Taalas turns AI models into custom silicon, claims 17K tokens/sec throughput

Finnish startup Taalas developed a platform that transforms any AI model into custom silicon ("Hardcore Models") within two months of receiving a previously unseen model. The company claims its approach delivers an order of magnitude improvement in speed, cost, and energy efficiency over GPU-based inference. The post argues that AI's current reliance on massive data centers mirrors ENIAC-era computing, and that custom silicon could bring the same democratization that transistors brought to general-purpose computing. Taalas was founded 2.5 years ago.

View full digest for February 21, 2026