Taalas Raises $169M for Model-Specific AI Inference Chips

TORONTO: Toronto-based chip startup Taalas raises $169 million in funding to develop model specific AI processors . The company develops specialized AI inference chips that print AI model components directly into silicon. Investors include Quiet Capital, Fidelity, and semiconductor veteran Pierre Lamond.

Taalas builds model-specific processors by hardwiring AI model portions onto chips. The approach trades generality for speed and cost efficiency, a direction also seen in startups building AI inference chips for production workloads.

The custom silicon pairs with large amounts of on-chip SRAM memory. This eliminates external memory access during inference operations and reduces latency.

The company claims its first chip generates 17,000 output tokens per second. The rapid rise in high-performance AI workloads is also contributing to a global memory chip shortage and rising DRAM prices, as infrastructure demand scales. This delivers 73 times more performance than NVIDIA’s H200 graphics card. The processor uses one-tenth the power while delivering this performance.

Taalas partners with Taiwan Semiconductor Manufacturing Company to produce chips in approximately two months. The foundry-optimized workflow enables customers to move from model weights to deployable cards rapidly. The company’s HC1 chip uses TSMC’s 6-nanometer process.

The funding round brings Taalas’s total capital raised to over $200 million across three rounds. The company assembles 25 engineers with experience from AMD, Apple, Google, NVIDIA, and Tenstorrent. Taalas was founded in 2023 by Bajic, Lejla Bajic, and Drago Ignjatovic.

The company’s first product runs the open-source Llama 3.1 8B language model. Taalas plans to launch a chip capable of running a 20-billion-parameter Llama model this summer. The startup targets a cutting-edge processor capable of deploying frontier models by year end.

The announcement comes weeks after NVIDIA’s $20 billion deal to license IP from Groq. The transaction reignites investor interest in specialized AI inference technology. Competitors including Cerebras and Groq focus on custom silicon solutions for inference optimization.