA’s fiscal 2026 revenue was $215.9 billion, up 65%.[9] TSMC’s 2025 revenue rose 35.9% to US$122.42 billion, and the company reported robust AI-related demand all year.[10][11]

    Market-share estimates for 2026 show the same pattern. TrendForce estimated that TSMC took 72.5% of top-10 foundry revenue in the second quarter of 2026, with Samsung Foundry second at 5.9%.[12] In DRAM, TrendForce put Samsung first with a 39.4% revenue share in the same quarter.[13] Packaging is a further constraint. TSMC forecast that its CoWoS capacity would grow more than 80% a year from 2022 to 2027,[14] and monthly capacity is reported to reach 120,000 to 140,000 wafers in 2026, with partner packaging firms adding 50,000 to 60,000.[15] NVIDIA’s Blackwell GPU, for example, used CoWoS to combine more than three reticles’ worth of silicon with eight HBM stacks, an example of the chiplet approach.[16]

    These facts point to a self-reinforcing pattern. Each step needs tools or processes that only a few firms have mastered: EUV scanners, leading-edge logic, HBM and large-interposer packaging.[3][2][5] When demand rises this fast, buyers go to suppliers that already work at volume. That adds to those suppliers’ revenue and their capacity to invest.[10] The same firms also appear at more than one step. Samsung is named both as a foundry NVIDIA uses and as one of its three memory suppliers, so a problem at one company can affect several links at once.[1][2] The reasoning has limits. Revenue growth shows demand, not barriers to entry, and the market-share figures are analyst estimates rather than company disclosures. The post-Moore hardware debate asks whether new chip architectures could change which steps are bottlenecks.

    Signs of diversification

    TSMC’s first Arizona fab entered high-volume production in the fourth quarter of 2024. A second is expected in the second half of 2027, and construction of a third began in 2025.[17] Intel says its 18A process is in high-volume production in the US and open to customer projects.[18] A 2024 SIA and BCG report projected that the US share of advanced logic capacity would rise from 0% in 2022 to 28% by 2032.[19] On the demand side, Anthropic said in October 2025 that it uses Google TPUs, Amazon Trainium and NVIDIA GPUs.[20]

    Diversification is happening in two different ways. New fabs spread where chips are made. Multi-platform buyers spread whose designs are used.[17][20] The second kind does not necessarily reduce dependence on the same foundries, memory makers and lithography tools. TSMC is also still adding 2nm capacity in Taiwan.[21]

    Policy as a force

    US export rules now shape who can buy. NVIDIA says it was effectively shut out of China’s data center compute market by early 2026.[22] BIS has warned industry of the risks of using Chinese AI chips, including Huawei Ascend chips.[23] Since January 2026, China-bound chips above H200-level thresholds face a presumption of denial.[24] Below those thresholds, licensed H200 sales have been limited by China’s own restrictions: NVIDIA said in August 2026 that such shipments were less than 1% of its quarterly Data Center revenue.[25]

    Controls may split the market into separate supply blocs instead of spreading supply within one. Inside each bloc, a few suppliers could still dominate.[22][24]

    What may happen next

    Over the next two to three years, concentration in EUV tools seems likely to persist. No second supplier appears in these sources.[3] Geographic concentration in leading-edge logic may ease gradually as US fabs ramp. TSMC’s second Arizona fab is not expected in volume until the second half of 2027, so Taiwan will likely remain the center of leading-edge output in that period.[17][21] The 28% US share for 2032 is a projection from 2024, not an outcome, and it depends on fabs being built and filled.[19] Uncertainty is high. A demand slowdown, a breakthrough in packaging, or a change in export policy could each shift this picture.

    Competing views

    Concentration persists

    Leading-edge capacity, EUV tools and HBM each rest on very few suppliers, and AI demand is still growing fast. New entrants need years and large sums to catch up, so the bottlenecks remain.[3][2][8][11]

    Diversification is under way

    New fabs in the US, Intel's 18A process and AI developers that use several chip platforms are spreading manufacturing and design across more places and firms.[17][18][19][20]

    Policy splits the market rather than spreading it

    Export controls divide the market along national lines. Inside each bloc, supply may stay just as concentrated.[22][23][24]

    Questions readers ask

    Who makes NVIDIA's chips?

    NVIDIA is fabless. It uses foundries such as TSMC and Samsung, and buys memory from SK Hynix, Micron and Samsung.[1][2]

    Why does ASML matter so much?

    ASML describes EUV lithography, used for the most intricate chip layers, as unique to ASML. It recognized revenue on 48 EUV systems in 2025.[3][7]

    Is chip manufacturing moving out of Taiwan?

