AMD acquires Taalas, a startup that embeds AI models directly in silicon

AMD announced on 6 August 2026 a definitive agreement to acquire Taalas, a Toronto-based startup that embeds model weights directly into the silicon of inference chips. The acquisition is still subject to customary closing conditions and regulatory approvals.

Taalas was founded in 2023 by Ljubisa Bajic, former CEO of Tenstorrent and former director at AMD itself. The company has since raised $219 million in venture capital. The acquisition price has not been disclosed. The team will report to Vamsi Boppana, Senior Vice President of AMD's Artificial Intelligence Group.

AMD intends to integrate Taalas's technology into its accelerator roadmap and combine it with existing products such as the Instinct GPUs, EPYC CPUs, the Helios rack-scale solutions and the ROCm software stack.

Model weights fixed in hardware

The core idea behind Taalas's approach is that an AI model does not need to be loaded from memory if the weights are already part of the chip itself. This eliminates a key bottleneck in inference: the constant data transfer between memory and compute units that costs significant time and energy in generic architectures.

Taalas encodes the weights into what the company calls a mask ROM recall fabric. The demo chip, designated HC1 and manufactured on TSMC's N6 process, stores the complete Llama 3.1 8B model across 53 billion transistors. In testing, this chip achieved more than 16,000 tokens per second per user at a power draw of approximately 200 watts.

The downside of this approach is as concrete as the performance gain: each chip is tied to one specific model. Anyone wanting to run a different model needs a different chip. Taalas attempts to keep this limitation practical by reducing the design time for a model-specific chip to around two months. This is possible because only two of the roughly one hundred metal layers need to be modified per design. The current ceiling for a single chip appears to be models of approximately 8 billion parameters.

From stealth to acquisition in six months

Taalas operated largely behind closed doors until February 2026, when the company publicly unveiled its HC1 demo chip for the first time. Six months later, AMD announced the acquisition.

That rapid sequence fits the pattern of the AI inference hardware market: once a startup emerges with credible performance results, strategic interest and acquisition talks tend to follow quickly. Founder Bajic also brought with him a network that likely facilitated the conversations with AMD: he was previously a director at the company before leading Tenstorrent and subsequently founding Taalas.

AMD is not the only major technology company keeping an eye on Taalas's approach. According to reports, Google is working on a comparable method for its Gemini models, similarly bringing model weights closer to the silicon.

Position within AMD's broader AI strategy

For AMD, the acquisition is a way to differentiate itself in the AI inference segment, where it competes with Nvidia as well as a growing number of specialised chip companies. The Instinct GPU line is AMD's primary product for AI training and inference workloads in data centres, but GPUs are inherently general-purpose and therefore architecturally distinct from Taalas's specialised approach.

AMD describes the integration as complementary: Taalas chips would be deployed alongside Instinct GPUs for specific, well-defined inference tasks, while the GPUs remain flexibly deployable for broader workloads. How that combination will look in practice will become clearer once the acquisition is completed.

For the European and Dutch AI scene, this acquisition illustrates a pattern that is also visible here: specialised hardware for AI inference is attracting serious investment, and the boundary between software optimisation and silicon-level modifications is shifting. For founders and investors focused on AI infrastructure, it is worth noting that a startup with a clearly defined technical approach, and a founder with deep sector experience, can move from founding to a strategic exit at one of the world's largest chip companies within three years.