AI chips for ultra-fast training and inference
HQ: San Francisco, CA, United States | Founded: 2026 | Employees: 3 | Stage: Seed (Y Combinator S26) | Website: https://baudlabs.ai
Baud is building AI hardware for training, fine-tuning and running frontier models, claiming large gains over existing chips. Its core idea is a new way of representing numbers in neural networks that involves no multiplication: chip designs drop multiplier circuits from both training and inference, and model weights shrink more than 10x without losing capability. Because the cores need no multipliers they are far smaller and simpler than GPU tensor cores or the processing elements inside TPU-style chips, so more computing power and fast memory fits on each piece of silicon, and more weights move for the same memory bandwidth. The stated catch is that models must be trained in this representation, or start from a base model already using it, which is why Baud targets training rather than acting purely as an inference accelerator, unlike Cerebras, Groq, Etched and D-Matrix, which it argues address only the silicon and still run existing weights. The first chip is validated on the Global Foundries 12nm process and on schedule to be manufactured by the end of 2026, and the training and inference service is already live on a cluster of FPGAs emulating the design. Founded in 2026 in San Francisco with 3 employees. Chief executive Sarang Zambare previously led machine learning for Peloton Guide from concept to more than 100,000 devices shipped; chief hardware architect Eric Taylor has more than a decade of chip design experience including four manufactured designs and two patents, with previous roles at NVIDIA and Enfabrica.
Baud is the highest-risk, highest-ceiling company in this set. Its claim is not a faster chip but a different mathematical representation of neural networks that involves no multiplication at all, plus more than 10x smaller model weights without losing capability. If that holds, the consequences are large: chip cores become far smaller and simpler than GPU tensor cores or the arrays inside Google's TPUs, so more computing and memory fits on each chip. The constraint is honest and self-disclosed: because the representation has to be trained into the model, this is not a drop-in replacement, so customers must be training a base model in Baud's format — a much harder sale than pure inference acceleration. Chip startups also fail on money and time.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki