Sublayers scored under the 5LC Importance Index list every company scored, in rank order, with the ten marked. Neither the scores nor the ranks are investment recommendations. A company can appear in several sublayers when it performs distinct activities; each placement is scored on its own.
Energy
Fuel
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Fuel for the electricity data centres use, weighted by each fuel's share of it.
- Parts
- 42% Coal for power: About 30% of data-centre electricity, mostly in China.
- 30% Natural gas production: 26% of data-centre electricity; production an estimated 82% of its cost.
- 7% Gas pipelines and LNG: An estimated 18% of delivered gas cost.
- 9% Uranium mining: Nuclear is 15% of data-centre electricity; uranium an estimated 43% of fuel cost.
- 2% Uranium conversion: An estimated 10% of nuclear fuel cost.
- 8% Uranium enrichment: An estimated 39% of nuclear fuel cost.
- 2% Fuel fabrication: An estimated 8% of nuclear fuel cost.
- Shares from
- Coal supplied to power plants (tonnes); China 58%, US 23%, other 19%; Share of the region's marketed gas production (volume) or of its gas transport and LNG revenue (US dollars); regions US 71%, China 3%, other 26%; Stage output (tonnes U or SWU); US 60%, China 8%, other 32%
- Hard to replace
- Coal and gas have many producers, but pipelines take years to permit and build. Nuclear fuel passes through a few conversion, enrichment and fabrication plants that take years to license.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Coal, natural gas and nuclear fuel supplied to the power plants that generate data-centre electricity: coal and gas production, gas pipelines and LNG, and uranium mining, conversion, enrichment and fuel fabrication. Wind, solar and hydro burn no fuel; bioenergy, oil and other minor sources are left out. Running the plants is Generation; making their equipment is Power equipment.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Coal for power | 42% | 25.4–59.2% | International Energy Agency · 10 April 2025 |
| Natural gas production | 30% | 17.9–41.9% | International Energy Agency · 10 April 2025 |
| Gas pipelines and LNG | 7% | 3.4–9.4% | U.S. Energy Information Administration · 30 September 2026 |
| Uranium mining | 9% | 5.4–12.7% | Cameco Corporation / U.S. Securities and Exchange Commission · 13 February 2026 |
| Uranium conversion | 2% | 1.3–3% | Cameco Corporation / U.S. Securities and Exchange Commission · 13 February 2026 |
| Uranium enrichment | 8% | 4.9–11.5% | Centrus Energy / U.S. Securities and Exchange Commission · 11 February 2026 |
| Fuel fabrication | 2% | 1–3.8% | World Nuclear Association · 24 March 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | UrencoTop 10 | 3.5% | 100 | 100 |
| 2 | RosatomTop 10 | 3.7% | 82 | — |
| 3 | OranoTop 10 | 2.9% | 43 | — |
| 4 | CamecoTop 10 | 2.5% | 38 | 5 |
| 5 | Coal IndiaTop 10 | 2% | 32 | 40 |
| 6 | KazatompromTop 10 | 1.6% | 28 | 95 |
| 7 | CHN EnergyTop 10 | 3.7% | 24 | 37 |
| 8 | National Iranian Oil CompanyTop 10 | 0.7% | 22 | — |
| 9 | CNNC | 1.3% | 19 | — |
| 10 | Gazprom | 1.2% | 14 | — |
| 11 | Cheniere Energy | 0.8% | 12 | — |
| 12 | Jinneng Holding Group | 2.1% | 12 | 21 |
| 13 | Peabody Energy | 2.2% | 10 | — |
| 14 | China National Petroleum Corporation (PetroChina, Kunlun) | 0.6% | 9.3 | — |
| 15 | QatarEnergy | 0.5% | 8.1 | — |
| 16 | Kinder Morgan | 0.4% | 7.0 | — |
| 17 | China General Nuclear Power Group (CGN) | 0.7% | 6.8 | — |
| 18 | Williams | 0.4% | 6.6 | — |
| 19 | Energy Transfer | 0.4% | 6.3 | — |
| 20 | Venture Global | 0.4% | 6.1 | — |
| 21 | EQT | 1.2% | 5.9 | — |
| 22 | Expand Energy | 1.3% | 5.7 | — |
| 23 | Core Natural Resources | 1.4% | 5.7 | 0 |
| 24 | Shandong Energy GroupRising | 1.3% | 5.3 | 70 |
| 25 | China Coal (China National Coal Group) | 1.3% | 5.2 | 60 |
| 26 | Westinghouse | 0.4% | 5.0 | — |
| 27 | General Atomics (Quasar, Heathgate; ConverDyn partner) | 0.5% | 4.8 | — |
| 28 | Shaanxi Coal and Chemical Industry GroupRising | 1.2% | 4.8 | 72 |
| 29 | BHP (Mt Arthur coal, Olympic Dam uranium) | 0.7% | 4.7 | — |
| 30 | Navoiyuran (Navoi Mining) | 0.6% | 4.5 | — |
| 31 | EDF | 0.4% | 4.4 | — |
| 32 | NACCO Industries (North American Coal) | 0.6% | 3.9 | — |
| 33 | GE Vernova | 0.3% | 3.7 | — |
| 34 | China Huaneng Group | 0.9% | 3.3 | — |
| 35 | Solstice Nuclear (Metropolis Works) | 0.2% | 3.3 | — |
| 36 | Navajo Transitional Energy Company | 0.9% | 3.2 | — |
| 37 | TC Energy | 0.2% | 2.6 | — |
| 38 | Enbridge | 0.2% | 2.5 | — |
| 39 | Eagle Specialty Materials (FM Coal) | 0.7% | 2.4 | — |
| 40 | Alliance Resource Partners | 0.7% | 2.4 | — |
| 41 | ExxonMobil | 0.7% | 2.3 | — |
| 42 | SPIC | 0.6% | 2.3 | — |
| 43 | Berkshire Hathaway Energy | 0.2% | 2.2 | — |
| 44 | Chevron | 0.7% | 2.1 | — |
| 45 | American Consolidated Natural Resources (ACNR Holdings) | 0.6% | 2.0 | — |
| 46 | PipeChina (China Oil & Gas Pipeline Network Corporation) | 0.1% | 1.8 | — |
| 47 | ConocoPhillips | 0.6% | 1.8 | — |
| 48 | Coterra Energy | 0.5% | 1.7 | — |
| 49 | ONEOK | 0.2% | 1.7 | — |
| 50 | Inner Mongolia Energy Group | 0.4% | 1.6 | — |
| 51 | EOG Resources | 0.5% | 1.6 | — |
| 52 | Sonatrach | 0.3% | 1.3 | — |
| 53 | BP | 0.4% | 1.3 | — |
| 54 | Targa Resources | 0.2% | 1.3 | — |
| 55 | Mitsubishi Corporation (Adamas Energy) | 0.4% | 1.2 | — |
| 56 | Petronas | 0.3% | 1.2 | — |
| 57 | Freeport LNG | 0.1% | 1.2 | — |
| 58 | Antero Resources | 0.4% | 1.2 | — |
| 59 | Ascent Resources | 0.4% | 1.2 | — |
| 60 | Shell | 0.3% | 1.2 | — |
| 61 | Range Resources | 0.3% | 1.1 | — |
| 62 | Western Midstream | 0.1% | 1.0 | — |
| 63 | Lu'an Chemical Group | 0.4% | 1.0 | — |
| 64 | MPLX | 0.1% | 1.0 | — |
| 65 | Diamondback Energy | 0.3% | 1.0 | — |
| 66 | Occidental Petroleum | 0.3% | 1.0 | — |
| 67 | China Huadian Corporation | 0.3% | 0.9 | — |
| 68 | Foresight Energy | 0.3% | 0.9 | — |
| 69 | Nuclear Fuel Complex (India, DAE) | <0.1% | 0.9 | — |
| 70 | Glencore (thermal coal) | 0.3% | 0.9 | — |
| 71 | SUEK (Siberian Coal Energy Company) | 0.3% | 0.9 | — |
| 72 | Novatek | 0.3% | 0.8 | — |
| 73 | Enterprise Products Partners | 0.1% | 0.8 | — |
| 74 | CNX Resources | 0.3% | 0.8 | — |
| 75 | Saudi Aramco | 0.3% | 0.8 | — |
| 76 | Huaihe Energy Holding Group (Huainan Mining) | 0.3% | 0.8 | — |
| 77 | Henan Energy Group | 0.3% | 0.8 | — |
| 78 | Inner Mongolia Yitai Group | 0.3% | 0.7 | — |
| 79 | Comstock Resources | 0.3% | 0.7 | — |
| 80 | Devon Energy | 0.3% | 0.7 | — |
| 81 | Hilcorp Energy | 0.3% | 0.7 | — |
| 82 | TotalEnergies | 0.2% | 0.7 | — |
| 83 | Bumi Resources (KPC, Arutmin) | 0.2% | 0.7 | — |
| 84 | Woodside Energy | 0.1% | 0.7 | — |
| 85 | Ovintiv | 0.2% | 0.7 | — |
| 86 | Vistra | 0.2% | 0.7 | — |
| 87 | KEPCO | <0.1% | 0.6 | — |
| 88 | Singareni Collieries Company | 0.2% | 0.6 | — |
| 89 | Turkmengaz | 0.2% | 0.6 | — |
| 90 | Banpu (Indo Tambangraya Megah, Centennial, BKV) | 0.2% | 0.6 | — |
| 91 | National Fuel Gas (Seneca, National Fuel pipelines) | 0.2% | 0.6 | — |
| 92 | Equinor | 0.2% | 0.5 | — |
| 93 | Adaro Andalan Indonesia | 0.2% | 0.5 | — |
| 94 | Shanxi Coking Coal Group | 0.2% | 0.5 | — |
| 95 | Boardwalk Pipelines (Loews) | 0.1% | 0.5 | — |
| 96 | BWX Technologies (BWXT Nuclear Energy Canada) | <0.1% | 0.5 | — |
| 97 | Bayan Resources | 0.2% | 0.5 | — |
| 98 | Westmoreland Mining Holdings | 0.2% | 0.5 | — |
| 99 | Prairie State Generating Company (Lively Grove mine) | 0.2% | 0.5 | — |
| 100 | Rosneft | 0.2% | 0.5 | — |
| 101 | Peter Kiewit Sons' (Kiewit Mining) | 0.2% | 0.5 | — |
| 102 | ADNOC (ADNOC Gas) | 0.2% | 0.5 | — |
| 103 | Global Mining Group | 0.2% | 0.5 | — |
| 104 | RWE | 0.2% | 0.5 | — |
| 105 | Eni | 0.2% | 0.5 | — |
| 106 | Paladin Energy | 0.1% | 0.4 | — |
| 107 | Sinopec Group | 0.2% | 0.4 | — |
| 108 | CNOOC | 0.1% | 0.4 | — |
| 109 | PGE Polska Grupa Energetyczna | 0.1% | 0.4 | — |
| 110 | Energy Fuels | 0.1% | 0.3 | — |
| 111 | NNPC | 0.1% | 0.3 | — |
| 112 | Elektroprivreda Srbije | 0.1% | 0.3 | — |
| 113 | LEAG (Lausitz Energie) | 0.1% | 0.3 | — |
| 114 | NLC India | 0.1% | 0.3 | — |
| 115 | Ur-Energy | 0.1% | 0.2 | — |
| 116 | INPEX | 0.1% | 0.2 | — |
| 117 | Nuclear Fuel Industries | <0.1% | 0.2 | — |
| 118 | Sempra | <0.1% | 0.1 | — |
| 119 | Centrus Energy | <0.1% | 0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 74.2% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.82.
Ratings more than a level apart were settled by a third, reconciling pass: Gazprom (Gas pipelines and LNG), Nuclear Fuel Complex (India, DAE) (Fuel fabrication).
Limits of the evidence
- The fuel mix is IEA's estimate for 2024; no later published mix was readable for this card.
- Regional factors in shareData are 5LayerCake estimates combining IEA's global, US and China figures with China's 2025 national power mix (Ember); each company's regional shares are combined with them.
- The 'other' region's coal is dominated by India's power coal, although India's data-centre load is small.
- The gas transport share is the US delivered price to power plants minus Henry Hub, which also contains regional price differences; LNG costs for gas used outside the United States are not measured separately.
- Nuclear stage costs use WNA quantities and its 2021 fabrication price with 2025 uranium and conversion term prices and the end-2025 enrichment spot price; term enrichment prices are not public.
- Enrichment output is rarely published by company; capacity (WNA table) is off basis.
- Company supply to power plants by region is often unpublished, especially for Chinese coal groups, which report total sales.
- Coal moved by general freight railways and ships is outside this sublayer.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Urenco, Rosatom, Orano, Cameco, Coal India, Kazatomprom, CHN Energy, National Iranian Oil Company. Rising places: Shaanxi Coal and Chemical Industry Group, Shandong Energy Group.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Generation
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Electricity generated by plants a company owns, in the regions where data centres draw power, weighted by each region's share of data-centre electricity use. A utility's power plants count here; its network counts in Utilities.
- Parts
- 45% United States: 45% of data-centre electricity use in 2024, the largest region.
- 25% China: 25% of data-centre electricity use in 2024.
- 15% Europe: 15% of data-centre electricity use in 2024.
- 15% Rest of world: The other 15%, including Japan, Korea, India, Southeast Asia and the Middle East.
- Shares from
- Company share of the region's net generation (TWh, by ownership share of each plant), from company reports and national totals
- Hard to replace
- A plant's output can only be replaced by other plants on the same grid. Where data centres cluster, spare capacity is thin, and new plants take years; gas turbine order books stretch past 2029.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Companies that own power plants, credited by their ownership share of the electricity the plants generate in the regions where data centres draw power, including plants owned by regulated utilities. Delivering power to data centres through networks is Utilities: a utility's plants count here and its wires count there. Fuel is Fuel and plant equipment is Power equipment.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| United States | 45% | 27–63% | Carbon Brief · 15 September 2025 |
| China | 25% | 15–35% | Carbon Brief · 15 September 2025 |
| Europe | 15% | 9–21% | Carbon Brief · 15 September 2025 |
| Rest of world | 15% | 9–21% | Carbon Brief · 15 September 2025 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | EDFTop 10 | 2.3% | 100 | 45 |
| 2 | ConstellationTop 10 | 3% | 96 | 3 |
| 3 | NextEra EnergyTop 10 | 2.4% | 55 | — |
| 4 | Duke EnergyTop 10 | 2.1% | 54 | 27 |
| 5 | Southern CompanyTop 10 | 1.9% | 48 | 68 |
| 6 | VistraTop 10 | 2.1% | 37 | 100 |
| 7 | Tennessee Valley AuthorityTop 10 | 1.4% | 34 | 45 |
| 8 | CHN EnergyTop 10 | 2.9% | 26 | 18 |
| 9 | Dominion EnergyRising | 1.2% | 23 | 73 |
| 10 | Berkshire Hathaway EnergyRising | 1.2% | 21 | 91 |
| 11 | Entergy | 1.2% | 18 | 23 |
| 12 | KEPCO | 0.5% | 18 | — |
| 13 | China Huaneng Group | 2.1% | 16 | 36 |
| 14 | AEP | 0.9% | 14 | — |
| 15 | Egyptian Electricity Holding Company | 0.3% | 13 | — |
| 16 | Xcel Energy | 0.7% | 12 | — |
| 17 | China Huadian Corporation | 1.7% | 11 | — |
| 18 | SPIC | 1.7% | 11 | 68 |
| 19 | Eskom | 0.2% | 10 | — |
| 20 | Talen Energy | 0.7% | 9.8 | — |
| 21 | NRG Energy | 0.8% | 9.5 | — |
| 22 | Taiwan Power Company | 0.2% | 9.1 | — |
| 23 | Hydro-Québec | 0.2% | 8.9 | — |
| 24 | NTPC | 0.5% | 8.6 | — |
| 25 | China Datang Corporation | 1.4% | 8.5 | — |
| 26 | China General Nuclear Power Group (CGN) | 0.8% | 8.1 | — |
| 27 | DTE Energy | 0.5% | 7.9 | — |
| 28 | Vattenfall | 0.4% | 7.9 | — |
| 29 | China Three Gorges Corporation | 1% | 7.8 | — |
| 30 | Engie | 0.8% | 7.6 | — |
| 31 | JERA | 0.3% | 6.9 | — |
| 32 | RWE | 0.7% | 6.9 | — |
| 33 | Saudi Electricity Company | 0.3% | 6.8 | — |
| 34 | Zhejiang Energy Group | 0.5% | 6.3 | — |
| 35 | Rosatom | 0.2% | 5.5 | — |
| 36 | Enel | 0.8% | 5.5 | — |
| 37 | Comisión Federal de Electricidad (CFE) | 0.2% | 4.8 | — |
| 38 | ČEZ | 0.2% | 4.7 | — |
| 39 | Statkraft | 0.3% | 4.3 | — |
| 40 | Iberdrola (incl. Avangrid, ScottishPower, Neoenergia) | 0.6% | 4.2 | — |
| 41 | PGE Polska Grupa Energetyczna | 0.2% | 4.2 | — |
| 42 | PLN (Perusahaan Listrik Negara) | 0.2% | 3.8 | — |
| 43 | CNNC | 0.5% | 3.5 | — |
| 44 | RusHydro | 0.2% | 2.9 | — |
| 45 | China Resources Power | 0.5% | 2.7 | — |
| 46 | Uniper | 0.2% | 2.5 | — |
| 47 | Fortum | 0.2% | 2.3 | — |
| 48 | LEAG (Lausitz Energie) | 0.2% | 2.3 | — |
| 49 | Energias de Portugal (EDP) | 0.4% | 2.3 | — |
| 50 | Gazprom (Gazprom Energoholding) | 0.2% | 2.1 | — |
| 51 | Eletrobras (Axia Energia) | 0.2% | 2.0 | — |
| 52 | Invenergy | 0.4% | 1.7 | — |
| 53 | AES | 0.4% | 1.6 | — |
| 54 | Evergy | 0.4% | 1.5 | — |
| 55 | Oglethorpe Power | 0.3% | 1.3 | — |
| 56 | SDIC Power | 0.3% | 1.3 | — |
| 57 | Ameren | 0.3% | 1.3 | — |
| 58 | PPL (LG&E and KU) | 0.3% | 1.3 | — |
| 59 | WEC Energy Group | 0.3% | 1.3 | — |
| 60 | PG&E | 0.3% | 1.2 | — |
| 61 | PSEG (PSEG Power nuclear) | 0.3% | 1.2 | — |
| 62 | Alpha Generation (ArcLight) | 0.3% | 1.2 | — |
| 63 | Salt River Project | 0.3% | 1.1 | — |
| 64 | Alliant Energy | 0.3% | 1.1 | — |
| 65 | New York Power Authority | 0.2% | 1.0 | — |
| 66 | Guangdong Energy Group | 0.2% | 0.9 | — |
| 67 | TotalEnergies (power) | 0.2% | 0.9 | — |
| 68 | Clearway Energy | 0.2% | 0.9 | — |
| 69 | Cogentrix Energy | 0.2% | 0.9 | — |
| 70 | Tenaska | 0.2% | 0.7 | — |
| 71 | Inner Mongolia Energy Group | 0.2% | 0.7 | — |
| 72 | Vietnam Electricity (EVN) | 0.1% | 0.5 | — |
| 73 | Axpo | 0.1% | 0.5 | — |
| 74 | Verbund | 0.1% | 0.5 | — |
| 75 | Tenaga Nasional | 0.1% | 0.5 | — |
| 76 | State Grid | 0.1% | 0.4 | — |
| 77 | Public Power Corporation (PPC) | 0.1% | 0.4 | — |
| 78 | Ontario Power Generation | 0.1% | 0.4 | — |
| 79 | Itaipu Binacional | 0.1% | 0.4 | — |
| 80 | En+ Group (EuroSibEnergo) | 0.1% | 0.4 | — |
| 81 | BC Hydro | 0.1% | 0.2 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 53.8% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.62.
Ratings more than a level apart were settled by a third, reconciling pass: China Three Gorges Corporation (China), Egyptian Electricity Holding Company (Rest of world), Taiwan Power Company (Rest of world).
Limits of the evidence
- Regional weights are IEA's 2024 estimates, cited through Carbon Brief; Chinese government estimates of data-centre use are higher than IEA's.
- Within a region every grid counts alike, although data-centre load is concentrated in a few grids (Northern Virginia, Texas, Dublin, Frankfurt, Tokyo).
- The rest-of-world region includes large systems with little data-centre load, which raises their companies' shares relative to data-centre use.
