AMD has expanded its data centre portfolio with a coordinated launch spanning rack-scale systems, processors, accelerators, networking and software. The company introduced its Helios architecture alongside sixth-generation EPYC CPUs, Instinct MI400-series GPUs and the ROCm.ai development platform at Advancing AI 2026.
The launch positions AMD to compete for a larger role in AI factories, hyperscale inference, frontier-model training and sovereign computing environments. Instead of supplying individual processors alone, AMD is presenting a complete infrastructure architecture designed to scale from single racks to gigawatt campuses.
Helios integrates the AI rack
Helios combines 72 Instinct MI455X GPUs, 18 sixth-generation EPYC “Venice” CPUs, Pensando networking and ROCm software within an open rack-scale design.
AMD says the system delivers exaflop-class performance and up to 30% more inference tokens per dollar than the leading competing architecture. Its scale-up fabric connects the GPUs within a single compute domain, reducing the need for operators to integrate separate processing, networking and software layers.
Technical highlights include:
- 72 AMD Instinct MI455X GPUs per rack
- 18 sixth-generation EPYC CPUs
- HBM4 high-bandwidth memory
- Pensando scale-up and scale-out networking
- ROCm open software
- Deployment from individual racks to large AI clusters
New processors address mixed workloads
The EPYC 9006 family targets enterprise computing, cloud services, high-performance computing and agentic AI orchestration. CPUs remain critical in AI facilities because they coordinate GPU workloads, connect models with applications and manage the numerous steps executed by autonomous software agents.
The Instinct MI400 portfolio includes the MI455X for large AI factories and the MI430X for sovereign AI and scientific computing. AMD is using a common software foundation across these products to reduce fragmentation between conventional HPC and generative-AI environments.
“As agentic AI reshapes how enterprise infrastructure is designed and deployed, EPYC is evolving with our customers once again.” said Dan McNamara, Senior Vice President and General Manager of Compute and Enterprise AI at AMD.
ROCm.ai strengthens the software layer
ROCm.ai introduces AI-assisted installation, deployment, profiling, debugging and workload optimisation. AMD reports performance improvements of up to 3.3 times for inference and 2.4 times for training compared with the preceding ROCm generation on equivalent hardware.
The software strategy is important because hardware adoption increasingly depends on developer tooling, framework compatibility and workload portability, rather than processor specifications alone.
Competition moves to full-system design
AMD’s launch reflects a broader shift from selling discrete chips to engineering complete AI systems. Compute density, memory bandwidth, networking efficiency, cooling readiness and software optimisation must now be developed as one platform.
Helios therefore represents both a product launch and a long-term market strategy. Its success will depend on production deployments, customer adoption and whether AMD’s open architecture can offer hyperscalers a credible alternative to vertically integrated competing systems.
Professionals at AMD: Philip Guido | Bill Orner | Carl Nyqvist | Andrew Ruffin | Onil Gunawardana | Anjana Srinivasan | Keith Toh | Anil Kumar Gopalakrishna | Awanish Verma | Aleks Shar, MBA 🇺🇦 🎗️ | Sachin Dhingra | Yousuf Khan | Simon Khozam, PMP®, SAFe® | Melanie Kiely | David Lam | Alex Ebeid | Chris Johnson, MBA | Rudy Torrijos | Raghav Dhir
