The companies, 2026

Scored by the panel on humans reached and purchasing power (1-19)

1. 1
   Nvidia USDoerShaper

   designs the accelerators that train and serve 80 to 85 percent of the world’s AI

   Humans reached: 4.5 billion, panel estimate, almost all of them other companies’ users counted once more; unique reach beyond the rest of the table is 500 to 800 million. Purchasing power touched: 40 trillion USD a year. AI revenue: 150 to 300 billion USD run-rate. Valuation: above 3 trillion USD on the panel’s figure, above 5 trillion on the market’s in 2026, the most valuable company in the world.

   Key work Grace Blackwell, Vera Rubin and the CUDA ecosystem; agreed to acquire Hugging Face on September 3rd 2026 (source). Company page.
2. 2
   Google / Alphabet USDoerShaper

   Search, YouTube, Android, Gemini and Google DeepMind

   Humans reached: 4.0 billion, panel estimate; AI Overviews alone reach more than 2 billion people a month, the largest direct AI surface on Earth. Purchasing power touched: 37 trillion USD a year; reaches about 80 percent of the world’s top income decile. AI revenue: 12 to 100 billion USD, unverifiable by AI share. Money mediated: about 260 billion USD of ad auctions ranked by its models.

   Key work Gemini 3, AI Overviews, AlphaGenome Atlas (source). Company page.
3. 3
   Meta USDoerShaper

   Facebook, Instagram, WhatsApp, Meta AI, Llama and Muse

   Humans reached: 3.2 billion, panel estimate, 3.2 to 3.4 billion daily people under its ranking models; Meta AI about 1 billion monthly users. Purchasing power touched: 26 trillion USD a year. Money mediated: about 160 billion USD of ad auctions. Third-party Llama deployments are the least certain figure on the whole matrix, 100 to 900 million people.

   Key work Muse Spark, Meta Superintelligence Labs, Llama open weights (source). Company page.
4. 4
   Microsoft USDoerShaper

   Windows, Office, Azure AI and Copilot; host of OpenAI’s models

   Humans reached: 2.2 billion, panel estimate. Purchasing power touched: 26 trillion USD a year; 72 percent of the top income decile through the workplace. Valuation: 3.3 trillion USD. About 40 percent of Microsoft’s AI reach is OpenAI intellectual property, and 40 to 70 percent of OpenAI’s reach arrives through Microsoft and Apple surfaces.

   Key work Microsoft 365 Copilot, Azure OpenAI, the reworked OpenAI partnership (2025) (source). Company page.
5. 5
   Apple USDoerShaper

   2.35 billion active devices, about 1.45 billion people, all running on-device machine learning

   Humans reached: 1.5 billion, panel estimate, 7th on headcount, 4th on purchasing power. Purchasing power touched: 29 trillion USD a year; 52 to 76 percent of the world’s top income decile, about 90 percent in the United States. Valuation: 3.5 trillion USD. AI revenue: near zero on its own account; Apple’s power is deciding which model gets the Siri handoff.

   Key work Apple Intelligence, Private Cloud Compute, the ChatGPT handoff (source). Company page.
6. 6
   Amazon USDoerShaper

   retail ranking and pricing, Alexa, and AWS Bedrock, which hosts Anthropic, Mistral and Cohere

   Humans reached: 1.55 billion, panel estimate. Purchasing power touched: 21 trillion USD a year. Money mediated: about 370 billion USD of commerce recommended by its models plus cloud AI spend. Bedrock end users, 500 million to 1.8 billion, are the least grounded figure on the matrix.

   Key work AWS Bedrock, Trainium and Project Rainier, Alexa+ (source). Company page.
7. 7
   ByteDance ChinaDoer

   TikTok, Douyin, CapCut, Toutiao, Doubao and the Seed models

   Humans reached: 2.45 billion, panel estimate, 4th on headcount and the largest Chinese footprint by far, 1.3 to 1.6 billion of it outside China. Purchasing power touched: 18 trillion USD a year; its users are young and spend about a seventh of an Apple user each, which is why it sits 7th here. Money mediated: about 210 billion USD of ad and commerce flow. Runs an estimated 60 percent of its inference on Nvidia silicon.

   Key work Doubao, Seed models, Volcano Engine (source). Company page.
8. 8
   TSMC TaiwanDoerShaper

   fabricates nearly every leading AI accelerator

   Humans reached: none of its own; every person Nvidia, Apple, Google and Amazon reach through custom silicon is TSMC’s by construction. Purchasing power touched: 17 trillion USD a year, panel estimate, with the widest range on the table because four models gave a fab zero and seven gave it Nvidia’s footprint. AI revenue: 30 to 90 billion USD. Valuation: about 930 billion USD. No TSMC, no frontier AI.

   Key work 3 nanometre and 2 nanometre nodes, CoWoS packaging (source). Company page.
9. 9
   OpenAI USDoer

   ChatGPT and the GPT models; distributed through Microsoft Copilot and Apple’s Siri handoff

   Humans reached: 1.45 billion, panel estimate, 400 to 800 million weekly ChatGPT users plus 600 to 900 million through partners; 40 to 70 percent of the total is borrowed distribution. Purchasing power touched: 18 trillion USD a year. AI revenue: 12 to 30 billion USD run-rate. Valuation: 80 to 600 billion USD on the panel’s range, unverifiable.

   Key work GPT-6 Astra, shipped September 3rd 2026 (source). Company page.
10. 10
    Tencent ChinaDoerShaper

    WeChat, Yuanbao, Hunyuan, games and Tencent Cloud

    Humans reached: 1.37 billion, panel estimate; WeChat’s 1.38 billion monthly users are the ceiling and 50 to 80 percent of them are exposed to Hunyuan ranking and search. Purchasing power touched: 9 trillion USD a year. Money mediated: about 110 billion USD. Served DeepSeek R1 inside Yuanbao for roughly half its 2025 answers.

    Key work Hunyuan HY4, WeChat AI search (source). Company page.
11. 11
    Alibaba ChinaDoerShaper

    Taobao, Tmall, Alibaba Cloud and Qwen, the most downloaded open-weight family after Llama

    Humans reached: 1.1 billion, panel estimate. Purchasing power touched: 8 trillion USD a year. Money mediated: about 700 billion USD of commerce recommended by its models, the largest figure on the table, which is why Alibaba holds 11th with a 15 percent top-decile share.

    Key work Qwen 3.8 Max, a 2.4-trillion-parameter multimodal model, July 2026 (source). Company page.
12. 12
    ASML NetherlandsDoerShaper

    the only maker of extreme-ultraviolet lithography machines

    Humans reached: none of its own; 100 percent of TSMC’s reach is ASML’s by construction. Purchasing power touched: 15 trillion USD a year, panel estimate, same split as TSMC. AI revenue: 6 to 30 billion USD attributable. Valuation: about 350 billion USD. On a substitution-cost matrix ASML ranks first: if it stops, chip advancement stops.

    Key work High-NA EUV (source). Company page.
13. 13
    Huawei ChinaDoerShaper

    Ascend accelerators, HarmonyOS, Pangu models and Huawei Cloud

    Scored for the first time in the money round; no humans-reached figure from the earlier rounds. Purchasing power touched: 5 trillion USD a year, panel estimate. AI revenue: 5 to 22 billion USD. Ascend is the partial escape from Nvidia that decides how much of China’s reach is sovereign: the panel puts ByteDance’s inference at 25 to 55 percent on domestic accelerators.

    Key work Ascend 910C, CloudMatrix (source). Company page.
14. 14
    Anthropic USDoer

    Claude, Claude Code and the enterprise pipelines on AWS Bedrock and Google Vertex

    Humans reached: 385 million, panel estimate, 50 million of them direct; 50 to 85 percent of the total is cloud-mediated. Purchasing power touched: 6.5 trillion USD a year; 16 percent of the top decile, skewed to enterprise. AI revenue: 2 to 15 billion USD run-rate. No Anthropic model sat on the panel, because a Claude model drafted this.

    Key work Claude Fable 5.1 and Claude Mythos 5.1, September 1st 2026 (source). Company page.
15. 15
    SpaceX, Tesla and xAI USDoer

    Grok and the X feed, Tesla FSD and Optimus, Starlink

    Humans reached: 285 million, panel estimate, most of it X’s ranking; FSD is a few million cars, robotics negligible today. Purchasing power touched: 6 trillion USD a year; 8 to 40 percent of the top decile. Valuation: 1.24 trillion USD after the February 2026 SpaceX and xAI merger, the fourth-highest on this list, which the panel read as pricing Starlink as a pipe and Optimus as a workforce in 2030 rather than AI reach in 2026.

