| DeepSeek | $52B | 2023 | General-purpose models — Math, Open source, Open weights | High-Flyer | Liang Wenfeng | Hangzhou China | Spun out of the quant fund High-Flyer; open-weight frontier models trained at a fraction of the compute cost US labs disclose. |
| Prometheus | $41B | 2025 | Physical AI & robotics | Amazon, Xaira Therapeutics, Stanford, Foresite Labs, GRAIL, UC Berkeley, Lawrence Berkeley National Laboratory, Verily, Google | Vik Bajaj, Jeff Bezos | San Francisco United States | Jeff Bezos' return to an operating role; AI for the physical economy. |
| SSI (Safe Superintelligence) | $32B | 2024 | General-purpose models | U. Toronto, Google, OpenAI, Apple, Stanford | Daniel Gross, Daniel Levy, Ilya Sutskever | Palo Alto United States | A straight-shot lab pursuing safe superintelligence, operating in total secrecy. |
| Reflection AI | $25B | 2024 | General-purpose models — Open weights, Sovereign AI | DeepMind, Claire AI, University of Chicago, UC Berkeley, UCL | Ioannis Antonoglou, Misha Laskin | New York United States | American open-weights frontier models, built by ex-DeepMind researchers. |
| Skild AI | >$14B | 2023 | Physical AI & robotics — Academic spinout, Robot foundation models | UC Berkeley, Carnegie Mellon, Meta (FAIR), Allen Institute for AI, U. Maryland | Abhinav Gupta, Deepak Pathak | Pittsburgh United States | Robot foundation models out of Carnegie Mellon. |
| Mistral AI | ~$14B | 2023 | General-purpose models — Agents, On-device, Open weights, Sovereign AI | DeepMind, Université Paris-Saclay, Meta (FAIR), Sorbonne Université, École des Ponts | Timothée Lacroix, Guillaume Lample, Arthur Mensch | Paris France | Europe's frontier lab; Apache-licensed open-weight models alongside proprietary ones, positioned as a sovereign alternative to the US labs. |
| Thinking Machines Lab | $12B | 2025 | General-purpose models — Open source, Open weights, Training infra / tooling | OpenAI, Leap Motion, Tesla, Anthropic, UC Berkeley, Google, Affirm, Indiana University, Meta (FAIR) | Luke Metz, Mira Murati, John Schulman, Andrew Tulloch, Lilian Weng, Barret Zoph | San Francisco United States | A research and product company building AI systems that are more understandable, customisable and capable. |
| Poolside | $12B * | 2023 | Coding — Agents, Open weights | GitHub, Heroku, Canonical, Athenian, source{d}, Tyba | Eiso Kant, Jason Warner | San Francisco United States | Founded by GitHub's former CTO; builds frontier foundation models purpose-built for software development. |
| Physical Intelligence | >$11B * | 2024 | Physical AI & robotics — Open source, Open weights, Robot foundation models | UC Berkeley, Google, DeepMind, Stanford, Stripe, Microsoft, Anduril, Tesla | Adnan Esmail, Chelsea Finn, Lachy Groom, Karol Hausman, Brian Ichter, Sergey Levine, Quan Vuong | San Francisco United States | General-purpose robot policies from the Stanford/Berkeley/Google robotics world. |
| StepFun | ~$10B | 2023 | General-purpose models — Multimodal, Open weights | Microsoft, Nanyang Technological University, Google, ByteDance, UC Santa Barbara | Jiao Binxing, Jiang Daxin, Zhu Yibo | Shanghai China | The Step model series; unwound its VIE structure to file in Hong Kong. |
| Cohere | ~$7B | 2019 | Applied — Agents, Multimodal, Open weights, Sovereign AI | Google, University of Oxford, FOR.ai | Nick Frosst, Aidan Gomez, Ivan Zhang | Toronto Canada | Enterprise LLMs sold on private deployment; the Command and Aya model families and the North agent platform. Announced a merger with Germany's Aleph Alpha in 2026. |
| Periodic Labs | ~$7B * | 2025 | AI for math & science — Materials | OpenAI, Google, DeepMind | Ekin Doğuş Çubuk, Liam Fedus | San Francisco United States | Autonomous labs for materials discovery; ex-OpenAI VP of Post-Training and ex-Google Brain/DeepMind materials lead. |
| Hark | $6B | 2025 | Applied — Agents | Archer Aviation, Figure AI | Brett Adcock | San Jose United States | Personalised multimodal AI and bespoke hardware designed as a new interface to computing. |
| Runway | $5.3B | 2018 | Generative media & voice — Multimodal, Video | Chartbeat | Anastasis Germanidis, Alejandro Matamala, Cristóbal Valenzuela | New York United States | Generative video for film and production, from Gen-1 to Gen-4.5; a co-author of the latent diffusion work behind Stable Diffusion. |
| Ineffable Intelligence | $5.1B | 2026 | Recursive self-improvement & continual learning — Continual learning | Flying Fish Partners, J.P. Morgan, onsemi, Tesla, Microsoft, DeepMind, Google, Fundamental, Isomorphic Labs, InstaDeep | Chris Apps, Wojciech Marian Czarnecki, Lasse Espeholt, Heather Gorham, Alex Laterre, Junhyuk Oh, David Silver | London United Kingdom | David Silver's lab building reinforcement-learning "superlearners" that learn from experience rather than human data. |
