At the “Advancing AI” conference on June 12, 2025 in San Jose, AMD CEO Lisa Su officially unveiled Helios, the company’s next-gen AI server. Powered by 72 MI400-series chips, Helios is slated for release in 2026 and aims to compete head-to-head with Nvidia’s NVL72 architecture
🔍 5 Key Highlights of Helios
1. Massive GPU Density
Helios will feature 72 MI400 GPUs per rack, directly rivaling Nvidia’s NVL72-based systems in both scale and compute density.
2. Open Networking Standards
Unlike Nvidia’s proprietary NVLink tech, Helios is built on open networking standards, making it interoperable with other vendors like Intel and improving flexibility.
3. Major Industry Adoption
AMD confirmed that OpenAI, along with Meta, Oracle, xAI, and Crusoe, will adopt the MI400 chips and Helios architecture—with OpenAI’s Sam Altman on stage and Crusoe committing $400 million in purchases.
4. Rack-Level AI Infrastructure
Helios is built on AMD’s evolving full-rack design—combining MI350 accelerators, 5th-gen Epyc CPUs, and Pensando NICs, with Helios representing the next step into MI400-powered AI data centers.
5. Competitive Performance & Cost
AMD claims the MI350X GPUs in Helios will provide up to 40% more tokens per dollar than Nvidia’s B200 equivalents, though stocks dipped ~2% post-announcement as analysts await real-world benchmarks.
🌐 Strategic Importance
- Open innovation vs. proprietary control: Helios helps AMD attract partners prioritizing flexible, standards-based solutions.
- Strengthening the AMD-OpenAI relationship: The collaboration suggests confidence in MI400’s potential as a powerhouse back-end for generative AI.
- Leveling the playing field: Delivering rack-scale AI systems puts AMD in direct competition with Nvidia’s data center dominance.
- Software ecosystem ambitions: AMD continues advancing ROCm to rival Nvidia’s CUDA stack and enhance Helios adoption
🔮 What’s Coming & Challenges Ahead
- Availability in 2026: MI400-based Helios is expected in production next year.
- Benchmark anticipation: Industry watchers await Helios performance data against NVL72 and Nvidia’s upcoming Vera Rubin.
- Software maturity: The ROCm ecosystem must close the gap with CUDA to maximize Helios impact in real-world deployments reuters.com
Summary
AMD’s Helios AI server, packed with 72 MI400 chips and built on open standards, marks a pivotal challenge to Nvidia’s AI infrastructure dominance. With backing from OpenAI and partners, Helios aims to power the next wave of generative AI workloads—provided AMD delivers software support and real-world performance by its 2026 release.