    Key work Grok 4.5, the Colossus buildout, the SpaceX IPO (June 2026) (source). Company page.
16. 16
    Xiaomi ChinaDoer

    HyperOS on about 660 million monthly devices, 70 percent outside China, plus the SU7 electric car

    Humans reached: 690 million, panel estimate; every model stripped IoT device counts, Xiaomi’s own included. Purchasing power touched: 3.5 trillion USD a year.

    Key work HyperOS 3, MiMo V2.5 (source). Company page.
17. 17
    DeepSeek ChinaDoer

    open-weight models that reset the cost floor of frontier AI

    Humans reached: 435 million, panel estimate, a small app and a large dependency layer through distills served on Azure, Bedrock, Perplexity and Chinese clouds. Purchasing power touched: 4 trillion USD a year. Revenue: near zero. Valuation: about 10 billion USD on the panel’s figure, unverifiable. Per unit of headcount it is the highest-leverage entity on the list; that leverage is a different dimension from reach.

    Key work DeepSeek V4 and V4.1 (source). Company page.
18. 18
    Zhipu AI / Z.ai ChinaDoer

    GLM models inside Chinese government, telecom and state-enterprise pipelines

    Humans reached: 200 million, panel estimate, low confidence on every model. Purchasing power touched: 1.5 trillion USD a year.

    Key work GLM 5.3 (source). Company page.
19. 19
    Moonshot AI ChinaDoer

    Kimi, the long-context assistant

    Humans reached: 110 million, panel estimate. Purchasing power touched: 0.8 trillion USD a year. Valuation: 35 billion USD (July 2026).

    Key work Kimi K3, a 2.8-trillion-parameter open-weight model with a 1-million-token context (source). Company page.

Chips, memory and fabrication20-33

1. 20
   AMD USDoer

   designs GPUs and CPUs that compete directly with Nvidia for AI training and inference workloads

   In 2026 AMD deepened its position as a core AI infrastructure supplier, agreeing on 22 July 2026 to invest up to 5 billion USD in Anthropic as part of a computing power deal, alongside its earlier commitment to deploy up to 6 gigawatts of Instinct GPUs for OpenAI.

   Key work up to 5 billion USD Anthropic investment and compute deal, announced 22 July 2026 (source). Company page.
2. 21
   Broadcom USDoerShaper

   designs custom AI accelerators (ASICs) and networking silicon that let hyperscalers and AI labs build compute capacity outside Nvidia’s platform. On 6 April 2026 Broadcom agreed expanded custom chip deals with Google and Anthropic, with analysts estimating Anthropic’s committed custom chip orders at around 21 billion USD for 2026, while Broadcom was separately reported to be raising up to 100 billion USD in debt to finance AI chip production for Anthropic and OpenAI.

   Key work expanded Google and Anthropic custom chip agreements, announced 6 April 2026 (source). Company page.
3. 22
   SK hynix South KoreaDoerShaper

   supplies the majority of the world’s high bandwidth memory (HBM), the component that feeds Nvidia’s AI accelerators

   SK hynix held a 56.4 percent share of the HBM market in the first quarter of 2026 and reported record first quarter profit on 23 April 2026 on AI memory demand, and a July 2026 partnership between Nvidia and parent SK Group included a long term AI memory supply agreement.

   Key work record Q1 2026 earnings on HBM demand, reported 23 April 2026 (source). Company page.
4. 23
   Samsung Electronics South KoreaDoer

   manufactures memory chips and semiconductors, including the HBM used in AI accelerators

   Samsung moved closer to Nvidia’s qualification for HBM4 chips, reported by Bloomberg on 26 January 2026, and unveiled its next generation HBM4E chip, delivering 16 gigabits per second per pin and 4.0 terabytes per second of bandwidth, at Nvidia’s GTC event on 16 March 2026.

   Key work HBM4E unveiled at Nvidia GTC 2026, 16 March 2026 (source). Company page.
5. 24
   Micron USDoer

   produces DRAM and NAND memory, including HBM4, for AI data centres

   Micron posted record fiscal third quarter revenue of 41.46 billion USD in results reported around 25 June 2026, more than four times the prior year, and disclosed 16 strategic customer agreements worth about 100 billion USD in minimum contracted revenue through 2030.

   Key work fiscal Q3 2026 results and 100 billion USD in strategic HBM supply agreements, reported 25 June 2026 (source). Company page.
6. 25
   Intel USDoer

   designs CPUs and operates chip foundries seeking a larger role in AI manufacturing

   Intel moved its 18A-P process into risk production, reported by CNBC on 16 June 2026, after Nvidia’s 5 billion USD equity investment in the company closed in December 2025, though Nvidia was later reported to have paused testing production on the 18A process.

   Key work 18A-P risk production start, reported 16 June 2026 (source). Company page (unverified).
7. 26
   Arm UKDoerShaper

   licenses the CPU architecture used in over 99 percent of smartphones and increasingly in AI data centre chips, setting a default other chipmakers build on. Arm reported record full year revenue of 4.92 billion USD, up 23 percent year on year, and on 16 September 2026 chief executive Rene Haas said he was increasingly confident in meeting a 2 billion USD customer demand target for its new AGI CPU, launched in March 2026 for AI inference.

   Key work AGI CPU demand confidence update, 16 September 2026 (source). Company page.
8. 27
   Qualcomm USDoerShaper

   chipmaker pushing from mobile SoCs into AI data centre inference

   Qualcomm unveiled its AI200 and AI250 rack-scale inference chips for 2026 and 2027 launch, built on its Hexagon NPU with a memory architecture pitched as a lower-cost alternative to Nvidia’s HBM approach. In September 2026 it confirmed a multi-year partnership with Amazon Web Services to co-develop custom AI inference silicon and optical networking gear for AWS data centres, and it is targeting more than 15 billion USD in annual data centre revenue by 2029.

   Key work AWS custom silicon and optical networking partnership, announced 2026-09 (source). Company page.
9. 28
   SMIC ChinaDoerShaper

   China’s largest contract chip foundry and sole domestic fabricator of advanced AI logic at scale

   SMIC entered pilot runs of a 5 nm-class process during 2026 to fabricate Huawei’s Ascend 950 AI chip family, with ByteDance, Alibaba Cloud and Tencent reported to have placed large procurement orders, while its fabs run above 93 percent utilisation despite yield rates on the 7 nm node reported between 20 and 40 percent. As the foundry Huawei and other domestic AI chip designers depend on, SMIC underpins China’s push to reduce reliance on foreign AI silicon.

   Key work pilot production of Huawei’s Ascend 950 family on SMIC’s 5 nm-class process, reported 2026 (source). Company page.
10. 29
    Cerebras USDoer

    maker of wafer-scale AI chips built for inference speed rather than GPU-style parallelism

    Cerebras priced its IPO at 185 USD per share on May 13th 2026, raising 5.55 billion USD in the largest US IPO of the year and reaching a market capitalisation near 100 billion USD, after earlier striking a 20 billion USD multi-year compute deal with OpenAI in early 2026. The company reported 510 million USD in 2025 revenue, up nearly 76 percent year on year, on the strength of its Wafer Scale Engine 3 chips.

    Key work IPO raising 5.55 billion USD, priced May 13th 2026 (source). Company page.
11. 30
    Groq USDoer

    designer of the LPU inference chip that pivoted into an AI inference cloud provider in 2026

    After Nvidia’s 20 billion USD licensing deal in late 2025 drew away much of Groq’s senior team including founder Jonathan Ross, the company raised 650 million USD in mid-2026 and then a further 350 million USD Series A round in August 2026 led by Disruptive with Nvidia participating, valuing it at 3.5 billion USD, roughly half its valuation a year earlier. Groq says nearly two million developers and teams have used its inference cloud service.

    Key work 350 million USD Series A at a 3.5 billion USD valuation, closed August 17th 2026 (source). Company page.
12. 31
    Cambricon ChinaDoer

    Beijing-based designer of AI training and inference chips and GPUs for China’s domestic AI buildout

    Cambricon’s first-half 2026 revenue surged 108 percent year on year to 6 billion yuan (about 890 million USD), with net profit up 122.6 percent, driven by government-led demand to replace foreign AI hardware. The company has set a target of more than 13.5 billion yuan in 2026 revenue and over 100 billion yuan (14.8 billion USD) cumulatively over the next three years, tied to an employee incentive plan.