- Not every company reports ownership-adjusted generation; some report consolidated or attributable output.
- Behind-the-meter generation at data-centre sites is counted only where national statistics include it.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: EDF, Constellation, NextEra Energy, Duke Energy, Southern Company, Vistra, Tennessee Valley Authority, CHN Energy. Rising places: Berkshire Hathaway Energy, Dominion Energy.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Grid
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- High-voltage lines, substations and their construction that connect data centres to the grid, up to the site; equipment on the campus is Infrastructure · Power and cooling. The US mix of transmission spending stands in for the world.
- Parts
- 45% Substations: transformers, switchgear and protection: 45% of US utilities' 2025 transmission equipment and construction spending (FERC filings).
- 33% Towers, poles and line construction: 33% of that spending.
- 18% Overhead conductors: 18% of that spending.
- 4% Underground and submarine cables: 4% of that spending; more where lines are buried.
- Shares from
- Company revenue from the function's equipment and installation, as a share of world spending on it
- Hard to replace
- Large power transformers and high-voltage breakers come from a few qualified makers; transformers take up to four years to procure and cables two to three. New lines take years to permit.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Equipment and construction for the high-voltage utility network up to the data-centre site: substations, towers and poles, overhead conductors and underground cables, including the site substation whether built for a utility or a data-centre developer. Medium-voltage switchgear, transformers and cables on the campus after the site substation, and power distribution inside buildings, are Infrastructure · Power and cooling. Owning and operating networks is Utilities; local distribution lines serving homes and businesses are left out.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Substations: transformers, switchgear and protection | 45% | 31.6–58.6% | U.S. Federal Energy Regulatory Commission · undated |
| Towers, poles and line construction | 33% | 23–42.7% | U.S. Federal Energy Regulatory Commission · undated |
| Overhead conductors | 18% | 12.6–23.5% | U.S. Federal Energy Regulatory Commission · undated |
| Underground and submarine cables | 4% | 2.8–5.2% | U.S. Federal Energy Regulatory Commission · undated |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | Hitachi EnergyTop 10 | 6.6% | 100 | 75 |
| 2 | Siemens EnergyTop 10 | 5.6% | 85 | 73 |
| 3 | GE VernovaTop 10 | 3.6% | 43 | 79 |
| 4 | China Electrical EquipmentTop 10 | 3.3% | 21 | 0 |
| 5 | State GridTop 10 | 3.2% | 17 | 11 |
| 6 | Quanta ServicesTop 10 | 2.8% | 17 | 34 |
| 7 | Hyosung Heavy IndustriesTop 10 | 1.3% | 14 | 70 |
| 8 | PrysmianTop 10 | 1.1% | 12 | 74 |
| 9 | HD Hyundai ElectricTop 10 | 1.2% | 10 | 39 |
| 10 | TBEATop 10 | 1.8% | 9.4 | 17 |
| 11 | Mitsubishi Electric | 0.8% | 5.9 | — |
| 12 | NKT | 0.4% | 5.0 | — |
| 13 | LS Corp (group: LS Electric, LS Cable) | 0.8% | 4.9 | — |
| 14 | Nexans | 0.4% | 4.1 | — |
| 15 | Virginia Transformer Corp | 0.5% | 3.8 | — |
| 16 | Toshiba (Toshiba Energy Systems & Solutions) | 0.5% | 3.6 | — |
| 17 | Power Construction Corporation of China (PowerChina) | 0.8% | 3.0 | — |
| 18 | VINCI (Omexom, Cobra IS) | 0.8% | 3.0 | — |
| 19 | Sieyuan Electric | 0.9% | 2.8 | — |
| 20 | Sumitomo Electric Industries | 0.7% | 2.7 | — |
| 21 | MasTec | 0.7% | 2.3 | — |
| 22 | KEC International | 0.6% | 2.2 | — |
| 23 | Zhongtian Technology (ZTT) | 0.6% | 2.1 | — |
| 24 | Valmont Industries | 0.4% | 2.1 | — |
| 25 | APAR Industries | 0.6% | 2.0 | — |
| 26 | Peter Kiewit Sons' | 0.5% | 1.7 | — |
| 27 | ABB | 0.5% | 1.7 | — |
| 28 | Schweitzer Engineering Laboratories | 0.5% | 1.7 | — |
| 29 | China Energy Engineering (CEEC) | 0.4% | 1.6 | — |
| 30 | Hellenic Cables (Cenergy Holdings) | 0.2% | 1.5 | — |
| 31 | MYR Group | 0.6% | 1.4 | — |
| 32 | Larsen & Toubro (Power Transmission & Distribution) | 0.5% | 1.2 | — |
| 33 | Southwire | 0.4% | 1.2 | — |
| 34 | Hengtong Group | 0.3% | 1.2 | — |
| 35 | Pike Electric | 0.5% | 1.1 | — |
| 36 | Kalpataru Projects International | 0.5% | 1.1 | — |
| 37 | Eaton | 0.4% | 1.1 | — |
| 38 | Hubbell | 0.4% | 1.0 | — |
| 39 | Siemens SI | 0.4% | 1.0 | — |
| 40 | Elsewedy Electric | 0.4% | 1.0 | — |
| 41 | Taihan Cable & Solution | 0.2% | 1.0 | — |
| 42 | WEG | 0.4% | 1.0 | — |
| 43 | SGB-SMIT Group | 0.4% | 0.9 | — |
| 44 | Burns & McDonnell | 0.4% | 0.8 | — |
| 45 | Schneider Electric | 0.2% | 0.8 | — |
| 46 | Končar | 0.3% | 0.8 | — |
| 47 | Chint Group | 0.3% | 0.8 | — |
| 48 | Ningbo Orient Wires & Cables | 0.1% | 0.8 | — |
| 49 | Jiangsu Huapeng Transformer | 0.3% | 0.7 | — |
| 50 | Sterlite Electric | 0.2% | 0.7 | — |
| 51 | Furukawa Electric | 0.1% | 0.7 | — |
| 52 | Trench Group | 0.3% | 0.7 | — |
| 53 | Meidensha | 0.3% | 0.6 | — |
| 54 | Sabre Industries | 0.2% | 0.6 | — |
| 55 | Balfour Beatty | 0.2% | 0.5 | — |
| 56 | Skipper | 0.2% | 0.5 | — |
| 57 | Arcosa (utility structures) | 0.2% | 0.5 | — |
| 58 | SPIE | 0.2% | 0.4 | — |
| 59 | Primoris Services (Power Delivery) | 0.1% | 0.3 | — |
| 60 | CG Power and Industrial Solutions | 0.1% | 0.3 | — |
| 61 | Elecnor | 0.1% | 0.3 | — |
| 62 | Linxon | 0.1% | 0.3 | — |
| 63 | Black & Veatch | 0.1% | 0.3 | — |
| 64 | Bharat Heavy Electricals (BHEL) | 0.1% | 0.2 | — |
| 65 | Tata Projects | 0.1% | 0.2 | — |
| 66 | Powell Industries | <0.1% | 0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 70.2% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.70.
No rating was more than a level apart between the two passes.
Limits of the evidence
- Weights use US utilities' 2025 transmission plant additions (FERC Form 1, accounts 352–358); grids in China and Europe have different mixes, with more ultra-high-voltage lines in China and more cable in Europe.
- FERC accounts include installation and utility overheads within each asset; contractor work cannot be separated from equipment cost.
- World spending by function is not published; share denominators must be estimated from grid-investment totals.
- Substations built and owned by data-centre developers are not in FERC data.
- Distribution substations are excluded; including them would raise the substation weight to about 57%.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Hitachi Energy, Siemens Energy, GE Vernova, China Electrical Equipment, State Grid, Quanta Services, Hyosung Heavy Industries, Prysmian. Rising places filled by the next Importance ranks: HD Hyundai Electric, TBEA.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Power equipment
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Generating equipment for data-centre load, by value.
- Parts
- 49% Gas turbines: About a quarter of Siemens Energy's 2025 turbine orders were for data centres; an estimated $10bn worldwide.
- 45% Engines and generator sets: Backup and on-site sets, an estimated $9.1bn.
- 6% Fuel cells: Bloom Energy's 2025 sales, an estimated three-quarters for data centres.
- Shares from
- Turbines over 10 MW for power, excluding oil and gas (MW ordered); Generator sets (1 MW and over) or fuel cells ordered for data centres (MW)
- Hard to replace
- Three makers build most large gas turbines, and Siemens Energy says its capacity is sold out until 2028; other equipment has more suppliers.
- Check
- Since 2025, GE Vernova's heavy-duty gas turbine output has been committed years ahead, with data-centre demand a named driver. Gas equipment backlog plus slot reservations rose from 62 to 83 GW during Q4 2025 and to 116 GW by mid-2026, against about 20 GW a year of output. Industry backlogs stretch past 2029. In November 2025, Siemens Energy said its gas turbine production capacity was sold out until fiscal 2028. It reported about 78 GW of new-unit backlog and the largest share of FY2025 gas turbine orders, with US data-centre programmes among the named orders.
Parts and weights
Generating equipment whose output serves data-centre load: gas turbines, both in utility-scale plants ordered for data-centre demand and on site; reciprocating engines, generator sets and gas turbines under 10 MW for backup and on-site power; and fuel cells. Wind turbines, solar modules, grid and on-site battery storage and nuclear reactor equipment are outside this sublayer; UPS systems and batteries inside the data centre are Infrastructure · Power and cooling, and running the plants is Generation.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Gas turbines | 49% | 33.9–80.9% | Siemens Energy · 20 November 2025 |
| Engines and generator sets | 45% | 16.6–61.1% | Grand View Research · October 2026 |
| Fuel cells | 6% | 1.4–13.8% | Bloom Energy / U.S. Securities and Exchange Commission · 5 February 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | GE VernovaTop 10 | 15.2% | 100 | 48 |
| 2 | Siemens EnergyTop 10 | 13.4% | 88 | 61 |
| 3 | Bloom EnergyTop 10 | 5.7% | 53 | 38 |
| 4 | MHITop 10 | 9% | 39 | 77 |
| 5 | CaterpillarTop 10 | 12.3% | 35 | — |
| 6 | CumminsTop 10 | 8.5% | 22 | 0 |
| 7 | INNIOTop 10 | 4.9% | 13 | — |
| 8 | Rolls-Royce Power Systems (mtu)Top 10 | 4.8% | 11 | — |
| 9 | RehlkoTop 10 | 3.6% | 6.6 | — |
| 10 | Doosan Enerbility | 2% | 6.6 | 33 |
| 11 | Dongfang ElectricRising | 2.6% | 6.1 | 91 |
| 12 | Weichai Power (incl. Baudouin, PSI) | 3.2% | 4.6 | — |
| 13 | Ansaldo Energia | 1.5% | 3.4 | — |
| 14 | Wärtsilä (Energy) | 1.4% | 3.0 | — |
| 15 | Generac | 1.3% | 1.7 | — |
| 16 | Shanghai Electric | 0.8% | 1.7 | — |
| 17 | Baker Hughes | 0.7% | 1.1 | — |
| 18 | Langley Holdings (Bergen Engines, Piller) | 0.5% | 1.0 | — |
| 19 | Yanmar (incl. Himoinsa) | 0.7% | 0.9 | — |
| 20 | HITEC Power Protection | 0.4% | 0.7 | — |
| 21 | China Yuchai International (Yuchai) | 0.5% | 0.6 | — |
| 22 | FuelCell Energy | 0.2% | 0.4 | — |
| 23 | Everllence (formerly MAN Energy Solutions) | 0.4% | 0.4 | — |
| 24 | Kirloskar Oil Engines | 0.4% | 0.4 | — |
| 25 | Harbin Electric | 0.2% | 0.4 | — |
| 26 | Kawasaki Heavy Industries | 0.3% | 0.3 | — |
| 27 | Bharat Heavy Electricals (BHEL) | 0.3% | 0.3 | — |
| 28 | ProEnergy | 0.2% | 0.3 | — |
| 29 | IHI (aeroderivative packages, Niigata Power Systems) | 0.2% | 0.2 | — |
| 30 | HD Hyundai Heavy Industries (HiMSEN engines) | 0.1% | 0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
Since 2025, GE Vernova's heavy-duty gas turbine output has been committed years ahead, with data-centre demand a named driver. Gas equipment backlog plus slot reservations rose from 62 to 83 GW during Q4 2025 and to 116 GW by mid-2026, against about 20 GW a year of output. Industry backlogs stretch past 2029.
GE Vernova ranks 1 of 30 by Importance; the test asks for a place in the top 3: passed.
In November 2025, Siemens Energy said its gas turbine production capacity was sold out until fiscal 2028. It reported about 78 GW of new-unit backlog and the largest share of FY2025 gas turbine orders, with US data-centre programmes among the named orders.
Siemens Energy ranks 2 of 30 by Importance; the test asks for a place in the top 3: passed.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 76.9% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.75.
No rating was more than a level apart between the two passes.
Limits of the evidence
- The data-centre share of turbine orders (about 26%) comes from Siemens Energy's FY2025 order mix and its February 2026 commitments; it counts orders tied directly to data centres, not all utility plants built for data-centre load growth.
- The 85 GW is Siemens Energy's estimate of 2025 gas capacity additions, including combined-cycle steam turbines, not orders; turbine value uses its average of about €400m per GW and about $1.13 per euro.
- The generator-set market is a market-research estimate covering all sizes, including units under 1 MW, while shares count units of 1 MW and above.
- Fuel-cell value is an estimate: Bloom Energy's 2025 product and installation revenue times an estimated 75% data-centre share, which Bloom does not disclose.
- Nuclear reactor equipment and on-site battery storage are not scored in any Energy sublayer, for lack of a published annual data-centre value.
- Order shares differ from deliveries, which lag orders by two to four years for large turbines.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: GE Vernova, Siemens Energy, Bloom Energy, MHI, Caterpillar, Cummins, INNIO, Rolls-Royce Power Systems (mtu). Rising places: Dongfang Electric. Rising places filled by the next Importance ranks: Rehlko.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Utilities
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Data-centre load connected to the wires a company owns, weighted by where data centres use power.
- Parts
- 45% United States: 45% of data-centre electricity use in 2024, the largest region.
- 25% China: 25% of data-centre electricity use in 2024.
- 15% Europe: 15% of data-centre electricity use in 2024.
- 15% Rest of world: The other 15%, including Japan, Korea, India, Southeast Asia and the Middle East.
- Shares from
- Company share of the region's operating data-centre capacity connected to its wires (MW of connected sites, not average load); contracted but unbuilt load is excluded
- Hard to replace
- A data centre can take power only from the network that reaches its site, so there is no second supplier for the connection, and adding capacity needs transmission work that takes years.
- Check
- In July 2022, Dominion Energy told large customers that transmission constraints in eastern Loudoun County (Northern Virginia, the largest data-centre market) would hold up new data-centre connections. It briefly paused, then rationed, new service, and some projects faced delays of up to four years, with reduced power allocations through January 2026.
Parts and weights
Companies that own the transmission or distribution network a data centre connects to, credited with the operating data-centre load connected to their wires, including the network side of integrated utilities; system operators that own no wires, such as PJM, ERCOT or CAISO, are not scored. Data-centre load supplied without the public network (behind the meter, at a power plant or by on-site generation only) stays in the regional total but is credited to no company. Electricity from utility-owned plants is Generation, grid equipment and construction are Grid, and retail suppliers without wires are left out.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| United States | 45% | 27–63% | Carbon Brief · 15 September 2025 |
| China | 25% | 15–35% | Carbon Brief · 15 September 2025 |
| Europe | 15% | 9–21% | Carbon Brief · 15 September 2025 |
| Rest of world | 15% | 9–21% | Carbon Brief · 15 September 2025 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | State GridTop 10 | 21.1% | 100 | — |
| 2 | Dominion EnergyTop 10 | 9.1% | 35 | 13 |
| 3 | Berkshire Hathaway EnergyTop 10 | 4.7% | 7.7 | — |
| 4 | AEPTop 10 | 4.4% | 7.5 | 75 |
| 5 | ESBTop 10 | 1.9% | 4.0 | — |
| 6 | China Southern Power GridTop 10 | 2.8% | 3.0 | — |
| 7 | SempraTop 10 | 2.3% | 3.0 | — |
| 8 | Southern CompanyTop 10 | 2.1% | 2.9 | 88 |
| 9 | TEPCOTop 10 | 1.3% | 2.1 | — |
| 10 | EDFTop 10 | 1.2% | 2.0 | — |
| 11 | SP Group (Singapore Power) | 1.1% | 2.0 | — |
| 12 | Alliander (Liander) | 1.1% | 1.9 | — |
| 13 | Exelon | 1.4% | 1.9 | — |
| 14 | Engie | 1% | 1.8 | — |
| 15 | KEPCO | 1% | 1.7 | — |
| 16 | Tenaga Nasional | 0.9% | 1.1 | — |
| 17 | Grant County PUD | 0.8% | 0.9 | — |
| 18 | Inner Mongolia Power (Group) | 1% | 0.8 | — |
| 19 | Pinnacle West (Arizona Public Service) | 0.9% | 0.7 | — |
| 20 | Portland General Electric | 0.8% | 0.7 | — |
| 21 | CPS Energy | 0.8% | 0.7 | — |
| 22 | Salt River Project | 0.8% | 0.7 | — |
| 23 | Mainova (NRM Netzdienste Rhein-Main) | 0.7% | 0.7 | — |
| 24 | Duke Energy | 0.8% | 0.6 | — |
| 25 | Enel | 0.8% | 0.6 | — |
| 26 | Omaha Public Power District | 0.7% | 0.6 | — |
| 27 | Iberdrola (Avangrid, ScottishPower, Electricity North West, i-DE, Neoenergia) | 0.7% | 0.5 | — |
| 28 | Taiwan Power Company | 0.5% | 0.5 | — |
| 29 | E.ON (Westnetz, Bayernwerk, Syna, E.ON Energidistribution) | 0.6% | 0.5 | — |
| 30 | SSE (SSEN) | 0.5% | 0.4 | — |
| 31 | CLP Holdings (CLP Power Hong Kong) | 0.5% | 0.4 | — |
| 32 | National Grid | 0.5% | 0.4 | — |
| 33 | Rosseti | 0.3% | 0.4 | — |
| 34 | PSEG (PSE&G) | 0.5% | 0.4 | — |
| 35 | NiSource (NIPSCO) | 0.5% | 0.4 | — |
| 36 | PG&E | 0.5% | 0.4 | — |
| 37 | Memphis Light, Gas and Water | 0.5% | 0.3 | — |
| 38 | Silicon Valley Power (City of Santa Clara) | 0.5% | 0.3 | — |
| 39 | TenneT | 0.4% | 0.3 | — |
| 40 | Ausgrid | 0.4% | 0.3 | — |
| 41 | Endeavour Energy | 0.4% | 0.3 | — |
| 42 | Saudi Electricity Company | 0.4% | 0.3 | — |
| 43 | Hydro-Québec | 0.4% | 0.3 | — |
| 44 | PLN (Perusahaan Listrik Negara) | 0.3% | 0.3 | — |
| 45 | Maharashtra State Electricity Distribution (MSEDCL) | 0.4% | 0.3 | — |
| 46 | Kansai Electric Power (Kansai Transmission and Distribution) | 0.3% | 0.2 | — |
| 47 | Dubai Electricity and Water Authority (DEWA) | 0.3% | 0.2 | — |
| 48 | TAQA (Abu Dhabi distribution and transmission) | 0.3% | 0.2 | — |
| 49 | Israel Electric Corporation | 0.2% | 0.2 | — |
| 50 | Comisión Federal de Electricidad (CFE) | 0.3% | 0.2 | — |
| 51 | A2A (Unareti) | 0.3% | 0.2 | — |
| 52 | Enexis | 0.2% | 0.2 | — |
| 53 | Stedin | 0.2% | 0.2 | — |
| 54 | FirstEnergy | 0.4% | 0.1 | — |
| 55 | Black Hills Corporation | 0.4% | 0.1 | — |
| 56 | Ellevio | 0.2% | 0.1 | — |
| 57 | Austin Energy | 0.4% | 0.1 | — |
| 58 | Northern Virginia Electric Cooperative (NOVEC) | 0.4% | 0.1 | — |
| 59 | Eskom | 0.2% | 0.1 | — |
| 60 | Elvia | 0.1% | 0.1 | — |
| 61 | Consolidated Edison | 0.1% | 0.1 | — |
| 62 | Xcel Energy | 0.3% | 0.1 | — |
| 63 | Adani Energy Solutions (Adani Electricity Mumbai) | 0.2% | 0.1 | — |
| 64 | Tamil Nadu Power Distribution Corporation | 0.2% | 0.1 | — |
| 65 | Edison International (Southern California Edison) | 0.3% | 0.1 | — |
| 66 | Evergy | 0.3% | 0.1 | — |
| 67 | Entergy | 0.2% | 0.1 | — |
| 68 | Energias de Portugal (E-Redes, EDP Brasil) | 0.2% | 0.1 | — |
| 69 | Puget Sound Energy | 0.2% | 0.1 | — |
| 70 | Fortis (ITC, Tucson Electric Power, Central Hudson, FortisAlberta) | 0.2% | 0.1 | — |
| 71 | AES | 0.2% | 0.1 | — |
| 72 | PPL (PPL Electric, LG&E and KU) | 0.2% | 0.1 | — |
| 73 | Alliant Energy | 0.2% | 0.1 | — |
| 74 | CenterPoint Energy | 0.2% | 0.1 | — |
| 75 | Ameren | 0.2% | 0.1 | — |
| 76 | WEC Energy Group | 0.2% | 0.1 | — |
| 77 | MDU Resources (Montana-Dakota Utilities) | 0.2% | <0.1 | — |
| 78 | CMS Energy (Consumers Energy) | 0.2% | <0.1 | — |
| 79 | Rappahannock Electric Cooperative | 0.1% | <0.1 | — |
| 80 | Tata Power | 0.1% | <0.1 | — |
| 81 | Telangana Southern Power Distribution Company | 0.1% | <0.1 | — |
| 82 | Vattenfall (Eldistribution) | 0.1% | <0.1 | — |
| 83 | EnBW (Netze BW, TransnetBW) | 0.1% | <0.1 | — |
| 84 | NextEra Energy | 0.1% | <0.1 | — |
| 85 | OGE Energy (OG&E) | 0.1% | <0.1 | — |
| 86 | Tennessee Valley Authority | 0.1% | <0.1 | — |
| 87 | TXNM Energy (PNM, TNMP) | 0.1% | <0.1 | — |
| 88 | DTE Energy | 0.1% | <0.1 | — |
| 89 | Santee Cooper | 0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
In July 2022, Dominion Energy told large customers that transmission constraints in eastern Loudoun County (Northern Virginia, the largest data-centre market) would hold up new data-centre connections. It briefly paused, then rationed, new service, and some projects faced delays of up to four years, with reduced power allocations through January 2026.
