Supply chain

From sand to model

What has to be built, in order, before anyone can train an AI model — 12 steps, and where they narrow to a chokepoint.

5 3 1 3

  1. Raw materials for Critical materials

    The materials at the start of the chain: rare earths for the magnets that move hard-drive arms and spin cooling motors, and the polished silicon wafers every chip starts as.

    The record Ga: China 99% · heavy-REE magnet controls in force · He: watchlist
    99 % China share, primary low-purity gallium USGS MCS 2026, gallium (no range — disclosed figure) pubs.usgs.gov

    MP Materials GlobalWafers

    GlobalWafers sells to TSMC · Samsung

    2 deals at this step newest first
  2. Chipmaking machines for EUV lithography

    The machines that print a chip's circuits. The ones that print the finest features — EUV lithography — come from a single company.

    The record ASML — sole supplier

    ASML

    ASML sells to TSMC · Samsung · Intel · SK Hynix · Micron

    No deal in the record sits at this step yet.

  3. Chip factories

    Where the chips are made: nearly every advanced AI chip, whoever designs it, is physically made in TSMC's fabs in Taiwan. Intel is one of the few companies that both designs chips and makes them in its own factories, a business it runs as Intel Foundry. Samsung, too, makes chips that other companies design, Tesla's AI chips among them.

    TSMC Intel Samsung

    Samsung sells to Penguin Solutions

    1 deal at this step newest first
  4. Memory for HBM memory

    The high-bandwidth memory stacked beside every AI chip, which only three companies make, and the ordinary DRAM that makers such as Nanya also produce.

    The record 3 makers make virtually all — SK ~56–63% rev share (IDC) · ~$11.4B/qtr · Samsung's ready capacity booked, Micron's 2026 supply agreed · HBM eats ~22% of DRAM wafers
    11.4 $B/qtr (10.26–12.54) Epoch AI Chip Components hub, 2026-05 epoch.ai

    In gigawatts17.3–24.7 GW/yr at the wall it could feed · Q4 2025, annualized · 3% of supply booked to other products

    0–35 GW a year

    How this is counted

    From 9.2 GW a year of chip power that HBM could feed (8.3–10.1), Q4 2025 annualized; 3% of the quarter's supply is booked to other products.

    The bar runs 0–35 GW a year at the wall, the axis the chip, HBM and packaging steps share.

    GW a year at the wall of accelerators that a year of HBM output could feed. The quarter's whole HBM supply × the accelerator TDP that its four tracked buyers (Nvidia, AMD, Google, Amazon) shipped per HBM dollar that quarter, annualized, then to the wall at ×2.09–2.44. Epoch assigns nearly all of the quarter's HBM to those buyers' chips, so this is the accelerators' number seen from the memory side: it shows where the stage binds, not a ceiling with room to spare.

    source: Epoch AI chip-components and chip-sales hubs × Epoch's data-center power chain epoch.ai

    SK Hynix Samsung Micron Nanya Technology

    SK Hynix sells to Penguin Solutions

    Samsung sells to Penguin Solutions

    Micron sells to Penguin Solutions

    6 deals at this step newest first
  5. Advanced packaging

    The step that bonds a chip to its memory stacks — at TSMC, the CoWoS process — on substrates that need Ajinomoto's ABF film and boards from makers like Ibiden and Kinsus; outside firms such as Amkor and SPIL package and test chips for other companies, KYEC tests them, and GlobalFoundries has agreed to make silicon interposers for CoWoS at its plant in New York.

    The record TSMC CoWoS ~198k wafers/qtr · Nvidia ~59% · ABF: Ajinomoto >95%
    197,813 wafers/qtr (183,750–211,875) Epoch AI Chip Components hub, 2026-05 epoch.ai

    In gigawatts20.5–27.5 GW/yr at the wall it could carry · Q4 2025, annualized · 16% of supply booked to other products

    0–35 GW a year

    How this is counted

    From 10.5 GW a year of chip power that CoWoS could carry (9.8–11.3), Q4 2025 annualized; 16% of the quarter's supply is booked to other products.

    The bar runs 0–35 GW a year at the wall, the axis the chip, HBM and packaging steps share.