| World Labs | ~$5B * | 2024 | World models & simulation — Multimodal | Google, Stanford, University of Illinois Urbana-Champaign, Caltech, Meta (FAIR), U. Michigan, Epic Games, Amazon, U. Tübingen, UC Berkeley | Justin Johnson, Christoph Lassner, Fei-Fei Li, Ben Mildenhall | San Francisco United States | Spatial intelligence and 3D world generation, from the creator of ImageNet. |
| River AI | <$5B * | 2026 | Applied — Training infra / tooling | DeepMind, OpenAI, xAI, Apple, Meta (FAIR), AMD, Qualcomm, Tesla | Igor Babuschkin, Aaron Rogers, Dmytro "Dima" Soboliev, Ievgen Soboliev | Palo Alto United States | Personal AI agents that learn from and are controlled by their users, built on an end-to-end stack spanning training, models, products and on-device hardware. |
| Zyphra | >$5B * | 2021 | General-purpose models — Open source, Open weights | Cambridge Quantum Computing, University of Illinois Urbana-Champaign, IBM, Xerion Advanced Battery Corp., Conjecture, University of Oxford, Edinburgh, Apple, Xilinx, Qualcomm, Johns Hopkins University | Tomás Figliolia, Danny Martinelli, Beren Millidge, Krithik Puthalath | San Francisco United States | Outsized results from small models; the Zamba and Zaya (ZAYA1) stack. |
| Recursive Superintelligence | $4.65B | 2025 | Recursive self-improvement & continual learning — Agents | Salesforce, Stanford, Other startups, DeepMind, UCL, Inceptive, Google, Moscow State, Intel, OpenAI, UC Berkeley, Meta (FAIR), Carnegie Mellon, University of British Columbia, Uber, Michigan State | Jeff Clune, Alexey Dosovitskiy, Tim Rocktäschel, Tim Shi, Richard Socher, Yuandong Tian, Josh Tobin, Caiming Xiong | San Francisco United States | Self-improving AI; raised $500M+ within four months of founding. |
| AMI Labs | ~$4.5B | 2025 | World models & simulation | Meta (FAIR), Other startups, DeepMind, NYU, HKUST, Columbia, McGill | Pascale Fung, Alexandre LeBrun, Yann LeCun, Michael Rabbat, Laurent Solly, Saining Xie | Paris France | LeCun's post-Meta bet on world models over LLMs. |
| Unconventional AI | $4.5B | 2025 | Compute & chips | Databricks, MosaicML, Intel, Nervana Systems, Qualcomm, Chipletz, Google, Atheros, MIT, Microsoft, Stanford | Sara Achour, Michael Carbin, MeeLan Lee, Naveen Rao | San Francisco United States | Serial hardware founder (Nervana→Intel, MosaicML→Databricks) building analog AI chips for biology-scale efficiency. |
| humans& | $4.48B | 2026 | General-purpose models | xAI, Stanford, Notbad AI, HSLABS, Google, Silicon Graphics, University of Illinois Urbana-Champaign, U. Michigan, Anthropic, Defense Innovation Unit, Microsoft, White House / NSC, MIT, OpenAI, Meta (FAIR), DeepMind, Uber | Noah Goodman, Georges Harik, Yuchen He, Andi Peng, Eric Zelikman | San Francisco United States | Human-centric frontier lab founded by alumni of xAI, OpenAI, Meta, Anthropic, Google, and Stanford; $480M seed within months of founding. |
| Core Automation | ~$4B * | 2026 | Recursive self-improvement & continual learning — Continual learning | OpenAI, DeepMind, Google, Anthropic | Rohan Anil, Anmol Gulati, Joanne Jang, Jerry Tworek, Julia Villagra | San Francisco United States | Ex-OpenAI VP building “the world's most automated AI lab”. |
| Inflection AI | ~$4B | 2022 | Applied | DeepMind, Google, University of Oxford, Other startups, Greylock | Reid Hoffman, Karén Simonyan, Mustafa Suleyman | Palo Alto United States | The Pi assistant, then a 2024 Microsoft reverse acquihire of its co-founders and much of its team; rebuilt since as an enterprise AI company. |
| Luma AI | ~$4B | 2021 | Generative media & voice — Agents, Video | Apple, Circle Medical, UC Berkeley, Codeplay Software | Amit Jain, Alberto Taiuti, Alex Yu | Palo Alto United States | Dream Machine and the Ray video-model family; evolved from neural rendering and smartphone 3D capture into generative video. |
| Ricursive Intelligence | $4B | 2025 | Compute & chips | Google, DeepMind, Anthropic, Stanford | Anna Goldie, Azalia Mirhoseini | Palo Alto United States | The researchers behind AI-designed chip floorplanning, now doing it commercially. |
| Decart | ~$4B | 2023 | World models & simulation — Video | Technion, Unit 8200 | Dean Leitersdorf, Moshe Shalev | Tel Aviv Israel | Real-time generative world models — Oasis and Mirage. |
| Chai Discovery | $3.8B | 2024 | AI for math & science — Drug discovery, Open weights | Meta (FAIR), OpenAI, Absci, Stripe, VantAI, University of Chicago, Aqemia | Jacques Boitreaud, Jack Dent, Matthew McPartlon, Joshua Meier | San Francisco United States | OpenAI-backed; Chai-1 structure prediction and Chai-2 de novo antibody design, used by Pfizer, Lilly and Novartis. |