Dominion Energy ranks 2 of 89 by Importance; the test asks for a place in the top 3: passed.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 57.4% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.46.
Ratings more than a level apart were settled by a third, reconciling pass: Enexis (Europe), Stedin (Europe).
Limits of the evidence
- Regional weights are IEA's 2024 estimates, cited through Carbon Brief; Chinese government estimates of data-centre use are higher than IEA's.
- Utilities report data-centre load inconsistently: connected MW, sales in TWh, contracted capacity or pipeline; only connected capacity of operating sites is on basis.
- Chinese grid companies do not publish data-centre load, so shares must be estimated from provincial data-centre electricity statistics.
- Where transmission and distribution owners differ, the company whose wires the site connects to is credited, and the input must say which.
- Behind-the-meter data-centre load must be estimated to complete the regional total.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: State Grid, Dominion Energy, Berkshire Hathaway Energy, AEP, ESB, China Southern Power Grid, Sempra, Southern Company. Rising places filled by the next Importance ranks: TEPCO, EDF.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Chips
Tools
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Worldwide 2025 sales of chip design software, fab and test equipment, and wafer materials; AI chips use the same tools as all other chips.
- Parts
- 8% Design software and IP: Every chip is designed with it.
- 12% Lithography machines: Prints each layer's pattern.
- 3% Lithography optics and light-source lasers: Each has a single supplier for EUV machines.
- 38% Wafer processing: Deposits, etches and polishes each layer.
- 9% Inspection and metrology: Finds defects in each layer.
- 7% Test equipment: Testers and probe cards check every chip.
- 6% Silicon wafers: The starting material of every chip.
- 17% Process materials: Resists, gases, chemicals and photomasks.
- Shares from
- EDA and non-processor IP revenue; Machine revenue, net of the subsystems part; Revenue in the function
- Hard to replace
- Switching supplier means requalifying in each fab or design flow, often over years. ASML alone makes EUV machines, with ZEISS optics and TRUMPF lasers.
- Check
- Throughout the AI build-out (2023 to date), ASML has been the world's only manufacturer of EUV lithography systems. Leading-edge (5 nm-class and below) logic, the node family used for current AI accelerators, is not mass-produced without EUV, so ASML is a sole-supplier choke point.
Parts and weights
Chip and package design software with non-processor design IP, front-end wafer fab equipment (with the optics and EUV lasers inside lithography machines as their own part), test equipment with probe cards, and wafer fab materials sold by suppliers; processor-core licensing belongs to Chip design, back-end assembly equipment, package substrates, build-up film and substrate glass cloth to Packaging, and photomasks made in chipmakers' own mask shops to Foundry and Memory, while other package materials, such as mould compound, underfill and solder, are counted in neither Tools nor Packaging.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Design software and IP | 8% | 7.5–10.5% | SEMI Electronic System Design Alliance (PR Newswire distribution) · 13 April 2026 |
| Lithography machines | 12% | 10.5–13.5% | ASML Holding N.V. / U.S. SEC · 28 January 2026 |
| Lithography optics and light-source lasers | 3% | 2.7–3.5% | Carl Zeiss AG · 18 December 2025 |
| Wafer processing | 38% | 35.5–40.5% | SEMI (PR Newswire Asia release, published by TNGlobal) · 8 April 2026 |
| Inspection and metrology | 9% | 7.5–10% | Electronics Weekly (David Manners), citing Yole Group · 22 June 2026 |
| Test equipment | 7% | 6.5–8.5% | SEMI (PR Newswire Asia release, published by TNGlobal) · 8 April 2026 |
| Silicon wafers | 6% | 5.5–6.5% | SEMI Silicon Manufacturers Group (PR Newswire distribution) · 10 February 2026 |
| Process materials | 17% | 15–18% | SEMI (PR Newswire distribution) · 12 May 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | ASMLTop 10 | 11.4% | 100 | 87 |
| 2 | ZEISS SMTTop 10 | 2.8% | 24 | 67 |
| 3 | Applied MaterialsTop 10 | 10% | 22 | 11 |
| 4 | KLATop 10 | 4.7% | 18 | 57 |
| 5 | Lam ResearchTop 10 | 7% | 16 | 93 |
| 6 | Tokyo ElectronTop 10 | 6.1% | 15 | 3 |
| 7 | AdvantestTop 10 | 3.3% | 10 | 87 |
| 8 | SynopsysTop 10 | 3.1% | 6.3 | 27 |
| 9 | CadenceRising | 2.6% | 5.1 | 69 |
| 10 | TRUMPF (EUV business field) | 0.4% | 4.1 | — |
| 11 | Shin-Etsu Chemical | 2.4% | 4.0 | — |
| 12 | NAURARising | 2.5% | 3.6 | 93 |
| 13 | ASM International | 1.4% | 2.9 | 39 |
| 14 | Lasertec | 0.6% | 2.8 | — |
| 15 | Siemens EDA (Siemens Digital Industries Software) | 1.2% | 2.5 | — |
| 16 | Teradyne (Semiconductor Test) | 1.3% | 2.4 | 47 |
| 17 | SUMCO | 1.4% | 2.3 | 43 |
| 18 | Qnity Electronics | 1.1% | 1.8 | 40 |
| 19 | Hitachi High-Tech | 0.9% | 1.7 | — |
| 20 | SCREEN Holdings (Semiconductor Production Equipment) | 1.2% | 1.6 | 10 |
| 21 | Merck KGaA Electronics (EMD Electronics) | 1.1% | 1.5 | 20 |
| 22 | Fujifilm Electronic Materials | 1.1% | 1.5 | — |
| 23 | Hoya (mask blanks) | 0.4% | 1.5 | — |
| 24 | JSR | 0.8% | 1.4 | — |
| 25 | Canon (Industrial Group: semiconductor lithography) | 0.9% | 1.3 | — |
| 26 | AMEC (Advanced Micro-Fabrication Equipment) | 0.9% | 1.3 | — |
| 27 | Tokyo Ohka Kogyo (TOK) | 0.6% | 1.2 | — |
| 28 | Ebara (Precision Machinery) | 0.8% | 1.1 | — |
| 29 | Air Liquide (Electronics) | 0.8% | 1.1 | — |
| 30 | Entegris | 0.7% | 1.0 | — |
| 31 | GlobalWafers | 0.9% | 1.0 | — |
| 32 | Linde (electronics gases) | 0.7% | 1.0 | — |
| 33 | Atlas Copco Vacuum Technique (Edwards) | 0.8% | 1.0 | — |
| 34 | IMS Nanofabrication (Intel) | 0.3% | 0.2 | — |
| 35 | NuFlare Technology (Toshiba) | 0.2% | 0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
Throughout the AI build-out (2023 to date), ASML has been the world's only manufacturer of EUV lithography systems. Leading-edge (5 nm-class and below) logic, the node family used for current AI accelerators, is not mass-produced without EUV, so ASML is a sole-supplier choke point.
ASML ranks 1 of 35 by Importance; the test asks for a place in the top 3: passed.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 89.7% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.83.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No public source splits these markets by AI use, so weights and shares use worldwide 2025 markets.
- AI chips are made on leading-edge logic and memory lines, which use more EUV lithography and inspection per wafer than the industry mix; worldwide shares also include mature-node lines, including China's.
- SEMI publishes 2025 equipment growth rates but not segment dollar values; 2025 values apply those rates to the 2024 values implied by its December 2025 release.
- Canon and Nikon report lithography inside wider segments and SMEE publishes no reachable figures, so other makers' lithography sales are estimated as half of Canon's Industrial Group and Nikon's Precision Equipment sales.
- ZEISS reports its semiconductor segment as one figure that also holds mask and process-control tools, and its year ends in September; its lithography optics sales are estimated as 85% of the segment, part of which serves upgrades rather than new machines.
- TRUMPF's years end in June, so its 2025 EUV laser sales average the years to June 2025 and June 2026; they include service and spares for ASML.
- Wafer processing is a remainder: SEMI's wafer fab equipment less lithography and process control, so it also holds fab-support and mask-making equipment.
- The process-control market size rests on one market-research estimate, cross-checked against KLA's revenue.
- The ESD Alliance does not list its contributors; Arm's revenue and an estimated $1.05B of other processor-core licensing are removed, and board design is counted with package design because the two are reported together.
- The 2025 probe card market extends a 2024 market-research figure at its forecast growth rate.
- No current public figure splits photomasks between chipmakers' own mask shops and merchant suppliers; the captive part is removed by estimate.
- ZEISS SMT, Siemens EDA, SK Siltron and other units or private suppliers report segment totals at best, so their shares rest on segment disclosures and estimates.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: ASML, ZEISS SMT, Applied Materials, KLA, Lam Research, Tokyo Electron, Advantest, Synopsys. Rising places: NAURA, Cadence.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Chip design
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Value of AI data-centre chips supplied in 2025, at designers' revenue or, for custom chips, owners' cost.
- Parts
- 71.1% Merchant accelerators: Most AI training and serving runs on them.
- 6.7% Custom accelerators: Cloud and AI companies' own chips, made with partners; cheaper per chip.
- 18.1% AI networking chips: Join accelerators into clusters.
- 3.9% Host processors: Run and feed every accelerator server.
- 0.2% Processor IP: Licensed designs inside Arm-based and custom chips.
- Shares from
- Accelerator revenue from filings, or shipment estimates; Programme value, split equally between the owner and its design partner; Networking-chip revenue from filings and market-share reports; Share of processors in AI servers, including NVIDIA Grace; Data-centre royalty and licence revenue
- Hard to replace
- Leading accelerators need years of design, mature software and reserved manufacturing capacity, and custom chips tie owners to one design partner per generation. Processors and networking chips have more alternatives.
- Check
- In 2023–24, demand for NVIDIA's data-centre GPUs (H100, then H200 and Blackwell) exceeded supply. NVIDIA said it was 'significantly increasing' supply to meet surging demand (May 2023) and remained supply-constrained on H200 and Blackwell (May 2024). Server makers reported GPU lead times of 20+ weeks and lost sales.
Parts and weights
AI data-centre chips at the companies that design them, counting both the owner and the design partner of a custom chip; device chips are out, and chip manufacturing, memory, packaging, networking-equipment makers, design software and non-processor IP belong to their own sublayers.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Merchant accelerators | 71.1% | 68.2–74.1% | NVIDIA / U.S. SEC · 25 February 2026 |
| Custom accelerators | 6.7% | 5.9–8.3% | The Motley Fool (transcript of Broadcom's earnings call) · 12 December 2025 |
| AI networking chips | 18.1% | 14.2–19% | NVIDIA / U.S. SEC · 25 February 2026 |
| Host processors | 3.9% | 2.8–5.2% | TrendForce · 6 January 2025 |
| Processor IP | 0.2% | 0.2–0.4% | Arm Holdings / U.S. SEC · 26 May 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | NVIDIATop 10 | 77.9% | 100 | 65 |
| 2 | BroadcomTop 10 | 5.1% | 4.0 | 47 |
| 3 | HuaweiTop 10 | 3.1% | 2.3 | — |
| 4 | AMDTop 10 | 3.5% | 1.5 | 13 |
| 5 | AlphabetTop 10 | 1.6% | 1.2 | — |
| 6 | AmazonTop 10 | 1.2% | 0.7 | — |
| 7 | MarvellTop 10 | 1.4% | 0.6 | 31 |
| 8 | IntelTop 10 | 1.4% | 0.5 | 0 |
| 9 | Arm | 0.2% | 0.3 | — |
| 10 | Alibaba | 0.5% | 0.3 | — |
| 11 | CredoRising | 0.5% | 0.2 | 81 |
| 12 | Meta | 0.3% | 0.1 | — |
| 13 | ByteDance | 0.2% | 0.1 | — |
| 14 | Astera Labs | 0.3% | 0.1 | 69 |
| 15 | Hygon | 0.3% | 0.1 | 29 |
| 16 | Baidu | 0.2% | 0.1 | — |
| 17 | CambriconRising | 0.4% | 0.1 | 94 |
| 18 | Cisco | 0.2% | 0.1 | — |
| 19 | Alchip Technologies | 0.1% | <0.1 | — |
| 20 | Cerebras | 0.1% | <0.1 | — |
| 21 | GUC | <0.1% | <0.1 | — |
| 22 | Groq | 0.1% | <0.1 | — |
| 23 | Microsoft | 0.1% | <0.1 | — |
| 24 | Tencent | <0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
In 2023–24, demand for NVIDIA's data-centre GPUs (H100, then H200 and Blackwell) exceeded supply. NVIDIA said it was 'significantly increasing' supply to meet surging demand (May 2023) and remained supply-constrained on H200 and Blackwell (May 2024). Server makers reported GPU lead times of 20+ weeks and lost sales.
NVIDIA ranks 1 of 24 by Importance; the test asks for a place in the top 3: passed.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 85.2% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.82.
No rating was more than a level apart between the two passes.
Limits of the evidence
- AMD, Huawei and most Chinese and private designers do not report accelerator revenue; the weights use estimates (AMD's data-centre segment less an estimate for server processors; Huawei from reported shipment estimates).
- Custom chips have no market price: their value is what owners pay design partners, which leaves out memory or wafers an owner buys directly and the owner's own design work.
- NVIDIA sells networking mostly as switches, adapters and cables, while Broadcom, Marvell and others sell chips, so revenue shares in networking mix product levels.
- NVIDIA Grace processors are sold inside combined processor-accelerator modules; their value is estimated, moved to host processors and removed from accelerators.
- No source reports each vendor's share of processors in AI servers; the AI-server share of processor revenue is estimated from shipment data.
- Arm does not report data-centre royalties in dollars; the processor IP total is estimated from its statements.
- Fiscal years differ (NVIDIA and Marvell end in January, Broadcom in early November, Credo in late April or early May), so 2025 values are taken from the fiscal year closest to calendar 2025.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: NVIDIA, Broadcom, Huawei, AMD, Alphabet, Amazon, Marvell, Intel. Rising places: Cambricon, Credo.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Foundry
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Value of front-end wafers for the chips in AI data-centre systems.
- Parts
- 74% AI accelerator wafers: GPU and custom AI chip dies plus HBM base dies: most AI wafer value, on the costliest processes.
- 22% CPU, network and interface wafers: Host CPUs and switch, network-card, optical-DSP and retimer chips: an estimated three-tenths of accelerator wafer value.
- 1% Power and analog wafers: Power delivery to AI chips, on specialty wafers that cost far less than leading-edge logic.
- 3% Photonics and SiGe wafers: Photonic and SiGe chips for optical links, made by a few specialty foundries.
- Shares from
- The foundry of each AI accelerator and HBM base die, valued against TSMC's AI-accelerator revenue.; The foundry of each AI-server CPU and network chip; Intel Foundry's internal revenue for Xeon.; Foundry and chip-company statements on AI-server power wafers.; Tower's and GlobalFoundries' photonics and SiGe revenue; transceiver makers' supply statements.
- Hard to replace
- Moving a chip to another foundry means redesign, new masks and requalification. Only three companies run 3nm-class processes and most AI accelerators come from one of them; power, analog and photonics processes have more sources, though requalifying takes months to a year.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Front-end wafers made by foundry businesses (merchant foundries and the foundry units of Intel and Samsung, including wafers for their own group) for chips used in AI data-centre systems; advanced packaging and interposers belong to Packaging and test, memory wafers to Memory, chips made in fabs that sell no foundry service to Chip design, equipment, materials and merchant photomasks to Tools, and chips for phones, PCs and other devices are out.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| AI accelerator wafers | 74% | 54–85% | TSMC · 15 January 2026 |
| CPU, network and interface wafers | 22% | 11.5–39% | NVIDIA · 25 February 2026 |
| Power and analog wafers | 1% | 0.3–7% | Infineon Technologies · 12 November 2025 |
| Photonics and SiGe wafers | 3% | 1–8% | Tower Semiconductor (call transcript published by MarketBeat) · 11 February 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | TSMCTop 10 | 89.6% | 100 | 75 |
| 2 | SMICTop 10 | 2.4% | 2.5 | 82 |
| 3 | IntelTop 10 | 2.8% | 1.3 | 0 |
| 4 | Tower SemiconductorTop 10 | 1.4% | 0.7 | 41 |
| 5 | GlobalFoundriesTop 10 | 1.3% | 0.6 | 18 |
| 6 | Samsung FoundryTop 10 | 0.6% | 0.1 | — |
| 7 | UMCTop 10 | 0.4% | 0.1 | — |
| 8 | VISTop 10 | 0.3% | 0.1 | 80 |
| 9 | United Nova TechnologyTop 10 | 0.3% | 0.1 | — |
| 10 | Sanan ICTop 10 | 0.1% | <0.1 | — |
| 11 | Hua Hong | 0.2% | <0.1 | — |
| 12 | WIN Semiconductors | 0.1% | <0.1 | — |
| 13 | SilTerra | 0.1% | <0.1 | — |
| 14 | DB HiTek | 0.1% | <0.1 | — |
| 15 | GCS Holdings | <0.1% | <0.1 | — |
| 16 | STMicroelectronics (PIC100 silicon photonics) | <0.1% | <0.1 | — |
| 17 | SK keyfoundry | <0.1% | <0.1 | — |
| 18 | PSMC | <0.1% | <0.1 | — |
| 19 | GTA Semiconductor | <0.1% | <0.1 | — |
| 20 | X-FAB | <0.1% | <0.1 | — |
| 21 | China Resources Microelectronics | <0.1% | <0.1 | — |
| 22 | SMART Photonics | <0.1% | <0.1 | — |
| 23 | Nexchip | <0.1% | <0.1 | — |
| 24 | CanSemi | <0.1% | <0.1 | — |
| 25 | Advanced Wireless Semiconductor (AWSC) | <0.1% | <0.1 | — |
| 26 | Episil | <0.1% | <0.1 | — |
| 27 | Ligentec | <0.1% | <0.1 | — |
| 28 | CUMEC (Chongqing United Microelectronics Center) | <0.1% | <0.1 | — |
| 29 | Silex Microsystems | <0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 89.7% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.78.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No public source splits AI-related foundry revenue by function; the weights combine company disclosures with stated 5LayerCake estimates.
- TSMC does not say whether its AI-accelerator revenue includes CoWoS and other packaging; the accelerator weight's range covers both readings.
