Skip to content
Shared AI Researchan open archive

Organizations45 organizations

The institutions of the AI era

Who does what: the labs training frontier models, the platforms distributing them, the supply chain beneath them, and the institutes and agencies trying to understand and govern them. The startup tracker records funding; this page records function.

Data compiled as of July 2026. Corrections by pull request are welcome.

Frontier labs13

The handful of organizations training models at the frontier of capability.

  • OpenAI

    est. 2015 · San Francisco

    Restructured in late 2025 as OpenAI Group PBC under the nonprofit OpenAI Foundation; makers of GPT, the o-series, and ChatGPT, the fastest-adopted consumer product in history.

  • Anthropic

    est. 2021 · San Francisco

    Safety-focused lab behind Claude; pioneered constitutional AI, interpretability at scale, and the leading agentic coding stack.

  • Google DeepMind

    est. 2010 · London / Mountain View

    Alphabet’s merged research arm (DeepMind + Google Brain, 2023); AlphaGo, AlphaFold, and the Gemini line.

  • Meta Superintelligence Labs

    est. 2025 · Menlo Park

    Meta’s frontier effort under chief AI officer Alexandr Wang; repeated reorganizations and a retreat from open-weight releases after Llama 4.

  • xAI

    est. 2023 · Palo Alto / Memphis

    Acquired by SpaceX in February 2026 and folded into a unified AI unit; its Memphis Colossus buildout now rents training capacity to rival labs.

  • Mistral AI

    est. 2023 · Paris

    Europe’s frontier champion; open-weight releases, a sovereignty pitch to European governments, and ASML as its anchor shareholder.

  • DeepSeek

    est. 2023 · Hangzhou

    Hedge-fund-funded lab whose efficient open models forced a global repricing of training-cost assumptions; V4 (2026) extended the run.

  • Alibaba (Qwen)

    est. 2023 · Hangzhou

    The most downloaded open-weight model family and the default base for fine-tunes worldwide, now split between Apache-licensed workhorses and a closed frontier flagship.

  • Moonshot AI

    est. 2023 · Beijing

    Maker of Kimi; K3 (July 2026), the largest open-weight model released to date, claims parity with the closed frontier.

  • ByteDance Seed

    est. 2023 · Beijing

    ByteDance’s frontier model group; its Seed models power Doubao, China’s most-used AI chatbot.

  • Z.ai (Zhipu AI)

    est. 2019 · Beijing

    Tsinghua spinoff behind the GLM open-weight line; became the first pure-play foundation-model lab to go public (Hong Kong, January 2026).

  • Safe Superintelligence

    est. 2024 · Palo Alto / Tel Aviv

    Sutskever’s deliberately productless lab pursuing aligned superintelligence directly; a July 2026 NVIDIA partnership scaled its compute an order of magnitude.

  • Thinking Machines Lab

    est. 2025 · San Francisco

    Murati’s lab betting on customizable, human-collaborative AI; shipped the Tinker fine-tuning API and its first open-weight model, Inkling, in July 2026.

Platforms & hyperscalers9

The companies that distribute AI to billions of users and rent the compute underneath it.

  • Microsoft

    est. 1975 · Redmond

    OpenAI’s anchor partner, holding 27% of the restructured PBC with model IP rights through 2032; Copilot across Windows, Office, and GitHub.

  • Alphabet

    est. 1998 · Mountain View

    The only company vertically integrated from silicon (TPU) through models (Gemini) to distribution; Gemini now also powers Apple’s rebuilt Siri.

  • Amazon (AWS)

    est. 1994 · Seattle

    Anthropic’s largest investor; Bedrock for model access and the Trainium-powered Project Rainier cluster for Claude training.

  • Apple

    est. 1976 · Cupertino

    The distribution giant that stayed cautious: on-device Apple Intelligence plus a 2026 deal putting Google’s Gemini behind the rebuilt Siri.

  • Oracle

    est. 1977 · Austin

    Reinvented as an AI landlord; hundreds of billions in contracted OpenAI capacity, though the Stargate buildout slowed and narrowed through 2026.

  • CoreWeave

    est. 2017 · Livingston, NJ

    The archetypal neocloud: crypto-mining castoff turned NYSE-listed GPU landlord; its bid for Core Scientific was voted down by shareholders in late 2025.

  • Nebius

    est. 2024 · Amsterdam

    Neocloud carved out of Yandex’s international assets; builds AI capacity for Microsoft and pitches sovereign infrastructure to European customers.

