Helios launch, OpenAI deal and new AI roadmap

• AMD launched Helios, with OpenAI, Meta and Anthropic planning deployments at AMD Advancing AI.
• AMD also expanded its hardware roadmap and AI partnerships.

AMD launched its Helios rack-scale AI system at Advancing AI 2026, with OpenAI, Meta and Anthropic preparing deployments that could span several gigawatts of infrastructure.

The company also detailed new Instinct accelerators, future EPYC processors and software tools, while expanding its work with Cerebras, Cisco and Hugging Face.

“The next phase of AI will span frontier models, agents and physical AI, creating new opportunities to bring intelligence everywhere,” AMD Chair and CEO Lisa Su said. “Realising that potential will take the entire industry working together.”

Helios and large-scale AI deployments

Helios combines 72 Instinct MI455X GPUs, 18 sixth-generation EPYC “Venice” processors, Pensando networking and AMD’s ROCm software stack. The system is built for model training, fine-tuning and large-scale inference, with deployments ranging from individual racks to multi-rack clusters.

A single rack provides up to 2.9 exaflops of peak FP4 performance, 1.4 exaflops of peak FP8 performance, 31TB of HBM4 memory and 1.7PB per second of memory bandwidth, according to AMD.

AMD claims Helios delivers 15% more peak FP4 performance, 50% more high-bandwidth memory capacity and up to 30% more tokens per dollar than a competing platform. The company did not disclose full benchmark conditions in the announcement.

“Frontier AI infrastructure is evolving rapidly, and customers need platforms that can deliver exceptional performance while scaling efficiently as workloads grow,” AMD Senior Vice President for AI Vamsi Boppana said. “AMD Helios brings together leadership compute, high-performance networking and open software in a unified rack-scale platform.”

The architecture uses UALink over Ethernet for scale-up connectivity and Ethernet technologies aligned with the Ultra Ethernet Consortium for scale-out networking. Pensando Vulcano 800 network interface cards connect systems across larger clusters.

Helios follows the Open Rack Wide specification introduced by Meta through the Open Compute Project. Hardware suppliers preparing systems include Bull, HPE, Lenovo, Supermicro, Sanmina and Wiwynn.

OpenAI expects to begin bringing Helios systems online in the fourth quarter of 2026, with deployments accelerating during 2027 through several infrastructure partners. The rollout will form the first phase of a six-gigawatt agreement announced by AMD and OpenAI in October 2025.

OpenAI has had access to Helios hardware for several months while the companies optimise GPT workloads for the MI455X. Their work covers ROCm, networking, compilers, workload debugging and the Triton and Gluon GPU programming tools.

OpenAI is also providing AMD with information about changing model requirements and infrastructure bottlenecks. That feedback has informed the MI400 generation and is continuing as AMD develops its MI500 accelerators.

OpenAI Vice President of Compute Strategy Sachin Katti said the companies are aligning their infrastructure roadmaps with future model requirements. OpenAI is identifying emerging bottlenecks, while AMD is using that information to guide decisions across silicon, memory, networking, software and rack design.

Meta is also testing Helios racks and sixth-generation EPYC platforms. It has deployed millions of EPYC processors across the Milan, Bergamo and Turin generations and is now validating Venice systems.

The companies are developing a customised accelerator based on the MI450 generation. Meta plans to deploy up to six gigawatts of infrastructure using AMD GPUs.

Meta Head of Infrastructure Santosh Janardhan said AI systems increasingly depend on compute, networking, memory, power, cooling and software operating together. Meta and AMD are therefore aligning engineering decisions earlier in the development process rather than selecting completed components at the end of a product cycle.

Anthropic represents another planned Helios deployment, with the company preparing to install up to two gigawatts of MI455X GPUs. Under a multiyear engineering agreement, Anthropic will use Claude to help optimise workloads and support ROCm development, while AMD plans to use the model within parts of its engineering operations.

Cerebras partnership and AMD’s hardware roadmap

AMD and Cerebras are developing a disaggregated inference system that combines Helios racks with Cerebras Wafer-Scale Engine hardware. The two platforms will process separate stages of the same inference workflow.

