Taalas raises $169 million to build custom AI chips that hardwire models into silicon

Taalas Secures Major Funding to Challenge Nvidia in AI Chip Market

Toronto-based chip startup Taalas announced it raised $169 million in a funding round that brings its total capital to $219 million1

1

. Investors include Quiet Capital, Fidelity, and prominent semiconductor venture capitalist Pierre Lamond1

1

. The startup's approach centers on hardwiring AI models into silicon to deliver faster and cheaper AI applications, positioning itself to take on Nvidia in the rapidly evolving AI inference market.

Revolutionary Approach to Model-Specific AI Chips

The chip startup Taalas has developed a distinctive method for custom-designing chips that involves printing portions of an AI model directly onto silicon. This creates bespoke processors tailored for specific models such as Meta's Llama 3.1 8B language model1

1

. CEO Ljubisa Bajic explained that "this hardwiring is partly what gives us the speed"4

4

. The customized silicon pairs with large amounts of on-chip SRAM, similar to designs from Groq and Cerebras.

Technical Innovation Through Mask ROM Recall Fabric

Taalas assembles a nearly complete chip with roughly 100 layers, then performs final customization on just two metal layers1

1

. These custom layers feature what the company describes as a mask ROM recall fabric, where each module can store four bits using only a single transistor to process matrix multiplications—the mathematical calculations that LLMs use to make decisions2

2

. This design eliminates the need for HBM modules, avoiding the delays caused by moving information to and from external memory while removing various auxiliary components2

2

.

Dramatic Performance Gains Over Current Solutions

The company's first product, called HC1, delivers remarkable results for AI inference tasks. Taalas claims its chip can generate 17,000 tokens per second (TPS) when running the Llama 3.1 8B model—73 times more than Nvidia's H200 graphics card while consuming one-tenth the power2

2

. According to performance data, Taalas delivers 10 times the TPS of today's high-end infrastructure while achieving 20 times lower production costs3

3

. The HC1 chip uses TSMC's 6nm node and features a chip size up to 815 mm², nearly matching Nvidia's H100 dimensions3

3

.

Faster Manufacturing Timeline Creates Competitive Edge

Using TSMC for manufacturing, Taalas can complete fabrication of a chip customized for a particular model in approximately two months1

1

. This contrasts sharply with the roughly six-month timeline required to fabricate an AI processor such as Nvidia's Blackwell1

1

. The company's platform can transform any AI model into custom silicon, creating what it calls "Hardcore Models" that are an order of magnitude faster, cheaper, and lower power than software-based implementations3

3

.

Addressing Response Latency in Agentic AI Environments

Taalas founded its approach 2.5 years ago around solving response latency, which has emerged as a massive constraint for modern compute providers3

3

. In agentic environments, the primary competitive advantage lies in tokens per second figures and task completion speed. The company's solution focuses on specialization of AI workloads at the hardware level and merging storage and computation to overcome memory walls3

3

. All computation happens at DRAM-level density to ensure faster intercommunication, avoiding the complex cooling, HBM integration, and packaging required by conventional ASICs3

3

.

Ambitious Roadmap Targets Frontier Models

Taalas can currently produce chips capable of running less sophisticated models and is developing a new chip that can run a Llama model with 20 billion parameters, expected to be ready this summer2

2

. The company has plans to build a processor called HC2 capable of deploying cutting-edge models such as GPT-5.2 by the end of this year1

1

. To address scaling challenges, Taalas has adopted a cluster-based approach, demonstrating 12,000 TPS per user with DeepSeek's R1 in a 30-chip configuration3

3

.

Market Context and Competitive Landscape

The announcement arrives weeks after Nvidia's deal to license intellectual property from chip startup Groq for $20 billion, which reignited interest in startups focused on AI inference—the process where an AI model responds to user queries1

1

. Groq's first generation processor used an SRAM-heavy approach to chip design, as does Cerebras, which signed a cloud computing deal with OpenAI in January, and d-Matrix1

1

. However, Taalas differentiates itself by pivoting away from general-purpose computing entirely toward model-specific implementations, though this means hardware remains specific to certain LLMs without the option to change model weights3

3

.