The world’s most valuable foundation model startups

Labs that train general-purpose AI models from scratch: large language and multimodal models, open-weight model makers, sovereign and regional language models, and world models that learn from video and simulation. Ranked worldwide by AI-estimated valuation, with listed pure players shown in grey at their market value for reference. The market sets the value; our AI estimates it.

Key metrics for this segment: key milestone. Figures as published by each company or the press, dated, not recalculated. The AIs weigh them alongside funding history and comparables.

Not in this segment. Listed, so not ranked: Zhipu AI and MiniMax (shown in grey), LightOn, whose platform sits above other makers’ models. Bought or being bought by a group: xAI (merged into SpaceX, listed), World Labs (AMD, agreed Sep 2026), Aleph Alpha (merging into Cohere), Silo AI (AMD), ELYZA (KDDI). Created by a group or a state, or without outside investors: Humain (PIF), Inception and Jais (G42), Kyutai (non-profit), Japan’s national foundation-model company (SoftBank, NEC, Sony, Honda), BharatGen. In other segments: coding models (Poolside, Magic) under Coding agents; image, video and voice models (Black Forest Labs, Runway, Luma AI, ElevenLabs) under Generative media; computer-use agents (H Company), AI21 Labs and Inflection AI, now focused on enterprise agents and assistants; Krutrim, now an AI cloud; robotics and science models (Physical Intelligence, Skild AI, Periodic Labs, Lila Sciences) under Specialized models. Chinese companies (DeepSeek, Moonshot AI, StepFun) are never ranked. A company appears in one segment only. Method