AMD and Anthropic have signed a strategic partnership to deploy up to 2 gigawatts of Instinct MI450 accelerators for training and running Claude. The first gigawatt will start installation during the first half of 2027 using Helios systems, as AMD prepares a capital investment of up to $5 billion in the AI company.
The key points of the AMD and Anthropic agreement in 20 seconds
- Anthropic will deploy up to 2 GW of AMD Instinct MI450 GPUs to train and run Claude.
- The first gigawatt will begin installation in the first half of 2027.
- The infrastructure will use Helios racks with MI455X GPUs, EPYC Venice CPUs, Pensando networking, and ROCm.
- AMD plans to invest up to $5 billion in Anthropic.
- Claude will also help improve AMD’s software.
The agreement makes Anthropic another major client of AMD’s upcoming AI platform and expands its strategy to avoid reliance on a single chip supplier. The company already uses Instinct MI355X accelerators, but this new commitment elevates the relationship to a scale comparable with some of the industry’s largest data center projects.
The 2 GW figure represents the electrical power of the planned infrastructure, not the computing capacity of the GPUs nor their isolated power consumption. It covers complete systems that also require processors, memory, networking, storage, cooling, and electrical equipment. AMD and Anthropic have not specified how many data centers will participate, where they will be located, nor the exact number of accelerators to be installed.
Anthropic will use Helios racks with MI455X GPUs
The infrastructure will be based on AMD Helios, a rack-scale architecture designed to cluster accelerators, processors, and networks into a single AI system.
Each Helios reference configuration integrates 72 Instinct MI455X GPUs, sixth-generation EPYC Venice CPUs, and AMD Pensando networking technology. The setup uses ROCm, AMD’s software platform aimed at providing an alternative to CUDA dominance in accelerated computing.
According to AMD’s published specifications, a Helios rack can have 31 TB of HBM4 memory and reach up to 2.9 exaFLOPS in FP4 operations. These figures describe the theoretical performance in low-precision numerical formats used in certain AI workloads and do not directly compare to the actual behavior of all models.
The MI455X GPU is part of the Instinct MI450 series and uses a new generation of the CDNA architecture. AMD states it has 432 GB of HBM4 memory per accelerator and a bandwidth of up to 19.6 TB/s. This memory is especially important for training large models and performing inference over many requests without constantly repartitioning data across devices.
Helios also adopts the Open Rack Wide format introduced by Meta within the Open Compute Project. Compared to conventional racks, its double-width design provides more space for high-density systems, power supplies, and liquid cooling.
The first gigawatt contracted by Anthropic will begin deployment in the first half of 2027. The announcement does not specify a timeline for the second gigawatt nor presents the 2 GW as already installed capacity. Deployment will depend on chip manufacturing, HBM4 availability, data center construction, and power supply.
Claude will help improve ROCm and AMD GPU workloads
The partnership goes beyond hardware purchasing. AMD and Anthropic have agreed on a multi-year engineering collaboration to adapt Claude’s workloads to Instinct accelerators and accelerate ROCm development.
Software compatibility remains one of AMD’s main challenges compared to NVIDIA. CUDA has years of adoption, libraries, and tools used by researchers, cloud providers, and developers. ROCm has progressed, but running large models on different architectures requires adapting compute cores, communications, and memory management.
Teams from both companies will use Claude to review code, suggest improvements, and assist in software optimization. AMD also plans to extend this assistance across its engineering and product development areas.
This creates a bilateral relationship. Anthropic gains compute capacity and participates in platform adaptation, while AMD can test ROCm with one of the most demanding commercial models. The improvements developed for Claude could later benefit other clients, although the companies have not specified which parts of the work will become part of the public ROCm project.
Diversification also benefits Anthropic. Tom Brown, co-founder and head of computing at the company, said that working with different architectures allows each workload to be assigned to the most suitable hardware. Besides technical reasons, using multiple suppliers reduces risks related to dependence on available capacity and platform pricing.
AMD secures another multi-gigawatt customer for MI450
This commitment follows other large-scale contracts AMD announced. In October 2025, OpenAI agreed to deploy up to 6 GW of Instinct GPUs, with the first gigawatt of MI450 expected in the second half of 2026. Meta announced another up to 6 GW deal in February 2026, and Microsoft confirmed use of Helios in Azure for inference workloads.
These commitments do not necessarily equate to firm sales recognized immediately. They are multi-year plans subject to delivery schedules, infrastructure availability, and meeting conditions. Still, they show AMD is securing contracts with some of the biggest AI accelerator buyers.
For Anthropic, the agreement coincides with increasing demand for Claude across enterprises and developer services. Training new generations of models requires extensive clusters over long periods, while inference continuously consumes capacity whenever a user or application makes a request.
The projected up to 2 GW highlights that inference is becoming a central part of investment. Once trained, a widely used model might need more infrastructure to handle requests than was needed during initial training.
A $5 billion investment
AMD has also committed to making a strategic capital investment of up to $5 billion in Anthropic. The figure indicates a maximum limit for future investments, not an immediate disbursement.
The companies have not disclosed the valuation, timelines, conditions, or the percentage AMD might control. Nor have they clarified whether the investment is tied to deployment or hardware purchase targets.
The agreement strengthens the provider-client relationship. AMD gains exposure to Anthropic’s growth and increases the chances of Claude continuing to use its infrastructure. Anthropic receives funding and a new source of computational capacity.
Such agreements also complicate the AI chip market landscape. Part of the invested capital could return to the provider through hardware purchases, so distinguishing between announced commitments and recognized revenue will be necessary as systems are delivered.
The technical challenge begins even before racks are operational. Anthropic must adapt its models and tools to ROCm, while AMD must produce enough MI455X units and demonstrate Helios can sustain performance, stability, and efficiency at gigawatt scale. The agreement sets the planned scope, but its success will depend on execution over the coming years.
Frequently Asked Questions
How many GPUs will Anthropic purchase from AMD?
Neither company has published a unit count. The agreement covers up to 2 GW of infrastructure based on Instinct MI450 GPUs, with the first gigawatt starting in the first half of 2027.
What components will Anthropic’s platform use?
The Helios systems will incorporate MI455X GPUs, EPYC Venice CPUs, Pensando networking technology, and AMD’s ROCm software.
Has AMD already invested $5 billion in Anthropic?
Not necessarily. AMD has committed to a future strategic investment of up to $5 billion, but has not disclosed the disbursements or conditions.
Why does Anthropic want to use AMD GPUs?
The company aims to expand its capacity for Claude and diversify hardware sources. It will also collaborate with AMD to adapt workloads and improve ROCm.
