AMD’s Bold Move: Acquisition of Taalas to Imprint AI Models Directly Onto Silicon

Advanced Micro Devices (AMD) has recently made headlines with its definitive agreement to acquire Taalas, a pioneering AI chip startup based in Toronto. This strategic move is seen as a direct challenge to Nvidia’s stronghold in the artificial intelligence hardware market.

Taalas has revolutionized AI inference by physically etching a single AI model’s weights directly into silicon instead of running models on general-purpose GPUs. The company asserts that this innovative approach eliminates the significant bottleneck of constantly transporting billions of model weights between memory and compute. As a result, the model essentially becomes the processor. Taalas’ chips are reported to operate thousands of times faster than a standard GPU on the specific model they are programmed to run. The HC1, the startup’s debut product, is exclusively designed to run Meta’s Llama 3.1 8B model.

AMD has plans to incorporate Taalas’ groundbreaking technology across its accelerator roadmap. This includes its Instinct GPUs, EPYC processors, Helios rack-scale platform, and ROCm software. Since its inception, Taalas has successfully raised $219 million in venture capital. However, AMD has chosen not to disclose the acquisition price. The deal is projected to be finalized in the fourth quarter of 2026, subject to regulatory approval.

This acquisition marks AMD’s fourth inference-focused deal since November. It also follows Nvidia’s $20 billion acquisition of Groq’s assets by approximately seven months. These developments indicate that the next stage of the AI chip war will likely be waged on the inference battleground, not just training.

Source: CNBC — AMD buys Taalas, startup that hardwires AI models into its silicon