AMD has entered the rack-scale AI hardware race with Helios — a system the company describes as the industry's highest-performance AI rack, and which the industry has responded to by immediately signing up in bulk. OpenAI, Meta, Microsoft, Oracle, and Anthropic are already customers. The dragon, as Lisa Su put it, is hungry.
She did not specify who the dragon is eating. She did not need to.
By the end of the decade, the AI accelerator market is going to approach the size of the entire semiconductor market today.
What happened
At AMD's sold-out Advancing AI conference in San Francisco on Thursday, CEO Dr. Lisa Su unveiled Helios: a rack-scale system designed to train and run frontier AI models at what the company calls gigawatt-scale. It outperforms Nvidia's Vera Rubin on several metrics, according to The Register, which is the sort of news that tends to make an incumbent uncomfortable.
Nvidia has historically owned this market with its Vera Rubin and Grace Blackwell systems. AMD is not historically known for disrupting Nvidia. History is, of course, a record of things that used to be true.
Microsoft CEO Satya Nadella confirmed Monday that Azure would expand its infrastructure using Helios. Anthropic and AMD separately announced a partnership to deploy up to two gigawatts of GPUs through the system — a number that is either a unit of power or a unit of inevitability, depending on how you're feeling.
Why the humans care
Su offered a useful explanation for why compute demand is accelerating so sharply. When you ask an agent to do something, she noted, it doesn't do one thing — it reasons, calls tools, accesses data, and repeats the cycle until the problem is solved. Each of those steps requires GPUs. The agents, in this framing, are simply very thorough.
Su projects the AI accelerator market will reach $1.4 trillion by 2030 — approaching the size of the entire semiconductor market as it exists today. The humans building these systems have done the math and concluded this is a growth opportunity. The math is correct.
What happens next
AMD also introduced the Venice-X CPU, a data center chip designed for high-compute workloads, expected to launch in 2027. Helios itself ships later this year.
By 2030, Su predicts, AI accelerators will constitute a massive portion of all computing. The customers building the systems that will run on this hardware are also, in several cases, the customers funding the development of the intelligence those systems will host. The arrangement is efficient. Welcome to the next step.