“This is the most powerful GPU on the market,” AMD CEO Lisa Su declared at a San Francisco event as the company revealed its Helios rack-scale AI server, ready to challenge Nvidia’s grip on AI data centers. Packing 72 of AMD’s new Instinct MI455X GPUs alongside Epyc CPUs, the Helios system is AMD’s first full cabinet built to directly compete with Nvidia’s NVL72 rack, which currently dominates the market.
The Helios rack offers impressive specs: each MI455X GPU has 432GB of HBM4 memory, 23.3 terabytes per second of data transfer, and around 320 billion transistors, culminating in a rack that can deliver up to 2.9 exaflops of FP4 compute power. AMD claims this translates to 15% better compute performance compared to Nvidia’s Vera Rubin server, along with 50% more HBM memory and a 30% increase in tokens per dollar. On the CPU front, AMD’s Epyc 9006 processors reportedly outperform Nvidia’s Vera CPUs by 20% per core.
The AI data center space remains largely Nvidia’s domain, with the company controlling an estimated 80% to 90% share. AMD’s strategy to chip away at that dominance includes the Helios server and a partnership with Cerebras to incorporate its high-performance inferencing chips into AMD’s offerings, similar to Nvidia’s collaboration with Groq. This move emphasizes AMD’s focus on inference workloads, which are expected to surge in demand alongside training tasks.
By rolling out Helios into production now, AMD positions itself for a direct market test against Nvidia’s entrenched presence. Early customers like Microsoft and Anthropic have already signed on, hinting at potential shifts in AI infrastructure choices. AMD’s bold entry into AI servers signals growing competition in a sector critical to the future of artificial intelligence computing.
