Taalas, a Toronto-based AI chip startup, has raised $169 million to develop specialized processors that embed AI model components directly into silicon, aiming for faster and more cost-efficient inference than general-purpose hardware.

Investors in the round include Quiet Capital, Fidelity, and Pierre Lamond, signaling strong investor confidence in model-embedded silicon.

Taalas operates in a competitive landscape alongside Nvidia, Groq, Cerebras, OpenAI, and D-Matrix, all pursuing SRAM-heavy or model-specific chip designs.

The company plans to release two more models in 2026 as part of its expansion.

Taalas was founded by Ljubisa Bajic (ex-Tenstorrent) with Drago Ignjatovic and Lejla Bajic, building a team of about 25 engineers and appointing Paresh Kharya as vice president of products.

Taalas aims to launch first chips to early customers soon, with a goal of hardcoding a 20-billion-parameter Llama 3.1 model into HC chips by summer and deploying a frontier-class LLM across HC cards under the HC2 architecture by year-end.

Models targeted include a small version of Meta’s Llama, with potential to scale to larger models as development progresses.

Taalas plans to advance toward cutting-edge models while leveraging existing manufacturing partners and a phased customization approach to accelerate deployment.

The funding and product announcements come amid industry activity, including Nvidia’s deal with Meta to supply millions of chips for Meta’s AI infrastructure.

The broader trend toward specialized AI accelerators for inference is evident, though specific financial terms or additional investors aren’t detailed.

Taalas’ first chip, HC1, reportedly achieves over 16,000 tokens per second per user on the Llama3.1-8B model.

Taalas has developed a hard-wired Llama 3.1 8B chip, offered as a chatbot demo and inference API, built by a 24-member team with roughly $30 million spent in under three years.