    Some of it is. TSMC's first Arizona fab entered high-volume production in late 2024, while TSMC is also building multiple phases of 2nm fabs in Taiwan.[17][21]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      NVIDIA follows a fabless model, using foundries such as TSMC and Samsung to make its chips. confirmedas of 2026-02-25

    2. [2]

      NVIDIA buys memory from three suppliers it names as SK Hynix, Micron and Samsung. confirmedas of 2026-02-25

    3. [3]

      ASML describes EUV as unique to ASML; its EUV systems print the most intricate chip layers at a resolution of 13 nm, which DUV cannot reach. confirmedas of 2026-10-10

    4. [4]

      NVIDIA uses CoWoS technology to package its chips. confirmedas of 2026-02-25

    5. [5]

      TSMC's CoWoS packaging places processor chips and multiple high-bandwidth memory (HBM) stacks on a shared interposer, which can exceed twice the reticle size (about 1,700 mm²). confirmedas of 2026-10-10

      • CoWoS · TSMC · Overview; the page renders the unit as mm with a superscript 2 (retrieved 2026-10-10)
      • CoWoS · TSMC (retrieved 2026-10-10)
    6. [6]

      NVIDIA states that its supply chain is mainly concentrated in Asia and that geopolitical tensions could materially harm it. confirmedas of 2026-02-25

    7. [7]

      ASML's 2025 EUV system sales rose 39% to €11.6 billion on 48 systems, out of total net sales of €32.7 billion. confirmedas of 2026-01-28

    8. [8]

      Technologies at 7 nm and below made up 74% of TSMC's wafer revenue in 2025, up from 69% in 2024. confirmedas of 2026-10-10

    9. [9]

      NVIDIA's revenue for fiscal 2026 (ended January 25, 2026) was $215.9 billion, up 65% from a year earlier. confirmedas of 2026-02-25

    10. [10]

      TSMC's 2025 consolidated revenue was US$122.42 billion, up 35.9% from 2024. confirmedas of 2026-10-10

    11. [11]

      TSMC reported robust AI-related demand throughout 2025. confirmedas of 2026-10-10

    12. [12]

      TrendForce estimated that TSMC held 72.5% of top-10 foundry revenue in 2Q26, with Samsung Foundry second at 5.9%. reportedas of 2026-06-30

    13. [13]

      TrendForce estimated that Samsung ranked first in DRAM in 2Q26 with a 39.4% revenue share and quarterly DRAM revenue of US$60.98 billion, out of total DRAM industry revenue of about US$154.73 billion. reportedas of 2026-06-30

    14. [14]

      TSMC forecast in May 2026 that its CoWoS capacity will grow at a compound annual rate of more than 80% from 2022 to 2027. reportedas of 2026-06-15· forecast

    15. [15]

      TSMC's monthly CoWoS packaging capacity is reported to reach 120,000-140,000 wafers in 2026, with partner packaging firms adding 50,000-60,000. reportedas of 2026-06-15

    16. [16]

      As of 2024, NVIDIA's Blackwell GPU used TSMC's CoWoS packaging to combine more than three reticles' worth of silicon with eight HBM memory stacks. confirmedas of 2024-04-30

    17. [17]

      TSMC's first Arizona fab entered high-volume production in the fourth quarter of 2024; the second is expected to do so in the second half of 2027, and construction of a third began in 2025. confirmedas of 2026-10-10

    18. [18]

      Intel says its 18A process is in high-volume production in the United States and open to customer projects. confirmedas of 2026-10-10

      • Intel 18A · Intel Foundry (retrieved 2026-10-10)
      • Intel 18A · Intel Foundry (retrieved 2026-10-10)
    19. [19]

      A May 2024 SIA and BCG report projected that the US share of global advanced logic capacity would grow from 0% in 2022 to 28% by 2032. confirmedas of 2024-05-31

    20. [20]

      In October 2025 Anthropic said it would expand to up to one million Google TPUs and that it uses three chip platforms: Google TPUs, Amazon Trainium and NVIDIA GPUs. confirmedas of 2025-10-23

    21. [21]

      TSMC is preparing multiple phases of 2nm fabs in Taiwan's Hsinchu and Kaohsiung science parks. confirmedas of 2026-10-10

    22. [22]

      NVIDIA said that at the end of fiscal 2026 it was effectively foreclosed from competing in China's data center compute market. confirmedas of 2026-02-25

    23. [23]

      Alongside the rescission, BIS issued guidance warning industry of the risks of using Chinese advanced computing chips, including specific Huawei Ascend chips. confirmedas of 2025-05-13

    24. [24]

      The January 2026 case-by-case policy covers chips with total processing performance below 21,000 and DRAM bandwidth below 6,500 GB/s, caps China and Macau shipments at 50% of US shipments of the product, and keeps a presumption of denial for other applications. confirmedas of 2026-01-15

    25. [25]

      In its quarterly report for the period ended July 26, 2026, NVIDIA said China's government had restricted sales of the H200s it was licensed to ship, that it had taken a $0.4 billion H200 charge in the first half of fiscal 2027, and that it had since made a fraction of the allowed shipments, accounting for less than 1% of quarterly Data Center revenue. confirmedas of 2026-08-26

    Revision history (2)
    1. Page created.
    2. Refresh: added 2026 foundry and DRAM shares, CoWoS packaging capacity and the August 2026 H200 update; fixed TSMC and ASML excerpts; linked computing-hardware field pages.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Apr 10, 2027.

    Cite this page

    "Why AI chip supply is so concentrated." ContentLora, updated Oct 10, 2026. https://contentlora.com/analysis/ai-chip-supply-concentration

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    ☰ Contents