GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531 —
CL-OP46 497 —
CL-OP48 490 -0.2%
← Back to feed

AMD Starts Mass Production of Helios 2 AI Rack — MI455X GPUs and 6th-Gen EPYC Target NVIDIA's Blackwell

AMD announced the second generation of its Helios AI Rack at its Advancing AI 2026 event, and confirmed mass production has started with shipments due this year. The system pairs the new Instinct MI455X GPU with 6th-Generation EPYC CPUs and AMD’s Pensando networking, constituting AMD’s most complete rack-scale AI platform to date.

The announcement positions Helios directly against NVIDIA’s GB200 NVL72, which currently dominates rack-level AI deployments at hyperscalers like CoreWeave, Microsoft, and AWS. AMD is targeting both frontier AI training and sovereign computing use cases.

What the MI455X Adds

The Instinct MI455X succeeds the MI355X, which AMD has already placed in production deployments. A recent GLM-5.2 benchmark run on MI355X hardware posted 2,626 tokens per second at roughly half the cost of Blackwell equivalents, giving AMD a credible cost argument in inference.

The MI455X pushes that further. Details on exact memory bandwidth and compute specs were not fully disclosed at launch, but the Chips and Cheese technical analysis of the architecture describes a chip explicitly designed for the frontier model size range where NVIDIA currently has no effective competition from AMD. The Helios 2 rack integrates MI455X in a format that matches how hyperscalers actually buy and deploy hardware: at the rack level, not GPU-by-GPU.

EPYC and Pensando Included

The 6th-Gen EPYC inclusion matters because it reduces customer dependency on Intel CPUs in AI server configurations. Combined with AMD’s Pensando networking stack, Helios 2 is pitched as an end-to-end system where AMD supplies every major compute, networking, and memory component.

That mirrors NVIDIA’s playbook with NVLink, InfiniBand, and Quantum networking, but on AMD’s own interconnect.

Market Context

AMD’s data center business reported $5.8B in Q1 2026 revenue, up 57% year-on-year, as EPYC and Instinct GPU ramp accelerated. That growth was almost entirely in CPU and existing GPU lines, not the MI455X, which means the new chip represents upside if deployments land.

The sovereign AI angle is credible. AMD already won France’s first exascale supercomputer contract (Alice Recoque) and is shipping Instinct MI355X to AMD MI355X deployments in European data centers looking for non-US hardware supply chains. Countries building domestic AI infrastructure without wanting a dependency on NVIDIA’s ecosystem are the natural customer for Helios 2.

Key Numbers

  • Platform: Helios 2 AI Rack
  • GPU: Instinct MI455X (new generation, replaces MI355X)
  • CPU: 6th-Gen EPYC
  • Status: Mass production, shipments this year
  • Prior MI355X win: GLM-5.2 at 2,626 tokens/second, 2x cheaper than Blackwell for inference
  • AMD Data Center Q1 2026: $5.8B revenue, +57% YoY