- No foundry discloses wafer revenue from host CPUs or networking chips; that weight is estimated from NVIDIA's networking-to-compute revenue ratio and TSMC's change of AI definitions between October 2024 and January 2025.
- Only a few chip companies (Infineon, Monolithic Power Systems) disclose AI-server power revenue; the market total, the part made at foundries and its wafer value are estimates.
- SMIC's and Samsung's AI-accelerator wafer output is not disclosed; Huawei Ascend volumes come from press and analyst estimates.
- Intel Foundry does not split its internal revenue by product, so Xeon wafers used in AI servers must be estimated.
- The weights describe 2025; HBM base dies move to foundry processes with HBM4 from 2026, and accelerator revenue is growing faster than the other functions, so the mix changes year to year.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: TSMC, SMIC, Intel, Tower Semiconductor, GlobalFoundries, Samsung Foundry, UMC, VIS. Rising places filled by the next Importance ranks: United Nova Technology, Sanan IC.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Memory
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Spending on memory and storage for AI servers over the four quarters to June 2026: HBM, host DRAM, data-centre SSDs, and their interface and controller chips.
- Parts
- 40.2% High-bandwidth memory (HBM): Stacked memory beside each AI accelerator.
- 19.7% Server DRAM: Main memory for each AI server's host processors.
- 39.2% Data-centre SSDs: Stores training data, checkpoints and inference caches.
- 0.9% Memory interface and SSD controller chips: Chips that buffer memory modules and run SSDs; small in value.
- Shares from
- HBM revenue shares (Counterpoint, TrendForce); Host DRAM revenue shares for AI servers (DDR5 RDIMM, LPDDR5X for Arm hosts); total DRAM share where needed (labelled); Enterprise SSD revenue shares (TrendForce); NAND bit share for flash-only sellers (labelled); Interface and SSD-controller chip revenue shares
- Hard to replace
- Only three companies make HBM at volume, and a new HBM or server DRAM source needs years of process work and qualification. SSDs and interface chips have more qualified suppliers.
- Check
- In 2024, SK hynix, the HBM market leader, said its planned 2024 HBM output was sold out (February 2024). By May 2024 it said 2025 supply was almost sold out too, which left AI accelerator makers supply-limited on high-bandwidth memory.
Parts and weights
Memory and storage bought for AI servers and the storage systems attached to them: HBM, conventional server DRAM, data-centre SSDs, and the merchant chips that interface or control them; HBM stacking by memory makers counts here and placing HBM beside the processor counts in Packaging; HBM base dies made by foundries also count in Foundry's wafer measure; hard drives, non-AI servers and module assembly from purchased chips are outside this sublayer.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| High-bandwidth memory (HBM) | 40.2% | 21–60% | Epoch AI · 21 May 2026 |
| Server DRAM | 19.7% | 12–30% | Counterpoint Research · 2 September 2026 |
| Data-centre SSDs | 39.2% | 20–52% | TrendForce · 1 September 2026 |
| Memory interface and SSD controller chips | 0.9% | 0.5–2% | Rambus Inc. (transcript published by The Motley Fool) · 2 February 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | SK hynixTop 10 | 32.9% | 100 | 0 |
| 2 | Samsung ElectronicsTop 10 | 34.3% | 95 | 78 |
| 3 | MicronTop 10 | 18.5% | 38 | 45 |
| 4 | KioxiaTop 10 | 4.4% | 5.8 | 36 |
| 5 | CXMTTop 10 | 2.1% | 3.3 | 74 |
| 6 | SandiskTop 10 | 2.9% | 3.0 | 70 |
| 7 | Union MemoryTop 10 | 0.8% | 0.7 | — |
| 8 | PhisonTop 10 | 0.6% | 0.6 | — |
| 9 | NanyaRising | 0.4% | 0.4 | 53 |
| 10 | DapuStorTop 10 | 0.5% | 0.4 | 28 |
| 11 | Huawei | 0.5% | 0.4 | — |
| 12 | YMTC | 0.4% | 0.2 | — |
| 13 | Microchip Technology | 0.1% | 0.2 | — |
| 14 | Marvell | 0.1% | 0.2 | — |
| 15 | Memblaze | 0.2% | 0.1 | — |
| 16 | Montage | 0.1% | 0.1 | — |
| 17 | Silicon Motion | 0.1% | 0.1 | — |
| 18 | Longsys | 0.1% | 0.1 | — |
| 19 | FADU | 0.1% | 0.1 | — |
| 20 | Rambus | <0.1% | 0.1 | — |
| 21 | InnoGrit | 0.1% | 0.1 | — |
| 22 | Kingston | 0.1% | 0.1 | — |
| 23 | DERA | 0.1% | 0.1 | — |
| 24 | Renesas Electronics | <0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
In 2024, SK hynix, the HBM market leader, said its planned 2024 HBM output was sold out (February 2024). By May 2024 it said 2025 supply was almost sold out too, which left AI accelerator makers supply-limited on high-bandwidth memory.
SK hynix ranks 1 of 24 by Importance; the test asks for a place in the top 3: passed.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 88.5% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.93.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No public source splits AI-server memory and storage spend by type; the weights combine Epoch AI's HBM cost series, NVIDIA's GB300 NVL72 memory configuration, DRAM price movements and enterprise SSD totals, with an estimated AI share of enterprise SSDs (45 to 72%).
- Quarterly HBM revenue for 2026 is not published; Q1 and Q2 2026 are 5LayerCake estimates from Micron's 2025 HBM market size, Micron's expected 40% yearly growth and TrendForce's expected 70% growth in 2026 HBM use.
- The weights move with memory prices, which changed sharply over the reference period while HBM prices are set in annual contracts; the mix needs checking at each quarterly review.
- Supplier shares are mostly published for total DRAM rather than server DRAM, and for enterprise SSDs as a whole rather than their AI portion; the card assumes AI buyers' supplier mix matches the whole function.
- Merchant SSD controller revenue and shares have no public tracker, and in-house controllers sit inside the SSD function.
- Memory for Chinese AI servers (CXMT, YMTC and domestic SSD makers) is thinly covered by market trackers.
- Hard drives used for AI data storage are not measured in this or any other sublayer.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: SK hynix, Samsung Electronics, Micron, Kioxia, CXMT, Sandisk, Union Memory, Phison. Rising places: Nanya, DapuStor.
- 22 October 2026DapuStor: input update (DapuStor's Rising score 28.1 is below the minimum, so its place is relabelled from Rising to Importance (rank 10); membership unchanged)
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Packaging and test
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Value of 2025 spending on packaging and testing AI data-centre chips, including substrates, film, glass cloth and equipment.
- Parts
- 54.2% 2.5D/3D integration: Joins each AI processor to its HBM; the 2023–25 bottleneck.
- 17.0% Assembly and test: Mounts, tests and burns in each package.
- 15.6% Package substrates: Boards under every AI package; few makers supply the largest.
- 1.4% Substrate build-up film: Insulating film in nearly all advanced substrates; one maker.
- 0.6% Substrate glass cloth: Core glass cloth of AI substrates; one company makes most.
- 11.2% Packaging equipment: Bonders, dicers and grinders, including for HBM stacking.
- Shares from
- AI-related revenue shares; CoWoS-class capacity shares where needed (labelled); AI-related revenue shares; whole-function shares where no AI split exists (labelled)
- Hard to replace
- Integration, build-up film and high-end glass cloth have few alternatives at scale; new capacity takes a year or more. Assembly, test and equipment have more.
- Check
- From mid-2023 into 2025, TSMC's CoWoS advanced-packaging capacity was the binding constraint on AI accelerator output. In July 2023 TSMC said it could not fully meet customer demand. In September 2023 its chairman put the share of CoWoS demand it could meet at about 80%. In July 2024 the company said constraints would last into 2025.
Parts and weights
Packaging and test of AI data-centre chips: 2.5D/3D integration, assembly and test services, package substrates with their build-up film and glass cloth, and the bonding, dicing and grinding equipment used for this work and for HBM stacking; HBM stacking itself counts in Memory, test equipment and front-end wafer equipment in Tools, other package materials (underfill, mould compound, solder) are excluded here and in Tools, and TSMC's packaging business is scored as part of TSMC.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| 2.5D/3D integration | 54.2% | 44–72% | Epoch AI · 21 May 2026 |
| Assembly and test | 17.0% | 6–25% | ASE Technology Holding Co., Ltd. (furnished to the US SEC) · 5 February 2026 |
| Package substrates | 15.6% | 11–22% | Taiwan Printed Circuit Association (TPCA) and ITRI ISTI, published by I-Connect007 · 10 September 2026 |
| Substrate build-up film | 1.4% | 0.8–2.4% | Ajinomoto Co., Inc. · 2026 |
| Substrate glass cloth | 0.6% | 0.3–1% | Nittobo (Nitto Boseki Co., Ltd.); English translation hosted by Infomart · 5 August 2026 |
| Packaging equipment | 11.2% | 6–16% | DISCO Corporation · 22 April 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | TSMCTop 10 | 47.8% | 100 | 71 |
| 2 | ASE TechnologyTop 10 | 8% | 7.1 | — |
| 3 | DiscoTop 10 | 3.7% | 4.4 | — |
| 4 | Ajinomoto Build-up Film (ABF)Top 10 | 1.3% | 3.5 | — |
| 5 | UnimicronTop 10 | 4.4% | 3.3 | — |
| 6 | AmkorTop 10 | 4.4% | 2.9 | — |
| 7 | IbidenTop 10 | 3.1% | 2.6 | 17 |
| 8 | SJ Semiconductor (SJSemi)Top 10 | 2.5% | 2.6 | — |
| 9 | KYECRising | 2% | 1.6 | 92 |
| 10 | NittoboTop 10 | 0.5% | 1.2 | — |
| 11 | Besi | 1.5% | 1.1 | — |
| 12 | Hanmi Semiconductor | 1.8% | 0.9 | — |
| 13 | Samsung Electro-Mechanics | 1.4% | 0.8 | — |
| 14 | ASMPT | 1.2% | 0.7 | — |
| 15 | Tongfu Microelectronics | 1.3% | 0.6 | — |
| 16 | Shinko Electric Industries | 0.9% | 0.6 | — |
| 17 | Samsung Electronics | 1.1% | 0.5 | — |
| 18 | AT&S | 0.8% | 0.5 | — |
| 19 | Nan Ya Printed Circuit Board | 0.8% | 0.4 | — |
| 20 | UMC | 0.5% | 0.4 | — |
| 21 | Intel | 0.6% | 0.3 | — |
| 22 | Sigurd Microelectronics | 0.6% | 0.3 | — |
| 23 | Zhen Ding Technology | 0.6% | 0.3 | — |
| 24 | JCET | 0.9% | 0.3 | — |
| 25 | EV Group | 0.4% | 0.2 | — |
| 26 | SUSS MicroTec | 0.4% | 0.2 | — |
| 27 | Powertech (PTI) | 0.4% | 0.2 | — |
| 28 | Sekisui Chemical (build-up film) | 0.1% | 0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
From mid-2023 into 2025, TSMC's CoWoS advanced-packaging capacity was the binding constraint on AI accelerator output. In July 2023 TSMC said it could not fully meet customer demand. In September 2023 its chairman put the share of CoWoS demand it could meet at about 80%. In July 2024 the company said constraints would last into 2025.
TSMC ranks 1 of 28 by Importance; the test asks for a place in the top 3: passed.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 87.5% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.91.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No public source splits AI packaging spend by function; the weights combine Epoch AI's CoWoS cost series, ASE's leading-edge packaging and test revenue, TPCA's ABF substrate output with an estimated AI share, Ajinomoto's and Nittobo's materials sales, and bonding and dicing equipment sales.
- Assembly companies that buy substrates (Amkor's filing lists laminate substrates among the materials it purchases) include substrate cost in their revenue; no company discloses this for AI packages, so it is not netted out and the assembly and test range is widened downward.
- TSMC does not disclose CoWoS revenue, so 2.5D/3D supplier shares mostly rest on capacity estimates, which sit off the card's basis.
- Ajinomoto does not report build-up film sales on their own; Functional Materials (about 102 billion yen in FY2025) also includes other materials, and the film's AI share is taken to match substrates.
- Nittobo reports glass yarn and cloth together (Electronic Materials, about 61 billion yen in FY2025, including cloth for printed circuit boards); the low-expansion cloth share (taken as 35 to 55%) and its AI share (35 to 70%) are estimates.
- The AI share of ABF substrate output (taken as 30 to 50%, from Unimicron's 40% in 2025) and of DISCO's sales (taken as about 30%) is not disclosed industry-wide.
- Outsourced test of AI chips is split between assembly companies, test houses and in-house lines, and no public figure gives the total.
- Domestic Chinese 2.5D packaging for AI chips is thinly documented.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: TSMC, ASE Technology, Disco, Ajinomoto Build-up Film (ABF), Unimicron, Amkor, Ibiden, SJ Semiconductor (SJSemi). Rising places: KYEC. Rising places filled by the next Importance ranks: Nittobo.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Infrastructure
Campuses
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Data-centre capacity in operation that operators offer to tenants, in megawatts of IT load, weighted by region.
- Parts
- 50% Americas: About half of world data-centre capacity, nine-tenths of it in the United States.
- 19% Europe, Middle East and Africa: About a fifth of world capacity, led by a few European metros and new Gulf and Nordic campuses.
- 31% Asia-Pacific, including mainland China: About three-tenths of world capacity, across China, Japan, India, Southeast Asia and Australia.
- Shares from
- Operating leasable MW over JLL's regional capacity times Synergy's 40% operator share, as in the weights
- Hard to replace
- New capacity takes years because grid connections and electrical equipment are scarce, so a tenant can move only to space already built nearby. Retail colocation in large metros is easier to switch.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Capacity in operation that data-centre operators run for tenants (colocation, wholesale and AI-campus developers), counted once with the operator that runs it. All self-built capacity is outside this card: data centres that clouds, AI companies and other businesses build for their own use, such as Meta's and xAI's, are not counted as campuses anywhere, and their output is counted where it is sold, in Cloud as rented computing or in Models and Applications; equipment and electricity are counted in Power & cooling and Energy, and capacity under construction or only planned is excluded.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Americas | 50% | 42–56% | JLL · 5 January 2026 |
| Europe, Middle East and Africa | 19% | 16–23% | JLL · 5 January 2026 |
| Asia-Pacific, including mainland China | 31% | 24–38% | JLL · 5 January 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | EquinixTop 10 | 6.6% | 100 | 17 |
| 2 | Digital RealtyTop 10 | 6.9% | 77 | 16 |
| 3 | China MobileTop 10 | 5% | 55 | — |
| 4 | China TelecomTop 10 | 4.3% | 42 | — |
| 5 | China UnicomTop 10 | 3.8% | 38 | — |
| 6 | QTSTop 10 | 3.4% | 37 | — |
| 7 | NTT GDCTop 10 | 3.6% | 30 | 35 |
| 8 | CyrusOneTop 10 | 2.7% | 29 | — |
| 9 | Aligned | 2.5% | 27 | — |
| 10 | Khazna Data Centers | 1.1% | 22 | — |
| 11 | STACK | 2.3% | 20 | — |
| 12 | Switch | 1.8% | 18 | — |
| 13 | Vantage | 2.1% | 18 | — |
| 14 | DayOne | 1.4% | 17 | — |
| 15 | EdgeConneX | 1.7% | 15 | — |
| 16 | DataBank | 1.8% | 14 | — |
| 17 | VNET | 1.5% | 14 | 75 |
| 18 | AirTrunk | 1.2% | 13 | — |
| 19 | GDS | 1.5% | 13 | 36 |
| 20 | Applied DigitalRising | 1.1% | 11 | 96 |
| 21 | STTGDC | 1.2% | 11 | — |
| 22 | Core ScientificRising | 1.1% | 10 | 94 |
| 23 | Iron Mountain Data Centers | 1.3% | 9.8 | 53 |
| 24 | Global Switch | 1.2% | 9.7 | — |
| 25 | CloudHQ | 1.1% | 9.4 | — |
| 26 | CoreSite (American Tower) | 1% | 9.2 | — |
| 27 | CDC Data Centres | 0.7% | 9.1 | — |
| 28 | Data4 | 0.9% | 8.4 | — |
| 29 | Crusoe | 0.8% | 8.3 | — |
| 30 | Chindata (WinTriX DC Group) | 0.9% | 6.8 | — |
| 31 | Compass Datacenters | 0.8% | 6.7 | — |
| 32 | Centersquare (Evoque and Cyxtera) | 0.8% | 6.1 | — |
| 33 | Flexential | 0.7% | 5.7 | — |
| 34 | Sabey Data Centers | 0.6% | 4.8 | — |
| 35 | Princeton Digital Group | 0.6% | 4.7 | — |
| 36 | Range Intelligent Computing (Runze) | 0.5% | 4.2 | — |
| 37 | NEXTDC | 0.4% | 3.4 | — |
| 38 | Scala Data Centers | 0.3% | 2.9 | — |
| 39 | Nxtra (Airtel) | 0.3% | 1.4 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 60% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.04.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No free source gives leasable capacity by region, so regions are weighted by total data-centre capacity (JLL); operators lease a larger share of capacity outside the US, which the ranges allow for.
- Regional denominators are JLL's regional capacity times Synergy Research's worldwide operator share (about 40%); if a public leasable total is used for any region, the weights must be re-derived from the same totals.
- Operators report capacity on different definitions (critical IT load or gross power; joint ventures at 100% or at ownership share); figures on other definitions are labelled off-basis.
- Megawatts do not capture interconnection, where a few operators run the densest network hubs.
- Large private operators (Vantage, STACK, Aligned, AirTrunk, CyrusOne, QTS) publish capacity irregularly.
- Capacity used for AI cannot be separated from capacity used for other cloud and enterprise computing.
- Chinese carriers' data-centre businesses count as operators, but they report capacity inside wider segments.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Equinix, Digital Realty, China Mobile, China Telecom, China Unicom, QTS, NTT GDC, CyrusOne. Rising places: Applied Digital, Core Scientific.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Power and cooling
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Equipment that powers and cools data-centre computers on the campus, by worldwide sales, weighted by 2025 spending.
- Parts
- 37% UPS and power distribution: Backup power and the busways that carry it to each rack; the largest category.
- 6% Switchgear and switchboards: Routes and protects power from the incoming lines to each hall.
- 17% Server and rack power supplies: Power supplies inside servers and AI racks, including power shelves.
- 18% Air cooling and heat rejection: Air handlers, chillers and dry coolers that move heat out of halls.
- 7% Liquid cooling distribution: Coolant distribution units and manifolds that new GPU racks need.
- 15% Server thermal parts: Cold plates, heat sinks and fans on each AI server.