    Key work first-half 2026 revenue of 6 billion yuan, reported 2026-08 (source). Company page.
13. 32
    MediaTek TaiwanDoerShaper

    fabless chip designer extending from smartphone SoCs into AI infrastructure and edge computing

    On August 31st 2026, Nvidia and MediaTek announced an expanded partnership in which Nvidia invested 3.5 billion USD in MediaTek convertible bonds, with MediaTek adopting Nvidia’s NVLink Fusion platform to let customers build custom AI accelerators integrated with Nvidia’s rack-scale systems, alongside continued work on RTX Spark, DGX Spark and automotive AI platforms.

    Key work 3.5 billion USD Nvidia investment and NVLink Fusion partnership, announced August 31st 2026 (source). Company page.
14. 33
    Tokyo Electron JapanDoerShaper

    semiconductor equipment maker whose etching, deposition and process-control tools are a chokepoint for advanced AI chip and HBM production. Tokyo Electron raised its full-year outlook in 2026 on AI-driven chip demand, with equipment for advanced chips projected to reach nearly 40 percent of total sales by fiscal 2026 and high-bandwidth memory demand driving strong orders for its etching tools. In June 2026 it won “AI Semiconductor Manufacturing Solution of the Year” at the AI Breakthrough Awards for its AI-based process control technology.

    Key work full-year outlook raised on AI-driven demand, reported July 31st 2026 (source). Company page.

Clouds, compute and data centres34-45

1. 34
   Oracle USDoerShaper

   hyperscale cloud provider that has become one of the primary landlords of AI training capacity

   Oracle Cloud Infrastructure now anchors some of the largest AI compute deals on record, including the Stargate data centre build-out with OpenAI, and its backlog of contracted cloud work has grown to a scale that few providers can match.

   Key work remaining performance obligations reached 638 billion USD at the end of fiscal Q4 2026, up 363 percent year on year, reported June 10th 2026 (source). Company page.
2. 35
   CoreWeave USDoer

   specialist AI cloud provider (a “neocloud”) built to run Nvidia GPU clusters at scale for AI labs and hyperscalers

   CoreWeave has become a primary supplier of dedicated GPU capacity to Meta and other large AI buyers, expanding contracted power well beyond its original scope through 2026.

   Key work CoreWeave and Meta expanded their AI infrastructure agreement to 21 billion USD, extending the partnership through December 2032 and including early Nvidia Vera Rubin platform deployments, announced April 9th 2026 (source). Company page.
3. 36
   Nebius NetherlandsDoer

   European-headquartered neocloud that rents Nvidia GPU infrastructure to major AI labs and hyperscalers, built from the remains of Yandex’s cloud business. Nebius has signed successive multi-billion dollar capacity agreements with Microsoft and Meta through 2026, positioning it as a significant non-US alternative to the large American cloud providers for AI compute.

   Key work Nebius signed an expanded AI infrastructure agreement with Meta worth up to approximately 27 billion USD, including 12 billion USD of dedicated capacity on the Nvidia Vera Rubin platform, announced March 16th 2026 (source). Company page.
4. 37
   Lambda USDoer

   GPU cloud provider offering on-demand and clustered access to Nvidia hardware for AI training and inference, positioning itself for an initial public offering. Lambda’s revenue has grown rapidly on AI demand and it has been raising large pre-IPO rounds to fund further data centre build-out, following the path set by CoreWeave’s 2025 listing.

   Key work Lambda was reported to be in talks to raise up to 3 billion USD in pre-IPO financing at a valuation of 12 billion USD or more, reported August 25th 2026 (source). Company page.
5. 38
   Equinix USDoerShaper

   the world’s largest neutral data centre and interconnection operator, whose facilities carry the network traffic connecting enterprises to every major cloud and AI platform. Equinix sets conditions for AI adoption beyond hyperscale training by giving enterprises low-latency, interconnected access to inference infrastructure, deepening its partnership with Nvidia through 2026.

   Key work Equinix launched the Equinix Inference Exchange with Nvidia and Together AI, giving enterprises low-latency access to more than 200 open-source models across its global footprint, announced September 2nd 2026 (source). Company page.
6. 39
   Digital Realty USDoer

   global data centre real estate investment trust that leases hyperscale and colocation capacity to cloud and AI companies. Digital Realty has been expanding its hyperscale portfolio through acquisitions and record leasing tied to AI workloads, including the largest hyperscale deal in its history in early 2026.

   Key work Digital Realty agreed to buy Blackstone’s interest in three fully leased Northern Virginia data centres totalling 288 megawatts for 3.5 billion USD, announced June 29th 2026 (source). Company page.
7. 40
   Vertiv USDoer

   maker of power and cooling systems that keep AI data centres running

   Its backlog reached 15 billion USD after orders climbed 252 percent year on year, and Q1 2026 net sales rose 30 percent to 2.65 billion USD, driven mainly by data centre demand in the Americas.

   Key work acquisition of Strategic Thermal Labs to strengthen cold-plate and server-side liquid cooling engineering for AI workloads, announced April 27th 2026 (source). Company page.
8. 41
   Schneider Electric FranceDoerShaper

   builds the power and cooling infrastructure that hyperscale AI data centres run on, working with Nvidia reference designs for GPU racks. It signed a 1.9 billion USD supply capacity agreement with Switch on November 19th 2025 to power AI factories, its largest cooling service engagement to date, and later launched 2.5 megawatt prefabricated power modules built around Nvidia’s GB300 NVL72 platform.

   Key work 2.5 megawatt AI-ready power modules for Nvidia-based high-density computing, announced September 17th 2026 (source). Company page.
9. 42
   Supermicro USDoer

   builds and ships Nvidia GPU-based AI servers at scale, competing on speed to market for each new chip generation

   Its GB300-powered Super AI Station reached general availability around GTC 2026 in March, but the company also faces a US Department of Justice indictment, unsealed March 19th 2026, alleging three individuals tied to the company illegally exported about 2.5 billion USD of Nvidia-equipped AI servers to China.

   Key work general availability of the Nvidia GB300-powered Super AI Station, announced at GTC 2026 (source). Company page.
10. 43
    Dell Technologies USDoer

    one of the largest sellers of Nvidia-based AI servers to enterprises and AI labs

    Dell marked the two-year anniversary of its Dell AI Factory partnership with Nvidia on March 16th 2026, reporting over 4,000 customers deploying the offering and early adopters seeing up to 2.6x return on investment within the first year. It was also reported to be nearing a deal worth more than 5 billion USD to supply Nvidia GB200-based servers to Elon Musk’s xAI.

    Key work Dell AI Factory with Nvidia two-year update, 4,000-plus customers, announced March 16th 2026 (source). Company page.
11. 44
    Hewlett Packard Enterprise USDoer

    supplies AI servers and, since its Juniper Networks acquisition, the networking gear tying together hyperscale AI data centres. Cloud and AI is now HPE’s largest segment, with Q3 FY26 revenue up 34 percent to 12.2 billion USD and a backlog near 6 billion USD.

    Key work expanded multi-year Oracle deal deploying HPE Juniper networking across Oracle’s gigawatt-scale AI data centres, announced September 2nd 2026 (source). Company page (unverified).
12. 45
    Crusoe USDoer

    builds and operates AI data centres and modular “AI factories,” including the 1.2 gigawatt Abilene, Texas campus that Oracle uses to host OpenAI workloads. It raised a 3.9 billion USD Series F on September 17th 2026 at a 30.9 billion USD post-money valuation, backed by Nvidia, Founders Fund, GIC and others, to fund further data centre and modular factory build-out.

    Key work 3.9 billion USD Series F funding round at 30.9 billion USD valuation, announced September 17th 2026 (source). Company page.

Enterprise platforms46-57

1. 46
   Salesforce USDoerShaper

   the dominant enterprise CRM platform now positioning agentic AI as its core product layer

   At Dreamforce 2026 (15-17 September 2026), Salesforce announced AIforce, the Koa CRM reasoning model, a Salesforce integration inside Claude in beta, and Gemini support in the Agentforce Reasoning Engine. Its own 2026 Agentic Enterprise Index found organisations nearly tripled the number of activated AI agents over the fiscal year while cutting average agent-creation time by 53 percent.

