AMD to Acquire AI Inference Chip Startup Taalas
AMD said Thursday that it had agreed to acquire Taalas, a Toronto-based startup that develops AI inference chips. Taalas’ chips are designed specifically for inference tasks and can be customized or hardwired for a single AI model rather than serving as general-purpose processors. The trade-off for giving up generality is lower chip costs, according to Taalas. The company says its chips can produce outputs thousands of times faster than traditional GPUs for a specific model. AMD did not disclose the purchase price. Since its founding in 2023, Taalas has raised a total of $219 million in venture capital. The deal comes more than seven months after Nvidia spent $20 billion to acquire assets related to Groq, a high-performance AI chip designer. That was Nvidia’s largest transaction ever. Taalas’ current chips can run a smaller-parameter version of Meta’s Llama 3.1 model, while the company is developing chips for larger and more advanced models. The chips are built using an older manufacturing process and include high-speed SRAM memory on the chip. Taalas CEO Ljubisa Bajic said on the company’s website that it had “built a platform that can turn any AI model into custom silicon.” Bajic wrote: “From the moment we receive a completely new and unknown model, it takes only two months to implement it in hardware.” Alternative chips from companies such as Taalas and Groq are particularly important for low-latency applications, where the time it takes an AI model to produce its first response is critical. At a product launch in July, AMD CEO Lisa Su said: “I firmly believe that there is no one-size-fits-all solution in chips.” She added that AMD still expects GPUs to hold the majority of the AI chip market because their flexibility allows them to support continually emerging AI models. The acquisition also reflects a broader trend: leading GPU suppliers increasingly need to offer integrated systems made up of multiple components and different types of chips, rather than selling processors alone. AMD recently began delivering Helios, its first rack-scale product, to Meta and other customers. Helios is positioned as a rival to Nvidia’s integrated server racks. AMD said it will incorporate Taalas’ chips and technology into its product roadmap and use them in complete systems that include AMD central processors and Instinct-series GPUs. To fill out the components and technology needed to build Helios systems, AMD has pursued a series of acquisitions. In 2024, AMD spent $665 million to acquire Silo AI, an AI model-development company, and $4.9 billion to acquire ZT Systems, which provided the technology foundation for AMD’s rack-scale products. In 2025, AMD also acquired several smaller AI companies, including MK1, which develops inference software.
The trade-off for giving up generality is lower chip costs, according to Taalas. The company says its chips can produce outputs thousands of times faster than traditional GPUs for a specific model. AMD did not disclose the purchase price. Since its founding in 2023, Taalas has raised a total of $219 million in venture capital.
The deal comes more than seven months after Nvidia spent $20 billion to acquire assets related to Groq, a high-performance AI chip designer. That was Nvidia’s largest transaction ever.
Taalas’ current chips can run a smaller-parameter version of Meta’s Llama 3.1 model, while the company is developing chips for larger and more advanced models. The chips are built using an older manufacturing process and include high-speed SRAM memory on the chip.
Taalas CEO Ljubisa Bajic said on the company’s website that it had “built a platform that can turn any AI model into custom silicon.” Bajic wrote: “From the moment we receive a completely new and unknown model, it takes only two months to implement it in hardware.”
Alternative chips from companies such as Taalas and Groq are particularly important for low-latency applications, where the time it takes an AI model to produce its first response is critical.
At a product launch in July, AMD CEO Lisa Su said: “I firmly believe that there is no one-size-fits-all solution in chips.” She added that AMD still expects GPUs to hold the majority of the AI chip market because their flexibility allows them to support continually emerging AI models.
The acquisition also reflects a broader trend: leading GPU suppliers increasingly need to offer integrated systems made up of multiple components and different types of chips, rather than selling processors alone. AMD recently began delivering Helios, its first rack-scale product, to Meta and other customers. Helios is positioned as a rival to Nvidia’s integrated server racks.
AMD said it will incorporate Taalas’ chips and technology into its product roadmap and use them in complete systems that include AMD central processors and Instinct-series GPUs.
To fill out the components and technology needed to build Helios systems, AMD has pursued a series of acquisitions. In 2024, AMD spent $665 million to acquire Silo AI, an AI model-development company, and $4.9 billion to acquire ZT Systems, which provided the technology foundation for AMD’s rack-scale products. In 2025, AMD also acquired several smaller AI companies, including MK1, which develops inference software.