AMD Advancing AI 2026: Helios Rack, Cerebras Partnership, and $5B Anthropic Deal — Key Specs and Timelines

At its Advancing AI 2026 event (July 22–23, San Francisco), AMD made three major announcements: the Helios rack-scale platform entering full production, a technical partnership with Cerebras for low-latency AI inference, and a strategic deal with Anthropic. Here is the breakdown with key specifications and timelines.

Helios Rack-Scale Platform — Entering Full Production

CEO Lisa Su confirmed the Helios servers are in full production and will begin shipping at the end of Q3 2026 . AMD describes Helios as the industry's highest-performance AI rack for training and running frontier models at massive scale .

Key specifications (per single Helios rack) — from AMD's official product page and brochure :

Component	Spec
GPUs	72x AMD Instinct MI455X (CDNA 5, TSMC N2)
CPUs	AMD EPYC "Venice" (Zen 6, 6th Gen)
Pooled memory	31 TB HBM4
Aggregate memory bandwidth	1.4 PB/s
Per-GPU memory	432 GB HBM4, 19.6 TB/s bandwidth
Per-GPU compute	40 FP4 PFLOPS / 20 FP8 PFLOPS
Rack-level compute	2.9 FP4 exaFLOPS (inference), 1.4 FP8 exaFLOPS (training)
Networking	AMD Pensando "Vulcano" with UALink
Form factor	18 compute trays (4 GPUs + 1 EPYC Venice per tray), double-wide rack
Estimated price	~$5.25M per rack

The MI455X delivers roughly 2x the FP4 compute of the previous MI350 series . The Helios system is AMD's primary rack-scale competitor to Nvidia's Vera Rubin platform and has already lined up customers including Microsoft, Meta, OpenAI, Oracle, and Anthropic .

Technical Partnership with Cerebras — Ultra-Low-Latency AI Inference

Announced on stage at Advancing AI 2026 (July 23), AMD and Cerebras are building a disaggregated inference platform that splits AI inference into two phases :

• Prefill (prompt processing): Handled by AMD's Helios rack (Instinct GPUs + EPYC CPUs)
• Decode (token generation): Handled by Cerebras' Wafer-Scale Engine (WSE), an SRAM-powered massive chip designed for ultra-low latency

The two systems operate as a single disaggregated solution via high-speed interconnect .

Claimed performance: Up to 5x higher tokens per second per watt compared to conventional architectures (based on July 2026 modeling using the Kimi 2.6 1T model) .

Availability: The joint AMD-Cerebras AI inference solution is expected to be available through cloud providers in H2 2026 . Cerebras will deploy Helios systems in its own data centers, with the solution available first through Cerebras Cloud .

Cerebras CEO Andrew Feldman said at the event that the architecture directly targets Nvidia's Groq LPUs and is positioned for agentic AI workloads where low-latency inference is critical .

Strategic Partnership with Anthropic — Up to $5B Investment

Announced July 22, 2026, via AMD official press release :

Deal terms:

• Chip deployment: Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs in AMD Helios rack-scale solutions .
• First phase: Deployment of the first 1 GW begins in H1 2027 .
• Equity investment: AMD committed a strategic equity investment of up to $5 billion in Anthropic .
• Software collaboration: The companies will work to optimize Anthropic's Claude models for AMD Instinct GPUs and accelerate the AMD ROCm software platform. AMD will also deploy Claude across its engineering and product development teams .

This adds Anthropic to AMD's growing roster of major AI customers alongside OpenAI, Meta, Oracle, and Microsoft . The MI450 Series mentioned in this deal is the generation of GPU that powers the Helios racks — the MI455X is the specific SKU used in the first Helios production units .

Summary of Timelines

Product / Program	Availability
Helios rack (MI455X + EPYC Venice)	Full production now; shipping end of Q3 2026
AMD-Cerebras disaggregated inference	Cloud availability H2 2026
Anthropic MI450 deployment (1st GW)	Begins H1 2027
MI455X engineering samples	H2 2026; mass production Q2 2027

All performance figures are AMD-reported and not yet independently benchmarked . The Helios pricing of ~$5.25M/rack and the Cerebras 5x throughput-per-watt claim are based on vendor modeling.