- Shares from
- Company data-centre sales in the part over the world total, from market reports and filings
- Hard to replace
- Several large makers supply UPS systems, switchgear and chillers, though backlogs have grown. Liquid cooling, cold plates and power shelves are qualified per GPU platform, leaving few suppliers.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Equipment that carries power around the data-centre campus and inside its buildings, and removes heat from IT equipment: UPS systems and their batteries, low- and medium-voltage switchgear and switchboards on the campus, busway and rack power distribution, server and rack power supplies including rack power shelves and battery backup units, air cooling and heat rejection, liquid cooling distribution, and thermal parts fitted to servers. On-campus transformers and medium-voltage cables belong here but are not counted, because public estimates of their data-centre sales differ widely and use no common definition; the utility network up to and including the site substation, generators and on-site generation belong to Energy, and IT racks, enclosures and management software are not counted.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| UPS and power distribution | 37% | 13–66% | Dell'Oro Group · 19 August 2026 |
| Switchgear and switchboards | 6% | 2–17% | P&S Intelligence · September 2026 |
| Server and rack power supplies | 17% | 5–42% | Business Next (數位時代) · 26 February 2026 |
| Air cooling and heat rejection | 18% | 7–42% | Intelligent CIO, reporting Omdia · 19 June 2024 |
| Liquid cooling distribution | 7% | 2–18% | Dell'Oro Group · 8 July 2025 |
| Server thermal parts | 15% | 4–38% | Commercial Times (工商時報) · 12 March 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | Delta ElectronicsTop 10 | 11.8% | 100 | — |
| 2 | Schneider ElectricTop 10 | 11.6% | 65 | — |
| 3 | VertivTop 10 | 10.6% | 62 | 50 |
| 4 | AVCTop 10 | 3.9% | 32 | — |
| 5 | EatonTop 10 | 5.4% | 23 | — |
| 6 | FlexTop 10 | 3.5% | 14 | — |
| 7 | HuaweiTop 10 | 2.7% | 9.2 | — |
| 8 | Siemens SITop 10 | 1.8% | 8.7 | — |
| 9 | LITEONTop 10 | 1.8% | 8.0 | — |
| 10 | Jentech Precision | 0.6% | 7.3 | — |
| 11 | ABB | 1.7% | 7.1 | — |
| 12 | Legrand | 2.1% | 7.0 | — |
| 13 | Mitsubishi Electric | 1.9% | 6.4 | — |
| 14 | Modine | 1.7% | 6.2 | — |
| 15 | Advanced Energy | 1.3% | 6.0 | — |
| 16 | CarrierRising | 1.4% | 5.3 | 58 |
| 17 | Johnson Controls | 1.4% | 5.2 | — |
| 18 | Trane Technologies | 1.2% | 4.6 | — |
| 19 | Daikin | 1.2% | 4.3 | — |
| 20 | Foxconn | 0.9% | 3.8 | — |
| 21 | EnerSys | 1% | 3.7 | — |
| 22 | nVent | 1.1% | 3.5 | 39 |
| 23 | STULZ | 1% | 3.5 | — |
| 24 | Munters | 0.9% | 3.0 | — |
| 25 | Auras Technology | 0.7% | 3.0 | — |
| 26 | Simplo Technology | 0.8% | 2.9 | — |
| 27 | Fuji Electric | 0.7% | 2.5 | — |
| 28 | AcBel Polytech | 0.7% | 2.5 | — |
| 29 | Kehua Data | 0.6% | 2.1 | — |
| 30 | Rittal | 0.6% | 1.8 | — |
| 31 | FläktGroup | 0.5% | 1.8 | — |
| 32 | Ecolab | 0.5% | 1.8 | — |
| 33 | Chicony Power | 0.4% | 1.7 | — |
| 34 | Envicool | 0.5% | 1.5 | — |
| 35 | Samsung SDI | 0.5% | 1.5 | — |
| 36 | AAON | 0.4% | 1.4 | — |
| 37 | Nidec | 0.4% | 1.1 | — |
| 38 | Cooler Master | 0.3% | 1.1 | — |
| 39 | Socomec | 0.4% | 1.1 | — |
| 40 | Nortek Air Solutions | 0.3% | 1.1 | — |
| 41 | Sunon | 0.4% | 1.1 | — |
| 42 | LS Electric | 0.4% | 1.0 | — |
| 43 | Kstar | 0.4% | 1.0 | — |
| 44 | Toshiba (UPS) | 0.3% | 0.9 | — |
| 45 | Shenling Environmental | 0.3% | 0.9 | — |
| 46 | Shenzhen Honor Electronic | 0.2% | 0.8 | — |
| 47 | Megmeet | 0.2% | 0.8 | — |
| 48 | ebm-papst | 0.3% | 0.8 | — |
| 49 | SPX Technologies | 0.3% | 0.8 | — |
| 50 | Furukawa Electric | 0.3% | 0.8 | — |
| 51 | Baltimore Aircoil Company | 0.2% | 0.7 | — |
| 52 | Dynapack | 0.2% | 0.7 | — |
| 53 | EVAPCO | 0.2% | 0.7 | — |
| 54 | Narada Power Source | 0.2% | 0.6 | — |
| 55 | LG Electronics | 0.2% | 0.6 | — |
| 56 | Wistron | 0.2% | 0.5 | — |
| 57 | Inventec | 0.2% | 0.5 | — |
| 58 | Alfa Laval | 0.2% | 0.5 | — |
| 59 | Danfoss | 0.2% | 0.5 | — |
| 60 | Midea (data-centre cooling) | 0.1% | 0.3 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 91.8% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.46.
Ratings more than a level apart were settled by a third, reconciling pass: Delta Electronics (Liquid cooling distribution), Jentech Precision (Server thermal parts).
Limits of the evidence
- No free source gives segment sizes for the whole market; UPS and power distribution is derived from Dell'Oro's total and segment ranking with Omdia's thermal figure, so it carries the widest tier range.
- Switchgear rests on two market-research estimates ($2.35B for 2025 and $3.6B for 2026) whose scopes differ.
- Server and rack power supplies and server thermal parts are outside the market trackers; their sizes are estimated from Delta Electronics' and AVC's revenue.
- On-campus transformers and medium-voltage cables are in scope but not counted, for lack of a reliable total.
- Weights reflect all data centres; AI facilities use more liquid cooling and denser power distribution than the average.
- Liquid-cooling totals from market trackers may include some cold-plate value, which this card counts under server thermal parts.
- Omdia's thermal-management total is a 2023 figure grown at Omdia's own forecast rate.
- Many suppliers (Schneider Electric, Eaton, ABB, Siemens, Johnson Controls, Daikin, Delta Electronics, Huawei) report data-centre sales inside wider segments, so their shares rest on estimates.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Delta Electronics, Schneider Electric, Vertiv, AVC, Eaton, Flex, Huawei, Siemens SI. Rising places: Carrier. Rising places filled by the next Importance ranks: LITEON.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Servers and systems
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- AI servers and rack-scale systems delivered to data centres, by sales value, weighted by AI server sales.
- Parts
- 74% GPU servers and racks outside mainland China: Most AI computing runs on GPU servers and racks; about three-quarters of AI server sales.
- 17% Servers for cloud companies' own AI chips: Servers built to order around chips such as Google TPUs and AWS Trainium; about a sixth of sales.
- 9% AI servers sold in mainland China: A separate market under export controls, with domestic makers and chips; about a tenth of sales.
- Shares from
- Company GPU server and rack revenue outside China over IDC's total (IDC, filings); NVIDIA only for complete systems it sells (DGX); Company revenue from servers for other accelerators outside China over IDC's total (filings); Company share of China's accelerated server market (IDC China)
- Hard to replace
- Rack-scale systems need qualified manufacturing, liquid-cooling integration and full-rack testing that few makers provide at volume. Standard GPU servers have many qualified makers.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Servers and rack-scale systems built around GPUs or other AI accelerators, credited to the company that sells the finished system to the data-centre operator, whether a branded system maker or a manufacturer selling directly to cloud companies; accelerators, CPUs and the boards and modules their designers sell belong to Chips, switches and cabling to Networking, power shelves and cold plates to Power & cooling, and general-purpose and storage servers are outside the measure.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| GPU servers and racks outside mainland China | 74% | 68–79% | IDC · 21 July 2026 |
| Servers for cloud companies' own AI chips | 17% | 12–21% | IDC · 15 June 2026 |
| AI servers sold in mainland China | 9% | 7–12% | IDC · 21 July 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | FoxconnTop 10 | 19% | 100 | 79 |
| 2 | QuantaTop 10 | 12.8% | 47 | 51 |
| 3 | DellTop 10 | 12.6% | 45 | 93 |
| 4 | WistronTop 10 | 9.3% | 37 | 56 |
| 5 | SupermicroTop 10 | 8% | 23 | 33 |
| 6 | Inspur InformationTop 10 | 3.6% | 11 | — |
| 7 | HuaweiTop 10 | 1% | 7.2 | — |
| 8 | LenovoTop 10 | 2.2% | 4.9 | — |
| 9 | GIGABYTETop 10 | 1.8% | 3.9 | — |
| 10 | ASUSTop 10 | 1.6% | 3.6 | — |
| 11 | H3C | 1.3% | 3.3 | — |
| 12 | xFusion | 1.1% | 2.8 | — |
| 13 | HPE | 1.2% | 2.7 | 0 |
| 14 | NVIDIA | 1.1% | 2.5 | — |
| 15 | Celestica | 0.6% | 2.3 | — |
| 16 | Inventec | 0.8% | 2.1 | — |
| 17 | Nettrix | 0.8% | 1.9 | — |
| 18 | ZTE (servers) | 0.7% | 1.6 | — |
| 19 | Jabil | 0.8% | 1.6 | — |
| 20 | Sanmina (incl. ZT Systems manufacturing) | 0.6% | 1.4 | — |
| 21 | MiTAC (TYAN) | 0.6% | 1.1 | — |
| 22 | Huakun Zhenyu | 0.4% | 1.0 | — |
| 23 | Flex | 0.4% | 0.9 | — |
| 24 | Sugon | 0.3% | 0.7 | — |
| 25 | Pegatron | 0.3% | 0.6 | — |
| 26 | Compal | 0.2% | 0.4 | — |
| 27 | Cerebras | 0.1% | 0.2 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 74.1% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.79.
Ratings more than a level apart were settled by a third, reconciling pass: Foxconn (GPU servers and racks outside mainland China), Huawei (AI servers sold in mainland China).
Limits of the evidence
- IDC does not split 3Q25 and 4Q25 AI servers by chip type; those two quarters are interpolated between 2Q25 and 1Q26.
- China's share comes from IDC's AI infrastructure total for 1Q26 and a matched first-half 2025 comparison, not from IDC's server tracker by chip type.
- Manufacturers that build systems sold under another company's brand are not credited for that work, because each system is credited once, to its seller.
- IDC reports direct sales by manufacturers to cloud companies as one total; each manufacturer's AI server revenue comes from filings that mix AI servers with other products.
- When a cloud company buys accelerators itself and supplies them to the manufacturer, the manufacturer's revenue is lower than IDC's system value.
- Fiscal years differ (Dell, HPE, Supermicro, Lenovo), so inputs use the latest four reported quarters.
- System prices include the accelerators inside, so sales shares follow the value of chips passed through as well as the system maker's own work.
- IDC does not split GPU systems into rack-scale systems and standard servers.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Foxconn, Quanta, Dell, Wistron, Supermicro, Inspur Information, Huawei, Lenovo. Rising places filled by the next Importance ranks: GIGABYTE, ASUS.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Networking and optics
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Switches, optical transceivers and cabling that connect data-centre computers, by worldwide sales of finished parts, weighted by 2025 spending.
- Parts
- 53% Data-centre switches: Every AI cluster links its accelerators through switch fabrics; the largest share of spending.
- 27% Optical transceivers: Links beyond a few metres run over optical modules, whose lasers have been a bottleneck.
- 10% Copper links and connectors: Short links inside and between racks run over copper cables and cable cartridges.
- 10% Fibre and passive connectivity: Every optical link runs over fibre cable, trunks and connectors laid through each hall.
- Shares from
- Data-centre switch revenue over the world total (IDC, Dell'Oro, filings); Finished transceiver and optical cable sales over the world total; resold modules go to the selling brand; Finished copper data cable sales over the estimated world total; Data-centre fibre and connectivity sales over the estimated world total
- Hard to replace
- Switches and transceivers are qualified for each network design, and a few makers supply most high-speed optics, whose lasers have run short. Cabling has more qualified makers.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Networking parts inside AI data centres, each credited to the company that sells the finished part (switch, module or cable): Ethernet and InfiniBand switch systems, including switches sold directly to cloud companies by their manufacturers, optical transceivers, and copper and fibre cabling and connectivity; switch, network-adapter and optical signal-processing chips and the adapter cards their designers sell belong to Chips · Chip design, switch trays inside rack-scale systems to Servers & systems, and routers and long-distance optical links between separate data centres are not counted.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Data-centre switches | 53% | 45–61% | IDC · 18 March 2026 |
| Optical transceivers | 27% | 22–35% | optics.org, reporting Yole Group · 1 July 2026 |
| Copper links and connectors | 10% | 5–15% | Amphenol Corporation / U.S. SEC · 11 February 2026 |
| Fibre and passive connectivity | 10% | 6–14% | Converge Digest, reporting Corning's fourth-quarter 2025 results · 28 January 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | NVIDIA Networking (ex-Mellanox)Top 10 | 17.3% | 100 | 73 |
| 2 | InnoLightTop 10 | 6.8% | 53 | 47 |
| 3 | AmphenolTop 10 | 6.5% | 34 | 56 |
| 4 | AristaTop 10 | 9.7% | 33 | 21 |
| 5 | EoptolinkTop 10 | 4.4% | 27 | 93 |
| 6 | CiscoTop 10 | 8.7% | 26 | 50 |
| 7 | CoherentTop 10 | 3.6% | 25 | 27 |
| 8 | CelesticaTop 10 | 4.7% | 16 | 77 |
| 9 | Corning | 2.1% | 13 | 49 |
| 10 | Fujikura | 2% | 11 | 37 |
| 11 | Huawei | 2.9% | 9.5 | — |
| 12 | Accton | 3.4% | 9.1 | 47 |
| 13 | CredoRising | 1.2% | 5.9 | 87 |
| 14 | AccelinkRising | 1.3% | 4.0 | 57 |
| 15 | Ruijie Networks | 1.3% | 3.9 | 30 |
| 16 | H3C | 1.3% | 3.9 | — |
| 17 | HPE | 1.4% | 3.5 | — |
| 18 | Hisense Broadband (Nazhen Technology) | 1% | 3.2 | 50 |
| 19 | Foxconn | 1.2% | 3.1 | — |
| 20 | Luxshare Precision | 1% | 2.9 | 13 |
| 21 | Sumitomo Electric | 0.7% | 2.8 | — |
| 22 | Source Photonics | 0.8% | 2.5 | — |
| 23 | TE Connectivity | 0.8% | 2.4 | — |
| 24 | HGTECH (HG Genuine) | 0.8% | 2.2 | — |
| 25 | Molex | 0.7% | 2.0 | — |
| 26 | BizLink | 0.6% | 1.6 | — |
| 27 | Delta Electronics | 0.6% | 1.3 | — |
| 28 | Furukawa Electric (Lightera) | 0.3% | 1.2 | — |
| 29 | Lumentum | 0.5% | 1.2 | — |
| 30 | Dell | 0.5% | 1.1 | — |
| 31 | ZTE | 0.4% | 0.8 | — |
| 32 | US Conec | 0.2% | 0.7 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 97.7% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.85.
No rating was more than a level apart between the two passes.
Limits of the evidence
- Totals cover all data-centre networking (IDC data-centre Ethernet, Yole datacom optics), not AI clusters alone.
- InfiniBand switch sales are not published; they are estimated from Dell'Oro's statement that Ethernet held more than two-thirds of AI back-end switch sales in 2025.
- IDC's public release does not say whether switches sold directly by manufacturers to cloud companies are inside its data-centre Ethernet total.
- Yole's transceiver total includes telecom optics; the data-centre portion is estimated at 85–90%.
- Yole counts transceivers at the module maker's price, while a brand that resells modules (such as NVIDIA LinkX) is credited at its own price, so its share is measured slightly high.
- No free source gives a market total for copper cable assemblies or data-centre fibre; both rest on the leading makers' reported sales (Amphenol, Corning) and estimates, and carry the widest ranges.
- Lasers and photonic chips inside modules are credited to the module seller, because no maker of them is a near-sole supplier.
- Some makers report data-centre sales inside wider segments (Cisco, Huawei, Corning, Fujikura, TE Connectivity), so their shares rest on estimates.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: NVIDIA Networking (ex-Mellanox), InnoLight, Amphenol, Arista, Eoptolink, Cisco, Coherent, Celestica. Rising places: Credo, Accelink.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Cloud
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- AI computing rented to customers as a service, measured by revenue and weighted by AI server spending over the latest four quarters.
- Parts
- 74% GPU computing outside mainland China: Rented GPU clusters carry much of AI training and inference; GPU servers outside China are about three-quarters of AI server spending.
- 17% Computing on clouds' own AI chips: Capacity on chips offered mainly by the cloud that designed them, such as Google TPUs and AWS Trainium; about a sixth of spending.
- 9% AI computing rented in mainland China: A separate market under export controls, on imported and domestic chips; about a tenth of spending.
- Shares from
- Company GPU rental revenue outside China over the total (company results, Synergy Research); Company revenue from renting its own chips outside China over the total (company statements); Company share of China's AI cloud market (Omdia, IDC China)
- Hard to replace
- GPU clusters take a year or more to build and are reserved for years, so lost capacity cannot quickly move elsewhere. A cloud's own chips are offered mainly by that cloud.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
AI computing capacity that providers rent to customers as a service (GPU and other accelerator instances, clusters and reserved capacity), whether the provider owns its data centres or leases them; leased buildings are scored in Campuses with their operator, servers in Servers & systems, and model APIs sold per token and dedicated model deployments in Models · Inference serving, so each layer of the same capacity is counted once, and computing a company keeps for its own products is not counted here.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| GPU computing outside mainland China | 74% | 66–87% | IDC · 21 July 2026 |
| Computing on clouds' own AI chips | 17% | 6–21% | IDC · 15 June 2026 |
| AI computing rented in mainland China | 9% | 7–12% | IDC · 21 July 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | Amazon Web Services (AWS)Top 10 | 17% | 100 | — |
| 2 | Google CloudTop 10 | 15.3% | 100 | — |
| 3 | Oracle Cloud Infrastructure (OCI)Top 10 | 17.2% | 84 | 92 |
| 4 | Microsoft Azure / Intelligent CloudTop 10 | 15.5% | 76 | 63 |
| 5 | CoreWeaveTop 10 | 8.8% | 39 | 25 |
| 6 | Alibaba Cloud (Cloud Intelligence Group)Top 10 | 3.8% | 22 | — |
| 7 | Volcano Engine (incl. BytePlus)Top 10 | 1.9% | 9.1 | — |
| 8 | NebiusTop 10 | 2% | 5.4 | 75 |
| 9 | LambdaTop 10 | 1.3% | 2.8 | 13 |
| 10 | Baidu AI CloudTop 10 | 0.9% | 2.4 | — |
| 11 | IREN | 0.9% | 1.9 | — |
| 12 | Tencent Cloud | 0.7% | 1.8 | — |
| 13 | Huawei Cloud | 0.6% | 1.6 | — |
| 14 | China Telecom (Tianyi Cloud) | 0.6% | 1.5 | — |
| 15 | Crusoe | 0.7% | 1.2 | — |
| 16 | Together AI | 0.6% | 1.0 | — |
| 17 | Kingsoft Cloud | 0.4% | 0.9 | — |
| 18 | Core42 (G42) | 0.4% | 0.7 | — |
| 19 | China Mobile Cloud | 0.3% | 0.7 | — |
| 20 | Fluidstack | 0.4% | 0.7 | — |
| 21 | Modal | 0.4% | 0.7 | — |
| 22 | Voltage Park | 0.3% | 0.4 | — |
| 23 | Nscale | 0.2% | 0.4 | — |
| 24 | China Unicom Cloud | 0.2% | 0.4 | — |
| 25 | Runpod | 0.2% | 0.4 | — |
| 26 | SenseTime (SenseCore) | 0.2% | 0.3 | — |
| 27 | NVIDIA | 0.2% | 0.3 | — |
| 28 | Rumble (Northern Data, Taiga Cloud) | 0.2% | 0.3 | — |
| 29 | Sakura Internet | 0.1% | 0.2 | — |
| 30 | Mistral AI | 0.1% | 0.1 | — |
| 31 | NAVER Cloud | 0.1% | 0.1 | — |
| 32 | TensorWave | 0.1% | 0.1 | — |
| 33 | Vultr | 0.1% | 0.1 | — |
| 34 | GMO Internet | 0.1% | 0.1 | — |
| 35 | Yotta Data Services (Shakti Cloud) | 0.1% | 0.1 | — |
| 36 | NexGen Cloud (Hyperstack) | 0.1% | 0.1 | — |
| 37 | SoftBank Corp. | <0.1% | 0.1 | — |
| 38 | KDDI (GPU cloud) | <0.1% | 0.1 | — |
| 39 | NTT (GPU cloud services) | <0.1% | 0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 75% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.68.
No rating was more than a level apart between the two passes.
Limits of the evidence
- Weights use AI server spending over four quarters (IDC), a proxy for the installed fleet that earns rental revenue; the older fleet has fewer custom-chip servers.
- IDC does not split 3Q25 and 4Q25 AI servers by chip type; those two quarters are interpolated between 2Q25 and 1Q26.
- IDC's 'other accelerated' servers include custom chips used only inside their owners' products, which overstates capacity on custom chips that customers can rent.
- Hyperscalers report AI revenue as run-rates that include model APIs and some applications, and none reports rental revenue by chip type; their shares rest on estimates.
- China's share comes from IDC's AI infrastructure total for 1Q26 and a matched first-half 2025 comparison, and its GPU and non-GPU split from IDC China's accelerator mix.
- Private GPU clouds (Lambda, Crusoe, Fluidstack, Together) report revenue rarely; Synergy Research's neocloud total is the main cross-check.
- Installed accelerators and contracted power (MW) are reported by specialist clouds only, so they serve as off-basis cross-checks.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Amazon Web Services (AWS), Google Cloud, Oracle Cloud Infrastructure (OCI), Microsoft Azure / Intelligent Cloud, CoreWeave, Alibaba Cloud (Cloud Intelligence Group), Volcano Engine (incl. BytePlus), Nebius. Rising places filled by the next Importance ranks: Lambda, Baidu AI Cloud.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Models
Frontier labs
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- API spending on closed general-purpose and coding models, counted for each model's developer.