   Key work Dreamforce 2026 AIforce and Agentforce announcements, September 15th 2026 (source). Company page.
2. 47
   Palantir USDoerShaper

   a data-integration and AI-platform provider whose Foundry and AIP layers set the operating standard for government and large-enterprise deployments. Federal spending on Palantir crossed 1 billion USD in 2026, with the Pentagon adding 244 million USD in funding in August 2026, while US commercial revenue grew 149 percent year-on-year to 764 million USD in Q2 2026. Stellantis also renewed and expanded its five-year partnership to broaden use of Foundry and begin deploying AIP across select business functions.

   Key work Pentagon funding surpassing 1 billion USD and Stellantis AIP expansion, August 2026 (source). Company page.
3. 48
   Databricks USDoerShaper

   the lakehouse and AI-platform company that many enterprises build their data and model pipelines on

   Databricks raised 5 billion USD in a strategic funding round at a 190 billion USD valuation on August 13th 2026, led by Coatue with Blackstone, MGX and T. Rowe Price participating, following an earlier 2026 round that valued it at 134 billion USD on a 5.4 billion USD revenue run-rate growing over 65 percent year-on-year.

   Key work 5 billion USD raise at 190 billion USD valuation, August 13th 2026 (source). Company page.
4. 49
   Snowflake USDoerShaper

   repositioning from cloud data warehouse to an “AI Data Cloud” that governs the context enterprise AI agents draw on

   At Snowflake Summit 2026, which drew over 20,000 attendees, the company unveiled more than 26 new capabilities around AI agents and governed context, with Anthropic president Daniela Amodei co-keynoting; it had earlier launched Project SnowWork, an autonomous enterprise AI platform, in a research preview in March 2026.

   Key work Snowflake Summit 2026 agentic-enterprise announcements (source). Company page.
5. 50
   ServiceNow USDoerShaper

   the workflow platform now positioning itself as the governance layer for AI agents across the enterprise

   On April 9th 2026 it replaced its five legacy licensing tiers with three AI-native tiers that bundle Now Assist, the Moveworks layer and its AI Control Tower into every package, and at Knowledge 2026 it expanded its “Autonomous Workforce” with new AI specialists for IT, CRM, employee service and security.

   Key work AI-native licensing overhaul, April 9th 2026 (source). Company page (unverified).
6. 51
   SAP GermanyDoerShaper

   the enterprise resource-planning incumbent embedding AI across its core business applications for a vast installed customer base. At SAP Sapphire 2026 it unveiled the “Autonomous Enterprise” concept and a unified SAP Business AI Platform, deepening partnerships with Anthropic, AWS, Google Cloud, Microsoft, NVIDIA and Palantir; its Joule assistant is now live across 35 solutions, with Financial Closing, Billing, Tax and Compliance, and Accounts Receivable assistants reaching general availability planned for Q2 2026.

   Key work SAP Sapphire 2026 Autonomous Enterprise and Business AI Platform launch, May 2026 (source). Company page.
7. 52
   IBM USDoerShaper

   enterprise AI platform and standards-setter through watsonx and open Granite models

   In 2026 Gartner named IBM a Leader in the Magic Quadrant for AI Platforms for Data Science and Machine Learning for the second year running, and separately a Leader in the inaugural Magic Quadrant for AI Governance Platforms on June 26th 2026.

   Key work watsonx.governance now monitors AI systems across IBM and third-party stacks including OpenAI, AWS and Meta models, recognised June 26th 2026 (source). Company page.
8. 53
   Accenture Ireland/USDoer

   the largest AI systems integrator embedding generative and agentic AI into enterprise client operations at scale

   Third-quarter fiscal 2026 results, reported June 18th 2026 for the quarter ended May 31st 2026, showed revenues of 18.72 billion USD, up 6 percent in US dollars, with new bookings of 19.3 billion USD.

   Key work advanced AI bookings, last broken out separately in Q1 FY2026, had nearly doubled year over year to 2.2 billion USD before AI became embedded across nearly all of Accenture’s delivery (source). Company page.
9. 54
   Adobe USDoer

   generative creative AI supplier through Firefly, embedded across the Creative Cloud suite

   On April 15th 2026 Adobe announced the Firefly AI Assistant, a conversational creative agent that orchestrates multi-step workflows across Photoshop, Premiere, Lightroom, Illustrator and Firefly itself, alongside Firefly Image Model 4 with 2K generation.

   Key work Firefly’s monthly active user base passed 6 million by mid-2026, up 65 percent year on year, with the AI Assistant announced April 15th 2026 (source). Company page (unverified).
10. 55
    Workday USDoer

    enterprise HR and finance software supplier building and governing AI agents for corporate back-office workflows

    On June 2nd 2026 Workday launched Developer Agent, Agent-Ready Tools over the Model Context Protocol, and Agent Passport, a system that tests, verifies and continuously monitors any AI agent against standards including the OWASP LLM Top 10 and NIST AI RMF before and after production deployment.

    Key work Agent Passport, launched with Cisco as testing partner on June 2nd 2026, with general availability targeted before end of 2026 (source). Company page.
11. 56
    Atlassian AustraliaDoer

    enterprise collaboration software maker embedding AI agents and search across Jira, Confluence and its Teamwork Graph via Rovo. Fourth-quarter fiscal 2026 results, reported August 6th 2026, showed revenue of 1.77 billion USD, up 28 percent year on year, with Rovo monthly active users doubling in a single quarter to 1 million and adoption at over 80 percent of the Fortune 500.

    Key work Rovo-assisted actions grew 50 percent quarter over quarter, with customers using Rovo growing annual recurring revenue at more than twice the rate of non-adopters, reported August 6th 2026 (source). Company page.
12. 57
    Stripe USDoerShaper

    payments infrastructure provider moving into AI model routing as a chokepoint between businesses and the AI token economy. On August 19th 2026 Stripe confirmed it had agreed to acquire OpenRouter, which routes and optimises token usage across more than 400 models from over 80 providers for customers including NVIDIA, Zoom and Lovable, in a deal reported by Bloomberg at over 7 billion USD.

    Key work agreement to acquire OpenRouter, announced August 19th 2026, framed by Stripe as helping businesses manage both revenue and cost sides of AI-era profitability (source). Company page.

Frontier and open labs58-71

1. 58
   Mistral AI FranceDoerShaper

   builds and sells frontier open-weight AI models and enterprise AI infrastructure

   Samsung Electronics led a funding round valuing Mistral at about 24 billion USD, reported September 8th 2026, and the company’s annual recurring revenue reached 400 million USD in 2026. Its open-weight models and Le Chat assistant are widely adopted by European enterprises and public-sector bodies seeking an alternative to US and Chinese AI providers.

   Key work Samsung-led funding round valuing Mistral at roughly 24 billion USD, September 8th 2026 (source). Company page.
2. 59
   Thinking Machines Lab USDoer

   AI research lab founded by former OpenAI CTO Mira Murati in February 2025 to build more customisable AI systems

   On July 15th 2026 it released its first model, Inkling, an open-weight 975-billion-parameter mixture-of-experts model, followed by the smaller Inkling Small on July 31st 2026. The company raised a 2 billion USD seed round in July 2025 at a 12 billion USD valuation and also runs Tinker, a platform for fine-tuning open models.

   Key work launch of the open-weight Inkling model, July 15th 2026 (source). Company page.
3. 60
   Cohere CanadaDoer

   enterprise-focused AI lab selling private, on-premises large language model deployments to regulated industries

   On April 24th 2026 Cohere announced a merger with Germany’s Aleph Alpha, later formalised at a roughly 20 billion USD valuation, to create a transatlantic AI provider serving government and regulated sectors from Berlin and Toronto. Cohere’s annual recurring revenue reached about 240 million USD in February 2026, with roughly 85 percent of revenue from private deployments to customers including Oracle, Fujitsu and RBC.

   Key work merger agreement with Aleph Alpha, announced April 24th 2026 (source). Company page.
4. 61
   Safe Superintelligence US/IsraelDoer

   research lab founded in June 2024 by former OpenAI chief scientist Ilya Sutskever, focused solely on developing safe superintelligence with no commercial product. On July 27th 2026 Nvidia committed to invest 5 billion USD in the company and to scale its compute on Nvidia’s upcoming Vera Rubin platform, part of roughly 8 billion USD in total funding raised. The company remains pre-revenue with about 50 employees and has not shipped a model.

