Baud

Baud (Baud Labs) is a hardware company building an AI accelerator chip intended to make training and serving frontier models far cheaper. Its approach has two parts: an arithmetic representation that removes multiplications from neural networks without loss and compresses them during both forward and backward passes, and an ASIC architecture designed to train and serve that representation. Because an adder grows linearly with bit width while a multiplier grows quadratically, the company argues that trading multiplications for additions fits an order of magnitude more compute and memory into the same die area. It is aimed at businesses and research teams that train or run their own models and want an alternative to incumbent accelerators. To avoid the usual software barrier for new chips, Baud is building a compiler that converts any PyTorch exportable model to its format, with bit exact results in most cases, including popular open source architectures such as Qwen, DeepSeek, GLM, Flux and Wan. The company reports a demo in which a small model trained on its ASIC, emulated on an FPGA, runs inference at more than 1,000 tokens per second on a single FPGA. Pricing is not published. Baud is opening early access to its first cluster in beta, with supported workloads covering pretraining and fine tuning, reinforcement learning post training, and inference, and early users reserve capacity through a login on the site. The chip itself is not yet generally available: the current offering is a beta cluster, and the published performance figure comes from FPGA emulation rather than production silicon.