    GW a year at the wall of accelerators that a year of CoWoS output could carry. The quarter's whole CoWoS supply × the accelerator TDP that its four tracked users (Nvidia, AMD, Google, Amazon) shipped per wafer that quarter, annualized, then to the wall at ×2.09–2.44. Epoch books the rest of the wafers to 'Other', products it doesn't split by designer, so this is a ceiling only if every wafer went to accelerators.

    source: Epoch AI chip-components and chip-sales hubs × Epoch's data-center power chain epoch.ai

    TSMC Ajinomoto Ibiden Kinsus Amkor SPIL KYEC GlobalFoundries

    Amkor sells to TSMC · Qualcomm

    7 deals at this step newest first
  6. AI chips for Accelerators

    The processors AI runs on. Nvidia's GPUs dominate and AMD is the only large GPU rival; Broadcom co-designs custom chips for buyers routing around Nvidia, Marvell designs them for the cloud giants, Qualcomm has begun to for Amazon, and MediaTek has built its first AI accelerator for an American cloud company; Google, Amazon, Cerebras and Groq have designs of their own, and Taalas, which AMD agreed to buy in 2026, builds one AI model into each chip.

    The record Nvidia dominant

    In gigawatts19.2–34.9 GW/yr at the wall · last four quarters, to Sep 2026 at the latest

    0–35 GW a year

    How this is counted

    From 11.5 GW a year of chip power shipped (9.2–14.3) in each maker's last four complete quarters, to Sep 2026 at the latest; 14.5 GW a year at the latest quarter's pace.

    The bar runs 0–35 GW a year at the wall, the axis the chip, HBM and packaging steps share.

    GW a year at the wall for the AI accelerators that the six makers Epoch tracks shipped in a year. Epoch's count of accelerator TDP shipped by Nvidia, Google, AMD, Amazon, Huawei and Cambricon over each maker's last four complete quarters: Nvidia's to September 2026, the other five's to June 2026, each maker's latest complete quarter in Epoch's data. Chips from makers Epoch doesn't track are not in it. To the wall with Epoch's own chain: ×1.53 for the rest of the server, ×1.14 for networking and other IT, then peak PUE of 1.2 (hyperscalers) to 1.4, which makes ×2.09–2.44. The band runs from Epoch's P5 at ×2.09 to its P95 at ×2.44; P5s and P95s are summed across makers, which widens it.

    source: Epoch AI chip-sales hub (accelerator TDP shipped) × Epoch's data-center power chain epoch.ai

    Nvidia AMD Broadcom Marvell Qualcomm MediaTek Google Trainium2/3 Cerebras Groq Taalas

    21 deals at this step newest first
    1. Nov 17, 2026 updated Nov 21
    2. Sep 8, 2026 updated Nov 20
    3. Aug 6, 2026 updated Nov 18
    4. Jul 24, 2026 updated Nov 21
    5. Jul 22, 2026 updated Nov 20
    6. Apr 19, 2026 updated Nov 21
    7. Apr 14, 2026 updated Nov 21
    8. Feb 24, 2026 updated Nov 21
    9. Jan 14, 2026 updated Nov 19
    10. Dec 24, 2025 updated Oct 8
    11. Nov 21, 2025 updated Nov 19
    12. Oct 29, 2025 updated Nov 18
    13. Oct 23, 2025 updated Nov 21
    14. Oct 13, 2025 updated Nov 19
    15. Oct 6, 2025 updated Nov 21
    16. Sep 22, 2025 updated Nov 18
    17. Sep 4, 2025 updated Nov 21
    18. Mar 7, 2025 updated Nov 20
    19. Feb 2025 updated Nov 21
    20. Dec 2, 2024 updated Nov 19
    21. Sep 30, 2024 updated Nov 19
  7. Networking & optics

    The switches, optical modules, lasers and fiber that wire thousands of chips into one cluster.

    The record Broadcom/Nvidia fabrics · 200G EML lasers the pinch

    Broadcom Nvidia Marvell Qualcomm Arista Cisco Nokia Lumentum Coherent Corning InnoLight Applied Optoelectronics Astera Labs Fabrinet

    Corning sells to Meta · Amazon / AWS

    11 deals at this step newest first
    1. Oct 28, 2026 updated Nov 21
    2. Sep 8, 2026 updated Nov 20
    3. May 6, 2026 updated Nov 18
    4. Mar 2, 2026 updated Nov 19
    5. Feb 2026 updated Nov 5
    6. Mar 18, 2025 updated Nov 19
    7. Mar 13, 2025 updated Nov 21
    8. Mar 13, 2025 updated Nov 20
    9. Dec 2, 2024 updated Nov 19
    10. Feb 20, 2024 updated Nov 18
    11. Feb 23, 2023 updated Nov 21
  8. Servers, storage & cooling

    The clusters the chips are built into, the hard drives and flash storage that hold the data, and the power-distribution and cooling gear inside each hall.

