AMD (AMD.US) Acquires AI Inference Chip Startup Taalas to Complete Data Center Product Matrix

The Zhitong Finance App learned that American chip manufacturer AMD (AMD.US) is further expanding its artificial intelligence (AI) chip layout. As the generative AI wave enters its fourth year, GPUs are still the core hardware for AI training and inference, but the industry is also gradually realizing that a single GPU cannot meet all AI application scenarios. On Thursday, AMD announced that it has reached an agreement to acquire Taalas, an AI inference chip startup headquartered in Toronto, Canada, to further strengthen its competitiveness in the field of inference chips and improve the data center AI product matrix.

AMD did not disclose the amount of this transaction. According to data, Taalas, which was founded in 2023, has raised a total of about US$219 million. The deal is only about seven months since Nvidia (NVDA.US) spent 20 billion US dollars to acquire the assets of the high-performance AI chip design company GroQ at the beginning of this year, reflecting that AI inference chips are becoming a new focus of competition in the industry.

With the rapid spread of generative AI applications, the industry's focus is gradually shifting from model training to model inference, that is, the computational process of AI models responding to user requests in actual business.

Unlike general-purpose computing chips such as GPUs, TaalAs develops a dedicated inference accelerator (ASIC) with hardware customization for a single AI model. The chip is “hardwired” optimized for a specific model during design, so although it is less flexible, it can significantly improve inference efficiency and reduce costs.

According to Taalas, its technology can quickly convert any AI model into a dedicated silicon chip. It only takes about two months from receiving a new AI model to completing hardware implementation.

Currently, Taalas' products mainly support the Llama 3.1 small model under Meta (META.US), and is developing new products suitable for larger models. The chip is manufactured using TSM.US's proven process and incorporates high-speed SRAM storage to reduce inference delays.

The industry believes that this type of dedicated inference chip is particularly suitable for low-latency scenarios, such as AI chatbots, real-time search, autonomous driving, and intelligent assistants, which require quick response.

AMD CEO Su Zifeng said earlier that she has always believed that there is no “one chip for all applications” in the AI chip market.

She pointed out that due to their high versatility, GPUs will still occupy a large share of the AI chip market because they can support a new generation of AI models that continue to evolve. But at the same time, different application scenarios also require more specialized accelerators to jointly build a complete AI computing ecosystem.

Currently, demand for GPUs continues to be strong, driving Nvidia's market capitalization to surpass 5 trillion US dollars and become the listed company with the highest market capitalization in the world.

The acquisition also reflects AMD's shift from simply selling GPUs to providing complete AI computing system solutions.

In recent years, AI data center competition has expanded from a single chip to a complete cabinet system competition. This year, AMD began delivering its first cabinet-level AI system Helios to customers such as Meta and Microsoft (MSFT.US) to compete with Nvidia's DGX system.

AMD said that in the future, Taalas' inference chips and related technologies will be integrated into the company's product roadmap and co-developed with EPYC CPUs, Instinct GPUs and complete machine systems to provide customers with a more complete data center AI infrastructure.

In fact, AMD has continued to improve the AI ecosystem through acquisitions recently. In July of this year, the company announced a partnership with AI chip company Cerebras to integrate its AI accelerator into the AMD system.

Previously, AMD also completed a number of AI-related mergers and acquisitions, including spending US$665 million to acquire AI model developer Silo AI in 2024 and the acquisition of server manufacturer ZT Systems for US$4.9 billion to lay the technical foundation for Helios cabinet products. In addition, the company also acquired a number of AI startups last year, including reasoning software developer MK1.

Analysts believe that as the AI industry gradually moves from model training to large-scale commercial deployment, demand for inference computing is growing rapidly, and AMD will further supplement its dedicated inference chip capabilities through the acquisition of Taalas, which is expected to enhance its comprehensive strength to compete with Nvidia in the enterprise AI infrastructure market.