A Toronto-based technology company announced Thursday it has secured $169 million in investment funding while unveiling a new processor designed to run artificial intelligence programs more efficiently and affordably than existing solutions.
The funding announcement from Taalas comes just weeks following Nvidia’s significant $20 billion Christmas Eve agreement to acquire intellectual property rights from competing chip developer Groq, a deal that has renewed investor attention toward emerging companies developing specialized AI inference technology – the systems that allow AI programs like ChatGPT to process and respond to user questions.
The company’s innovative manufacturing method involves embedding specific AI model components directly into silicon wafers, creating processors tailored for particular applications such as smaller versions of Meta’s Llama system. These specialized processors incorporate substantial amounts of high-speed SRAM memory directly on the chip, an approach that mirrors Groq’s design philosophy.
However, according to company officials, the custom engineering for individual AI models provides Taalas with its competitive edge.
“This hard wiring is partly what gives us the speed,” CEO Ljubisa Bajic told Reuters in an interview.
According to Bajic, the manufacturing process involves creating a nearly finished processor with approximately 100 layers, then completing customization work on the final two metal layers. Working with TSMC for production, the company can complete a model-specific chip in roughly two months, he explained.
By comparison, manufacturing an AI processor like Nvidia’s Blackwell requires approximately six months for completion.
Company representatives say they can currently manufacture chips suitable for less complex AI models, with plans to produce processors capable of running advanced systems like GPT 5.2 before year’s end.
Several other startups including Groq, Cerebras – which recently signed a cloud computing partnership with OpenAI in January – and D-Matrix have adopted similar SRAM-focused design strategies for their first-generation processors.