| Isomorphic Labs | Not independently valued | 2021 | AI for math & science — Drug discovery | DeepMind | Demis Hassabis | London United Kingdom | Alphabet’s AI drug-discovery company, building predictive and generative models beyond AlphaFold. |
| Xaira Therapeutics | Undisclosed | 2023 | AI for math & science — Drug discovery | Stanford, Genentech, Rockefeller, UCSF, UCL, U. Washington, Meta (FAIR), Carnegie Mellon, Foresite Labs, GRAIL, UC Berkeley, Lawrence Berkeley National Laboratory, Verily, Google | Vik Bajaj, David Baker, Hetu Kamisetty, Marc Tessier-Lavigne | South San Francisco United States | An AI-native drug-discovery company combining model research, data generation and therapeutic development. |
| Black Forest Labs | $3.25B | 2024 | Generative media & voice — Multimodal, Open weights, Video | Stability AI, LMU Munich, Runway, Heidelberg University | Andreas Blattmann, Patrick Esser, Dominik Lorenz, Robin Rombach | Freiburg Germany | The Stable Diffusion authors' lab; the FLUX family of image models, part of it Apache-licensed, now extending into video and robot manipulation. |
| Genesis AI | ~$3B * | 2025 | Physical AI & robotics — Robot foundation models | Carnegie Mellon, Mistral, Skild AI | Théophile Gervet, Zhou Xian | Paris France | Robotics foundation models (the GENE series) on a proprietary physics-simulation engine; a Franco-American team out of CMU, Mistral and Skild AI. |
| Galbot | >$3B | 2023 | Physical AI & robotics — Humanoid, Robot foundation models | Peking University, Stanford, ABB | Wang He, Yao Tengzhou | Beijing China | Embodied foundation models and autonomous retail stores; over $900M raised. |
| Baichuan AI | ~$2.9B | 2023 | Applied — Agents | Sogou, Other startups, Tsinghua | Ru Liyun, Wang Xiaochuan | Beijing China | Sogou's founder; shifted strategic focus from the general-purpose LLM race toward AI-for-healthcare, while continuing to develop its base model series. |
| Sakana AI | $2.65B | 2023 | General-purpose models — Sovereign AI | Google, Stability AI, U. Tokyo, Goldman Sachs, Mercari | David Ha, Ren Ito, Llion Jones | Tokyo Japan | Japan's first AI unicorn, co-founded by a Transformer co-author; nature-inspired model merging and AI for science, betting on sample efficiency over raw compute. |
| CuspAI | ~$2.6B | 2024 | AI for math & science — Materials | Karlsruhe Institute of Technology, Other startups, Caltech, U. Amsterdam, Utrecht, U. Toronto, CIFAR, Microsoft, Qualcomm, Scyfer, UC Irvine | Chad Edwards, Max Welling | Cambridge United Kingdom | AI materials-discovery platform — expanding from carbon capture into semiconductors, batteries, coatings and catalysts. |
| General Intuition | $2.3B | 2025 | World models & simulation — Video | Medal, University of Geneva, Ubisoft, Edinburgh, Dataiku | Eloi Alonso, Adam Jelley, Vincent Micheli, Pim de Witte | New York United States | Action and world models trained on gameplay video to perceive, predict and act in virtual and physical environments. |
| Liquid AI | >$2B | 2023 | General-purpose models — Academic spinout, On-device, Open weights | MIT, Cornell, Dartmouth College, The Vanguard Group, TU Wien, Infineon Technologies, ISTA (Austria), Themis AI | Alexander Amini, Ramin Hasani, Mathias Lechner, Daniela Rus | Cambridge, MA United States | MIT CSAIL spinout; liquid neural networks for on-device inference. |
| Simile | $2B | 2025 | World models & simulation | Stanford, MIT, Meta (FAIR), UC Berkeley, Valence, Hebbia | Michael Bernstein, Percy Liang, Joon Sung Park, Lainie Yallen | Palo Alto United States | The Stanford generative-agents lineage, commercialised. |
| Aleph Alpha | ~$2B * | 2019 | Applied — Multimodal, Sovereign AI | Apple, Pallas Ludens, Deloitte | Jonas Andrulis, Samuel Weinbach | Heidelberg Germany | Europe's early frontier lab, from the Luminous models to PhariaAI — a sovereign AI platform built for government, defence and regulated industry. |
| Generalist AI | $2B | 2024 | Physical AI & robotics — Robot foundation models | DeepMind, MIT, Princeton, Broad Institute, Boston Dynamics | Andrew Barry, Pete Florence, Andy Zeng | San Mateo United States | GEN-0 and GEN-1, embodied foundation models pretrained from scratch on hundreds of thousands of hours of raw physical interaction collected with handheld grippers rather than in simulation; the DeepMind robotics researchers behind PaLM-E and RT-2, selling cross-embodiment robot intelligence rather than robots. |
| Rhoda AI | $1.7B | 2024 | Physical AI & robotics — Robot foundation models | Other startups, Stanford, MIT, University of British Columbia, UT Austin, Meta (FAIR) | Eric Chan, Changan Chen, Jagdeep Singh, Gordon Wetzstein, Andrew Wooten | San Francisco United States | Serial founder Jagdeep Singh's robotics startup; FutureVision, a video-pretrained foundation model ('Direct Video-Action') that predicts the next frames of a scene and converts them into motor commands, driving a bimanual manipulation platform on factory lines. |