    Penguin Solutions Dell Supermicro Seagate Western Digital Sandisk Backblaze Vertiv Fabrinet

    7 deals at this step newest first
    1. Oct 27, 2026 updated Nov 21
    2. Jun 23, 2026 updated Nov 18
    3. May 7, 2026 updated Nov 21
    4. Jan 29, 2026 updated Nov 18
    5. Mar 13, 2025 updated Nov 20
    6. Jul 2024 updated Nov 19
    7. Jun 21, 2024 updated Nov 18
  9. Power for Power equipment & grid

    The electricity the data centers run on: new gas turbines, some of them rented to run a data center off the grid, existing nuclear plants and small reactors still being built, fuel cells that make power on site, geothermal plants still to be built for a data-center customer — and the backup generators that keep a data center running if its grid power fails.

    The record GEV turbines sold out through 2028 · <10 GW of 2029 slots · transformers 2–4 yr · queues >5 yr
    122 GW turbine backlog+SRAs GEV 3Q'26 8-K ex-99.1 (no range — disclosed figure; see pw-gev-sra) sec.gov

    In gigawatts15.3–20.0 GW/yr of GE Vernova gas turbines, from shipped to target

    How this is counted

    GW a year of GE Vernova gas turbines, from what it shipped in 2025 to the annual output it targets from 3Q26. Low is what GE Vernova shipped in 2025. High is the annual output it says it is 'on track' to reach in the third quarter of 2026. That is a target neither 3Q26 filing restates: shipments were 4 GW, 3 GW and 5 GW in the first three quarters of 2026 (the releases' whole-GW figures), and 20 GW a year would be 5 GW a quarter. At its April call about a fifth of the gigawatts it had under contract were explicitly for data centers. One maker of three — Siemens Energy, which has its own record, and Mitsubishi Power are not counted here — and gas turbines are only one route to power.

    From the record

    source: the high end, the target, from GE Vernova's 2Q'26 release (linked here); the low end from its 2025 10-K, and what shipped since from its 1Q–3Q'26 releases, each linked from its own line in pw-gev-sra sec.gov

    GE Vernova Siemens Energy Generac Bloom Energy Solaris Energy Infrastructure Stateline Power Constellation Talen Energy NextEra Energy Vistra Entergy Louisiana Entergy Arkansas Georgia Power DTE Electric NiSource AES Indiana Xcel Energy NV Energy AEP Black Hills Corp. OG&E Fervo Energy Kairos Power X-energy TerraPower Oklo Ormat Technologies Quaise Energy

    28 deals at this step newest first
    1. Nov 17, 2026 updated Nov 21
    2. Nov 11, 2026 updated Nov 21
    3. Oct 16, 2026 updated Nov 21
    4. Oct 14, 2026 updated Nov 19
    5. Oct 5, 2026 updated Nov 21
    6. Sep 30, 2026 updated Nov 21
    7. Sep 16, 2026 updated Nov 18
    8. Sep 1, 2026 updated Nov 21
    9. Jul 22, 2026 updated Nov 21
    10. May 7, 2026 updated Nov 18
    11. Apr 22, 2026 updated Nov 21
    12. Mar 17, 2026 updated Nov 21
    13. Mar 2, 2026 updated Nov 21
    14. Feb 24, 2026 updated Nov 21
    15. Feb 17, 2026 updated Nov 18
    16. Jan 9, 2026 updated Nov 21
    17. Nov 20, 2025
    18. Oct 30, 2025 updated Nov 21
    19. Oct 27, 2025 updated Nov 21
    20. Oct 13, 2025 updated Nov 21
    21. Sep 5, 2025 updated Nov 21
    22. Jul 24, 2025 updated Nov 18
    23. Feb 20, 2025 updated Nov 21
    24. Oct 30, 2024 updated Nov 20
    25. Oct 16, 2024 updated Nov 19
    26. Oct 14, 2024 updated Nov 19
    27. Sep 20, 2024 updated Nov 21
    28. Mar 2024 updated Nov 19
  10. Data centers

    The buildings that hold the chips — built by developers and landlords, several of them former bitcoin miners, and leased to the clouds and labs.

