AMD’s open approach also extends to the physical rack and its partner implementations. Helios is based on the Open Compute Project’s Open Rack Wide specifications, with AMD defining the core 72-GPU architecture and networking topology. OEMs, ODMs and other manufacturing partners can differentiate around rack integration, power and cooling, management, deployment, and support, while hyperscalers and large service providers can integrate Helios with their own cluster networking, storage, orchestration and cloud services. This gives large buyers more influence over how Helios is incorporated into their infrastructure and allows multiple partners to commercialize the platform without requiring AMD to sell a single fixed, branded rack, leveraging the same commercialization approach NVIDIA has proved to be successful but with an added layer of openness.
 

Customer commitments validate Helios and the breadth of AMD’s data center portfolio

Microsoft offers the broadest validation of AMD’s data center strategy, with the companies’ expanded partnership both validating Helios as an integrated architecture and showing that AMD can win CPU and networking placements independently. Azure plans to deploy Helios at scale for frontier-model inference, Azure AI services and customer applications. Microsoft will also introduce new Azure virtual machines based on 6th Gen EPYC and broaden its use of Pensando DPUs across Azure networking services. Azure availability also gives AMD a route to enterprise demand beyond the small group of customers capable of deploying Helios directly.
 
AMD complemented Microsoft’s cloud-platform endorsement with commitments from leading model developers. OpenAI expects to bring Helios online beginning in 4Q26 and accelerate deployments during 2027, while Meta is testing Helios workloads and validating 6th Gen EPYC platforms. The Anthropic agreement goes further, with a planned deployment of up to 2 gigawatts of MI455X GPUs in Helios systems — with the first gigawatt beginning deployment in 1H27 — in addition to the companies establishing a multiyear engineering collaboration through which Claude will be used to optimize Instinct workloads and accelerate ROCm development. AMD also committed to investing up to $5 billion in Anthropic, using its balance sheet to deepen technical and commercial alignment with a major prospective infrastructure customer. The move follows a playbook NVIDIA has also used, pairing capital investment with technical collaboration and commitments to deploy its infrastructure.
 
Taken together, these engagements show that AMD is participating earlier in system design, workload optimization and capacity planning with several of the world’s largest AI infrastructure buyers. The announcements span different stages, from testing and planned availability to dated deployment commitments, but collectively establish Helios as a credible second source to NVIDIA at the rack level. These anchor customers also provide workloads and engineering input that can help AMD harden its overall AI platform before bringing it to a broader market.
 

Agentic AI broadens AMD’s opportunity across EPYC, ROCm and distributed AI

AMD framed agentic AI as a change in infrastructure demand that extends beyond model inference. The company estimates that inference will account for roughly 60% of global AI compute capacity in 2026 as usage shifts from training models toward putting them to work. Agents intensify that change because a single request can trigger repeated reasoning, tool calls, data retrieval and code execution. GPUs perform the model inference, while CPUs host and orchestrate the surrounding workflows.
 
AMD divided this CPU opportunity into three areas. AI host processors coordinate data movement and help keep accelerators fed; high-density agent sandboxes provide isolated environments for code execution and tool use; and general-purpose servers run the applications, databases and data services that agents access. The 6th Gen EPYC portfolio addresses these roles through different Venice configurations, ranging from high-frequency CPUs for GPU host nodes to processors with as many as 256 cores for agent sandboxes. AMD also emphasized x86 compatibility as an advantage because much of the enterprise software that agents will use already runs on x86 infrastructure.
AMD paired this CPU expansion with a software strategy aimed at its most persistent competitive limitation: the engineering work required to move and optimize applications for its GPUs. The strategy combines two ecosystem-oriented principles — openness and abstraction — with a new AMD-led automation layer.
 
ROCm’s open-source model means AMD does not have to build its software ecosystem alone. Customers, cloud providers and software developers can inspect how ROCm works, adapt it for their workloads and contribute fixes or performance improvements. NVIDIA also supports widely used open AI frameworks, but the core CUDA tool kit remains proprietary and NVIDIA-controlled. In contrast, ROCm gives outside engineers greater ability to improve the underlying GPU software and lets AMD draw on its ecosystem to close functionality and performance gaps. However, AMD remains responsible for integrating those contributions into a reliable platform that performs competitively.
 
Abstraction reduces the amount of application code tied directly to either ROCm or CUDA. Frameworks and programming tools such as PyTorch and Triton provide common interfaces, while AMD and NVIDIA optimize the hardware-specific software beneath them. This makes applications more portable, although platform-specific kernels, libraries and performance tuning remain necessary, meaning AMD must continue improving its compilers, runtimes, communications libraries and kernels even as abstraction makes them less visible to application developers.
 
ROCm.AI adds automation to these ecosystem-oriented principles through AMD Skills, the ROCm CLI and Console, and the Hyperloom optimization system. AMD Skills brings validated ROCm knowledge into coding agents such as Claude, Codex and Cursor; the CLI and Console provide repeatable setup, execution and diagnostic paths; and Hyperloom profiles inference workloads, identifies bottlenecks and tests targeted changes to host code and GPU kernels. Together, these tools can reduce the manual setup, troubleshooting and tuning previously performed by specialized engineers, albeit they still depend on the quality of AMD’s underlying compilers, runtimes, libraries and kernels.
 
Beyond data center CPUs and software, AMD extended the same portfolio-expanding strategy across enterprise servers, local systems and physical AI. For example, the company announced Instinct MI350P, a dual-slot PCIe accelerator designed to add inference capacity to existing air-cooled enterprise servers, and Ryzen AI Halo, a deskside developer platform powered by Ryzen AI Max PRO 400 Series processors for developing and running models and agents locally. Additionally, the company highlighted how it is working with Cisco to develop fleetwide observability, governance, security and network controls so enterprises can manage Ryzen AI Halo systems as part of their broader infrastructure. Finally, AMD announced the Kria AI Robotics Developer Platform, extending AMD’s portfolio into physical AI by combining CPU, GPU, neural processing unit (NPU) and FPGA (field-programmable gate array) compute. Together, these offerings support AMD’s argument that agentic workloads will span deployment locations and require different types of compute.