  • Lambda

    est. 2012 · San Francisco

    Developer-focused GPU cloud for training and inference; a multibillion-dollar Microsoft supply deal moved it into hyperscale territory.

  • Crusoe

    est. 2018 · Denver

    Energy-first datacenter developer that builds power alongside compute; anchor builder of the Abilene, Texas Stargate campus.

Semiconductors & supply chain9

The physical chokepoints of the AI economy.

  • NVIDIA

    est. 1993 · Santa Clara

    Supplies the overwhelming majority of AI training compute; CUDA remains the industry’s deepest moat.

  • TSMC

    est. 1987 · Hsinchu, Taiwan

    Fabricates essentially every leading-edge AI chip; the single point of failure in the global supply chain.

  • AMD

    est. 1969 · Santa Clara

    The credible second source; OpenAI’s six-gigawatt Instinct commitment begins deploying MI450 at scale in late 2026.

  • ASML

    est. 1984 · Veldhoven, Netherlands

    Monopoly maker of the EUV lithography machines below every advanced node; export-control linchpin.

  • SK Hynix

    est. 1983 · Icheon, South Korea

    Dominates high-bandwidth memory, the binding constraint on accelerator production since 2023; holds roughly 60% of the market into the HBM4 era.

  • Samsung

    est. 1969 · Suwon, South Korea

    Memory giant and the only leading-edge foundry alternative to TSMC; racing to regain HBM ground with HBM4 and HBM4E.

  • Micron

    est. 1978 · Boise, ID

    The third high-bandwidth memory supplier; overtook Samsung in HBM share during the HBM4 transition, easing the memory bottleneck.

  • Intel

    est. 1968 · Santa Clara

    The US national-champion foundry bet: roughly 10% government-owned since 2025, with the 18A node reaching volume production in early 2026.

  • Broadcom

    est. 1961 · Palo Alto

    The quiet winner of custom silicon: co-designs TPUs, OpenAI’s in-house accelerators, and the wider wave of lab-specific chips.

Research institutes & nonprofits8

Organizations producing public research, evaluations, and open infrastructure.

  • Allen Institute for AI (Ai2)

    est. 2014 · Seattle

    Fully open models (OLMo) with released training data and code — the transparency benchmark.

  • EleutherAI

    est. 2020 · distributed

    The volunteer collective that kept open-source LLMs alive pre-Llama (GPT-J, GPT-NeoX, the Pile).

  • Epoch AI

    est. 2022 · distributed

    The field’s statisticians: compute trends, training-run databases, and FrontierMath.

  • METR

    est. 2022 · Berkeley

    Independent evaluator of autonomous capabilities; its task-horizon measurements track agentic progress.

  • Redwood Research

    est. 2021 · Berkeley

    Originator of the AI control agenda: techniques for deploying capable models safely even if they turn out to be misaligned.

  • Apollo Research

    est. 2023 · London / San Francisco

    Evaluates frontier models for scheming and deception; a fixture of pre-deployment safety testing at the major labs.

  • MATS

    est. 2021 · Berkeley

    The field’s main safety-talent pipeline; its research fellows and alumni staff much of the alignment and evaluations ecosystem.

  • Stanford HAI

    est. 2019 · Stanford

    Publisher of the annual AI Index, the standard reference for the field’s trajectory.

Governance bodies6

Public institutions attempting to measure, standardize, and regulate frontier AI.

  • UK AI Security Institute

    est. 2023 · London

    The first state frontier-evaluation body; renamed from “Safety” to “Security” in 2025 as its remit narrowed.

  • US Center for AI Standards & Innovation

    est. 2023 · Washington, DC

    Successor to the US AI Safety Institute (renamed 2025); voluntary security testing of frontier models and a 2026 AI-agent standards initiative.

  • EU AI Office

    est. 2024 · Brussels

    Enforces the AI Act; GPAI obligations began August 2025, while the 2026 Digital Omnibus deferred high-risk deadlines to 2027–28.

  • Cyberspace Administration of China

    est. 2014 · Beijing

    China’s internet regulator; administers the layered AI rulebook of generative-AI measures, model filings, and AI-content labeling mandates.

  • UN Global Dialogue on AI Governance

    est. 2025 · Geneva / New York

    The first universal-membership AI forum, paired with an independent scientific panel; held its inaugural session in Geneva in July 2026.

  • Frontier Model Forum

    est. 2023 · Washington, DC

    Industry body of the frontier labs; coordinates safety frameworks, threat information sharing, and anti-distillation intelligence.