AMD takes on Nvidia with its Helios AI rack-scale system

In a significant move poised to reshape the artificial intelligence hardware landscape, AMD has officially launched its highly anticipated Helios AI rack-scale system. Unveiled at the Advancing AI 2026 event, Helios represents AMD's most ambitious effort yet to directly compete with rival chipmaker Nvidia in the rapidly expanding market for high-performance AI infrastructure.

The company announced that the Helios system is now in full production, with orders slated to begin shipping to customers by the end of the third quarter of 2026. This launch marks a critical milestone for AMD as it seeks to capture a larger share of the booming AI accelerator market, which it projects to reach an astounding $1.4 trillion by 2030.

Helios: A Powerhouse for Frontier AI Workloads

The AMD Helios AI rack-scale system is engineered from the ground up to tackle the most demanding AI workloads, including the training of trillion-parameter models and large-scale AI inference. This integrated solution combines AMD's latest and most powerful hardware components, meticulously designed to work in concert.

The Core Components: GPUs, CPUs, and Networking

• AMD Instinct MI455X GPUs: At the heart of each Helios rack are 72 AMD Instinct MI455X GPUs. These accelerators are based on AMD's advanced CDNA architecture, providing exceptional compute performance.
• 6th Gen AMD EPYC "Venice" CPUs: The system is powered by 6th Gen AMD EPYC "Venice" CPUs, featuring up to 256 high-performance cores and 1.6 TB/s of memory bandwidth. These CPUs provide the robust general-purpose computing foundation necessary for complex AI operations.
• AMD Pensando "Vulcano" AI NICs: High-speed, efficient data movement is crucial for AI at scale. Helios incorporates AMD Pensando "Vulcano" AI NICs, offering 800 Gbps for AI scale-out bandwidth and supporting PCIe Gen 6.
• ROCm Software Platform: Tying this powerful hardware together is AMD's open ROCm software platform, which provides the necessary programming models, tools, compilers, and libraries for AI development and deployment.

Unprecedented Scale and Performance

The Helios system boasts impressive specifications designed to deliver a generational leap in AI capabilities. Each rack offers a staggering 31 terabytes of HBM4 memory, which is fully coherent and addressable across the internal network fabric. This massive memory pool allows for fewer model replicas during training, leading to better "tokens-per-dollar" efficiency, as highlighted by AMD.

In terms of raw compute power, Helios delivers 2.9 Exaflops of FP4 compute and 1.4 Exaflops of FP8 compute, catering to both AI inference and training needs. The internal interconnects are equally robust, providing 260 TB/s of aggregate scale-up bandwidth within the rack and 43 TB/s of scale-out bandwidth for inter-rack communication. This architecture ensures that all 72 GPUs can operate as a single, coherent compute unit, with any GPU able to reach any other in a single network hop, simplifying software optimization.

AMD claims that Helios offers up to 30% more inference tokens per dollar than competing systems, along with 50% more HBM4 memory capacity and bandwidth, and 50% more scale-out bandwidth. The system itself is substantial, weighing approximately 5000 pounds and consuming between 225-245 kW of power per rack. Pricing for a Helios rack is estimated to be around $5-$5.5 million.

An Open Standard Approach to AI Infrastructure

A key differentiator for AMD's Helios system is its commitment to open standards. Unlike some proprietary interconnect solutions, Helios leverages UALink over Ethernet (UALoE) and other open standards across its entire fabric, from the front-end network to the GPU-to-GPU interconnect. This approach aims to provide customers with greater flexibility and avoid vendor lock-in, which is a significant concern for large-scale AI deployments.

Challenging Nvidia's AI Dominance

AMD's launch of Helios is a direct challenge to Nvidia, which currently holds a dominant position, estimated at around 80%, in the AI accelerator market. AMD, with an estimated 5-7% market share in AI accelerators, is aggressively positioning itself as a credible second source for hyperscalers and enterprises seeking diversification and competitive options.

The company has already secured significant customer commitments for Helios. Major AI players like Anthropic, OpenAI, Meta, Microsoft, and Oracle are planning to deploy Helios systems at scale. For instance, Anthropic has committed to deploying up to 2 gigawatts of MI450-series GPUs in Helios racks, with AMD also investing up to $5 billion in Anthropic equity as part of the strategic pact. Microsoft has announced plans to deploy AMD's Helios AI rack systems on Azure to power advanced AI models. OpenAI expects to deploy Helios at massive scale starting towards the end of 2026 and accelerating through 2027.

Furthermore, AMD has partnered with AI chipmaker Cerebras Systems to deliver a new disaggregated AI inference solution, combining AMD Helios rack-scale solutions with the Cerebras Wafer-Scale Engine for ultra-low-latency AI applications. This joint offering is expected to be available through Cerebras Cloud in the second half of 2026.

Industry Implications and Future Outlook

The introduction of Helios signals AMD's strong intent to become a major force in the AI infrastructure market. By offering a fully integrated, high-performance, and open-standard rack-scale solution, AMD aims to provide a compelling alternative to existing offerings. The company's aggressive roadmap includes plans to release a new Helios system every year, ensuring continuous innovation and a dependable path forward for its customers.

As AI workloads continue to grow in complexity and demand, the need for powerful and efficient hardware solutions will only intensify. AMD's strategic partnerships and the robust capabilities of the Helios system position the company to significantly impact the future of AI development and deployment across data centers, cloud environments, and specialized AI applications. This move is expected to foster greater competition, potentially leading to faster innovation and more choices for organizations investing in AI.

Frequently Asked Questions

What is the AMD Helios AI rack-scale system?

The AMD Helios AI rack-scale system is a powerful, integrated hardware platform designed by AMD for large-scale artificial intelligence workloads, including training massive AI models and performing high-throughput inference. It combines 72 AMD Instinct MI455X GPUs, 6th Gen AMD EPYC "Venice" CPUs, and AMD Pensando AI NICs into a single, coherent system.

When will the AMD Helios system be available?

AMD has announced that the Helios AI rack-scale system is now in full production, with initial orders expected to ship by the end of the third quarter of 2026. Volume deployments are anticipated to ramp up through the second half of 2026.

How does AMD Helios compare to Nvidia's AI offerings?

AMD positions Helios as a strong competitor to Nvidia's high-end AI systems, such as the Vera Rubin NVL72. Helios boasts significant advantages in memory capacity (31 TB HBM4), offering 50% more memory and bandwidth, and claims up to 30% more inference tokens per dollar compared to competing systems. AMD also emphasizes its use of open standards like UALink over Ethernet, contrasting with Nvidia's proprietary interconnects.

What kind of AI workloads is Helios designed for?

AMD Helios is specifically designed for demanding AI workloads such as training frontier AI models with trillions of parameters, large-scale AI inference, and fine-tuning. Its massive memory capacity and high-performance interconnects make it suitable for complex generative AI and agentic AI applications.