AMD's Helios rack beats Nvidia's Vera Rubin on key metrics, ships this year

AMD is done playing catch-up. At the company’s packed Advancing AI conference in San Francisco on Thursday, CEO Lisa Su unveiled the Helios rack-scale system and made it clear she’s gunning directly for Nvidia’s data center crown. Su called Helios “the highest-performance AI rack” the industry has seen, built to train frontier models at gigawatt scale. The system doesn’t just have swagger — it has receipts. Early performance metrics show Helios beating Nvidia’s Vera Rubin on multiple benchmarks, as reported by The Register.

The customer list tells you this is serious. OpenAI, Meta, Oracle, Anthropic, and Microsoft have all signed on to deploy Helios. Microsoft CEO Satya Nadella confirmed Monday that Azure infrastructure will expand using the system, and Anthropic inked a strategic partnership to deploy up to two gigawatts of GPUs through it. That’s not a pilot program. That’s infrastructure commitment at the scale of small power plants.

Su also previewed the Venice-X CPU, a data center chip targeting the heaviest compute workloads, slated for 2027. But the headline was her market forecast: she expects AI accelerators to hit $1.4 trillion by 2030 — roughly the size of today’s entire semiconductor market. Her reasoning hinges on agentic AI, where a single request spawns dozens of reasoning steps, tool calls, and data lookups. “You need lots of GPUs to do all that,” Su said. Programmable silicon, she argued, still wins because the algorithms are “very much in their infancy.”

Nvidia isn’t standing still — Grace Blackwell and Vera Rubin are formidable platforms with an entrenched software ecosystem. But Helios arriving later this year with performance advantages and a blue-chip customer roster changes the conversation. The question now isn’t whether AMD belongs in the AI infrastructure race. It’s whether Nvidia’s lead is softer than anyone wanted to admit.