| Axiom Math | >$1.6B | 2025 | AI for math & science — Math, Open source | Stanford, Meta (FAIR), Baidu, Shazam, Intel, Sun Microsystems, UC Davis | Carina Hong, Shubho Sengupta | Palo Alto United States | AxiomProver, an AI system for mathematical research that generates formally verified Lean proofs. |
| Flapping Airplanes | $1.5B | 2025 | General-purpose models | Stanford, Prod, Thiel Fellowship, Neuralink, Chess.com | Aidan Smith, Asher Spector, Ben Spector | San Francisco United States | A research lab pursuing radically more data-efficient AI, especially in data-constrained fields such as robotics and science. |
| Magic | ~$1.5B * | 2022 | Coding | Meta (FAIR), Other startups | Sebastian De Ro, Eric Steinberger | San Francisco United States | AI models for software development, including ultra-long-context systems. |
| Sarvam AI | $1.5B | 2023 | General-purpose models — Open weights, Sovereign AI, Voice | UIDAI / Aadhaar, Other startups, Carnegie Mellon, Microsoft, IBM, IIT Madras | Pratyush Kumar, Vivek Raghavan | Bengaluru India | Foundation models and speech built for Indian languages; selected under the IndiaAI Mission to build an indigenous foundational model. |
| Harmonic | $1.45B | 2023 | AI for math & science — Math | Other startups, Stanford, Helm.ai, Quora | Tudor Achim, Vlad Tenev | Palo Alto United States | Co-founded by Robinhood's Vlad Tenev (chairman) and CEO Tudor Achim, building formally verified mathematical superintelligence (the Aristotle model). |
| AI21 Labs | $1.4B | 2017 | Applied — Open weights | Other startups, Stanford, Google, Mobileye, Hebrew University | Ori Goshen, Amnon Shashua, Yoav Shoham | Tel Aviv Israel | The original Israeli LLM company, predating the current wave by five years; its Jamba models pair Mamba state-space layers with transformer blocks for long context at lower memory cost. |
| Fundamental | ~$1.4B | 2024 | Applied — Tabular / structured data | Bridgewater Associates, J.P. Morgan, Other startups, DeepMind, Imperial College London, Greenfield Partners | Jeremy Fraenkel, Marta Garnelo, Gabriel Suissa | San Francisco United States | NEXUS, a deterministic large tabular model pretrained on billions of enterprise spreadsheets and database tables, aimed at fraud detection, pricing and forecasting rather than free-text generation. |
| Lila Sciences | >$1.3B | 2023 | AI for math & science — Agents, Drug discovery, Materials | MIT, Other startups, WashU | Noubar Afeyan, Molly Gibson, Geoffrey von Maltzahn | Cambridge, MA United States | Flagship Pioneering's scientific superintelligence venture: the Lila Iris reasoning model driving AI Science Factory autonomous labs, which have closed the loop across hundreds of thousands of experiments spanning antibodies, genetic medicines, catalysts and coatings. |
| Deepgram | $1.3B | 2015 | Generative media & voice — Agents, Voice | U. Michigan, UC Davis | Noah Shutty, Scott Stephenson, Adam Sypniewski | San Francisco United States | Founded by Michigan particle physicists who repurposed dark-matter waveform analysis for speech; the Nova speech-to-text models and a voice-agent stack sold as enterprise API infrastructure. |
| Goodfire | $1.25B | 2024 | AI safety & interpretability | Other startups, DeepMind, Imperial College London | Dan Balsam, Eric Ho, Tom McGrath | San Francisco United States | One of the first companies built entirely around mechanistic interpretability; Ember, an API for inspecting and steering model internals. |
| Prime Intellect | $1B | 2024 | Compute & chips — Agents, Open source, Open weights, Training infra / tooling | Molecule, Bunch, Aleph Alpha, VitaDAO | Johannes Hagemann, Vincent Weisser | San Francisco United States | Aggregated compute plus a training, sandbox and inference stack, sold on the pitch that any company can be its own lab; ships the INTELLECT open-weight models and an open RL toolchain. |
| Discovery Loop | Undisclosed | 2026 | Recursive self-improvement & continual learning | Google, DeepMind, U. Washington, MIT, Stanford, UC Berkeley | Jeff Dean, Sanjay Ghemawat, Quoc Le, Oriol Vinyals | Palo Alto United States | Three of the most-cited researchers in AI and two of the most-cited in distributed systems. Search, Spanner, MoE, distillation, Chinchilla laws and more. Now working on automating the scientific method. |
| Aaru | $1B * | 2024 | World models & simulation | MIT | Cameron Fink, John Kessler, Ned Koh | New York United States | AI-built populations and worlds that simulate how groups respond to new products, policies and other changed conditions. |
