AMD Acquires Taalas to Add Custom AI Inference Chips

AMD expands beyond GPUs with Taalas acquisition

Advanced Micro Devices announced Thursday it will acquire Taalas, a Toronto-based startup developing specialized chips for AI inference. The move signals AMD's recognition that general-purpose graphics processing units alone won't capture the full spectrum of AI computing demand.

Taalas designs accelerators that are hard-wired for individual AI models rather than built for general flexibility. The company claims its technology can produce output for specific models thousands of times faster than traditional GPUs, though at the cost of versatility. AMD declined to disclose the purchase price. Taalas has raised $219 million in venture funding since launching in 2023, according to CNBC, which first reported the acquisition details.

The startup's current chip runs a smaller version of Meta's Llama 3.1 model and uses an older TSMC manufacturing process with on-chip SRAM memory. Taalas CEO Ljubisa Bajic has stated the company can transform any AI model into custom silicon within two months of receiving it.

Why it matters

The Taalas deal underscores a broader industry shift toward heterogeneous AI infrastructure. While GPUs remain dominant for training and running new models, specialized chips optimized for inference—particularly in low-latency applications where response time is critical—are carving out significant market niches. For AMD, which trails Nvidia substantially in AI chip revenue, acquiring complementary technologies offers a path to differentiate its data center offerings beyond competing directly on GPU performance alone.

Building integrated systems

The acquisition follows Nvidia's $20 billion purchase of Groq assets seven months earlier, marking the graphics giant's largest transaction to date. Both deals reflect the rising importance of offering complete, integrated systems rather than standalone processors.

AMD CEO Lisa Su addressed this strategy at a July product launch, saying she's "a big believer that there's no one-size-fits-all as it comes to chips." Su noted that AMD still expects GPUs to dominate the AI chip market due to their flexibility with newly developed models, but acknowledged the need for specialized alternatives.

AMD recently began shipping Helios, its first rack-scale system designed to compete with Nvidia's integrated server racks, to customers including Meta and Microsoft. The company plans to integrate Taalas chips and technology into its roadmap alongside its central processors and Instinct GPUs.

Acquisition strategy accelerates

Taalas represents the latest in a series of strategic purchases aimed at building out AMD's AI infrastructure capabilities. In 2024, AMD paid $665 million for Silo AI, an AI model developer, and spent $4.9 billion on ZT Systems, which provided the technical foundation for its rack-scale products. Last year's acquisitions included MK1, which made inference software, along with several other smaller AI companies.

AMD also announced a partnership with Cerebras in July to integrate that company's AI chips into AMD systems later this year, further expanding its portfolio of specialized accelerators beyond the Taalas technology.

The details were first reported by CNBC.