AMD Acquires Custom Chip Startup for AI Inference…Accelerating Pursuit of Nvidia

AMD Acquires Custom Chip Startup for AI Inference…Accelerating Pursuit of Nvidia

U.S. semiconductor company AMD has decided to acquire Taalas, a Canadian startup that manufactures semiconductors for artificial intelligence (AI) inference.

According to U.S. economic broadcaster CNBC, AMD announced on the 6th (local time) that it has signed a definitive agreement to acquire Toronto-based Taalas, which possesses unique technology to embed AI models directly into silicon chips.

Taalas's accelerators are characterized by custom-designed circuits tailored to a single AI model, unlike general-purpose graphics processing units (GPUs).

Although less versatile, the company stated that they can produce computational results for specific models thousands of times faster than existing GPUs at a lower cost.

The transaction amount was not disclosed. Since its establishment in 2023, Taalas has raised a total of $219 million (approximately 312.1 billion KRW) in venture investments.

This acquisition is compared to rival Nvidia's earlier acquisition of high-performance AI chip designer Groq's assets for $20 billion (approximately 28.5 trillion KRW) in January.

Lisa Su, CEO of AMD, said at a new product launch event last month, "I am confident that there is no one-size-fits-all solution for chips," but added that GPUs will still account for most of the AI chip market because they can flexibly support new AI models.

AMD stated that it plans to incorporate Taalas's technology into its roadmap as a system combining its central processing units (CPUs) and Instinct GPUs.

AMD had previously partnered with Cerebras in July to integrate AI chips into its systems within the year, thereby strengthening its integrated system competitiveness rather than focusing on standalone processor products.

In 2024, AMD acquired AI model developer Silo AI ($665 million) and ZT Systems ($4.9 billion), which provided rack-scale product technology.

Industry analysis suggests that as the center of gravity of the AI industry shifts from model training to the inference stage, interest in alternative chips that can complement GPUs is growing.