| Ndea | Undisclosed | 2025 | Recursive self-improvement & continual learning | Google, ARC Prize Foundation, Zapier | François Chollet, Mike Knoop | San Francisco United States | Keras and Zapier founders betting on program synthesis over scaling. |
| Imbue | >$1B | 2022 | Applied — Agents, Open source | Dropbox, Other startups | Josh Albrecht, Kanjun Qiu | San Francisco United States | Agent lab building Sculptor, a sandboxed environment for running coding agents in parallel, alongside other tools. |
| doubleAI | >$1B * | 2023 | General-purpose models | Mobileye, Hebrew University, AI21 Labs, Caltech, Google | Gal Beniamini, Yoav Levine, Shai Shalev-Shwartz, Or Sharir, Amnon Shashua, Noam Wies | Tel Aviv Israel | A once-stealth lab pursuing 'Artificial Expert Intelligence' — deep domain-expert AI rather than AGI — now public with its first system, WarpSpeed, for GPU kernel optimisation. |
| Reka | $1B | 2022 | Physical AI & robotics — Multimodal | DeepMind, Baidu, Carnegie Mellon, University of Oxford, Meta (FAIR), University of the Basque Country, Google, Nanyang Technological University | Mikel Artetxe, Cyprien de Masson d'Autume, Qi Liu, Yi Tay, Dani Yogatama | San Francisco United States | Models for physical-world intelligence: reasoning, simulation and action across robots, wearables, media and edge devices. |
| Arcee AI | >$1B * | 2023 | Applied — Open source, Open weights, Training infra / tooling | Hugging Face, Roboflow, Tecton | Brian Benedict, Mark McQuade, Jacob Solawetz | Miami United States | Small enterprise language models; absorbed mergekit (created by Charles Goddard, now Arcee's Chief of Frontier Research) via a 2024 merger. |
| Mirendil | $1B | 2026 | Recursive self-improvement & continual learning | Anthropic, Google, DeepMind, xAI, Twitter, OpenAI, Prod | Harsh Mehta, Behnam Neyshabur, Tara Rezaei, Shayan Salehian | San Francisco United States | Founded by ex-Anthropic researchers Behnam Neyshabur and Harsh Mehta, building self-improving AI for scientific R&D; backed by a16z and Kleiner Perkins. |
| Nous Research | $1B | 2023 | Applied — Agents, Open source, Open weights | Other startups, Stability AI | Karan Malhotra, Shivani Mitra, Jeffrey Quesnelle, Ryan Teknium | New York United States | Open-source Hermes models and decentralised training, from an internet-native collective. |
| 01.AI | Undisclosed | 2023 | Applied — Open weights | Sinovation Ventures, Google, Microsoft, Apple, Silicon Graphics, Baidu, Other startups, SAP, Cisco, Alibaba, Huawei, Didi, Peking University | Anita Huang, Ma Jie, Kai-Fu Lee, Shen Pengfei, Qi Ruifeng, Li Xiangang, Gu Xuemei, Dai Zonghong | Beijing China | The Yi open-weight model family through early 2025; now enterprise and sovereign AI for industry and government. |
| Merge Labs | $850M | 2026 | AI for math & science — Neurotech / BCI | Brown University, MIT, Cyberkinetics, Third Rock Ventures, Caltech, Princeton, Forest Neurotech, University of Utah, UC Irvine, AE Studio, FAU Erlangen–Nürnberg, Max Planck Institute, Tools for Humanity / World, Stanford, Loopt, Y Combinator, OpenAI | Tyson Aflalo, Sam Altman, Alex Blania, Sandro Herbig, Sumner Norman, Mikhail Shapiro | San Francisco United States | Less-invasive, high-bandwidth brain-computer interfaces combining biology, devices and AI. |
| Inferact | $800M | 2026 | Inference — Open source | UC Berkeley, Anyscale, Character.AI, Seoul National University, DeepMind, Thinking Machines Lab, Tsinghua, Apple, Roblox, University of Waterloo, U. Washington | Woosuk Kwon, Simon Mo, Roger Wang, Kaichao You | Berkeley United States | Founded by vLLM’s creators and core maintainers to build and commercialise open-source LLM inference infrastructure. |
| Prior Labs | Not independently valued | 2024 | Applied — Academic spinout, Open source, Open weights, Tabular / structured data | University of Freiburg, University of British Columbia, Bosch Center for AI | Sauraj Gambhir, Noah Hollmann, Frank Hutter | Freiburg Germany | Creator of TabPFN, a tabular foundation model that predicts on structured data by in-context learning over synthetic priors instead of per-dataset gradient training; spun out of the AutoML lab at Freiburg. |
| Isara Laboratories | $650M | 2025 | Applied — Agents | OpenAI, MIT, Harvard, Oxford / Cambridge | Henry Gasztowtt, Eddie Zhang | San Francisco United States | OpenAI-backed; demoed ~2,000 agents forecasting commodity prices for finance and biotech analysis. |
| EvolutionaryScale | Undisclosed | 2024 | AI for math & science — Drug discovery, Open weights | Meta (FAIR) | Sal Candido, Alexander Rives, Tom Sercu | New York United States | Creator of ESM3, a generative model for protein sequence, structure and function that produced a novel fluorescent protein; its team later joined Biohub. |
