Taalas

Inference-chip startup that bakes one model's weights directly into chip circuitry so the model can be served cheaply. AMD confirmed its purchase in October 2026 to add model-specific silicon to the Instinct roadmap, mirroring Nvidia's $20 billion purchase of Groq in December 2025.