AMD has entered into a definitive agreement to acquire Taalas, a Toronto-based startup that has been building AI chips designed to run language models with unusually low latency. The announcement was made through a press release on AMD’s investor relations page, and confirms a deal that had been the subject of speculation in chip circles for weeks.
Taalas was founded in 2023 and has built its reputation around a fairly unconventional idea for how AI hardware should work. Rather than designing general-purpose accelerators that can run any model handed to them, the company etches specific model architectures directly into silicon. It calls this approach “Hardcore Models,” and the company’s own pitch line sums it up bluntly: “The Model is The Computer.” Earlier this year, it released ChatJimmy, an app built to demonstrate just how fast a model can respond when the compute and memory bottlenecks of conventional chips are designed away.
That speed is the entire point of the acquisition. AMD’s own release frames AI inference as one of the fastest-growing segments of the AI market, with workloads becoming increasingly specialized by the month. Taalas’s technology is aimed squarely at reducing the compute and memory bottlenecks that come with running inference on general-purpose architectures, and AMD wants that expertise folded into its broader AI stack, alongside AMD Instinct GPUs, EPYC CPUs, the Helios rackscale systems and the ROCm software layer.
Vamsi Boppana, senior vice president of AMD’s Artificial Intelligence Group, said in the release that AMD is building a full-stack AI platform that gives customers the flexibility to deploy the right compute for every workload, and that Taalas’s team and technology strengthen that portfolio on the inference side specifically. Ljubisa Bajic, Taalas’s co-founder and CEO, said the company set out to rethink AI inference from the ground up by building hardware around the model itself, and that joining AMD gives the team the scale and engineering resources to push that idea further.
The timing is worth noting. AMD has been on something of a hardware shopping and partnership spree this year as it tries to close the gap with Nvidia in AI compute. Just weeks ago, AMD’s had signed a multi-billion-dollar chip deal with Anthropic, which saw Anthropic commit to buying up to 2 gigawatts of AMD’s Instinct MI450 chips, with AMD investing up to $5 billion into Anthropic in return. Taalas fits a different part of the same strategy — instead of buying scale through a customer commitment, AMD is buying a specialized inference technology outright and pulling it in-house.
The bet that model-specific silicon is worth owning isn’t unique to AMD or Taalas either. OpenAI built its own inference chip called Jalapeño with Broadcom, purpose-built around the models it actually runs in production rather than general-purpose flexibility. And Google is reportedly developing a chip internally called “Frozen v2” that would bake Gemini’s architecture directly into silicon, following a philosophy strikingly close to what Taalas has been pitching since its stealth debut. When a company with Google’s own decade-plus head start in custom silicon starts moving in the same direction as a two-year-old startup, it tends to validate the approach rather than undercut it.
AMD says it plans to integrate Taalas’s technology into its accelerator roadmap and develop system-level solutions alongside Instinct GPUs, rather than run it as a standalone product line. The release also frames the deal as an extension of AMD’s existing presence in Canada, describing it as a commitment to retaining and growing Canadian AI talent. The acquisition is still subject to customary closing conditions and regulatory approval, and financial terms were not disclosed.