| Engram | $600M | 2026 | Recursive self-improvement & continual learning — Agents, Continual learning | Columbia, Stanford, Cornell, Meta (FAIR), Google, UC Berkeley, Harvard, U. Washington | Dan Biderman, Sabri Eyuboglu, Jessy Lin, Scott Linderman, Jack Morris, Chris Ré | San Francisco United States | A learned memory layer that trains models on an organization's knowledge in advance, creating compact, continuously improving memories that use up to 100× fewer tokens. |
| Moonvalley | Not independently valued | 2024 | Generative media & voice — Video | U. Toronto, Toggl, Zapier, ContentFly / Draft, Max Planck Institute, Saarland University, DeepMind, Imperial College London, Google, Mila, IBM, Jackman, RYOT, Verizon, XTR, Asteria | Mikołaj Bińkowski, Mateusz Malinowski, Bryn Mooser, Naeem Talukdar, John Thomas | Toronto Canada | Marey, a filmmaking-focused generative video model trained on licensed footage and built for precise creative control; later joined forces with Reka. |
| Inception Labs | ~$500M | 2025 | General-purpose models | Stanford, UCLA, Cornell | Stefano Ermon, Aditya Grover, Volodymyr Kuleshov | Palo Alto United States | Develops and deploys diffusion-based language models, including Mercury, for production applications. |
| Standard Intelligence | $500M | 2024 | General-purpose models — Agents, Video | Other startups | Galen Mead, Devansh Pandey | San Francisco United States | FDM-1, a video-first computer-use model. |
| Adaption Labs | Undisclosed | 2025 | Recursive self-improvement & continual learning — Continual learning | Cohere, Google | Sara Hooker, Sudip Roy | San Francisco United States | Cohere For AI leadership betting that adaptation beats scale. |
| Irregular | $450M | 2023 | Applied — Security | IBM, Unit 81, Google, Unit 8200 | Dan Lahav, Omer Nevo | Tel Aviv Israel | Frontier AI security lab that stress-tests models for offensive cyber capability inside simulated networks; its evaluations ship inside OpenAI and Anthropic system cards, alongside the SOLVE vulnerability-scoring framework. |
| RadixArk | $400M | 2026 | Inference — Open source, Training infra / tooling | xAI, Databricks, Stanford, U. Washington, Nvidia, Nexusflow AI | Ying Sheng, Banghua Zhu | Berkeley United States | Commercialising SGLang; out of the LMSYS and xAI orbit. |
| Oak Lab | Undisclosed | 2026 | Recursive self-improvement & continual learning — Continual learning | Keen Technologies, DeepMind, U. Alberta | Khurram Javed, Richard Sutton | Toronto Canada | The Turing laureate who wrote the book on reinforcement learning, arguing today's methods are a dead end and betting on agents that learn from their own experience — targeting a trillion-parameter agent that plans in real time on 20 watts. |
| Cartesia | >$400M | 2023 | Generative media & voice — Academic spinout, Agents, Voice | Stanford, Apple, Snorkel AI, Google, U. Washington | Arjun Desai, Karan Goel, Albert Gu, Chris Ré, Brandon Yang | San Francisco United States | Real-time voice built on state space models rather than transformers, out of the Stanford lab that invented them; the Sonic text-to-speech, Ink speech-to-text and Line agent stack, tuned for sub-100ms latency. |
| Orbital | Undisclosed | 2022 | AI for math & science — Agents, Materials, Open source, Open weights | DeepMind, Bloomsbury AI, DataSine, Pluto Data Analytics, Imprint AI | James Gin-Pollock, Jonathan Godwin, Daniel Miodovnik | London United Kingdom | Orb, an Apache-licensed universal interatomic potential fast enough to simulate 100,000 atoms on one GPU, and CurieOS, the agent OS built on top of it; renamed from Orbital Materials as it began selling what those models designed — a PFAS-free two-phase refrigerant for 2,000W+ GPUs, pitched as the first AI-designed molecule to reach commercial market. |
| QUTWO | $380M | 2026 | Compute & chips — Sovereign AI | AMD, IQM | Kaj-Mikael Björk, Peter Sarlin, Kuan-Yen Tan | Helsinki Finland | Qutwo OS and AI/quantum tooling for enterprises running classical, hybrid and quantum workloads. |
| H (The H Company) | Undisclosed | 2024 | Applied — Agents, Open weights | Paris Dauphine–PSL, CentraleSupélec, Stanford, Harvard, École Polytechnique, École des Ponts, DeepMind, IDSIA, Vrije Universiteit Brussel, Maastricht University, University of Liverpool, University of Lille | Charles Kantor, Julien Perolat, Laurent Sifre, Karl Tuyls, Daan Wierstra | Paris France | French agentic/computer-use AI company, founded by a large ex-DeepMind contingent. |
| Sooth Labs | $335M | 2026 | World models & simulation | Meta (FAIR), Electronic Arts, Carnegie Mellon, Apple, Carnegie Robotics, Uber | Chuck Hoover, David Larose, Ruslan Salakhutdinov, Yaser Sheikh, Shih-En Wei | Pittsburgh United States | Ex-Meta Codec Avatars leadership, now on probabilistic forecasting. |
