The companies training the largest models.
| Company | What it makes |
|---|---|
| OpenAI | ChatGPT, and the GPT models underneath it |
| Anthropic | Claude, and Claude Code for programming |
| Google DeepMind | Gemini, built into Google’s products |
| Meta | The Muse models. Meta was the major open-weight lab in the West until it made its best models closed |
| xAI | Grok, built into X. Now a SpaceX subsidiary |
| Mistral | French lab, open-weight. The main European player |
| DeepSeek | Chinese lab, open-weight, near the frontier at a fraction of the price |
| Alibaba Qwen | Chinese lab, open-weight, including its top-tier model |
| Z.ai | Chinese lab, first pure-play model company to go public |
| Moonshot | Chinese lab. Its Kimi model holds about a book’s worth of text in context |
| Cohere | Enterprise-only: sells to companies, no consumer chatbot |
| AI21 | Israeli lab. Uses a hybrid architecture rather than a pure transformer |
| Reka | Multimodal: video, images, audio, and text in a single model |
| Ai2 | Non-profit. Publishes training data alongside weights, which almost nobody does |
Four of those labs are Chinese, three of them open-weight. As of September 2026, the best open-weight models (GLM-5.3 and Kimi K3) score 44 on the Artificial Analysis Intelligence Index against 53 for Claude Fable 5.1 and GPT-6 Astra.
| Company | What it makes |
|---|---|
| Cursor | Code editor with a model built in. Acquired by SpaceX |
| Cognition | Devin, an agent that takes a ticket and works it end to end |
| Replit | Browser-based development, no local setup at all |
| Lovable | Describe an app in a sentence, get a deployed one. “Vibe coding” |
| Vercel | Web hosting, plus v0, which generates interfaces from a description |
| Modal | Rent a GPU by the second instead of the month |
| Parallel | Search and retrieval APIs built for agents rather than people |
| Browserbase | Hosted browsers an agent can click through |
| Magic.dev | Long context, aimed at reading whole codebases |
| StackBlitz | Bolt.new: prompt to running app in the browser |
| Company | What it makes |
|---|---|
| NVIDIA | The GPUs, plus CUDA, the software layer that locked the industry in |
| Databricks | Where enterprises store and prepare training data |
| CoreWeave | A neocloud: rents GPU clusters, unlike general-purpose clouds |
| Cerebras | Wafer-scale chips: one dinner-plate-sized chip replaces a rack |
| Groq | Custom inference chips built for low latency |
| SambaNova | Inference hardware deployed on-premises, inside a company’s own walls |
| Fireworks | Fast, cheap inference for open-weight models |
| Together | GPU clusters plus inference for open-source models |
| Baseten | Takes a model you trained and keeps it serving production traffic |
| Lambda | GPU cloud and workstations for ML engineers |
| Hugging Face | The public repository for open models and datasets. NVIDIA is acquiring it |
| Scale AI | Human-labeled training data and evaluation |
| LangChain | Framework for chaining model calls into agents |
| Company | What it makes |
|---|---|
| Midjourney | Image generation. Never raised outside money, which is rare for a lab this size |
| Black Forest Labs | FLUX, open-weight image models other products build on |
| Runway | Video generation, moving toward world models: systems that simulate how scenes evolve rather than just producing frames |
| Luma AI | Video with stronger spatial and temporal consistency |
| Synthesia | Avatar presenters reading a script. Corporate training and comms |
| ElevenLabs | Speech synthesis, dubbing, and voices for phone agents |
| Deepgram | Speech-to-text APIs for developers |
| Suno | Full songs, vocals and instrumentation, from a prompt |
| Company | What it makes |
|---|---|
| Perplexity | Answer engine: synthesizes a response instead of ranking links |
| Glean | Enterprise search across a company’s internal tools |
| Company | What it makes |
|---|---|
| Waymo | Driverless robotaxis, commercially operating today |
| Figure AI | Humanoid robots on BMW assembly lines |
| Skild AI | One foundation model controlling many different robot bodies |
| Physical Intelligence | Robot foundation models for ordinary manipulation tasks |
| 1X | NEO, a humanoid intended for homes |
| Wayve | End-to-end driving software licensed to carmakers |
| Nuro | Autonomy stack licensed to manufacturers |
| Company | What it makes |
|---|---|
| Anduril | Autonomous systems sold as products, not cost-plus contracts |
| Shield AI | Autonomy software that has flown F-16s |
| Applied Intuition | Simulation and validation for autonomous vehicle programs |
| Company | What it makes |
|---|---|
| Isomorphic Labs | DeepMind spinout applying AlphaFold’s approach to drug design |
| Abridge | Ambient scribe: listens to a visit, produces the clinical note |
| Recursion | Automated wet lab generating biological data to train on |
| Company | What it makes |
|---|---|
| Harvey | Legal work: contracts, diligence, research |
| Sierra | Customer service agents that resolve issues end to end |
| Decagon | Support agents across chat and voice |