   Key work Nvidia’s 5 billion USD investment and compute partnership, July 27th 2026 (source). Company page.
5. 62
   Recursive Superintelligence USDoer

   a startup founded in 2026, based in San Francisco and London, building AI systems designed to recursively improve their own research process. Founded by former Google DeepMind and Salesforce researchers including Tim Rocktäschel and Richard Socher, the company raised 650 million USD at a 4.65 billion USD valuation and on July 28th 2026 signed a 410 million USD multi-year compute agreement with AWS to run its self-improving AI research system. The roughly 25-person team has not yet launched a public product.

   Key work 410 million USD multi-year AWS compute agreement, announced July 28th 2026 (source). Company page.
6. 63
   Sakana AI JapanDoerShaper

   Tokyo-based AI lab building Japan-focused, nature-inspired AI systems and, since June 2026, a multi-model orchestration layer. On June 22nd 2026 Sakana launched Fugu and Fugu Ultra, an orchestration model that routes requests across a swappable pool of frontier models behind a single API, positioned as a hedge against export-control disruptions to providers such as Anthropic. Sakana reached a 2.635 billion USD valuation after a 135 million USD Series B round and is expanding its enterprise business from finance into industrial and government sectors in 2026.

   Key work launch of the Fugu orchestration model, June 22nd 2026 (source). Company page.
7. 64
   AI21 Labs IsraelDoer

   Tel Aviv-based AI lab that has shifted from selling standalone language models to an enterprise AI-agent orchestration platform, Maestro. On May 18th 2026 AI21 cut about 110 staff, roughly 61 percent of its workforce, saying standalone language models were “not a sufficiently sustainable revenue stream,” while Maestro contracts worth tens of millions of dollars were signed with customers including Wix. Acquisition talks with Nvidia and, more recently, cloud provider Nebius have been reported but not confirmed as completed.

   Key work workforce cut and pivot to the Maestro platform, May 18th 2026 (source). Company page.
8. 65
   Reka US/UKDoer

   builds multimodal foundation models aimed at enterprise and, increasingly, physical-world AI applications

   Reka reached unicorn status in 2025 on 168 million USD raised across two rounds, and on June 11th 2026 it joined forces with video-model maker Moonvalley, bringing in researchers formerly of Google DeepMind, Meta, Amazon, Microsoft and Wayve to build models and infrastructure for robotics, wearables and autonomous systems.

   Key work the Reka-Moonvalley combination for “physical AI”, announced June 11th 2026 (source). Company page.
9. 66
   Baidu ChinaDoerShaper

   runs China’s dominant search engine and AI cloud platform and builds the Ernie family of foundation models that other Chinese developers build on top of. Baidu’s Ernie assistant reached 200 million monthly active users as Ernie 5.0 launched in January 2026, and in May 2026 the company rolled out Ernie 5.1, a mixture-of-experts model that placed first among Chinese models and fourth globally on the LMArena search leaderboard while using a fraction of prior training compute.

   Key work Ernie 5.0 launch and the 200 million monthly active user milestone, reported 2026 (source). Company page.
10. 67
    StepFun ChinaDoer

    one of China’s “AI Six Tigers”, builds multimodal foundation models and AI infrastructure

    StepFun raised more than 5 billion yuan (717 million USD) in a Series B+ round announced January 26th 2026, one of the largest private AI financings in China, and has since pursued a Hong Kong IPO.

    Key work the 717 million USD Series B+ round, announced January 26th 2026 (source). Company page.
11. 68
    MiniMax ChinaDoer

    builds multimodal AI models and consumer AI products including the Hailuo video generator and Talkie companion app, serving over 200 million users. MiniMax listed in Hong Kong on January 9th 2026, raising 619 million USD in an IPO that was about 1,836 times oversubscribed and saw shares close up 109 percent on debut, valuing the firm at roughly 13 billion USD.

    Key work Hong Kong IPO debut, January 9th 2026 (source). Company page.
12. 69
    SenseTime China/Hong KongDoer

    a computer-vision and generative-AI provider, sells AI software and cloud infrastructure to enterprise and government customers. For the first half of 2026 SenseTime reported revenue of CNY 2.91 billion, up 23.4 percent year on year, with generative AI accounting for nearly 80 percent of revenue and a net profit of CNY 620 million under IFRS, its first such profit since listing, announced August 26th 2026.

    Key work first-half 2026 results and first IFRS profit, announced August 26th 2026 (source). Company page.
13. 70
    iFlytek ChinaDoer

    builds intelligent speech recognition, translation and AI hardware, with a trailing twelve-month revenue of 3.9 billion USD as of March 31st 2026. At IFA Berlin 2026 (September 4th 2026 to September 8th 2026) iFlytek showcased AINOTE 2, a Guinness World Record-holding 4.2mm e-ink tablet that transcribes meetings in 18 languages and translates across 14, alongside AI Glasses for hands-free translation and note-taking.

    Key work AINOTE 2 and AI Glasses showcase at IFA 2026, announced September 5th 2026 (source). Company page.
14. 71
    Nous Research USDoer

    develops open-source, human-centric language models and the Hermes AI agent, which has around 214,000 GitHub stars

    On July 13th 2026 TechCrunch reported the company was finalising a round of at least 75 million USD led by Robot Ventures at a 1.5 billion USD valuation, following a 50 million USD Series A and prior funding from Paradigm and OSS Capital.

    Key work Hermes agent funding round at 1.5 billion USD valuation, reported July 13th 2026 (source). Company page.

Applications and agents72-87

1. 72
   Hugging Face US/FranceDoerShaper

   the leading open source hub for AI models and datasets

   On September 3rd 2026, Nvidia confirmed a definitive agreement to acquire Hugging Face for 12.93 billion USD, folding the platform behind 3 million models, 500,000 datasets and 18 million developers into the world’s most valuable chipmaker. Nvidia said Hugging Face would keep operating as an open platform for the wider AI ecosystem.

   Key work Nvidia’s 12.93 billion USD acquisition agreement, confirmed September 3rd 2026 (source). Company page.
2. 73
   Anysphere USDoer

   maker of the Cursor AI coding assistant

   On August 14th 2026, SpaceX completed an all-stock acquisition of Anysphere valued at 60 billion USD, the largest startup acquisition on record, months after Cursor had crossed 1 billion USD in annualised revenue. The deal folds one of the most widely used AI coding tools into a much larger industrial group.

   Key work SpaceX’s 60 billion USD all-stock acquisition of Cursor, completed August 14th 2026 (source). Company page.
3. 74
   Cognition USDoer

   maker of the Devin AI software engineer and the Windsurf IDE

   Bloomberg reported on September 2nd 2026 that Cognition was raising about 1 billion USD at a 47 billion USD valuation, nearly double the 26 billion USD mark set three months earlier, as annualised revenue from Devin and Windsurf passed 900 million USD. The repricing shows how fast autonomous coding agents moved from demo to enterprise spend during 2026.

   Key work new funding round valuing Cognition at 47 billion USD, reported September 2nd 2026 (source). Company page.
4. 75
   Perplexity USDoer

   AI answer engine and maker of the Comet browser

   By late August 2026, Perplexity’s annualised revenue had tripled to more than 750 million USD, and Nvidia entered talks to invest in a round that would value the company above 30 billion USD, over 50 percent above its prior valuation. The talks, reported August 24th 2026, reflect Perplexity’s push from search into agentic browsing ahead of a targeted 2028 listing.

   Key work Nvidia investment talks valuing Perplexity above 30 billion USD, reported August 24th 2026 (source). Company page.
5. 76
   Scale AI USDoerShaper

   data labelling and evaluation infrastructure for frontier AI labs and governments

   On May 6th 2026, Bloomberg reported the Pentagon awarded Scale AI a 500 million USD defence contract, five times its prior award, months after Meta took a 49 percent stake valuing the company at about 29 billion USD. The deal underlines Scale’s position as a chokepoint in the training and evaluation data pipeline, even as several AI lab customers reduced their reliance on it after the Meta investment.

   Key work 500 million USD Pentagon data contract, reported May 6th 2026 (source). Company page.
6. 77
   Midjourney USDoer

   independent, bootstrapped AI image and video generator

   Midjourney shipped version 8.2 as its default model on July 24th 2026, adding a new edit model with image-to-image editing, inpainting and outpainting, while continuing to operate without external venture funding. The company remains a rare case of sustained independence in generative AI, run without outside investors even as rivals raise large rounds.