| Kyutai | Not applicable | 2023 | Generative media & voice — Open source, Open weights, Voice | Valeo.ai, DeepMind, Google, Meta (FAIR), Other startups | Alexandre Défossez, Edouard Grave, Hervé Jégou, Laurent Mazare, Xavier Niel, Patrick Perez, Rodolphe Saadé, Neil Zeghidour | Paris France | Non-profit European open-science lab; Moshi shipped real-time voice before the big labs. |
| Elorian AI | $300M | 2025 | General-purpose models — Vision | DeepMind, Google, Apple | Andrew Dai, Yinfei Yang | Palo Alto United States | Ex-DeepMind and Apple leads working on vision-grounded reasoning. |
| Deep Cogito | Undisclosed | 2024 | General-purpose models — Open weights | Google, DeepMind, Microsoft | Drishan Arora, Dhruv Malrana | San Francisco United States | Open-weight hybrid reasoning models trained by iterated distillation and amplification; the model's own search traces are folded back into its weights. |
| Essential AI | Undisclosed | 2023 | General-purpose models — Open weights | Google, Adept | Niki Parmar, Ashish Vaswani | San Francisco United States | An open platform for deep-learning research and engineering; released Rnj-1, an open-weight language-model family for code and STEM work. |
| Inherent | Undisclosed | 2026 | Recursive self-improvement & continual learning — Agents | DeepMind, White House / NSC, London School of Economics, Queen Mary University of London, Google, IDSIA, UCL, Hasso Plattner Institute, Reka, Microsoft, University of Oxford | Kaloyan Aleksiev, Tantum Collins, Edward Hughes, Louis Kirsch | London United Kingdom | London public-benefit AI lab developing inventive research agents and redesigning the lab itself around recursive collective self-improvement. |
| Latent Labs | Undisclosed | 2023 | AI for math & science — Agents, Drug discovery | DeepMind, German Cancer Research Center, Karlsruhe Institute of Technology | Simon Kohl | London United Kingdom | Generative protein, peptide and antibody design models and autonomous drug-design agents, paired with an in-house wet lab for experimental validation. |
| Edison Scientific | ~$250M | 2025 | AI for math & science — Agents, Drug discovery | FutureHouse, Francis Crick Institute, MIT, University of Rochester, University of Chicago, U. Washington, Max Planck Institute | Sam Rodriques, Andrew White | San Francisco United States | FutureHouse's for-profit spinout, commercialising Kosmos — an agent that runs multi-day research campaigns across literature and proprietary data to propose therapeutic targets. |
| Logical Intelligence | Undisclosed | 2025 | AI for math & science — Agents, Math | UC Santa Barbara | Eve Bodnia | San Francisco United States | Energy-based reasoning models and formal-verification agents that enforce constraints and produce machine-checkable answers for critical systems. |
| NeoCognition | Undisclosed | 2025 | Recursive self-improvement & continual learning — Continual learning, Agents, Academic spinout | Ohio State University, Microsoft, UC Santa Barbara, Scale AI, Google | Xiang Deng, Yu Gu, Yu Su | Palo Alto United States | Ohio State's AI agent group commercialized: the team behind Mind2Web, MMMU and SeeAct, now building agents that learn a “world model of work” on the job and specialize into domain experts rather than one general super-agent. |
| Grafton Sciences | Undisclosed | 2024 | Physical AI & robotics | Other startups | Anubhav Dubey | Redwood City United States | Robot-run factories and laboratories, and the models to drive them, so an AI can run its own physical experiments; funded so far by a $42.5M ARPA-H award for at-home cancer screening. |
| UniversalAGI | Undisclosed | 2025 | Physical AI & robotics | Brain Co., Anyscale, UC Berkeley | Ameer Haj-Ali | San Francisco United States | Large physics models in place of the CAE loop: SUV-PT predicts surface pressure, wall shear stress and drag straight from 3D geometry in seconds where high-fidelity CFD takes days, on an in-house Latent Interaction Field Transformer trained on millions of generated simulations. |
| Moonlake AI | $120M | 2025 | World models & simulation | Stanford, Nvidia | Sharon Lee, Fan-Yun Sun | San Francisco United States | Vibe-coding interactive worlds and games from text; $28M seed backed by Ian Goodfellow, Jeff Dean, and other angels. |
| Trajectory | $115M | 2026 | Recursive self-improvement & continual learning — Agents, Continual learning, Training infra / tooling | Windsurf, DeepMind, Stanford, Google, Apple | Michael Elabd, Arjun Karanam, Ronak Malde | San Francisco United States | A continual-learning layer that turns user feedback on agent trajectories into post-training data, helping deployed models improve from real-world use. |
