Advanced Micro Devices (AMD) CEO Lisa Su announced at the Advancing AI event in San Francisco on Thursday that the company's Helios rack-scale system is now in full production, positioning the chipmaker to compete for the AI data center business that Nvidia has largely dominated.

The launch marks the first time AMD has produced a complete server cabinet designed to go head-to-head with Nvidia's flagship rack offerings. Rack-scale systems like Helios and Nvidia's NVL72 represent a growing trend in the AI hardware market, where vendors deliver fully integrated cabinets rather than individual components, simplifying deployment for cloud providers and enterprises racing to scale AI infrastructure.

Helios Server Rack Components

Helios integrates 72 of AMD's new Instinct MI455X GPUs alongside the company's Epyc server CPUs within a single rack. This configuration is comparable to Nvidia's NVL72, which similarly houses 72 GPUs based on Grace Blackwell and Vera Rubin components.

During her keynote, Su described the MI455X as the most powerful GPU available on the market — a claim squarely directed at the current industry leader.

According to Constellation Research, reporting from the event, each MI455X features 432GB of HBM4 memory, data transfer speeds of 23.3 TB/s, and approximately 320 billion transistors. A fully configured Helios rack can achieve up to 2.9 exaflops of peak FP4 compute, 31 terabytes of HBM4 memory, and 1.7 petabytes per second of memory bandwidth. HBM4, the latest generation of high-bandwidth memory, is critical for AI workloads because training and running large models requires moving massive amounts of data between processors and memory, making memory bandwidth as important as raw compute in determining real-world system performance.

Competitive Positioning Against Nvidia's Vera Rubin

Su emphasized that the Helios rack delivers advantages beyond raw compute power. She stated it provides 15% better compute performance than Nvidia's Vera Rubin, carries 50% more HBM memory, and returns 30% more tokens per dollar. AMD also highlighted its Epyc 9006 CPUs, which Su said deliver 20% higher per-core performance compared to Nvidia's Vera CPU.

AMD is pursuing a market that Nvidia currently commands, with the latter holding an estimated 80% to 90% share of the AI data center space. Beyond hardware specs, Nvidia's dominance has been reinforced by its CUDA software ecosystem, which developers have relied on for years and which AMD's ROCm platform continues to challenge. To help close this gap, AMD announced a partnership with Cerebras to integrate the firm's inferencing chips into its data center lineup — a move that mirrors Nvidia's own arrangement with chip designer Groq. Both Cerebras and Groq specialize in inference-focused processors that take different architectural approaches to running AI models, reflecting the industry's growing emphasis on cost-efficient deployment.

AMD's Bet on Inferencing

AMD is positioning inference as the next major computing workload. Su told attendees that approximately 60% of compute capacity will be dedicated to running models rather than training them, citing AI agents as a key catalyst for this shift. This reflects the industry's evolution from one-time model training to ongoing inference costs that scale directly with user adoption and application deployment.

According to AMD's launch announcement, monthly token consumption has surged 158-fold over the past two years. On the DeepSeek-V4-Flash model, the MI455X reportedly achieves up to 34 times higher token throughput at high interactivity levels.

Reports indicate that AMD had already secured Helios deals with Anthropic and Microsoft, and the company featured both OpenAI and Anthropic on stage during the keynote. Su described Helios demand as "extremely strong" and estimated the AI accelerator market at $1.4 trillion by 2030.

AMD shares declined more than 2% during Su's presentation, according to Yahoo Finance. Over the broader 12-month period, however, AMD's stock has gained 222%, compared to Nvidia's 117% advance over the same timeframe. AMD has lagged behind its rival for years and only began narrowing the gap this year.