AMD launches new AI server in direct challenge to Nvidia's AI dominance

Advanced Micro Devices (AMD) CEO Lisa Su told a San Francisco audience on Thursday that AMD’s Helios rack scale system is in full production, setting the chipmaker up to assess the AI data center business that Nvidia has controlled almost by itself.

It is the first time AMD has fielded a complete server cabinet built to go head-to-head with Nvidia’s top rack.

AMD Helios server rack components

Helios packs 72 of AMD’s new Instinct MI455X GPUs and pairs them with the company’s Epyc server CPUs all in a single rack. This configuration puts it up there with Nvidia’s NVL72, which also runs 72 GPUs and draws on the Grace Blackwell and Vera Rubin parts.

Su stated during her keynote at the Advancing AI event that the MI455X was the most powerful GPU on the market, a claim directly aimed at the current leader.

Constellation Research, reporting from the event, said each MI455X carries 432GB of HBM4 memory, moves data at 23.3 TB/s, and holds about 320 billion transistors. A full server rack can get up to 2.9 exaflops of peak FP4 compute, 31 terabytes of HBM4 memory, and 1.7 petabytes per second of memory bandwidth.

AMD pitches against Nvidia’s Vera Rubin

CEO Su explained that the Helios server rack brings a lot of value in addition to its raw power. She said the rack delivers 15% better compute performance than Nvidia’s Vera Rubin, carries 50% more HBM, and returns 30% more tokens per dollar. AMD also pushed its Epyc 9006 CPUs, which Su said offer 20% higher per-core performance than Nvidia’s Vera CPU.

The company is chasing a market Nvidia currently owns. Nvidia’s share of the AI data center space is reported to be at about 80% to 90%.

To close this gap, AMD said it struck a deal with Cerebras to fold the firm’s inferencing chips into its data center lineup, which resembles the Nvidia deal with designer Groq.

AMD bets on inferencing

AMD is betting on inference as the computing workload that hits high levels next. Su told the event that about 60% of compute capacity will go to running models instead of training them, and she pointed to AI agents as the next driver of this change.

AMD’s launch post said monthly token consumption has increased 158 times in two years, and that on the DeepSeek-V4-Flash model the MI455X hits up to 34 times higher token throughput at high interactivity.

Reports claim AMD had already lined up Helios deals with Anthropic and Microsoft, and featured both OpenAI and Anthropic on stage during the keynote. Su said Helios demand is “extremely strong” and put the AI accelerator market at $1.4 trillion by 2030.

AMD stock dipped by more than 2% while Su spoke, according to Yahoo Finance. However, looking at the bigger picture, AMD’s shares are up 222% over the past 12 months, compared to 117% for Nvidia over the same period of time. AMD has trailed its rival for years and only began closing the gap this year.

If you're reading this, you’re already ahead. Stay there with our newsletter.