   Key work version 8.2 model release, July 24th 2026 (source). Company page.
7. 78
   Runway USDoerShaper

   builds video-generation and world models for film, brands and robotics

   Runway raised 315 million USD in a Series E round led by General Atlantic at a 5.3 billion USD valuation on February 10th 2026, up from 3.3 billion USD ten months earlier, following the release of its Gen 4.5 video model with native audio and multi-shot generation.

   Key work 315 million USD Series E at a 5.3 billion USD valuation to pre-train next-generation world models, dated February 10th 2026 (source). Company page.
8. 79
   ElevenLabs US/UKDoerShaper

   provides voice AI infrastructure for text-to-speech, dubbing and voice agents used across media and software products

   The company raised 500 million USD in a Series D round led by Sequoia Capital at an 11 billion USD valuation on February 4th 2026, more than tripling its valuation from a year earlier, and is reported to be eyeing an IPO.

   Key work 500 million USD Series D at an 11 billion USD valuation, dated February 4th 2026 (source). Company page.
9. 80
   Synthesia UKDoer

   builds an AI video avatar platform used by companies for training and internal communications

   Synthesia raised 200 million USD in a Series E round led by Google Ventures at a 4 billion USD valuation on January 26th 2026, nearly double the 2.1 billion USD valuation it held a year earlier, with annual recurring revenue at 150 million USD and expected to pass 200 million USD during 2026.

   Key work 200 million USD Series E at a 4 billion USD valuation, dated January 26th 2026 (source). Company page.
10. 81
    Harvey USDoer

    sells AI tools for contract analysis, due diligence and litigation support to law firms and professional services teams

    Harvey raised 550 million USD co-led by Lightspeed Venture Partners and Diffusion at a 15.6 billion USD valuation on September 9th 2026, up from an 11 billion USD valuation in March 2026, with the new capital earmarked for building its own AI models.

    Key work 550 million USD round at a 15.6 billion USD valuation to build proprietary legal AI models, dated September 9th 2026 (source). Company page.
11. 82
    Glean USDoerShaper

    operates a permissions-aware enterprise search and AI agent platform that other internal tools plug into

    Glean surpassed 300 million USD in annual recurring revenue on May 28th 2026, just 15 months after reaching 100 million USD, while nearly doubling its Fortune 500 customer count year on year.

    Key work surpassed 300 million USD ARR and introduced its Agent Development Lifecycle framework, dated May 28th 2026 (source). Company page.
12. 83
    Duolingo USDoer

    runs a consumer app that uses AI to teach languages, maths and other subjects

    Duolingo reported second-quarter 2026 revenue of 298.5 million USD on August 5th 2026, with daily active users up 23 percent year on year to 58.7 million, after cutting AI compute cost per video call from 0.30 USD to under 0.01 USD by moving to open-source models.

    Key work Q2 2026 results showing DAUs up 23 percent to 58.7 million, dated August 5th 2026 (source). Company page.
13. 84
    Character.AI USDoer

    consumer AI companion and roleplay chatbot platform

    Character.AI remains the largest AI companion platform, with over 20 million monthly users; on January 7th 2026 Character.AI and Google agreed to settle five lawsuits brought by families in Florida, Colorado, New York and Texas over harm to teenagers, and the company has since enforced age verification and removed open-ended chat for under-18 users.

    Key work settlement of five family lawsuits over teen harm, dated January 7th 2026 (source). Company page.
14. 85
    Notion USDoerShaper

    workspace and AI-agent platform for knowledge work

    Notion reached an 11 billion USD valuation via an employee tender offer in January 2026, with more than half of its roughly 500 million USD ARR now coming from AI-enabled customers. It has pushed to become an agent platform rather than just a docs tool, opening a developer platform for building AI agents and, on September 15th 2026, launching Notion 3.7 with shareable “Agent skills” usable across Notion, Claude Code, Cursor, Gemini and Grok.

    Key work Notion 3.7 Agent skills launch, dated September 15th 2026 (source). Company page.
15. 86
    Canva AustraliaDoerShaper

    visual design and content-creation platform

    Canva now has 260 million monthly active users and has passed 3.5 billion USD in annualised revenue, growing about 35 percent year on year while staying profitable, as it prepares for a possible second-half-2026 listing. It has been building out a professional creative suite to rival Adobe, acquiring motion-design tool Cavalry and video-ad tool MangoAI on February 23rd 2026.

    Key work acquisition of Cavalry and MangoAI to expand the creative suite, dated February 23rd 2026 (source). Company page.
16. 87
    Manus, Butterfly Effect Singapore/ChinaDoer

    general-purpose AI agent product out of China

    Manus, built by Singapore-based, China-founded Butterfly Effect, became a flashpoint in AI geopolitics after Meta’s planned acquisition of the company, announced in December 2025, was blocked by Beijing, with the two sides completing a full operational split around mid-2026. As of September 17th 2026, Manus was reported to be raising 500 million USD at a 4 billion USD valuation, which would make it China’s most valuable agentic AI startup.

    Key work reported 500 million USD raise at a 4 billion USD valuation, dated September 17th 2026 (source). Company page.

Robotics and embodiment88-93

1. 88
   Figure AI USDoer

   humanoid robotics for industrial and logistics work

   Figure closed more than 1 billion USD in Series C funding at a 39 billion USD post-money valuation in autumn 2025, and has been proving the money out with real deployments: a BMW Group press release dated February 27th 2026 confirmed its Figure 02 robot had supported production of more than 30,000 X3 vehicles at the Spartanburg plant, moving over 90,000 components across roughly 1,250 operating hours.

   Key work BMW Spartanburg production deployment confirmed in a BMW Group release, dated February 27th 2026 (source). Company page.
2. 89
   Physical Intelligence USDoer

   general-purpose robot foundation models

   Physical Intelligence, founded by former Google DeepMind and academic researchers, builds AI models meant to control any robot for any task. On March 27th 2026 it was reported to be in talks to raise about 1 billion USD at a valuation above 11 billion USD, roughly doubling its 5.6 billion USD valuation from four months earlier, with Founders Fund and Lightspeed Venture Partners joining returning backers Thrive Capital and Lux Capital.

   Key work reported 1 billion USD round at an 11 billion USD-plus valuation, dated March 27th 2026 (source). Company page.
3. 90
   Skild AI USDoer

   “omni-bodied” robot foundation models

   Skild AI closed a 1.4 billion USD Series C led by SoftBank on January 14th 2026, valuing the Pittsburgh company above 14 billion USD, more than triple its valuation seven months earlier, with Nvidia, Samsung, LG and Bezos Expeditions among the backers. On September 14th 2026 it unveiled S1, a robot foundation model that learns new manipulation tasks from a single video demonstration without task-specific fine-tuning.

   Key work launch of the S1 foundation model, dated September 14th 2026 (source). Company page.
4. 91
   Unitree ChinaDoer

   humanoid and quadruped robot manufacturer

   Unitree shipped roughly 18,000 cumulative bipedal humanoids by mid-2026 and targeted 20,000 units for the year, making it the leading humanoid producer by volume, and it listed on Shanghai’s STAR Market on August 19th 2026, closing up 487 percent on debut after the fastest CSRC approval in the exchange’s history.

   Key work autonomous humanoid-versus-human sparring demo powered by the UnifoLM-X2-1.0 world-action model, September 7th 2026 (source). Company page.
5. 92
   Waymo USDoerShaper

   Alphabet’s autonomous ride-hailing operator

   Waymo passed 500,000 paid rides a week in 2026 and is targeting 1 million weekly rides by year end, backed by a 16 billion USD round at a 126 billion USD valuation, the largest ever in an autonomous vehicle company, and it is laying groundwork for over 20 more cities in 2026 including Tokyo and London.

   Key work robotaxi launches in Denver, San Diego and Tampa extending commercial service to 14 US cities, September 1st 2026 (source). Company page.
6. 93
   Waabi CanadaDoer

   autonomous trucking software developer

   Waabi closed a 750 million USD Series C on January 28th 2026, with milestone-based capital from Uber to deploy 25,000 or more Waabi Driver-powered robotaxis on Uber’s platform, the largest fundraise in Canadian history.

   Key work transfer of its Waabi Driver system from a Peterbilt 579 to Volvo’s VNL Autonomous without new engineering or retraining, July 12th 2026 (source). Company page.

