Helios AI Rack System

Configuration: Each Helios rack integrates 72 AMD Instinct MI455X GPUs alongside 18 6th-gen EPYC "Venice" CPUs, organized across 18 compute trays and connected by AMD Pensando front-end, scale-up, and scale-out networking .

GPU: The MI455X is AMD's most advanced AI accelerator and the key silicon powering Helios .

Claimed position: Su called Helios "simply the best AI rack in the world" and "the industry's highest-performance AI rack for training and running frontier models at massive scale" .

Production status: Helios was announced as in full production, shipping in the same procurement window as Nvidia's Vera Rubin NVL72 .

Pricing: Confirmed at $5–5.5 million per rack — approximately $5.25 million at the midpoint .

EPYC Venice CPU (6th Gen, Zen 6)

The 6th-gen EPYC "Venice" CPU is the industry's first x86 server processor built on TSMC's 2nm node . It was launched commercially at the event and serves as the CPU backbone inside each Helios compute tray, scaling up to 256 Zen 6 cores .

Memory Advantages vs. Nvidia

AMD highlighted a memory advantage for Helios over Nvidia's Vera Rubin NVL72:

• Each MI455X GPU offers 288 GB of HBM3E memory, with AMD's rack delivering a higher aggregate memory pool across 72 GPUs . According to AMD, Helios provides 15% more compute, 50% more HBM4 memory capacity and bandwidth, and 50% more scale-out bandwidth than competing systems .
• AMD claimed superior memory bandwidth and larger total HBM capacity per rack, giving an edge in inference workloads that benefit from keeping large models in memory .
• However, analysts noted Nvidia still leads in training performance and software ecosystem maturity (CUDA) .

Confirmed Anchor Customers

Multiple major customers and adopters were confirmed at the keynote :

The official AMD press release also listed Cerebras as a collaborator .

Strategic Partnership: AMD & Anthropic

Announced on July 22 (day one of the event) :

• Investment: AMD committed a strategic equity investment of up to $5 billion in Anthropic (the AI lab behind Claude) .
• Compute deployment: Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs delivered via Helios rack-scale systems .
• Timeline: The first gigawatt of deployment begins in the first half of 2027 .
• Anthropic co-founder Tom Brown cited ease of integration and performance as reasons for the partnership, and the companies will also collaborate on AI safety and security .

Strategic Partnership: AMD & Cerebras

Announced on July 23 (keynote day) :

• AMD and Cerebras formed a technical partnership to deliver a disaggregated AI inference architecture .
• AMD Helios and Cerebras's Wafer-Scale Engine will operate as a single disaggregated inference workflow — combining AMD Instinct GPUs' ultra-high throughput with Cerebras's ultra-low-latency token generation .
• Customers can split AI inference workloads across both systems, targeting complex, latency-sensitive AI applications . The joint solution is expected to be available through Cerebras Cloud in the second half of 2026 .

Pricing Comparison: Helios vs. Nvidia Vera Rubin

Sources noted that Helios ships into the same procurement cycle as Vera Rubin, marking the first time AMD has a rack-scale system competing head-to-head with Nvidia's latest generation .

Helios AI Rack System

Configuration: Each Helios rack integrates 72 AMD Instinct MI455X GPUs alongside 18 6th-gen EPYC "Venice" CPUs, organized across 18 compute trays and connected by AMD Pensando front-end, scale-up, and scale-out networking .

GPU: The MI455X is AMD's most advanced AI accelerator and the key silicon powering Helios .

Claimed position: Su called Helios "simply the best AI rack in the world" and "the industry's highest-performance AI rack for training and running frontier models at massive scale" .

Production status: Helios was announced as in full production, shipping in the same procurement window as Nvidia's Vera Rubin NVL72 .

Pricing: Confirmed at $5–5.5 million per rack — approximately $5.25 million at the midpoint .

EPYC Venice CPU (6th Gen, Zen 6)

The 6th-gen EPYC "Venice" CPU is the industry's first x86 server processor built on TSMC's 2nm node . It was launched commercially at the event and serves as the CPU backbone inside each Helios compute tray, scaling up to 256 Zen 6 cores .

Memory Advantages vs. Nvidia

AMD highlighted a memory advantage for Helios over Nvidia's Vera Rubin NVL72:

• Each MI455X GPU offers 288 GB of HBM3E memory, with AMD's rack delivering a higher aggregate memory pool across 72 GPUs . According to AMD, Helios provides 15% more compute, 50% more HBM4 memory capacity and bandwidth, and 50% more scale-out bandwidth than competing systems .
• AMD claimed superior memory bandwidth and larger total HBM capacity per rack, giving an edge in inference workloads that benefit from keeping large models in memory .
• However, analysts noted Nvidia still leads in training performance and software ecosystem maturity (CUDA) .

Confirmed Anchor Customers

Multiple major customers and adopters were confirmed at the keynote :

The official AMD press release also listed Cerebras as a collaborator .

Strategic Partnership: AMD & Anthropic

Announced on July 22 (day one of the event) :

• Investment: AMD committed a strategic equity investment of up to $5 billion in Anthropic (the AI lab behind Claude) .
• Compute deployment: Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs delivered via Helios rack-scale systems .
• Timeline: The first gigawatt of deployment begins in the first half of 2027 .
• Anthropic co-founder Tom Brown cited ease of integration and performance as reasons for the partnership, and the companies will also collaborate on AI safety and security .

Strategic Partnership: AMD & Cerebras

Announced on July 23 (keynote day) :

• AMD and Cerebras formed a technical partnership to deliver a disaggregated AI inference architecture .
• AMD Helios and Cerebras's Wafer-Scale Engine will operate as a single disaggregated inference workflow — combining AMD Instinct GPUs' ultra-high throughput with Cerebras's ultra-low-latency token generation .
• Customers can split AI inference workloads across both systems, targeting complex, latency-sensitive AI applications . The joint solution is expected to be available through Cerebras Cloud in the second half of 2026 .

Pricing Comparison: Helios vs. Nvidia Vera Rubin

Sources noted that Helios ships into the same procurement cycle as Vera Rubin, marking the first time AMD has a rack-scale system competing head-to-head with Nvidia's latest generation .