- Parts
- 35% Coding and software agents: API use of general and code-specialised models for writing and running code, the largest single workload; a few models lead it.
- 63% General reasoning, text and agents: Chat, analysis, search, customer service and other agents: most of the spending, with more alternatives.
- 2% Closed models in mainland China: A separate market without the main foreign models; small in revenue at Chinese prices.
- Shares from
- Coding spend shares from Menlo Ventures surveys and gateway data, scaled to each developer's reported API revenue; units of larger groups from spend-share or usage data, never the parent's revenue; Enterprise and gateway spend shares net of coding, and reported API revenue; usage share for units of larger groups; IDC platform shares, closed models only, and company disclosures
- Hard to replace
- Customers can switch models within weeks through an API. Top-level capability is hard to replace, above all in coding agents, where one or two developers lead; a new frontier model takes years and very large computing budgets.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Paid API access to closed general-purpose and coding models (models whose weights are not published), including their embedding and real-time voice versions, counted for the model's developer whether bought from the developer, through a cloud platform such as Amazon Bedrock, Google Vertex AI or Microsoft Foundry, or as a licence to another platform. Consumer apps and subscriptions count in Applications · Consumer, business seats in Enterprise agents, coding-tool plans in Developer tools, open-weight models in Open-weight, serving by companies that did not build the model in Inference serving, and image, video, music and speech-only models in Specialist models.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Coding and software agents | 35% | 25.5–54.5% | Menlo Ventures · 9 December 2025 |
| General reasoning, text and agents | 63% | 41.3–73.5% | Anthropic · 15 January 2026 |
| Closed models in mainland China | 2% | 0.9–4.2% | IDC · 7 May 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | AnthropicTop 10 | 63.4% | 100 | 92 |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 2 | OpenAITop 10 | 20% | 20 | 0 |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 3 | Google DeepMindTop 10 | 9.8% | 7.4 | 42 |
| 4 | ByteDance SeedTop 10 | 1% | 1.1 | — |
| 5 | Alibaba Qwen (AI Labs and Applications)Top 10 | 0.7% | 0.7 | — |
| 6 | xAI (SpaceXAI)Top 10 | 1% | 0.7 | — |
| 7 | Amazon AGI (Nova)Top 10 | 0.5% | 0.3 | — |
| 8 | Mistral AITop 10 | 0.2% | 0.2 | — |
| 9 | Baidu ERNIETop 10 | 0.2% | 0.1 | — |
| 10 | Meta Superintelligence LabsTop 10 | 0.1% | 0.1 | — |
| 11 | Tencent Hy | 0.1% | 0.1 | — |
| 12 | Cohere | 0.1% | 0.1 | — |
| 13 | iFlytek (Spark) | 0.1% | 0.1 | — |
| 14 | TypeSafe AI | 0.1% | <0.1 | — |
| 15 | Sber (GigaChat) | 0.1% | <0.1 | — |
| 16 | Perplexity | 0.1% | <0.1 | — |
| 17 | NAVER HyperCLOVA X | 0.1% | <0.1 | — |
| 18 | Moonshot AI | <0.1% | <0.1 | — |
| 19 | SenseTime (SenseNova) | <0.1% | <0.1 | — |
| 20 | Microsoft AI (MAI) | <0.1% | <0.1 | — |
| 21 | China Mobile | <0.1% | <0.1 | — |
| 22 | Zhipu AI (Z.ai) | <0.1% | <0.1 | — |
| 23 | Huawei Cloud | <0.1% | <0.1 | — |
| 24 | StepFun | <0.1% | <0.1 | — |
| 25 | Kuaishou Kwaipilot (KAT-Coder) | <0.1% | <0.1 | — |
| 26 | Writer | <0.1% | <0.1 | — |
| 27 | NTT (tsuzumi) | <0.1% | <0.1 | — |
| 28 | Upstage | <0.1% | <0.1 | — |
| 29 | MiniMax | <0.1% | <0.1 | — |
| 30 | Inception | <0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 100% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.78.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No public source splits model API spending by workload; the coding weight combines Menlo Ventures' coding shares with Anthropic's own API traffic mix (November 2025 data), before most of 2026's growth in coding agents. Menlo Ventures is an investor in Anthropic.
- Developers count sales through cloud partners differently: Anthropic includes them in revenue and OpenAI does not, so shares must be rebuilt on the card's basis.
- Sales through cloud platforms are counted in full for the model's developer here, and the platform's retained part is counted again in Inference serving.
- OpenAI, Google and SpaceX (xAI) do not report API revenue separately from subscriptions; SpaceX's Super Intelligence segment also includes Grok subscriptions and Cursor. Their shares come from spend-share and usage data.
- Spending on OpenAI models sold through Microsoft Azure is not disclosed; the weights use a 5LayerCake estimate (US$3–12bn a year) built from reported revenue-share payments.
- IDC's China totals mix closed models with open models served on the same platforms; the closed-model portion is estimated (40 to 70%).
- Spend-share data come from samples (US enterprises, developer gateways), not the whole market.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Anthropic, OpenAI, Google DeepMind, ByteDance Seed, Alibaba Qwen (AI Labs and Applications), xAI (SpaceXAI), Amazon AGI (Nova), Mistral AI. Rising places filled by the next Importance ranks: Baidu ERNIE, Meta Superintelligence Labs.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Open-weight and regional
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- How much open-weight language models are used: tokens served through APIs, downloads for running on companies' own servers, and copies run on local devices.
- Parts
- 26% Hosted use through APIs: Open models bought per token from their developers or from inference providers.
- 60% Self-hosted in companies' own cloud or data centres: Companies run downloaded weights on servers they control; the most common use.
- 14% Local and on-device use: Smaller models run on laptops, phones and edge devices.
- Shares from
- Token shares from OpenRouter, Vercel AI Gateway and the ATOM Report; IDC China token volumes re-attributed to each model's base developer (platform figures are off-basis); Download shares of full weights (Hugging Face, ModelScope; ATOM Report), checked against derivative-model shares; Downloads of local-format copies (GGUF, MLX) and Ollama pulls
- Hard to replace
- Published weights stay available if their developer stops, and users can move to another open model within weeks. What is hard to replace is a steady supply of new, capable releases and the tools and fine-tunes built around one model family.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Use of open-weight language models (weights published for download under any licence), including code, reasoning and vision-language models, counted for the developer that trained the base weights; fine-tuned, quantised and merged copies count for the base model's developer, a developer's own API sales of its open models count here as hosted use, closed models, including regional ones, count in Frontier labs, serving revenue counts in Inference serving, open image, video and speech models count in Specialist models through their developers' revenue, and embedding models are not scored.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Hosted use through APIs | 26% | 15–50% | The Linux Foundation · 19 November 2024 |
| Self-hosted in companies' own cloud or data centres | 60% | 35–70% | The Linux Foundation · 19 November 2024 |
| Local and on-device use | 14% | 7–25% | The Linux Foundation · 19 November 2024 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | Alibaba Qwen (AI Labs and Applications)Top 10 | 40.6% | 100 | 72 |
| 2 | DeepSeekTop 10 | 16% | 32 | 56 |
| 3 | Google DeepMindTop 10 | 8.6% | 12 | 50 |
| 4 | Zhipu AI (Z.ai)Top 10 | 6.3% | 9.2 | 72 |
| 5 | Meta Superintelligence LabsTop 10 | 5.3% | 7.4 | 0 |
| 6 | NVIDIA NemotronTop 10 | 3.5% | 4.5 | 89 |
| 7 | Tencent HyTop 10 | 3% | 3.9 | 78 |
| 8 | XiaomiTop 10 | 2.4% | 3.9 | 50 |
| 9 | OpenAITop 10 | 1.6% | 2.1 | 17 |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 10 | Mistral AITop 10 | 1.2% | 1.6 | 17 |
| 11 | Moonshot AI | 0.9% | 1.5 | — |
| 12 | Hugging Face | 1% | 1.3 | — |
| 13 | ModelBest (MiniCPM) | 0.8% | 1.2 | — |
| 14 | Microsoft (Phi) | 0.7% | 0.9 | — |
| 15 | Poolside (Laguna) | 0.6% | 0.8 | — |
| 16 | MiniMax | 0.6% | 0.8 | — |
| 17 | StepFun | 0.5% | 0.7 | — |
| 18 | Thinking Machines | 0.5% | 0.7 | — |
| 19 | IBM Granite | 0.5% | 0.6 | — |
| 20 | Baidu ERNIE | 0.4% | 0.5 | — |
| 21 | Liquid AI | 0.3% | 0.3 | — |
| 22 | LG AI Research (EXAONE) | 0.2% | 0.3 | — |
| 23 | ByteDance Seed | 0.1% | 0.2 | — |
| 24 | Ant Group (Ling and Ring) | 0.1% | 0.2 | — |
| 25 | Cohere | 0.1% | 0.2 | — |
| 26 | Reflection | <0.1% | <0.1 | — |
| 27 | Meituan (LongCat) | <0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 100% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 1.00.
No rating was more than a level apart between the two passes.
Limits of the evidence
- The split of use comes from one 2024 survey that counts organisations, not volume, and covers all generative-AI models (its managed-API answers include closed models); the card uses organisations whose AI infrastructure is mostly open source as the proxy. Hosted use of open models has grown fast since (Vercel AI Gateway: open-weight models carried 71.5% of its tokens and about 23% of its spend in September 2026).
- Rising growth uses one measure for every developer: monthly downloads of its open language models, including derivatives, on Hugging Face and ModelScope, from the same source in both periods.
- Current commercial activity, for free models, means the developer's open models are available through at least one paid hosting provider, or a commercial deployment of them is documented.
- Downloads are noisy: automated pipelines and small models inflate counts, so the ATOM Report's size-adjusted adoption measure is used as a check.
- Router and gateway data cover developer traffic, not large enterprise contracts; China's hosted use is measured separately by IDC, whose totals mix open and closed models.
- Use of self-hosted models inside companies is not observed directly; downloads stand in for it.
- Attribution of fine-tuned copies depends on publishers tagging the base model correctly on Hugging Face.
- Free use of open image, video, speech and embedding models is not measured in any sublayer.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Alibaba Qwen (AI Labs and Applications), DeepSeek, Google DeepMind, Zhipu AI (Z.ai), Meta Superintelligence Labs, NVIDIA Nemotron, Tencent Hy, Xiaomi. Rising places filled by the next Importance ranks: OpenAI, Mistral AI.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Specialist models
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Revenue that specialist models' developers earn from their output: voices, music, images, video, robot skills and molecule designs.
- Parts
- 24% Speech and voice: Text-to-speech, transcription and voice-agent models.
- 6% Music: Song-generation models, sold mostly as creator subscriptions.
- 60% Images and video: Image and video generation for creators, advertisers and studios; most of the revenue.
- 7% Avatar video: Presenter and translated videos made for business training, sales and marketing.
- 3% Robot and science models: Models that control robots or design molecules; small revenue so far.
- Shares from
- Company ARR statements and press reports; labelled estimates for large companies' speech services; Company ARR and subscriber statements; Kling and Firefly disclosures, company statements and press reports; labelled estimates for large companies' units; Company ARR statements and press reports; Company revenue statements and disclosed partnership payments
- Hard to replace
- Most specialist outputs have several capable models, and an app can switch within weeks. Harder to replace are voice and avatar models built into business workflows, and robot and science models, where few developers have the data and industry partners at scale.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Models built for one kind of output (speech and voice, music, images and video, avatar video, robot control, and molecule and materials design), counted for their developer, including such models from large companies such as Google, OpenAI, ByteDance, Microsoft, Amazon and Adobe. Stand-alone specialist products and specialist-model APIs and licences count (for example Kling, Midjourney, Suno, Firefly plans and the Veo API); specialist features inside general assistants or creative suites (ChatGPT, the Gemini app, Doubao, Creative Cloud) count in Applications. General-purpose and coding models count in Frontier labs or Open-weight, hosting of others' models in Inference serving, apps built on other companies' models in Applications, and text-translation models are not scored in this version.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Speech and voice | 24% | 12.9–43.3% | TechCrunch · 24 September 2026 |
| Music | 6% | 3.4–11.9% | TechCrunch · 27 February 2026 |
| Images and video | 60% | 36.7–74.9% | Kuaishou Technology (HKEXnews) · 19 August 2026 |
| Avatar video | 7% | 4.3–13.2% | HeyGen · 25 June 2026 |
| Robot and science models | 3% | 1.5–8.3% | Skild AI · 10 September 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | ByteDance SeedTop 10 | 14.7% | 100 | — |
| 2 | ElevenLabsTop 10 | 7.9% | 75 | 75 |
| 3 | SunoTop 10 | 4.4% | 53 | — |
| 4 | Google DeepMindTop 10 | 8.7% | 47 | — |
| 5 | Kling AITop 10 | 7.9% | 43 | 92 |
| 6 | MidjourneyTop 10 | 6.4% | 39 | — |
| 7 | Adobe FireflyTop 10 | 4.4% | 37 | — |
| 8 | OpenAITop 10 | 5% | 33 | — |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 9 | HeyGenRising | 3% | 26 | 50 |
| 10 | Skild AIRising | 1.4% | 24 | 83 |
| 11 | Synthesia | 2.3% | 22 | 33 |
| 12 | Runway | 3.2% | 17 | — |
| 13 | Microsoft AI (MAI) and Azure AI Speech | 2.7% | 15 | — |
| 14 | SoundHound AI | 1.5% | 15 | 22 |
| 15 | iFlytek (speech) | 1.3% | 13 | 8 |
| 16 | Amazon AGI (Nova) | 2.2% | 12 | — |
| 17 | Black Forest Labs | 1.5% | 8.3 | — |
| 18 | Field AI | 0.6% | 7.7 | — |
| 19 | Deepgram | 1.3% | 6.9 | — |
| 20 | Topaz Labs | 0.7% | 6.2 | — |
| 21 | AssemblyAI | 1% | 5.3 | — |
| 22 | xAI (SpaceXAI) | 1% | 5.3 | — |
| 23 | Meshy | 0.8% | 4.9 | — |
| 24 | Photoroom | 0.9% | 4.8 | — |
| 25 | AIsphere (PixVerse) | 0.8% | 4.3 | — |
| 26 | Alibaba Qwen (AI Labs and Applications) | 0.8% | 4.3 | — |
| 27 | MiniMax | 0.7% | 4.0 | — |
| 28 | Stability AI | 0.7% | 3.8 | — |
| 29 | Mirage (Captions) | 0.7% | 3.7 | — |
| 30 | Hume AI | 0.6% | 3.1 | — |
| 31 | Chai Discovery | 0.3% | 3.1 | — |
| 32 | Leonardo.Ai | 0.5% | 2.6 | — |
| 33 | Lightricks (LTX) | 0.5% | 2.6 | — |
| 34 | Krea | 0.4% | 2.2 | — |
| 35 | Luma AI | 0.4% | 2.2 | — |
| 36 | Shengshu (Vidu) | 0.4% | 2.2 | — |
| 37 | Cartesia | 0.4% | 2.1 | — |
| 38 | Decart | 0.3% | 1.7 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 95% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.90.
No rating was more than a level apart between the two passes.
Limits of the evidence
- Large companies do not disclose revenue for their specialist models (Google Veo, Imagen and speech services; OpenAI's image and audio APIs; Microsoft and Amazon speech services; ByteDance Seedance and Seedream), so the weights include 5LayerCake estimates for them; Adobe reports Firefly ARR.
- Midjourney and Black Forest Labs publish no current revenue; their sizes are estimates built from older company and press figures.
- Midjourney's low end (0.30) rests on TIME's reported $300m for 2024; the stated estimate method alone would give 0.24.
- Science models earn mostly through partnership payments that are not disclosed.
- Companies define ARR differently, and the latest figures range from February to September 2026.
- Code models have no part here: closed code models count in Frontier labs' coding part, open ones in Open-weight, and coding tools in Applications · Developer tools.
- Text-translation models (for example DeepL) have no current public revenue data and are not scored in this version.
- Creation apps are also measured in Applications · Consumer, by monthly users (an allowed separate placement); HeyGen, Synthesia and Runway are counted here and not in Vertical AI.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: ByteDance Seed, ElevenLabs, Suno, Google DeepMind, Kling AI, Midjourney, Adobe Firefly, OpenAI. Rising places: Skild AI, HeyGen.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Inference and serving
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Revenue providers keep for running others' models, after any share paid to the model's developer.
- Parts
- 69% Cloud model platforms: Amazon, Microsoft, Google, Oracle, Databricks, Snowflake and other platforms serve other developers' models inside them.
- 22% Independent inference providers: Specialists serving open and custom language models, often faster or cheaper.
- 3% Image, video and audio model hosts: Independent hosts that run image, video and sound models for app builders.
- 6% Model serving in mainland China: Chinese clouds and specialists serving others' models in a separate market.
- Shares from
- Platform disclosures, developers' channel splits and fees, and press estimates; never a platform's total AI run-rate; Company run-rate statements and filings, less GPU rental and lab-bought capacity; Company revenue statements net of payments to model developers; IDC platform shares, third-party models only
- Hard to replace
- Most serving runs on standard chips and open software, so customers can move within weeks. Harder to replace are clouds holding enterprise contracts, data and security approvals, and the few providers with their own fast inference chips.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Running models for others, sold per token, per request or as dedicated deployments (including deployments billed by the hour when the provider runs and manages the model), when the provider did not develop the model. Any model served by a cloud or data platform (Amazon, Microsoft, Google, Oracle, Databricks, Snowflake and others) counts in Cloud model platforms; independent providers count in the language-model or the image, video and audio part, and all serving in mainland China in its own part. A developer selling its own models counts in Frontier labs, Open-weight or Specialist models; bare accelerator instances and capacity a model developer buys to serve its own users count in Infrastructure · Cloud; routing gateways that do not run models are outside this sublayer.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Cloud model platforms | 69% | 45–85.2% | Reuters (published by Khaleej Times) · 30 September 2026 |
| Independent inference providers | 22% | 10.3–42.3% | Fireworks AI · 15 July 2026 |
| Image, video and audio model hosts | 3% | 0.9–9.2% | TechCrunch · 9 December 2025 |
| Model serving in mainland China | 6% | 2–15.9% | IDC · 7 May 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | Microsoft Azure / Intelligent CloudTop 10 | 35.3% | 100 | — |
| 2 | Amazon Web Services (AWS)Top 10 | 20.1% | 57 | — |
| 3 | Google CloudTop 10 | 9.6% | 22 | — |
| 4 | Fireworks AITop 10 | 6.4% | 11 | — |
| 5 | falTop 10 | 2.1% | 7.4 | — |
| 6 | BasetenTop 10 | 3.6% | 5.8 | — |
| 7 | Cerebras InferenceTop 10 | 1.6% | 5.1 | — |
| 8 | Together AITop 10 | 2.9% | 4.1 | — |
| 9 | ModalTop 10 | 2.5% | 4.0 | — |
| 10 | Volcano Engine (incl. BytePlus)Top 10 | 1.3% | 2.6 | — |
| 11 | Oracle Cloud Infrastructure (OCI) | 1.3% | 2.0 | — |
| 12 | Groq | 1% | 1.7 | — |
| 13 | Databricks | 1% | 1.7 | — |
| 14 | Snowflake | 0.8% | 1.4 | — |
| 15 | Alibaba Cloud (Cloud Intelligence Group) | 0.8% | 1.2 | — |
| 16 | DeepInfra | 0.6% | 0.7 | — |
| 17 | DigitalOcean | 0.5% | 0.6 | — |
| 18 | Baidu AI Cloud | 0.5% | 0.6 | — |
| 19 | China Mobile Cloud | 0.5% | 0.6 | — |
| 20 | Huawei Cloud | 0.4% | 0.5 | — |
| 21 | Tencent Cloud | 0.4% | 0.5 | — |
| 22 | CoreWeave | 0.4% | 0.5 | — |
| 23 | Crusoe | 0.4% | 0.5 | — |
| 24 | Workers AI | 0.4% | 0.5 | — |
| 25 | Hugging Face | 0.3% | 0.4 | — |
| 26 | Nebius | 0.3% | 0.4 | — |
| 27 | PPIO | 0.3% | 0.4 | — |
| 28 | IBM | 0.3% | 0.3 | — |
| 29 | China Telecom Cloud (eSurfing) | 0.2% | 0.3 | — |
| 30 | Runpod | 0.2% | 0.3 | — |
| 31 | SambaNova (SambaCloud) | 0.2% | 0.2 | — |
| 32 | SAP | 0.2% | 0.2 | — |
| 33 | SiliconFlow | 0.2% | 0.2 | — |
| 34 | Novita AI | 0.1% | 0.2 | — |
| 35 | FriendliAI | 0.1% | 0.2 | — |
| 36 | Thinking Machines | 0.1% | 0.2 | — |
| 37 | Infinigence AI | 0.1% | 0.1 | — |
| 38 | Chutes | <0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 96.8% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.65.