Regions and sovereign programmes94-100

1. 94
   G42 UAEDoerShaper

   Abu Dhabi’s national AI and cloud infrastructure group

   G42 anchors Stargate UAE, the planned gigawatt-scale Nvidia cluster built with OpenAI and Oracle, and on July 6th 2026 it deepened its partnership with Microsoft to keep sensitive AI workloads and data inside the UAE. Its Core42 subsidiary closed a 4.9 billion USD AI infrastructure fund above target in mid-2026 and signed a sovereign AI cloud agreement with UAE telecom operator e&.

   Key work Microsoft partnership to localise sensitive AI data in the UAE, July 6th 2026 (source). Company page.
2. 95
   HUMAIN Saudi ArabiaDoerShaper

   Public Investment Fund subsidiary building Saudi Arabia’s sovereign AI stack

   Saudi Arabia’s Cabinet declared 2026 the “Year of Artificial Intelligence” on March 10th 2026, with HUMAIN positioned as the execution engine for gigawatt-scale data centres, sovereign frontier models and enterprise AI deployment, backed by over 100 billion USD in planned investment.

   Key work national AI platform built on an open model from China’s MiniMax, announced September 3rd 2026 (source). Company page.
3. 96
   Sarvam AI IndiaDoer

   Indian sovereign multilingual AI model developer

   Sarvam raised 234 million USD on June 15th 2026 at a 1.5 billion USD valuation, led by HCLTech with Bessemer, Khosla Ventures and Peak XV, becoming India’s newest AI unicorn, with the Indian government weighing a minority stake through the IndiaAI Mission.

   Key work 234 million USD Series B round making Sarvam a unicorn, June 15th 2026 (source). Company page.
4. 97
   Naver South KoreaDoerShaper

   South Korea’s dominant search and cloud platform, now building sovereign AI

   Naver Cloud and NC AI were eliminated from the first round of Korea’s national sovereign-AI model competition on January 15th 2026, after a panel found the technical originality of its HyperCLOVA X models insufficient, and Naver has since pushed sovereign AI through other channels.

   Key work sovereign defence AI partnership with Korea Aerospace Industries for drones and next-generation combat systems, July 7th 2026 (source). Company page.
5. 98
   Preferred Networks JapanDoer

   Japan’s leading independent AI chip and deep-learning developer, now scaling to compete with global GPU makers

   It is Japan’s most valuable AI startup, developing the MN-Core L1000 processor for generative AI inference and moving fabrication from TSMC to Samsung’s 2-nanometre process. On September 8th 2026 it confirmed it is courting foreign investors for the first time and eyeing an IPO in three to five years to fund mass production of its custom AI chips.

   Key work MN-Core L1000 generative-AI inference chip and IPO/foreign-funding push, reported September 8th 2026 (source). Company page.
6. 99
   SoftBank Group JapanDoerShaper

   the conglomerate financing and structuring much of the world’s AI infrastructure build-out through Arm, OpenAI and Stargate. Its roughly 90 percent stake in Arm underpins chip designs used across AI data centres, and it holds a 60 billion USD pledge to OpenAI alongside Stargate data-centre commitments running into the hundreds of billions. On September 18th 2026 it raised its Arm-backed margin loan by 5 billion USD to 25 billion USD to keep funding these AI bets, the loan’s third expansion since 2023.

   Key work 25 billion USD Arm margin loan expansion to fund AI investment, dated September 18th 2026 (source). Company page.
7. 100
   InstaDeep Tunisia/UK/Germany, part of BioNTechDoer

   a decision-making AI company built into BioNTech’s drug-discovery and life-sciences pipeline since its 2023 acquisition

   It runs the Kyber supercomputer for BioNTech’s compute needs and continues to sign applied-AI partnerships outside pharma, extending its reach into industrial and infrastructure sectors. On March 4th 2026 it announced a long-term partnership with water-technology group Wilo to build a scalable AI platform for smart water infrastructure.

   Key work strategic AI partnership with Wilo for smart water infrastructure, announced March 4th 2026 (source). Company page.

Two numbers per company, both medians from a twelve-house model panel run on September 19th 2026 and published in full as the Structural Reliance Matrix: humans reached, the distinct people whose daily tools run on the company’s AI, directly or through a downstream product, and purchasing power touched, the annual household consumption of those people. A third figure, money mediated, counts the third-party commerce, ad auctions, cloud spend or chip sales that flow through the company’s AI. The order weighs humans and money together. None of the figures is an audited statistic; where the panel disagreed, the entry gives the range. The method’s own caveat: headcount saturates at the internet population for the top three, so what separates them is substitutability, and on that measure the order at the top inverts to ASML, TSMC, Nvidia. Ranks 1 to 19 carry the panel’s two figures. Ranks 20 to 100 were not scored by the panel: they are grouped by role and ordered by judgement within each group, each with a dated 2026 fact and a source, and they will be scored in the January 2027 revision.

Tags for companies: [Doer] builds and sells AI; [Doer+Shaper] also sets conditions other companies must accept, a platform, a chokepoint, an app store or a default. No company carries a Thinker tag.

The models, September 2026

This tab is the models’ counterpart to the people list, and it leans on four public trackers. OpenRouter is the routing layer this review ran through and publishes usage rankings (licensed CC BY 4.0) and a benchmark view; Artificial Analysis runs independent evaluations and the Intelligence Index; Scale AI and the Center for AI Safety created and maintain Humanity’s Last Exam; LLM Stats aggregates the scores the labs report. Usage measures adoption, not quality; benchmarks measure a slice of quality, and each tracker’s method differs, so the same model can carry different numbers on different pages. Click a column header to sort a table. Data as of September 18th 2026 unless a table says otherwise; every figure is copied from the source, none is ours.

6.1 Text models most used on OpenRouter, trailing 30 days

Tokens processed through the OpenRouter API, prompt plus completion, over the trailing thirty days ending September 18th 2026. A free variant is ranked on its own. Prices are the list price on OpenRouter in USD per million tokens, input and output. Source: OpenRouter rankings.

6.2 The leading model for each task

OpenRouter classifies paid requests by task and reports each task’s share of spend and the models that take the largest share of it, over thirty days. The three leaders per task are shown with their share of that task’s spend. Source: top models by task.

6.3 Text models by intelligence and by price

The Artificial Analysis Intelligence Index (v4.3, ten evaluations combined), with its coding and agentic sub-indices, as published on OpenRouter’s benchmarks view. Weighted input price is what OpenRouter users actually paid per million input tokens over the period, cache discounts included; cost per request is the average paid per request. Sort by index for capability, by price or cost for economy, and read both together: the cheapest model at the intelligence you need is the practical answer.

6.4 Humanity’s Last Exam, on four trackers

Humanity’s Last Exam is 2,500 closed-ended questions written by nearly 1,000 subject experts from over 500 institutions, assembled by the Center for AI Safety and Scale AI, published in Nature in January 2026, and designed to be hard for models to search their way through. Each tracker runs or collects it differently, so the tables are shown side by side rather than merged. Calibration error, where reported, measures how far a model’s stated confidence sits from its accuracy; lower is better.

Scale AI SEAL leaderboard. Independent run; dataset updated September 17th 2026; accuracy with 95% confidence interval and calibration error; the rank is statistical, so tied ranks are real ties. Source.

Artificial Analysis. Independent run, 31 of the 616 models it tracks. Source.

LLM Stats. Aggregates the scores the labs report themselves, 104 models, none independently verified, which is why its numbers run higher; updated September 19th 2026. Source.

Center for AI Safety. The canonical page: paper, dataset, the HLE-Rolling fork and the original table (judge o3-mini, dataset of April 3rd 2025), kept here because it shows how far the frontier moved in eighteen months. Source.

6.5 Image, video, speech, transcription, embeddings and rerank

Text models are not comparable with generators of other media, so each modality has its own table: the models with most requests on OpenRouter over the trailing thirty days. Source: OpenRouter rankings, modality tabs.

Image generation.

Video generation.

Speech (text to speech).

Transcription.

Embeddings.

Rerank.

Quality for generated media is judged by people, not benchmarks. Design Arena collects blind pairwise votes; the leaders it reports, as relayed on OpenRouter’s benchmarks view, with Elo-style score and win rate:

6.6 The roster: flagship, newest and cheapest per house

One hundred models from OpenRouter’s catalogue on September 18th 2026. Houses are ordered by their share of OpenRouter token traffic in the week of publication; within a house, the highest-priced recent model stands in for the flagship, then the newest release, then the cheapest. Prices are USD per million tokens; context is in tokens.

