AMD Acquires Taalas, Betting Model-Specific Chips That Etch Weights Into Silicon Can Boost AI Inference Speed
AMD is buying Toronto-based Taalas, whose chips etch model weights directly into silicon instead of storing them in memory, to strengthen its AI inference lineup.
Overview
AMD said it has reached a definitive agreement to acquire Taalas, a Toronto-based startup that builds chips etching AI model weights directly into silicon, according to an AMD press release announced August 6, 2026. AMD’s release describes Taalas as “a pioneer in specialized AI inference silicon” and says the deal is meant to strengthen “AMD’s long-term AI roadmap with differentiated inference technology and world-class engineering expertise.” The Register reported the deal was announced at market close on Thursday and that AMD “didn’t disclose the terms of the deal.”
What We Know
AMD’s release states it plainly: “Founded in 2023 and headquartered in Toronto, Canada, Taalas’ technology optimizes inference dataflows, significantly reducing compute and memory bottlenecks associated with general-purpose architectures.” The company’s chips take a different approach than conventional accelerators: rather than storing model weights in high-bandwidth memory, Taalas etches them directly into the chip itself. The Register describes the resulting hardware as “model-specific integrated circuits or MSICs,” built around “the mask-ROM recall fabric where model weights are etched, and the SRAM recall fabric where KV caches and fine-tuning adapters are stored.”
Taalas’s approach is not purely conceptual. Per The Register, the startup “revealed its first test chip fabbed on TSMC’s 6nm process tech” in February 2026, calling it the HC1. At the time, the outlet reported, the chip served Meta’s Llama 3.1 8B model at “16,960 tokens a second” — a figure Taalas said was “48x faster than Nvidia’s GPUs and 8.5x faster than Cerebras’ accelerators” when the company first announced those results. Taalas’s planned second-generation chip, the HC2, is “due out this summer” and targets 20 billion parameters, according to The Register; at that density, the outlet noted, “you’d need just 50 accelerators to support a trillion-parameter model.”
AMD’s release says Taalas’s technology “will complement AMD’s full-stack AI platform, including AMD Helios rackscale solutions, AMD Instinct GPUs, AMD EPYC CPUs, AMD ROCm software and the company’s expanding AI ecosystem,” and that AMD “plans to integrate the technology into its accelerator roadmap and develop system-level solutions with AMD Instinct GPUs.” The Register reported that AMD intends to pair Instinct-based Helios racks with Taalas-derived chips in what it described as “a disaggregated architecture where compute-heavy prompt processing is done on GPUs while token generation is offloaded to Taalas-based accelerators.”
AMD quoted two executives on the deal. “AMD is building a full-stack AI platform that gives customers the flexibility to deploy the right compute solutions for every AI workload,” said Vamsi Boppana, senior vice president of the Artificial Intelligence Group at AMD, adding that “Taalas’ technology and world-class engineering team strengthen our AI portfolio by delivering differentiated inference performance and efficiency.” Ljubisa Bajic, co-founder and CEO of Taalas, said: “We founded Taalas to rethink AI inference from the ground up by building the hardware around the model. Our Canada-based team has combined deep technical expertise with a willingness to challenge conventional approaches. Joining AMD will give us the scale, engineering resources and global reach to accelerate our innovation.”
The acquisition, AMD said, “is subject to customary closing conditions and regulatory approvals.” The Register reported the deal “is expected to close in the fourth quarter.” The Register also noted that OpenAI, Anthropic, and Meta “are all major Instinct customers” of AMD already.
What We Don’t Know
Neither AMD nor The Register disclosed the acquisition’s purchase price. The Register reported that, based on its understanding, “this is an actual acquisition rather than an acquihire,” but did not specify what will happen to Taalas as an organization beyond folding into AMD’s AI roadmap. AMD’s release does not name a specific closing date beyond the fourth-quarter timeframe reported by The Register, nor does it independently confirm the HC1’s February performance figures or Taalas’s claim, relayed by The Register from a February interview with sibling outlet The Next Platform, that etching a model’s weights into silicon is “100x less expensive than training a frontier model” — a claim attributed to Taalas itself rather than independently verified.
Why It Matters
The deal comes as AMD works to compete with Nvidia in AI inference hardware. The Register drew a comparison to Nvidia’s roughly concurrent push into specialized inference infrastructure, noting that Nvidia’s recently unveiled LPX systems would need “a few dozen GPUs and at least 2,000 Groq LPUs” to serve a trillion-parameter model — a workload The Register said Taalas-based chips could theoretically handle with “50 accelerators.” The tradeoff, per The Register, is inflexibility: “Once the chips are deployed you’re stuck with that model. Any change bigger than something like a LoRA adapter is going to require a re-spin of the chips, which is not only expensive but time-consuming.” The outlet added that Taalas has said such a re-spin doesn’t mean starting over, since “just two layers of metal need to be changed.”