No rating was more than a level apart between the two passes.
Limits of the evidence
- Amazon, Microsoft and Google do not disclose model-serving revenue; the cloud weight rests on Anthropic's 2025 channel split and 16% platform fee, carried to 2026, plus a 5LayerCake estimate for OpenAI models sold through Microsoft Azure (US$3–12bn a year gross, built from reported revenue-share payments).
- Microsoft's reported 20% revenue share with OpenAI is not confirmed by either company.
- Sales through cloud platforms are counted in full for the model's developer in Frontier labs, and the platform's retained part is counted again here.
- Most independent providers are private and report run-rates that mix serving with GPU rental, training, batch jobs or sandboxes; Baseten publishes growth but no revenue figure, so its size is an estimate.
- fal and similar hosts report gross revenue, including payments to the developers of the proprietary models they host; their net figures are estimates.
- China's MaaS totals mix platforms' own models with third-party models; the third-party portion is estimated (20 to 45%).
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Microsoft Azure / Intelligent Cloud, Amazon Web Services (AWS), Google Cloud, Fireworks AI, fal, Baseten, Cerebras Inference, Together AI. Rising places filled by the next Importance ranks: Modal, Volcano Engine (incl. BytePlus).
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Data and evaluation
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Revenue from the data, training tasks and tests that model developers buy.
- Parts
- 63% Expert data for post-training: Specialists write answers, grade outputs and build reasoning examples; most of the spending.
- 16% Large-scale labelling and multimodal data: High-volume labelling of images, video, speech and sensor data.
- 15% RL environments and agent tasks: Simulated software and graded tasks for training agents; a newer part of lab spending.
- 4% Independent model evaluation: Leaderboards, expert test sets and red-teaming that labs pay for.
- 2% Labelling and data software: Software for labelling, curating and generating training data; small next to the services.
- Shares from
- Company run-rate statements and press reports, split by each vendor's services; Filings of listed vendors (TaskUs, Appen, Innodata, DataOcean AI) and company statements; Environment contracts in company statements and press reports (labelled estimates); Company run-rate statements and press reports; Software revenue statements of labelling-platform companies
- Hard to replace
- Labs use several data vendors at once and can move work within months. Harder to replace are large pools of vetted experts, environment builders with particular skills, and community evaluations with years of comparison data.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Human data, training tasks and tests sold to model developers: expert data for post-training, large-scale labelling and multimodal data, reinforcement-learning environments, independent model evaluation, and labelling, data-curation and synthetic-data software; content licensing by publishers and platforms is outside this sublayer, evaluation and monitoring tools for companies' own AI applications count in Applications · Data platforms, AI applications that data vendors sell to enterprises and governments count in Applications, and labs' in-house data teams are not counted.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Expert data for post-training | 63% | 44–80.5% | TechCrunch · 8 July 2026 |
| Large-scale labelling and multimodal data | 16% | 6.8–28.1% | TaskUs (slides hosted by Investing.com) · 5 August 2026 |
| RL environments and agent tasks | 15% | 5.9–27.2% | Epoch AI · 12 January 2026 |
| Independent model evaluation | 4% | 2.2–7.9% | Arena · 29 June 2026 |
| Labelling and data software | 2% | 0.8–5.6% | Snorkel AI · 22 September 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | MercorTop 10 | 19% | 100 | 89 |
| 2 | Surge AITop 10 | 15.5% | 72 | — |
| 3 | Handshake AITop 10 | 9.2% | 34 | 74 |
| 4 | micro1Top 10 | 8.7% | 33 | 81 |
| 5 | Scale AITop 10 | 6.6% | 20 | — |
| 6 | ArenaTop 10 | 0.8% | 9.2 | — |
| 7 | SnorkelTop 10 | 3.3% | 8.9 | 67 |
| 8 | TuringTop 10 | 3% | 8.5 | — |
| 9 | InnodataTop 10 | 2.9% | 8.1 | 30 |
| 10 | Prolific | 2.6% | 7.5 | — |
| 11 | Appen | 2.4% | 7.0 | 9 |
| 12 | TaskUs | 2.3% | 6.6 | 28 |
| 13 | Invisible Technologies | 2.2% | 6.0 | — |
| 14 | TELUS Digital | 1.8% | 5.3 | — |
| 15 | RWS TrainAI | 1.8% | 5.2 | — |
| 16 | ProximalRising | 1.7% | 4.9 | 60 |
| 17 | Labelbox | 1.3% | 3.6 | — |
| 18 | TransPerfect DataForce | 1.1% | 3.2 | — |
| 19 | Toloka | 1% | 2.6 | — |
| 20 | Fleet | 0.8% | 2.4 | — |
| 21 | Uber AI Solutions | 0.7% | 1.9 | — |
| 22 | Welo Data | 0.7% | 1.9 | — |
| 23 | iMerit | 0.6% | 1.7 | — |
| 24 | Lionbridge (AI training services) | 0.6% | 1.7 | — |
| 25 | LXT | 0.6% | 1.7 | — |
| 26 | DataOcean AI | 0.5% | 1.4 | — |
| 27 | Encord | 0.4% | 1.1 | — |
| 28 | Datatang | 0.4% | 1.1 | — |
| 29 | Sama | 0.4% | 1.0 | — |
| 30 | Halluminate | 0.3% | 0.8 | — |
| 31 | Gray Swan AI | 0.2% | 0.7 | — |
| 32 | SuperAnnotate | 0.3% | 0.7 | — |
| 33 | Irregular | 0.1% | 0.7 | — |
| 34 | CloudFactory | 0.2% | 0.6 | — |
| 35 | AfterQuery | 0.2% | 0.6 | — |
| 36 | Artificial Analysis | 0.1% | 0.6 | — |
| 37 | Datacurve | 0.2% | 0.5 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 94% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.71.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No vendor discloses its revenue split across these functions; the weights split each vendor's run-rate by the services it describes (estimates), and the ranges allow each vendor's smaller parts to be half or one and a half times the stated split.
- Most vendors are private: run-rates are company statements or press reports, and some publish gross figures before contributor pay while others publish net (micro1 gave both in August 2026).
- Surge AI and Scale AI have not published current revenue; their sizes are estimates, and Scale's enterprise and government applications are left out.
- Units of larger groups (TELUS Digital, EXL's iMerit, Uber AI Solutions) are measured from their own AI-data revenue, never the parent's; TELUS Digital does not report it separately.
- Labs' in-house data teams and direct contractor hiring are not measured, so the totals understate lab spending on data.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Mercor, Surge AI, Handshake AI, micro1, Scale AI, Arena, Snorkel, Turing. Rising places: Proximal. Rising places filled by the next Importance ranks: Innodata.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Applications
Data platforms
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Revenue from the platforms that store, move, organise and monitor the data AI applications use, measured across each whole market because vendors do not report the AI-related part.
- Parts
- 79% Databases, warehouses and lakehouses: Store and query the data AI applications learn from and look up, including vector search; by far the largest market.
- 6% Data integration and streaming: Move and clean data from business systems into AI applications.
- 2% Data context and decision platforms: Map business data into shared definitions and rules that models and people act on.
- 13% Observability and AI evaluation: Monitor and test running applications and agents.
- Shares from
- Revenue each company reports in the function; estimates for cloud providers, private companies and undisclosed products, labelled
- Hard to replace
- Harder to replace than most applications, though easier than hardware: moving a production database or warehouse to another vendor takes months to a couple of years of migration and testing. Integration and observability tools can usually be switched within months.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
Databases, warehouses and lakehouses, data integration and streaming, data context and decision platforms, and observability used to build and run AI applications; dashboards and BI tools are outside this sublayer, model-serving platforms count in Models, cloud compute in Infrastructure, and applications sold to one industry or to government in Vertical AI.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Databases, warehouses and lakehouses | 79% | 73–84% | MongoDB, Inc. (SEC EDGAR), citing IDC · 11 March 2026 |
| Data integration and streaming | 6% | 5–8% | Blocks & Files (Chris Mellor), citing Gartner · 12 December 2025 |
| Data context and decision platforms | 2% | 1.5–4% | Palantir Technologies Inc. (SEC EDGAR) · 17 February 2026 |
| Observability and AI evaluation | 13% | 9–16% | Dash0 (Mirko Novakovic) · 13 December 2024 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | AmazonTop 10 | 23.5% | 100 | — |
| 2 | MicrosoftTop 10 | 20.6% | 88 | — |
| 3 | OracleTop 10 | 12.4% | 67 | — |
| 4 | AlphabetTop 10 | 9.4% | 32 | — |
| 5 | SAPTop 10 | 2.5% | 20 | — |
| 6 | IBMTop 10 | 3.8% | 19 | — |
| 7 | PalantirTop 10 | 0.9% | 8.3 | — |
| 8 | DatadogTop 10 | 2.9% | 8.1 | 44 |
| 9 | Alibaba | 1.9% | 7.0 | — |
| 10 | SnowflakeRising | 2.8% | 6.9 | 72 |
| 11 | MongoDB | 1.7% | 6.5 | 17 |
| 12 | DatabricksRising | 1.7% | 5.2 | 100 |
| 13 | Dynatrace | 1.5% | 3.7 | 33 |
| 14 | Tencent | 1.1% | 3.3 | — |
| 15 | Huawei | 0.9% | 2.7 | — |
| 16 | Salesforce | 0.8% | 2.5 | — |
| 17 | Cisco | 0.8% | 1.9 | — |
| 18 | Elastic | 0.8% | 1.8 | — |
| 19 | New Relic | 0.7% | 1.7 | — |
| 20 | Cloudera | 0.6% | 1.3 | — |
| 21 | Teradata | 0.6% | 1.2 | — |
| 22 | InterSystems | 0.5% | 1.0 | — |
| 23 | SolarWinds | 0.3% | 0.8 | — |
| 24 | Fivetran + dbt Labs | 0.4% | 0.8 | — |
| 25 | Cribl | 0.2% | 0.8 | — |
| 26 | Boomi | 0.3% | 0.7 | — |
| 27 | Grafana Labs | 0.3% | 0.7 | — |
| 28 | ByteDance | 0.2% | 0.4 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 94% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.81.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No public source splits data-platform spending by AI use, so both the weights and the company shares use whole markets; counting only AI-related spend would lower databases and raise decision platforms.
- The 2025 database figure extends IDC's 2024 total at its forecast growth rate; the observability figure lies between a vendor's bottom-up estimate grown at its stated rate and Gartner's 2029 forecast taken back to 2025.
- The decision-platform market is sized from Palantir's commercial revenue plus an estimate for other vendors.
- Most vendors do not report revenue by function, and the largest cloud database vendors (AWS, Microsoft, Google, Oracle) report none, so their shares are labelled estimates.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Amazon, Microsoft, Oracle, Alphabet, SAP, IBM, Palantir, Datadog. Rising places: Databricks, Snowflake.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Developer tools
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Revenue from AI tools that write, review and test code, and from tools that build apps from prompts.
- Parts
- 88% Coding assistants and agents: Assistants in the code editor and agents that write and change code on their own; most of the spend.
- 10% App builders: Tools that turn a written request into a working app, for developers and non-developers.
- 2% Code review and testing: AI review of code changes, test writing and release checks; small so far.
- Shares from
- Run-rate revenue each company reports for its product; for model developers' own coding tools, plan and seat revenue only; estimates for undisclosed products, labelled
- Hard to replace
- Easy to replace: developers can move to another coding assistant in days or weeks, and many teams pay for several. The scarcer part is the models behind the tools, which are counted in Models.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
AI products that write, review and test code or build working apps from prompts, sold to developers and other builders; coding tools sold by model developers, such as Claude Code, Codex and Google's coding tools, count for their plan and seat revenue, while use billed per token through the developer's API counts in Models · Frontier labs; databases count in Data platforms, hosting and cloud compute are outside this sublayer, and design tools count only for their app-building features.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Coding assistants and agents | 88% | 75–95% | Space Exploration Technologies Corp. (SEC EDGAR) · 20 October 2026 |
| App builders | 10% | 4–21% | TechCrunch (Ivan Mehta), quoting Lovable co-founder Fabian Hedin · 24 September 2026 |
| Code review and testing | 2% | 0.5–5% | Menlo Ventures · 9 December 2025 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | SpaceXTop 10 | 27.1% | 100 | 36 |
| 2 | AnthropicTop 10 | 17.5% | 92 | 86 |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 3 | MicrosoftTop 10 | 14.6% | 61 | — |
| 4 | OpenAITop 10 | 8.8% | 33 | — |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 5 | CognitionTop 10 | 5.9% | 19 | 76 |
| 6 | LovableTop 10 | 3.1% | 10 | 16 |
| 7 | ReplitTop 10 | 2.7% | 9.9 | 66 |
| 8 | Zhipu AI (Z.ai)Top 10 | 2.9% | 9.6 | — |
| 9 | AlphabetTop 10 | 2.1% | 6.9 | — |
| 10 | Wix (Base44) | 1% | 3.4 | — |
| 11 | Amazon | 0.9% | 2.9 | — |
| 12 | Factory | 0.9% | 2.9 | — |
| 13 | Vercel | 0.8% | 2.5 | — |
| 14 | EmergentRising | 0.6% | 2.0 | 56 |
| 15 | JetBrains | 0.6% | 2.0 | — |
| 16 | StackBlitz (Bolt) | 0.5% | 1.7 | — |
| 17 | MiniMax | 0.5% | 1.5 | — |
| 18 | Alibaba | 0.4% | 1.3 | — |
| 19 | Warp | 0.4% | 1.3 | — |
| 20 | Tricentis | 0.3% | 1.2 | — |
| 21 | Moonshot AI | 0.4% | 1.2 | — |
| 22 | CodeRabbit | 0.3% | 1.1 | — |
| 23 | GitLab | 0.3% | 1.0 | — |
| 24 | Retool | 0.2% | 0.8 | — |
| 25 | Atlassian | 0.2% | 0.8 | — |
| 26 | Figma | 0.2% | 0.7 | — |
| 27 | Amp | 0.2% | 0.6 | — |
| 28 | ByteDance | 0.2% | 0.6 | — |
| 29 | Mistral AI | 0.2% | 0.6 | — |
| 30 | Sonar (SonarSource) | 0.1% | 0.5 | — |
| 31 | Harness | 0.1% | 0.5 | — |
| 32 | Snyk | 0.1% | 0.5 | — |
| 33 | Baidu | 0.1% | 0.4 | — |
| 34 | Kilo Code (Anaconda) | 0.1% | 0.4 | — |
| 35 | Bubble | 0.1% | 0.4 | — |
| 36 | BrowserStack | 0.1% | 0.4 | — |
| 37 | Qodo | 0.1% | 0.4 | — |
| 38 | Cline | 0.1% | 0.4 | — |
| 39 | Tencent | 0.1% | 0.4 | — |
| 40 | SmartBear | 0.1% | 0.3 | — |
| 41 | Builder.io | 0.1% | 0.3 | — |
| 42 | Greptile | 0.1% | 0.3 | — |
| 43 | LambdaTest | 0.1% | 0.3 | — |
| 44 | Blitzy | 0.1% | 0.3 | — |
| 45 | Anything | 0.1% | 0.3 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 96.7% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was -0.03.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No published source splits 2026 spending across these functions; the weights add company run-rates, several of them estimated.
- Claude Code and Codex plan revenue sits inside Anthropic's and OpenAI's totals and is not reported separately; Claude Code's last stated run-rate (US$2.5bn, February 2026) also includes use billed through the API, so both plan figures are estimates, as is GitHub Copilot's revenue.
- Cursor's run-rate is annualised from 48 days of revenue SpaceX reported after the acquisition.
- Review and testing revenue has no public tracker; its weight is an estimate.
- Run-rates in this market change by tens of percent a quarter, so the mix should be re-checked at each quarterly review.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: SpaceX, Anthropic, Microsoft, OpenAI, Cognition, Lovable, Replit, Zhipu AI (Z.ai). Rising places: Emergent. Rising places filled by the next Importance ranks: Alphabet.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Enterprise software and agents
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Revenue from AI assistants and agents that businesses buy for their employees and business processes.
- Parts
- 72% Work assistants: Assistants for writing, research and search across company files; about three-quarters of spend.
- 7% Agent and automation platforms: Tools for building and running agents that work across business systems.
- 7% IT and service operations: Agents that resolve IT tickets, run internal service desks and assist security teams.
- 8% Sales and marketing: Agents and content tools for prospecting, campaigns and marketing material.
- 6% Finance, HR and operations: Agents inside finance, payroll, HR and supply-chain software.
- Shares from
- AI product revenue each company reports; seats times list price as a labelled estimate; a product also used for customer service, such as Agentforce, is split by use and never counted twice
- Hard to replace
- Mostly easy to replace: a company can change assistant or agent vendor within months, and many run several at once. Switching is slower where the agent is built into a system of record, such as CRM, ERP or IT service software, that holds the company's data and workflows.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
AI assistants and agents that businesses buy for employees' general work and for IT, sales and marketing, finance, HR and operations processes, with suite vendors counted only for their AI products; coding tools count in Developer tools, customer-service agents and tools built for one industry in Vertical AI, consumer plans in Consumer, model API sales in Models, and design and media-creation tools are outside this sublayer: app-building features count in Developer tools, and media models sold through their makers' own apps in Models · Specialist models.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Work assistants | 72% | 60–80% | Menlo Ventures · 9 December 2025 |
| Agent and automation platforms | 7% | 4–12% | Menlo Ventures · 9 December 2025 |
| IT and service operations | 7% | 4–12% | Menlo Ventures · 9 December 2025 |
| Sales and marketing | 8% | 5–12% | Menlo Ventures · 9 December 2025 |
| Finance, HR and operations | 6% | 3–10% | Menlo Ventures · 9 December 2025 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | OpenAITop 10 | 28.6% | 100 | — |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 2 | MicrosoftTop 10 | 20.9% | 80 | — |
| 3 | AlphabetTop 10 | 9.5% | 32 | — |
| 4 | AnthropicTop 10 | 8.1% | 27 | — |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 5 | ServiceNowTop 10 | 1.9% | 7.5 | — |
| 6 | WorkdayTop 10 | 1.5% | 6.4 | 50 |
| 7 | SalesforceTop 10 | 2.1% | 6.4 | 100 |
| 8 | SAPTop 10 | 1.4% | 4.8 | — |
| 9 | GongTop 10 | 1.1% | 3.9 | 0 |
| 10 | IBMTop 10 | 0.8% | 2.0 | — |
| 11 | Palo Alto Networks | 0.5% | 1.9 | — |
| 12 | CrowdStrike | 0.5% | 1.8 | — |
| 13 | Adobe | 0.6% | 1.4 | — |
| 14 | Glean | 0.7% | 1.4 | — |
| 15 | UiPath | 0.5% | 1.4 | — |
| 16 | Oracle | 0.4% | 1.3 | — |
| 17 | Intuit | 0.5% | 1.2 | — |
| 18 | Superhuman | 0.6% | 1.2 | — |
| 19 | Amazon | 0.6% | 1.2 | — |
| 20 | Perplexity | 0.5% | 0.9 | — |
| 21 | Clay | 0.4% | 0.8 | — |
| 22 | ADP | 0.3% | 0.7 | — |
| 23 | SpaceX | 0.4% | 0.7 | — |
| 24 | Outreach | 0.3% | 0.7 | — |
| 25 | Kingdee International | 0.2% | 0.7 | — |
| 26 | Docusign | 0.2% | 0.6 | — |
| 27 | Yonyou Network Technology | 0.2% | 0.6 | — |
| 28 | Atlassian | 0.2% | 0.5 | — |
| 29 | Notion | 0.3% | 0.5 | — |
| 30 | Automation Anywhere | 0.2% | 0.5 | — |
| 31 | Salesloft (with Clari) | 0.2% | 0.5 | — |
| 32 | ByteDance | 0.2% | 0.5 | — |
| 33 | Darktrace | 0.2% | 0.5 | — |
| 34 | Mistral AI | 0.2% | 0.5 | — |
| 35 | SentinelOne | 0.2% | 0.5 | — |
| 36 | Otter.ai | 0.2% | 0.4 | — |
| 37 | HighRadius | 0.2% | 0.4 | — |
| 38 | n8n | 0.2% | 0.4 | — |
| 39 | HubSpot | 0.2% | 0.4 | — |
| 40 | Infor | 0.1% | 0.4 | — |
| 41 | Writer | 0.2% | 0.4 | — |
| 42 | Freshworks | 0.1% | 0.3 | — |
| 43 | monday.com | 0.2% | 0.3 | — |
| 44 | Torq | 0.1% | 0.3 | — |
| 45 | Cohere | 0.2% | 0.3 | — |
| 46 | Alibaba | 0.1% | 0.3 | — |
| 47 | Tencent | 0.1% | 0.3 | — |
| 48 | Zoom Communications | 0.1% | 0.2 | — |
| 49 | Celonis (with Make) | 0.1% | 0.2 | — |
| 50 | Fireflies.ai | 0.1% | 0.2 | — |
| 51 | BILL Holdings | 0.1% | 0.2 | — |
| 52 | Zapier | 0.1% | 0.2 | — |
| 53 | 6sense | 0.1% | 0.2 | — |
| 54 | Eightfold AI | 0.1% | 0.2 | — |
| 55 | Paychex | 0.1% | 0.2 | — |
| 56 | UKG | 0.1% | 0.2 | — |
| 57 | Rippling | 0.1% | 0.2 | — |
| 58 | Granola | 0.1% | 0.2 | — |
| 59 | Phenom | 0.1% | 0.2 | — |
| 60 | Dropbox | 0.1% | 0.2 | — |
| 61 | Apollo.io | 0.1% | 0.2 | — |
| 62 | ZoomInfo (GTM) | 0.1% | 0.2 | — |
| 63 | Klaviyo | 0.1% | 0.1 | — |
| 64 | BMC Helix | 0.1% | 0.1 | — |
| 65 | Coupa | 0.1% | 0.1 | — |
| 66 | Airtable | 0.1% | 0.1 | — |
| 67 | Zscaler | 0.1% | 0.1 | — |
| 68 | Sage Group | <0.1% | 0.1 | — |
| 69 | o9 Solutions | <0.1% | 0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 93.3% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.40.