    In gigawatts19.5 GW at the frontier sites Epoch tracks (13.9–27.3, Nov 2026; 96 sites)

    then 22.1 GW end-2026 (15.8–30.9) → 32.6 GW end-2027 (23.3–45.6) → 50 GW end-2028 (35.7–70.0)

    How this is counted

    GW of facility power at the frontier AI sites Epoch tracks worldwide. Facility power — what the whole site draws at its peak, cooling included, not its IT load — summed over every site in Epoch's hub, each at its latest timeline point on or before the date. Points after today are Epoch's projections from permits, construction and announcements. Its timelines end by 2028 for all but two sites, most of them in 2026 or 2027, and a site counts at its last point after that. The band, ×1.4 either way, is the accuracy Epoch states for one site's IT power, 80% of the time: an 80% interval, where the chip stages use P5–P95. It is applied here to facility power, and kept for the sum because many sites rest on the same models and defaults (Epoch's cooling model, its peak-PUE defaults), so their errors need not cancel.

    source: Epoch AI Frontier Data Centers hub (CC-BY), fetched 2026-11-17; the band from its methodology page epoch.ai

    Crusoe Core Scientific Aligned DC Digital Realty Khazna SB Energy TeraWulf Riot Platforms Hut 8 Cipher Digital CleanSpark Digi Power X Zone Frontier AIB Data Centers EverWind Fermi America Kazakhtelecom Blackstone Digital Infrastructure Trust Applied Digital TCS HyperVault WhiteFiber Blockfusion

    Sites Stargate Abilene · Colossus 1 · Colossus 2 · Project Rainier · Meta Prometheus · Meta Hyperion

    31 deals at this step newest first
    1. Nov 16, 2026 updated Nov 21
    2. Oct 16, 2026 updated Nov 21
    3. Sep 30, 2026 updated Nov 21
    4. Aug 10, 2026 updated Nov 21
    5. Aug 10, 2026
    6. Jul 29, 2026
    7. Jul 27, 2026 updated Nov 21
    8. Jul 6, 2026 updated Nov 20
    9. Jul 2026 updated Nov 21
    10. Jun 30, 2026
    11. by Jun 6, 2026 updated Nov 21
    12. May 6, 2026 updated Nov 19
    13. May 5, 2026 updated Nov 21
    14. May 4, 2026 updated Nov 21
    15. Dec 18, 2025
    16. Dec 17, 2025 updated Nov 20
    17. Dec 4, 2025
    18. Nov 12, 2025 updated Nov 21
    19. Nov 3, 2025 updated Nov 21
    20. Oct 28, 2025 updated Nov 21
    21. Oct 21, 2025 updated Nov 21
    22. Sep 25, 2025 updated Nov 20
    23. Sep 23, 2025 updated Nov 20
    24. Aug 14, 2025 updated Nov 20
    25. Aug 2025 updated Nov 21
    26. Jul 31, 2025 updated Nov 19
    27. Jun 2, 2025 updated Nov 20
    28. May 22, 2025 updated Nov 21
    29. Mar 28, 2025 updated Nov 20
    30. Oct 15, 2024 updated Nov 18
    31. Jun 21, 2024 updated Nov 18
  11. Clouds & compute providers for Cloud & capacity

    The companies that run the data centers and sell the computing — the big clouds and the newer GPU-rental 'neoclouds'.

    The record hyperscalers + neoclouds

    Microsoft Amazon / AWS Google Oracle CoreWeave Nebius IREN Nscale Akamai Lambda Fluidstack TensorWave RUM Group ChronoScale Boost Run Crusoe HIVE Digital SpaceX SK Telecom Bell Canada NAVER KT G42 Firebird HUMAIN

    37 deals at this step newest first
    1. Nov 16, 2026
    2. Oct 30, 2026 updated Nov 21
    3. Oct 6, 2026 updated Nov 19
    4. Sep 24, 2026 updated Nov 21
    5. Sep 23, 2026 updated Nov 18
    6. Aug 31, 2026 updated Nov 19
    7. Aug 26, 2026 updated Nov 19
    8. Aug 24, 2026
    9. Aug 19, 2026 updated Nov 21
    10. Jun 22, 2026 updated Nov 18
    11. Jun 18, 2026 updated Nov 18
    12. Jun 15, 2026 updated Nov 18
    13. Jun 5, 2026 updated Nov 21
    14. May 28, 2026 updated Nov 19
    15. May 7, 2026 updated Nov 21
    16. May 6, 2026 updated Nov 21
    17. Apr 2026 updated Nov 20
    18. Apr 2026 updated Nov 19
    19. Nov 18, 2025 updated Nov 20
    20. Nov 3, 2025 updated Nov 21
    21. Nov 3, 2025 updated Nov 19
    22. Nov 3, 2025 updated Nov 19
    23. Nov 2025 updated Nov 21
    24. Oct 28, 2025 updated Nov 21
    25. Oct 28, 2025 updated Nov 18
    26. Oct 2025
    27. Sep 25, 2025 updated Nov 19
    28. Sep 16, 2025 updated Nov 20
    29. Sep 2025 updated Oct 7
    30. Sep 2025 updated Nov 21
    31. Sep 2025 updated Nov 19
    32. Jul 2, 2025 updated Nov 21
    33. Jun 2025 updated Nov 21
    34. Mar 2025 updated Nov 19
    35. Jun 4, 2024 updated Nov 18
    36. Apr 15, 2024 updated Nov 18
    37. Jun 1, 2023 updated Nov 20
  12. AI labs for Frontier labs