Method

Influence, not fame, not merit. A person is on the list if removing them from 2026 (or from history) would visibly change what AI is, who controls it, or how people go about living with it. Wealth and public recognition are evidence of influence, not influence itself.

Stance and significance. This list is written from the co-evolutionary stance: the view that humans and AI are two forms of intelligence that can grow through each other, against the three other positions in the field, that AI must be restrained, that it is only a tool, or that it will replace us. The stance is declared so the reader can discount it. It shapes the ranking in one way. Among people of comparable influence, we rank higher the ones whose work makes a large positive change possible in human lives and does so with calm, respect and trust rather than fear or leverage. Significance is scored on the person’s own work, realized or credibly in progress with evidence, and applied symmetrically: potential counts for AlphaGenome and for Starlink or neither. Company leaders own their company’s conduct; scientists own their science. Heads of state are ranked on the magnitude of their instruments alone. Criticism appears only with a source and a stated sign; where a placement rests on harm, the entry carries the source. Each entry in the top 20, the heads of state excepted, carries the person’s public stance as a tag. Where the editor or the drafting agent has a personal stake, the record says so. The 2026 list also weighs what the person added to the field this year, not the size of the company behind them. Open weights are treated on this list as a co-evolutionary instrument: they distribute the agents, the audit and the responsibility.

Reach, two figures. Each entry in the top 20 ends with the reach of the company or cluster the person answers for: humans whose daily tools run on its AI, directly or through a downstream product, and the annual purchasing power of those people. Both are medians from a twelve-house model panel run on September 19th 2026 (the Structural Reliance Matrix, published as its own page), not audited statistics; scientists, researchers and heads of state carry the figure of the science, the report or the cluster their instruments govern, and the entry says which.

Four channels of influence, weighted roughly equally across the list:

1. Building the models and the science (labs, researchers).
2. Controlling the compute and capital (chips, clouds, data centres, sovereign funds).
3. Setting the rules (regulators, standards bodies, courts, unions).
4. Shaping how the world understands and uses it (educators, critics, safety researchers, creators).

World representation. The 2026 list is checked against region: it contains people based in North America, China, Europe, the UK, India, the Gulf, Japan, Korea, Taiwan, Africa and Latin America. Concentration in the US and China is real and reflected, not corrected away. Heads of state and government are included when they personally moved AI levers in the year: export controls, procurement, sovereign compute, binding law. AI is now a matter of world politics, and pretending otherwise would leave out the people with the largest budgets and the only binding rules. The people who run AI policy for them are included on their own merits.

Three categories, one list. Every entry is tagged. Doers build and deploy: models, chips, companies, infrastructure. Thinkers produce the science, the theory, the safety research and the critique that the doers act on. Shapers set the conditions: law, procurement, capital at sovereign scale, standards, courts, labour, religion. A fourth group, advocates and popularisers, is kept out of the 100 and listed separately in Part 3: they change what the public feels about AI, which matters, but they do not decide what AI is. Lists that mix the two groups without saying so read as cultural snapshots; this one separates them. A fourth badge, stance, appears on the top twenty and on a few entries below it: Co-evolutionist, Instrumentalist, Accelerationist, Retreatist or Not stated. Retreatists see AI as an existential risk and argue for restriction or pause; Instrumentalists see it as a product to be governed, risk managed without asking what AI could help humans become; Accelerationists see it as humanity’s inevitable successor and argue for unrestricted development; Co-evolutionists see it as a partner in intellectual collaboration in which both parties grow through the exchange, the position this list is written from. The tag is read from the person’s own words and work, and it is a description, not a grade; each badge explains itself on mouse-over. Region buttons (Americas, EMEA, Asia and Pacific) filter the people lists on every country in the label, so a person with two countries appears under both. Heads of state carry no stance tag: they are ranked on their instruments, not on their views.

Timing. Written 8.5 months into 2026. At the current pace the remaining 3.5 months can change the ranking materially, and the 2026 list will be revisited in January 2027. We are confident no major contribution has been forgotten; the ones too new to rank sit in Part 4.

Titles. In the 2026 list, Sir marks a knighthood and heads of state and government carry their office. Academic titles are left off for everyone, so that no one is ranked by a credential. The all-time list uses names alone, as histories do.

Country labels give the country where the person mainly works first, then origin where it differs. They are not verified citizenships.

Counting. Each list has 100 entries. The 2026 list names 101 people, because the Amodeis share one entry. Five entries name people who did one thing together (McCulloch and Pitts; Vapnik and Chervonenkis; Russell and Norvig; Sohl-Dickstein and Ho; the six remaining transformer authors), so the all-time list names 109 people.

No commander. The top of the 2026 list is a cluster, not a podium. Anthropic held the frontier early in the year with Opus 4.6, OpenAI may hold it for the rest of the year with GPT-6 Astra, and the chips, the science and the law sit with other people entirely. The order among the first seven is a judgement and the entries say what each one moved.

Ranking. The 2026 list is ranked. The default order is impact, the judgement described above; the page also offers two other orders, humans reached and purchasing power touched, built from the panel figures in the entries, so the reader can see the same people by scale. Entries without a figure keep their impact order below the ones that have one. The all-time list is grouped by era and lightly ordered within each group, because ranking Turing against Hinton is not a meaningful exercise.

Key works and company pages. Each 2026 entry ends with one key work (a paper, model, law or product) and, for doers, the company’s own page. Every link was fetched and checked by a research agent on September 18th 2026; where a site blocked automated access the link is marked unverified rather than guessed. Links are the best single pointer, not a bibliography.

Reviewers. One request per house, the latest model each had on OpenRouter on September 18th 2026, all given the same falsification prompt. Cutoff is what the model reported about itself.

Grok 4.6 joined for the wording review of the top ten (v7). Round one, on v2, used Nemotron 3.5 Lightning in Nvidia’s seat; it returned noise and was replaced by Nemotron 3 Ultra in the later rounds. Total spend about 4 USD for the list rounds, plus 5.04 USD for the four rounds of September 19th 2026 (structural reliance, ByteDance, purchasing power, top-20 validation), which ran without an Anthropic model because a Claude model drafted the texts they judged.

Sources for 2026 facts. TIME100 AI 2026 (published August 26th 2026), Fortune, Bloomberg, TechCrunch, CNBC, the International AI Safety Report 2026, company announcements up to September 16th 2026. Where a fact could not be confirmed, the entry says so.

Ten mistakes this list tries not to make

Written so the method outlives the list. Each is a failure we found in a published list or in our own earlier drafts, not a theory.

1. Confusing storylines with levers. A list built from “the year’s key storylines” rewards novelty and visibility. The test here is counterfactual. Removing the person would need to change what AI is, who controls it, or how people go about living with it. A CEO who stepped back to chair still passes it; a celebrity with a good year usually does not.
2. Mixing shapers with advocates without saying so. Advocacy is real influence of a different kind. Kept in Part 3 with its own label, it is honest; mixed into the 100, it turns a map of power into a cultural snapshot.
3. Excluding politicians by reflex, or including them by fame. Export controls, procurement, sovereign compute and binding law are the largest levers of 2026. Heads of state are ranked by the levers they personally pulled this year, not by office.
4. Counting the famous and missing the load-bearing. TSMC, ASML, SK hynix, SMIC and the EU AI Office decide more than most people who get quoted. Each round of review moved a bottleneck owner up and a spokesperson down.
5. Rewarding one’s own house. The agent that drafted this was built by Anthropic, and the first draft over-weighted Anthropic. Twelve houses reviewed it precisely because a model has no stake in claims it did not write. The count is now on the page, per lab, so the reader can judge.
6. Trusting a reviewer’s knowledge cutoff. Models with older training called verified 2026 facts hallucinations, and one invented three people. Every objection was checked against a primary source before it changed a line. Knowing which reviewer knows what is now part of the method.
7. Geography as decoration. One US-based founder does not represent a continent. The gaps that remained after review (Latin America, Korea outside memory chips) were named as gaps until real entries were found, and are still named where they remain.
8. Superlatives from press releases. Superlatives such as “largest fund” and “first ever” were caught by reviewers and softened to what the record supports.
9. Silent counting tricks. Pairs and groups are declared; the all-time list says it names 109 people in 100 entries.
10. Pretending September is December. The year is not over. The watch list holds what is too new to rank, and the list will be re-run in January 2027.