AMD challenges NVIDIA's AI dominance with open Helios rack-scale design
AMD has introduced the Helios open rack-scale reference design, which packs 72 Instinct GPUs, EPYC CPUs, and Pensando networking into a single system. The architecture aims to support on-premises enterprise AI workloads and agentic AI without requiring custom proprietary systems.
Key Takeaways
- Helios integrates 72 AMD Instinct GPUs and AMD EPYC CPUs into a single system based on Meta's Open Rack Wide (ORW) specification.
- Reference design utilizes AMD Pensando networking technology to optimize data flow between compute nodes and minimize latency.
- System is explicitly tuned for agentic AI, which AMD identifies as a dual workload requiring high-performance CPU and GPU coordination.
- Architecture supports 'sovereign AI' by enabling enterprises and national governments to host large-scale intelligence on private infrastructure.
Why It Matters
AMD is pivoting from a component supplier to a systems provider to compete with NVIDIA’s tightly integrated Blackwell/Rubin clusters. By embracing the Open Compute Project's standards, AMD offers a lower-friction alternative for hyperscalers and enterprises wary of proprietary vendor lock-in. For the streaming industry, this represents a significant structural shift in data center modernization; as agentic AI begins to drive content recommendation and real-time video processing, the flexibility of open rack-scale designs may lower the total cost of ownership for token-intensive workloads. Watch for adoption rates of Helios among tier-two cloud providers and global telecommunications firms treating AI as a local resource.
Additional Context
The launch of Helios follows a significant year for AMD in the data center. Per AMD’s October 2024 announcements, the company’s 5th Gen 'Turin' EPYC processors secured over 950 public cloud instances, providing up to 192 cores via Zen 5c architecture. This established a foundation of CPU dominance that AMD is now leveraging to fuel GPU-heavy AI racks. Per Silicon Analysts and IDC in July 2026, while NVIDIA currently commands approximately 75% to 81% of AI accelerator revenue, AMD has emerged as a credible second source, capturing an estimated 5% to 7% market share with its Instinct line. Technically, Helios is positioned to rival NVIDIA’s high-density NVL clusters. Per Network World in July 2026, the Helios design features up to 31TB of HBM4 memory, roughly 50% more capacity than NVIDIA's Vera Rubin systems. This memory advantage is critical for inference performance on frontier models. Furthermore, AMD has secured high-profile validation for this strategy; per Reuters in July 2026, Anthropic has entered a 2-gigawatt compute agreement with AMD, and Microsoft Azure is preparing a wide deployment of Helios systems to support its primary AI customer base. Competition is also intensifying within the networking layer as AMD integrates its Pensando technology. Per AMD and TechPowerUp reporting in July 2026, the Helios rack incorporates the Pensando 'Salina' 400 DPU and 'Vulcano' 800 AI NIC. These components are designed to meet the connectivity demands of 200 Gbps and 400 Gbps line rates, addressing the 'networking bottleneck' that often degrades performance in massive GPU clusters. This full-stack approach signals AMD’s intent to match NVIDIA's Infiniband ecosystem with an open, Ethernet-based alternative.
Read full article at siliconangle.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source