Helios will process prompts and large context windows, while the Cerebras system will handle token generation. The approach separates the throughput-intensive prompt stage from the memory-bandwidth-intensive generation stage.

AMD and Cerebras claim the system can deliver up to five times more tokens per second per watt. The companies did not provide complete benchmark conditions for the comparison.

“AI inference is becoming one of the largest infrastructure opportunities in AI, and its growing diversity requires a more flexible approach,” Su said. “Together with Cerebras, we are extending that leadership into the most latency-sensitive applications.”

The companies are targeting coding assistants, real-time copilots, autonomous agents, robotics and scientific workloads. Cerebras plans to deploy Helios systems in its data centres and offer the joint service through Cerebras Cloud in the second half of 2026.

“The demand for ultra-fast inference is growing at an unprecedented pace,” Cerebras CEO and co-founder Andrew Feldman said. “Partnering with AMD gives us an incredible opportunity to bring that performance to even more customers.”

Beyond the Helios launch, AMD outlined several additions to its Instinct range and longer-term processor roadmap. The MI455X is the primary accelerator used in the first Helios systems, with AMD claiming 34 times the token throughput of the previous MI355X.

AMD also introduced the MI430X for high-performance computing and sovereign AI deployments. It provides up to 288 teraflops of hardware-based FP64 performance and is planned for exascale-class systems in the United States and Europe.

The MI350P is designed to add AI acceleration to existing server infrastructure. AMD claims it delivers up to 4.2 times more tokens per second per dollar than a competing product.

The Instinct MI500 series is scheduled for 2027, followed by MI600 in 2028. Helios 500 will combine MI500 GPUs, EPYC “Verano” processors and Pensando “Como” and “Monza” networking.

Helios 600 will use MI600 accelerators, EPYC “Ferrara” processors and Pensando “Palma” and “Levanzo” networking. AMD has not disclosed detailed specifications for either rack system.

AMD plans to introduce three Zen 7-based server processors in 2028. The Florence, Ferrara and Fidenza products will target different combinations of system density, performance, power consumption and cost.

A Zen 8-based processor code-named Ravenna is scheduled for 2030. AMD did not disclose core counts, manufacturing processes or other platform details.

Software, local AI and robotics

AMD introduced ROCm.ai, a development platform that allows coding agents including Claude, Codex and Cursor to assist with software development for AMD GPUs. The platform is being used to prepare software for MI455X systems.

AMD said PyTorch, Hugging Face, vLLM and SGLang have already been enabled for the accelerator. The company is also working with OpenAI and Anthropic to use AI tools in kernel development, compiler optimisation and workload tuning.

The company is extending the same portfolio into local and hybrid AI systems through collaborations with Cisco and Hugging Face. AMD and Cisco are exploring systems that combine Ryzen AI Halo hardware with Cisco networking, monitoring and security products.

The collaboration is focused on managing local and hybrid AI deployments across endpoint systems, enterprise data centres and cloud infrastructure.

“By combining AMD Ryzen AI Halo systems and our broader local AI software capabilities with Cisco’s enterprise networking, observability and security technologies, we are helping customers deploy AI in a way that is performant, secure, observable and manageable at scale,” AMD Senior Vice President and Computing and Graphics Group General Manager Jack Huynh said.

Ryzen AI Halo includes 128GB of unified memory and supports models containing up to 200 billion parameters. AMD is expanding its Hugging Face partnership to provide models, libraries and development tools configured for the platform.

Systems based on Ryzen AI Max Pro 400 processors will increase unified memory to 192GB and support models containing up to 300 billion parameters. AMD and its hardware partners expect to release the platforms later in 2026.

“AI is moving from a cloud-only experience to something deeply personal, local and always available,” Huynh said. “With Ryzen AI platforms, AMD is helping developers, enterprises and creators bring powerful AI closer to where work happens.”

AMD also introduced Kria AI system-on-modules based on Ryzen AI Embedded X100 processors. A related robotics development platform combines CPU, GPU, neural processing unit and FPGA resources, although AMD did not announce pricing or general availability dates.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.