No rating was more than a level apart between the two passes.
Limits of the evidence
- The weights use Menlo Ventures' estimates of 2025 spend, built from a survey of US enterprises; no published source splits worldwide spend by function.
- Sales and finance were not sized; the unsized 6% of departmental spend is split equally between Sales and marketing and Finance, HR and operations.
- Most vendors report AI activity as company-defined run-rates, contract value, seats or customer counts rather than revenue; converting these to revenue gives estimates.
- Suite vendors (Microsoft, Salesforce, SAP, ServiceNow, Workday, Oracle) bundle AI into wider products, so their AI revenue by function is mostly estimated.
- Business plans of design tools built on other companies' models are not counted by revenue in this version.
- Spend grew several-fold in 2026; company figures suggest a similar mix, but it should be re-checked at each quarterly review.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: OpenAI, Microsoft, Alphabet, Anthropic, ServiceNow, Workday, Salesforce, SAP. Rising places filled by the next Importance ranks: Gong, IBM.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Vertical AI
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Revenue from AI applications, including predictive analytics and computer vision, built for one industry or profession.
- Parts
- 10% Healthcare and life sciences: Clinical notes, imaging, billing and drug-development tools.
- 2% Legal: Legal research, drafting and document review.
- 22% Financial services and insurance: The sector with the highest share of AI in its software spending.
- 7% Customer service: Agents that answer a business's customers; counted as a profession.
- 8% Government and public sector: Defence, intelligence and public-service applications.
- 25% Retail, property and other services: Retail planning, property and professional-service tools.
- 26% Industry, energy and infrastructure: Factory, construction, energy, telecom and transport tools.
- Shares from
- AI product revenue each company reports in the vertical; estimates for undisclosed products and incumbents' AI lines, labelled
- Hard to replace
- Mostly easy to replace: several vendors serve each profession, and a hospital or legal practice can change tool within a contract cycle. The exceptions are tools tied to systems of record or proprietary content, such as hospital record systems and legal databases, where switching takes years.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
AI applications built for one industry or profession, including predictive analytics, computer vision and generative tools, in healthcare and life sciences, legal, financial services and insurance, customer service, government, retail and other services, and industry and infrastructure; horizontal assistants and agents count in Enterprise agents, coding tools in Developer tools, consumer products in Consumer, search and ads, commercial data platforms in Data platforms, and image, video, voice and avatar tools built on their maker's own models in Models · Specialist models; AI-ranked advertising is not counted in this layer.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| Healthcare and life sciences | 10% | 7–18% | IDC (Mariya Yahnyuk) · 1 July 2026 |
| Legal | 2% | 1–5% | IDC (Mariya Yahnyuk) · 1 July 2026 |
| Financial services and insurance | 22% | 17–28% | IDC (Mariya Yahnyuk) · 1 July 2026 |
| Customer service | 7% | 5–12% | IDC (Mariana Fang) · 19 August 2024 |
| Government and public sector | 8% | 5–11% | IDC (Mariya Yahnyuk) · 1 July 2026 |
| Retail, property and other services | 25% | 17–32% | IDC (Mariya Yahnyuk) · 1 July 2026 |
| Industry, energy and infrastructure | 26% | 18–32% | IDC (Mariya Yahnyuk) · 1 July 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | PalantirTop 10 | 10.2% | 100 | 36 |
| 2 | Thomson ReutersTop 10 | 2.6% | 16 | — |
| 3 | PagayaTop 10 | 2.3% | 14 | — |
| 4 | Blue YonderTop 10 | 2.2% | 13 | 25 |
| 5 | HitachiTop 10 | 2.1% | 13 | — |
| 6 | UpstartTop 10 | 2.2% | 12 | — |
| 7 | SiemensTop 10 | 1.6% | 12 | — |
| 8 | AmadeusTop 10 | 1.4% | 11 | — |
| 9 | Wolters KluwerTop 10 | 2% | 11 | — |
| 10 | CCC Intelligent Solutions | 0.9% | 8.6 | — |
| 11 | Verisk | 0.7% | 8.6 | — |
| 12 | Applied Intuition | 1.3% | 8.3 | — |
| 13 | Visa | 1.2% | 8.1 | — |
| 14 | Mastercard | 1.2% | 7.6 | — |
| 15 | SamsaraRising | 2.2% | 7.5 | 64 |
| 16 | FICO | 1% | 7.4 | — |
| 17 | RELX (LexisNexis) | 1% | 6.7 | — |
| 18 | RealPage | 1% | 6.7 | — |
| 19 | Sabre | 0.9% | 6.6 | — |
| 20 | PROS | 1% | 6.1 | — |
| 21 | EliseAI | 1.1% | 6.0 | — |
| 22 | iFLYTEK | 1.3% | 6.0 | — |
| 23 | SLB | 0.8% | 5.9 | — |
| 24 | IQVIA | 0.6% | 5.5 | — |
| 25 | Hikvision | 0.9% | 5.5 | — |
| 26 | Experian | 0.9% | 5.3 | — |
| 27 | Manhattan Associates | 0.9% | 5.3 | — |
| 28 | Honeywell | 0.9% | 5.0 | — |
| 29 | Momenta | 0.6% | 4.9 | — |
| 30 | Huawei | 0.6% | 4.7 | — |
| 31 | Emerson (AspenTech) | 0.7% | 4.7 | — |
| 32 | Equifax | 0.7% | 4.2 | — |
| 33 | Autodesk | 0.7% | 4.0 | — |
| 34 | Schneider Electric | 0.7% | 4.0 | — |
| 35 | RELEX Solutions | 0.7% | 4.0 | — |
| 36 | Cognex | 0.8% | 3.9 | — |
| 37 | NICE (NiCE) | 0.7% | 3.8 | — |
| 38 | AlphaSense | 0.7% | 3.4 | — |
| 39 | TransUnion | 0.6% | 3.3 | — |
| 40 | PTC | 0.3% | 3.2 | — |
| 41 | Dassault Systèmes (Medidata, 3DEXPERIENCE) | 0.4% | 3.1 | — |
| 42 | Microsoft | 0.7% | 3.1 | — |
| 43 | Bloomberg L.P. | 0.4% | 3.1 | — |
| 44 | HeartFlow | 0.3% | 3.0 | — |
| 45 | Hexagon | 0.5% | 3.0 | — |
| 46 | AppFolio | 0.6% | 3.0 | — |
| 47 | Rockwell Automation | 0.5% | 2.9 | — |
| 48 | Tempus | 0.5% | 2.8 | — |
| 49 | GE Vernova | 0.4% | 2.7 | — |
| 50 | Axon Enterprise | 0.3% | 2.7 | — |
| 51 | SoundHound AI | 0.6% | 2.6 | — |
| 52 | Entrata | 0.5% | 2.5 | — |
| 53 | Genesys | 0.5% | 2.5 | — |
| 54 | Amdocs | 0.3% | 2.5 | — |
| 55 | Salesforce | 0.6% | 2.5 | — |
| 56 | Motive | 0.6% | 2.5 | — |
| 57 | Anduril Industries | 0.2% | 2.4 | — |
| 58 | Ant Group (Ant Digital Technologies) | 0.5% | 2.4 | — |
| 59 | Yardi | 0.5% | 2.4 | — |
| 60 | Trimble | 0.4% | 2.2 | — |
| 61 | Dahua Technology | 0.4% | 2.1 | — |
| 62 | Bentley Systems | 0.3% | 2.1 | — |
| 63 | S&P Global | 0.4% | 2.1 | — |
| 64 | Socure | 0.5% | 2.1 | — |
| 65 | Motorola Solutions | 0.3% | 2.0 | — |
| 66 | Verint | 0.4% | 2.0 | — |
| 67 | UnitedHealth Group (Optum) | 0.4% | 2.0 | — |
| 68 | GE HealthCare | 0.4% | 2.0 | — |
| 69 | ABB | 0.3% | 1.9 | — |
| 70 | IBM | 0.3% | 1.9 | — |
| 71 | Flock Safety | 0.3% | 1.7 | — |
| 72 | Baker Hughes | 0.3% | 1.7 | — |
| 73 | Epic Systems | 0.2% | 1.7 | — |
| 74 | Veeva Systems | 0.2% | 1.7 | — |
| 75 | Moody's | 0.3% | 1.6 | — |
| 76 | Hithink RoyalFlush | 0.3% | 1.5 | — |
| 77 | Philips | 0.3% | 1.5 | — |
| 78 | MRI Software | 0.3% | 1.5 | — |
| 79 | R1 RCM | 0.2% | 1.4 | — |
| 80 | Abridge | 0.4% | 1.4 | — |
| 81 | Bloomreach | 0.4% | 1.4 | — |
| 82 | Zendesk | 0.4% | 1.4 | — |
| 83 | Waystar | 0.3% | 1.3 | — |
| 84 | Alphabet | 0.4% | 1.3 | — |
| 85 | Amazon | 0.3% | 1.3 | — |
| 86 | Scale AI | 0.2% | 1.2 | — |
| 87 | Shopify | 0.3% | 1.2 | — |
| 88 | C3 AI | 0.3% | 1.0 | — |
| 89 | SenseTime | 0.3% | 1.0 | — |
| 90 | Algolia | 0.3% | 0.9 | — |
| 91 | Constructor | 0.2% | 0.7 | — |
| 92 | Halliburton | 0.2% | 0.7 | — |
| 93 | Horizon Robotics | 0.2% | 0.7 | — |
| 94 | Itron | 0.2% | 0.7 | — |
| 95 | Nemetschek | 0.2% | 0.7 | — |
| 96 | Tyler Technologies | 0.1% | 0.7 | — |
| 97 | Oracle | 0.1% | 0.6 | — |
| 98 | Cognite | 0.2% | 0.6 | — |
| 99 | Procore | 0.2% | 0.6 | — |
| 100 | LSEG | 0.2% | 0.6 | — |
| 101 | Quantexa | 0.2% | 0.6 | — |
| 102 | ServiceNow | 0.2% | 0.6 | — |
| 103 | Feedzai | 0.2% | 0.5 | — |
| 104 | Guidewire | 0.2% | 0.5 | — |
| 105 | Intapp | 0.2% | 0.5 | — |
| 106 | Glodon | 0.1% | 0.5 | — |
| 107 | Hundsun Technologies | 0.1% | 0.4 | — |
| 108 | FactSet | 0.1% | 0.4 | — |
| 109 | MSCI | 0.1% | 0.3 | — |
| 110 | Megvii | 0.1% | 0.3 | — |
| 111 | nCino | 0.1% | 0.2 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 75.4% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.53.
No rating was more than a level apart between the two passes.
Limits of the evidence
- The weights use IDC's sector mix of AI growth in digital-transformation software for 2024 to 2029, which covers all kinds of AI, including in-house builds; no public source measures vertical AI vendors' revenue by industry.
- Legal is not sized by IDC; its weight is an estimated part of Retail and Services. IDC's sector groupings are not published in detail, so which industries sit in Retail and Services or Industry and Infrastructure is partly estimated.
- Customer service is carved out of the industry totals at its 2024 share of all AI spending, the latest use-case figure IDC published openly.
- Incumbents such as Epic, Thomson Reuters, RELX, Wolters Kluwer, Veeva, FICO and Siemens rarely report AI revenue separately, so their AI lines are estimated.
- Spending on generative-AI applications alone has a very different mix (healthcare and legal much larger); see the notes.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Palantir, Thomson Reuters, Pagaya, Blue Yonder, Hitachi, Upstart, Siemens, Amadeus. Rising places: Samsara. Rising places filled by the next Importance ranks: Wolters Kluwer.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →
Consumer, search and ads
How we pick the ten
Same rule everywhere: Importance = share of AI output × years to replace; top 10 = 8 by Importance + 2 Rising.
- What counts here
- Consumer use of AI products, counted in monthly active users: assistants, AI answers in search, AI creation tools and specialised assistants. AI-ranked ads and feeds are not counted, because they have no comparable usage measure.
- Parts
- 54% General assistants: Chat assistants such as ChatGPT, Gemini and Meta AI; most consumer AI use happens here.
- 36% AI answers in search: AI summaries and answer modes in search engines, which reach billions of people each month.
- 6% AI creation tools: Image, video, design, music and voice generation; audiences far smaller than assistants'.
- 4% Companion, learning and shopping assistants: AI companions, tutors and shopping helpers; small next to general assistants.
- Shares from
- Monthly active users each company reports for its product or feature; panel estimates (QuestMobile, Sensor Tower) where none is reported, labelled
- Hard to replace
- Easy to replace: people can switch assistant, search engine or creation app in minutes, and many already use several. What is slow to copy is distribution, such as assistants built into phones, messaging apps and search pages that billions already use.
- Check
- No documented bottleneck case for this sublayer.
Parts and weights
AI products people use for themselves: general assistants (including ChatGPT, Gemini and Meta AI), AI answers in search, AI creation tools, and companion, learning and shopping assistants; business plans of general assistants count in Enterprise agents, model sales through APIs count in Models, and AI ranking of feeds and ads is not counted, because it has no usage measure comparable with the others.
| Part | Weight | Range | Source of the weight |
|---|---|---|---|
| General assistants | 54% | 45–68% | OpenAI · 7 October 2026 |
| AI answers in search | 36% | 22–42% | Google (The Keyword), Sundar Pichai · 19 May 2026 |
| AI creation tools | 6% | 3–11% | Bloomberg News (Natalie Lung), via Bloomberg Law · 1 April 2026 |
| Companion, learning and shopping assistants | 4% | 2–7% | Menlo Ventures · 16 September 2026 |
Every company scored
Importance is shown relative to the sublayer's leader, which scores 100. Share of AI output is the part of the sublayer's AI-related output that passes through the company. The Rising score, from 0 to 100, averages where the company stands among eligible companies on growth, change in share and committed expansion.
| Rank | Company | Share of AI output | Importance | Rising score |
|---|---|---|---|---|
| 1 | AlphabetTop 10 | 35% | 100 | 30 |
| 2 | OpenAITop 10 | 13.9% | 15 | 20 |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 3 | ByteDanceTop 10 | 7.4% | 8.3 | 60 |
| 4 | BaiduTop 10 | 3.1% | 4.2 | — |
| 5 | AmazonTop 10 | 3.3% | 3.9 | 10 |
| 6 | AlibabaTop 10 | 3.1% | 2.7 | 85 |
| 7 | AnthropicTop 10 | 3% | 2.4 | 95 |
Replaceability rated by two independent AI passes of a model family made by a company the index also scores; the ratings were applied as computed, without editorial change. | ||||
| 8 | DeepSeekTop 10 | 2.6% | 2.0 | — |
| 9 | OPPOTop 10 | 2% | 1.8 | — |
| 10 | MicrosoftTop 10 | 2% | 1.7 | — |
| 11 | Huawei | 1.9% | 1.6 | — |
| 12 | Apple | 1.9% | 1.5 | — |
| 13 | Meta | 1.7% | 1.3 | — |
| 14 | Samsung Electronics | 1.5% | 1.1 | — |
| 15 | Xiaomi | 1.2% | 1.0 | — |
| 16 | Snap | 1.1% | 0.8 | — |
| 17 | Yandex | 0.8% | 0.8 | — |
| 18 | Meituan | 0.6% | 0.7 | — |
| 19 | vivo | 0.8% | 0.7 | — |
| 20 | Honor | 0.8% | 0.7 | — |
| 21 | Tencent | 0.7% | 0.6 | — |
| 22 | SpaceX | 0.7% | 0.6 | — |
| 23 | Canva | 0.6% | 0.5 | — |
| 24 | Perplexity | 0.7% | 0.5 | — |
| 25 | Meitu | 0.6% | 0.5 | — |
| 26 | Adobe | 0.4% | 0.3 | — |
| 27 | JD.com | 0.2% | 0.2 | — |
| 28 | NAVER | 0.3% | 0.2 | — |
| 29 | Qihoo 360 (360 Security Technology) | 0.2% | 0.2 | — |
| 30 | Ant Group (AQ, Lingguang) | 0.2% | 0.2 | — |
| 31 | Picsart | 0.2% | 0.2 | — |
| 32 | Zuoyebang (Question.AI) | 0.2% | 0.1 | — |
| 33 | Genspark | 0.2% | 0.1 | — |
| 34 | Walmart (Sparky) | 0.2% | 0.1 | — |
| 35 | MiniMax | 0.2% | 0.1 | — |
| 36 | 0.1% | 0.1 | — | |
| 37 | Kuaishou | 0.1% | 0.1 | — |
| 38 | Superhuman | 0.1% | 0.1 | — |
| 39 | Moonshot AI | 0.1% | 0.1 | — |
| 40 | iFLYTEK | <0.1% | <0.1 | — |
Top 10 one of the eight highest Importance scores, or a Rising place no eligible company qualified for. Rising fastest documented growth in AI-related activity: one of the two highest Rising scores outside the eight, at least 50.
Test against a known bottleneck
This sublayer has no documented bottleneck to test the ranking against.
Agreement between the two rating passes
Two independent AI rating passes scored how hard each company's parts are to replace, on five levels. They agreed within half a level on 94% of the company parts both rated.
Weighted kappa, a standard measure of agreement beyond chance on which 1 means full agreement and 0 no more than chance, was 0.66.
No rating was more than a level apart between the two passes.
Limits of the evidence
- No public source splits consumer AI use by kind of product; the weights add the monthly users companies report for their main products, with estimates where none is reported. Many people use several products, so the total counts product users, not people.
- Monthly users of AI built into search and messaging (AI Overviews, Meta AI) include anyone who saw or used it once in a month, so they are not equivalent to active use of a standalone assistant; the weights count both the same.
- OpenAI reports weekly rather than monthly users, and Meta's latest Meta AI figure (more than 1 billion monthly actives) dates from mid-2025; both monthly figures are estimates.
- Few companies report users of AI features inside creative suites such as Canva, CapCut, Adobe and Meitu, so the creation weight rests on estimates.
- Chinese app figures come from QuestMobile panel data, which is not deduplicated across apps.
- Creation apps are counted here by users and in Models · Specialist models by revenue; the two measures are scored separately.
Changes
- 22 October 2026First list under the 5LC Importance Index. Importance places: Alphabet, OpenAI, ByteDance, Baidu, Amazon, Alibaba, Anthropic, DeepSeek. Rising places filled by the next Importance ranks: OPPO, Microsoft.
Ledger: every input, source and rating, with the rank stability analysis (JSON) →Previous method (21 October 2026) →