An AMD worker installs a the primary Helios rack-scale AI system in a knowledge middle lab in Rockdale, Texas, on June 24, 2026. Helios is available in 4 customizable configurations, and that is the one Meta will deploy later this 12 months.
Andrew Evers | CNBC
Superior Micro Units is relying on its graphics processing items to drive the majority of its information middle progress as cloud corporations snap up all of the superior AI chips they’ll discover.
However because the generative synthetic intelligence increase approaches its fourth anniversary, it is turning into clear that GPUs do not do all the pieces.
On Thursday, AMD mentioned it is entered into an settlement to accumulate Taalas, a Toronto-based startup that makes chips for inference. Taalas’ accelerators are custom-made, or hard-wired for a single AI mannequin, relatively than being basic function.
In change for that lack of flexibility, Taalas’ know-how guarantees a less-expensive chip that it says can produce output for particular fashions hundreds of occasions sooner than a standard GPU. An AMD consultant declined to supply a purchase order worth for the transaction. Taalas has raised a complete of $219 million in enterprise funding since its 2023 founding.
The deal comes a little bit over seven months after Nvidia spent $20 billion shopping for property from Groq, a designer of high-performance AI chips. It’s Nvidia’s largest transaction on file.
Taalas’ present chip runs a small model of Meta’s Llama 3.1 mannequin, although the corporate is engaged on chips for larger and extra superior fashions. It is manufactured utilizing an older Taiwan Semiconductor Manufacturing Co. course of, and makes use of speedy SRAM reminiscence on the chip itself.
Taalas CEO Ljubisa Bajic says on the startup’s web site that the corporate “developed a platform for remodeling any AI mannequin into customized silicon.”
“From the second a beforehand unseen mannequin is obtained, it may be realized in {hardware} in solely two months,” Bajic wrote.
Various chips like these from Taalas and Groq are notably vital for “low-latency” purposes, the place time to first response from an AI mannequin is vital.
“I am a giant believer that there isn’t any one-size-fits-all because it involves chips,” AMD CEO Lisa Su mentioned at a product launch in July. She added that AMD nonetheless expects GPUs to make up nearly all of the AI chip market as a result of they’re versatile sufficient to help newly developed AI fashions.
Demand for GPUs has turned Nvidia into the world’s most respected firm with a market cap of over $5 trillion.
The acquisition additionally displays the rising significance for main GPU makers to supply built-in methods with a number of totally different parts and chips as a substitute of simply processors. AMD just lately began to ship Helios, its first rack-scale rival to Nvidia’s built-in server racks, to prospects together with Meta and Microsoft.
AMD mentioned it will combine the Taalas chips and know-how into its street map, together with in methods with its central processors and Intuition GPUs. It isn’t the one complementary accelerator that AMD is supporting: In July, AMD introduced a partnership with Cerebras to combine its AI chips into its methods later this 12 months.
AMD has been on a shopping for spree to fill out among the parts and applied sciences it must construct its Helios racks. In 2024, it paid $665 million for Silo AI, which develops AI fashions, and bought ZT Methods, which offered the technical foundation of its rack-scale merchandise, for $4.9 billion. Final 12 months, AMD purchased a number of smaller AI corporations, together with MK1, which made software program for inference.
WATCH: CNBC’s full interview with AMD CEO Lisa Su
