Taalas

Taalas is a semiconductor company founded in Toronto in 2023 by Ljubisa Bajic, previously the founder and chief executive of Tenstorrent. It builds silicon dedicated to AI inference rather than general-purpose accelerators.

Its approach is to etch model weights directly into the chip. Rather than streaming parameters from high bandwidth memory on every forward pass, a Taalas part holds them in a mask read-only memory fabric, with a separate on-chip memory fabric reserved for key-value caches and fine-tuning adapters. Removing that memory round-trip attacks the bottleneck that dominates the cost and latency of serving a model at scale. The company reported in February 2026 that its HC1 test chip served an 8 billion parameter Llama 3.1 at 16,960 tokens per second, and said a following chip would target models up to 20 billion parameters.

The trade-off is flexibility: a chip built around a particular model is fast precisely because it is not general, which makes it awkward when the model changes. AMD announced a definitive agreement to acquire Taalas on August 6, 2026, without disclosing terms, and said it expected to close by the fourth quarter of 2026 subject to regulatory approval. AMD plans to sell the technology as system-level solutions alongside its Instinct accelerators.