Key Points
- 01AMD (AMD) unveils Instinct MI400 GPUs, Helios racks and EPYC Venice at its Advancing AI 2026 event
- 02Helios rack integrates 72 MI455X GPUs with 31 TB HBM4 and 1.4 PB/s bandwidth
- 03Per-rack Helios performance is rated at 2.9 exaflops FP4 inference and 1.4 exaflops FP8 training
- 04AMD (AMD) announces Anthropic, Microsoft Azure and Oracle as key Helios customers
AMD sets out full-stack AI data-center platform
At the Advancing AI 2026 event held in San Francisco on July 23, 2026, AMD introduced a coordinated set of products targeting AI and cloud data centers. The launch centered on the Instinct MI400 GPU family, the Helios rack-scale system, and the EPYC “Venice” server CPU. The company positioned these components as a unified platform designed for large-scale training and inference workloads.
The Instinct MI400 line includes the MI430X, MI440X and MI455X accelerators. These GPUs are intended to serve as the compute foundation within the new Helios rack, which is optimized around the MI455X model for high-density deployments.
Helios rack architecture and performance
AMD described Helios as a fully integrated rack built around 72 MI455X GPUs. The system pools 31 terabytes of HBM4 memory and delivers an aggregate memory bandwidth of 1.4 petabytes per second. This design is aimed at supporting very large AI models that require both substantial compute and high memory bandwidth.
In performance terms, a single Helios rack is rated at 2.9 exaflops of FP4 throughput for inference workloads. For training, AMD cited 1.4 exaflops of FP8 performance per rack. Analysts covering the event estimated Helios pricing between $5 million and $5.5 million per rack, implying an average level near $5.25 million.
AMD named Microsoft Azure and Oracle as early adopters of Helios. These hyperscale cloud providers are expected to use the racks to expand their AI infrastructure, reinforcing the system’s positioning in large-scale data centers.
EPYC Venice CPU and product rollout timing
Alongside the GPU and rack announcements, AMD introduced EPYC “Venice,” a server processor aimed at AI and cloud workloads. Venice is a Zen 6 design manufactured on TSMC’s 2-nanometer process and is specified to scale up to 256 CPU cores in a single package. This positions the chip as a high-core-count companion to the MI400 accelerators in AI servers.
AMD stated that EPYC Venice and the MI400-based Helios systems are scheduled to begin shipping in the second half of 2026. This timing aligns CPU and GPU availability for customers planning next-generation data-center deployments.
Anthropic and large-scale GPU commitments
AMD announced a commercial commitment with Anthropic for up to 2 gigawatts of Instinct MI450-series GPU capacity to be deployed in Helios racks. The first gigawatt of this deployment is expected to begin in the first half of 2027, indicating a multi-year rollout of the infrastructure.
In addition to the hardware supply agreement, AMD said it could make a strategic equity investment in Anthropic of up to $5 billion. The combination of large-scale GPU commitments and a potential capital investment underscores the depth of the relationship between the two companies.
Key Takeaways
- 01AMD is moving to deliver a tightly integrated stack of GPUs, CPUs and rack-scale systems aimed at large AI data centers.
- 02Helios racks combine dense MI455X accelerator deployments with high-bandwidth HBM4 memory to support demanding training and inference workloads.
- 03EPYC Venice’s 256-core, 2 nm design is intended to complement the MI400 family and align with Helios shipments in the second half of 2026.
- 04Large-scale commitments from Anthropic and early wins at major cloud providers suggest significant anticipated demand for AMD’s new AI infrastructure platform.