    Where the chain ends: the labs that train and run the models, and buy the computing everything above produces.

    The record 10+ players

    In gigawatts17.2–39.9 GW signed or in prospect for three labs

    OpenAI 14.9–22.2 GW · Anthropic 1.1–10.6 GW · Meta 1.2–7.1 GW · chip deals not added: 38.5 GW announced, 3.3 GW binding

    How this is counted

    Chip deals, beside the sites and never added: 38.5 GW announced, 3.3 GW binding.

    GW of sites and compute that OpenAI, Anthropic and Meta have signed for, or have in prospect, as our records state them. Low counts signed contracts only. High adds options, announced scope, deals not yet signed or only under a letter of intent, and contested or press figures. A site counts when a lab leases it, through a subsidiary too (Meta's at Sandersville), or even from a venture it part-owns and builds for: Meta's deals at Hyperion and El Paso, though Hyperion counts in the high end only, because the Hyperion record's one gigawatt figure is the contested expanded size. A site a lab owns outright has no contract to count, which is why xAI is not here (the next stage counts Colossus). A megawatt counts once: the compute Anthropic reportedly buys at Hut 8's Beacon Point through Lambda is not added to the optional capacity Hut 8 offered Anthropic across its pipeline, which may already include it. Nor is the 'more than 1 gigawatt' of Anthropic compute 'in Fluidstack-based sites' added, the capacity that the first $35B of the financing platform Broadcom set up with Apollo and Blackstone serves (the Anthropic–Fluidstack record): Broadcom's release calls it 'previously-announced', and it reads as the Google TPU gigawatt for 2026 counted here, though no party says they are one. No party names those sites either, but River Bend, counted firm, is dated outside that 2026 gigawatt: Hut 8 targets its first data hall for the second quarter of 2027, and its November 2026 results say construction is on track. Oracle's newer sites for OpenAI count in the high end only: their figure is an 'up to', and the parties' own texts disagree on whether a later total includes Abilene. So do Anthropic's 'up to' 5 GW of Amazon Trainium capacity and 'up to' 1 GW of Nvidia systems from its deal with Microsoft and Nvidia. That announcement also has Anthropic contract 'additional compute capacity up to one gigawatt' beside its Azure purchase without saying whether it is the same gigawatt, so one is counted. Nor does any document say whether Anthropic's Nvidia-powered rentals fill part of it: its prospectus names the Nvidia partnership together with its arrangements with SpaceXAI and other clouds. So Colossus 1, rented from SpaceX, is listed but not added, as Beacon Point is not. Critical-IT load goes to the wall at ×1.2 (low) and ×1.4 (high), Epoch's peak-PUE defaults; a grid or utility draw, or a figure that doesn't say which, counts as filed. Chip deals are listed beside the sites they would fill, never added to them. Nvidia's letter of intent with OpenAI is listed, its outcome disputed: no filing says whether it still stands, and the press reports it superseded by the $30B OpenAI counts from Nvidia in its 2026 round. Broadcom's TPUs for Anthropic from 2027 are disputed at 3.5–5 GW, and the chip total takes the 3.5 GW Broadcom filed. Not counted: contracts our records state only in dollars (OpenAI's with Microsoft, the labs' with CoreWeave, Meta's cloud rentals).

    Firm — counted in the low and the high

    Upside — counted in the high only

    Chip deals — beside the sites, never added

    From the record

    source: our deal records — every part names its own

    OpenAI Anthropic Google Meta SpaceX Reflection AI Poolside World Labs Thinking Machines Lab Cohere

    3 deals at this step newest first