The Startup Flipping Nvidia’s Playbook on its Head

And now for something completely different, as British comedy troupe Monty Python used to say. It’s no secret in the AI world that Nvidia’s success in selling AI chips lies partly in the chips’ flexibility. Developers can (relatively) easily tweak the chips to run different types of AI models. Taalas, a new AI chip developer, wants to go in the opposite direction.

Taalas is developing the exact opposite of customizable chips: rigid chips which are each specialized for a different AI model, whether it's Meta Platforms' Llama models or Stable Diffusion. The Toronto-based startup, which was founded in August last year by former Nvidia and AMD veteran Ljubisa Bajic, raised $12 million in September and $38 million in February from Quiet Capital and Pierre Lamond, an advisor at Eclipse Ventures who was previously a general partner at Khosla Ventures and Sequoia Capital.

With Taalas, Bajic hopes to tackle one of the most pressing obstacles in AI today: the sky-high costs of hardware. By creating highly specialized chips that require little additional software to work, Bajic is preparing for a world where AI models exist in many of our everyday devices, from fridges to cars—many of which may not even be connected to the internet. Those devices will need cheap, lightweight hardware rather than the powerful, highly-programmable chips Nvidia and others produce.

Bajic knows a thing or two about AI hardware—after stints at AMD and Nvidia, he founded AI chip developer Tenstorrent, which went on to raise $385 million from investors like Fidelity and Eclipse Ventures before Bajic left last March. (Bajic made way for then-CTO Jim Keller, who Bajic describes as better at managing large teams. You can check out Tenstorrent and other AI hardware developers in our database here.)