Summary
As of 2026: four frontier US labs (OpenAI, Anthropic, Google DeepMind, xAI), one Chinese frontier lab (DeepSeek), a strong Meta-led open-weights ecosystem (Llama), a growing European effort (Mistral), and hundreds of applied companies. The frontier is measured in dollars of pretraining and quality of post-training; the middle is measured in product.
Objectives
- 01Name the frontier labs and their signature strengths.
- 02Distinguish open weights from open source.
- 03Explain the strategic role of inference-time compute.
The Lesson
The frontier
OpenAI (product-first, reasoning models), Anthropic (safety-first, Constitutional AI, strong Claude series), Google DeepMind (multimodal-first, Gemini + AlphaFold), xAI (compute-scale, Grok), DeepSeek (open frontier, MoE).
The rest of the stack
Open weights: Llama, Mistral, Qwen, DeepSeek. Inference: Groq, Cerebras, SambaNova, Together, Fireworks. Applied: thousands of firms shipping on top.
Key Ideas
- The frontier is a small number of labs; the applied layer is enormous.
- Open weights ≠ open source; the training data is rarely released.