| PrismML | Undisclosed | 2025 | General-purpose models — Academic spinout, Multimodal, On-device, Open source, Open weights | Neural Propulsion Systems, Caltech, Bell Labs, Stanford, UCLA, Instacart | Babak Hassibi, Sahin Lale, Omead Pooladzandi, Reza Sadri | Pasadena United States | Caltech spinout built on a compression theory that pushes model weights down to 1 bit and ternary; the Apache-2.0 Bonsai family and its custom llama.cpp and MLX kernels fit a 27B-class model into 3.9GB, small enough to run on a phone. |
| Autoscience | Undisclosed | 2024 | Recursive self-improvement & continual learning — Agents | Google, MIT | Eliot Cowan | San Mateo United States | Autonomous AI scientists that read literature, propose and run experiments, and turn verified discoveries into improved production machine-learning models. |
| Unreasonable Labs | Undisclosed | 2026 | AI for math & science — Materials | DeepMind, Google, Baidu, KITT.AI, Johns Hopkins University, MIT, Caltech, Max Planck Institute | Markus J. Buehler, Yuan Cao | Palo Alto United States | Unreasonable.DISCOVERY, a cross-domain discovery engine that pairs LLMs with neurosymbolic abstractions over a unified world model of physics, biology, chemistry and materials — built to compose new hypotheses rather than retrieve known facts. |
| Physical Superintelligence | Undisclosed | 2025 | AI for math & science — Agents, Open source | Bitcoin Policy Institute, SentinelOne, Other startups, Harvard | Matthew Pines, Alex Wissner-Gross | Boston United States | Get Physics Done, an open-source agentic AI physicist that scopes a problem, plans the research, runs its own derivations and numerical checks and verifies the results against physical constraints; a public benefit corporation industrialising physics discovery rather than building robots. |
| Math, Inc. | Undisclosed | 2025 | AI for math & science — Agents, Math, Open source | OpenAI, University of Pittsburgh, xAI, Google, Cadence Design Systems, University of Bonn | Jesse Michael Han, Christian Szegedy | Palo Alto United States | Gauss, an autoformalization agent that turns published mathematics into machine-checked Lean; it closed Terence Tao and Alex Kontorovich's 18-month Prime Number Theorem challenge in three weeks, then formalised Viazovska's Fields Medal sphere-packing proofs. |
| Oumi | Undisclosed | 2024 | Applied — Multimodal, Open source, Open weights, Training infra / tooling | Google, Microsoft, Meta (FAIR), Princeton, MIT, Cornell, University of Cincinnati, Stanford, Apple, Twitter, Georgia Tech, Steel Perlot, Snap | Panos Achlioptas, Kostas Aisopos, Oussama Elachqar, Jeremiah Greer, Manos Koukoumidis, Matthew Persons, William Zeng | Bellevue United States | Public-benefit lab and open platform for evaluating, fine-tuning and deploying specialized foundation models that organizations can own. |
| Poetiq | Undisclosed | 2025 | Recursive self-improvement & continual learning — Agents | Google, DeepMind, Other startups | Shumeet Baluja, Ian Fischer | Mountain View United States | Model-agnostic, self-improving reasoning systems that learn from solving new tasks. |
| LawZero | Not applicable | 2025 | AI safety & interpretability | Mila, Université de Montréal, CIFAR, IVADO, McGill, MIT, Bell Labs, Element AI | Yoshua Bengio | Montreal Canada | Yoshua Bengio’s nonprofit developing safe-by-design advanced AI, centered on a transparent, non-agentic Scientist AI without goals of its own. |
| Cursive | Undisclosed | 2025 | Recursive self-improvement & continual learning | DeepMind, Imperial College London, UCL, University of Oxford, NYU, École Polytechnique | Talfan Evans, Olivier J. Hénaff, Oliver Vikbladh | London United Kingdom | London frontier lab building foundation models and low-latency generative infrastructure for adaptive software and agents that improve continuously from use. |
| FutureHouse | Not applicable | 2023 | AI for math & science — Agents, Open weights | Francis Crick Institute, MIT, University of Rochester, University of Chicago, U. Washington, Max Planck Institute | Sam Rodriques, Andrew White | San Francisco United States | Nonprofit building and wet-lab-validating AI scientists for basic research in biology and health; its commercial platform spun out as Edison Scientific in 2025. |
| Transluce | Not applicable | 2024 | AI safety & interpretability — Open source, Open weights | UC Berkeley, OpenAI, Open Philanthropy, Stanford, MIT, IBM, Baylor College of Medicine, UTHealth Houston, Rice University | Sarah Schwettmann, Jacob Steinhardt | San Francisco United States | Nonprofit building open infrastructure for scalable oversight of frontier AI through automated interpretability, model-behavior analysis and evaluations. |