AMD and Anthropic have announced a major strategic partnership that will bring up to 2 gigawatts of AMD Instinct MI450 series computing capacity into Anthropic’s artificial intelligence infrastructure.
The agreement will use AMD Helios rack scale systems to support future training and inference workloads for Claude. The first gigawatt of capacity is scheduled for deployment during the first half of 2027, with further expansion planned under the wider agreement.
AMD has also committed to a possible equity investment of up to $5 billion in Anthropic. The final value will depend on the terms and progress of the partnership, but the commitment shows that AMD wants a deeper relationship with one of the leading developers of frontier AI models.
The deal is significant because Anthropic is not buying individual accelerators alone. It is adopting a broader AMD platform that includes graphics processors, server CPUs, networking hardware, and software.
Anthropic will use AMD Helios systems for Claude
Anthropic plans to deploy AMD Helios rack scale infrastructure to increase the computing capacity available for Claude.
The company needs large amounts of hardware for two main tasks. Training requires powerful clusters to process enormous datasets and adjust model parameters. Inference uses those trained models to generate responses for customers and applications.
| Partnership detail | Confirmed information |
|---|---|
| AI company | Anthropic |
| Hardware partner | AMD |
| Planned capacity | Up to 2 gigawatts |
| First deployment | First half of 2027 |
| Main platform | AMD Helios |
| Accelerator family | Instinct MI450 series |
| Planned GPU | Instinct MI455X |
| Server processor | Sixth generation EPYC Venice |
| Networking | AMD Pensando |
| Software platform | ROCm |
| Potential AMD investment | Up to $5 billion |
Anthropic has already used AMD Instinct MI355X accelerators, so the agreement builds on an existing relationship rather than starting from the beginning.
The next phase will move toward MI455X GPUs and the wider Helios system design. This should give Anthropic a more integrated platform for running demanding models at a much larger scale.
The deal covers the full infrastructure stack
The agreement reflects a wider change in how major AI companies buy computing hardware.
Large customers increasingly evaluate complete systems rather than selecting GPUs separately and building the rest of the platform around them. Processor performance still matters, but networking, memory, software, cooling, power delivery, and deployment speed can be equally important.
AMD’s Helios platform combines several parts of its portfolio.
The Instinct MI455X GPUs will provide the main AI acceleration. Sixth generation EPYC Venice CPUs will manage general server workloads and support the accelerators. Pensando networking technology will help connect systems and move data across the cluster, while ROCm will provide the software environment needed to run AI frameworks and models.
This full stack approach is important at gigawatt scale. Small inefficiencies in networking, power use, or software can create major costs when thousands of racks are operating together.
A 2 gigawatt deployment would be enormous
A planned capacity of up to 2 gigawatts places the agreement among the largest AI infrastructure commitments announced by a model developer.
The figure refers to the potential electrical capacity of the systems rather than their computing performance alone. It includes the power required to operate accelerators, processors, networking equipment, memory, storage, and cooling systems.
| Scale consideration | Why it matters |
|---|---|
| Power supply | Large deployments need dependable electricity |
| Cooling | Dense AI racks generate substantial heat |
| Networking | Models must exchange data across many systems |
| Memory capacity | Large models require fast access to weights and data |
| Software maturity | Hardware must work reliably with AI frameworks |
| Deployment speed | Delays can limit model development and customer growth |
The first gigawatt is expected during the first half of 2027. Reaching that target will require coordination between AMD, Anthropic, data centre operators, energy providers, equipment manufacturers, and construction partners.
The remaining capacity may be installed in later phases depending on demand, site readiness, and AMD’s ability to supply the required systems.
AMD is gaining another major AI customer
The Anthropic agreement follows other large infrastructure wins for AMD.
The company has been working to expand its position in a market dominated by NVIDIA. Major deployments are important because they give AMD more opportunities to prove that its accelerators and software can support frontier models in production.
Anthropic’s decision also indicates growing interest in hardware diversity.
Relying on several suppliers can reduce dependence on one architecture and give AI companies more flexibility when assigning workloads. Different models or tasks may run more efficiently on different systems.
Anthropic said access to computing capacity is essential for keeping Claude competitive and meeting customer demand. A wider range of hardware could allow the company to choose the most suitable platform for training, inference, and specialised workloads.
The possible $5 billion stake deepens the relationship
AMD’s potential investment in Anthropic adds a financial dimension to the hardware agreement.
An equity position could align the two companies more closely over several years. AMD would benefit if Anthropic grows, while Anthropic would have a hardware partner with a direct interest in supporting its infrastructure expansion.
The arrangement may also help both companies plan future products together. Anthropic can provide feedback about the needs of large model developers, while AMD can use that information when designing accelerators, server platforms, networking systems, and software.
The maximum investment has been described as up to $5 billion, which means the final amount may be lower or divided into stages. The supplied information does not explain the valuation, ownership percentage, or conditions attached to the potential stake.
Helios now faces a major execution test
The agreement gives AMD an opportunity to establish Helios as a serious alternative for large AI deployments.
Winning the contract is only the first step. AMD must manufacture enough accelerators, deliver complete systems on schedule, maintain reliable software support, and help Anthropic operate the infrastructure efficiently.
ROCm will be especially important. Hardware performance has limited value if developers face compatibility problems, missing software features, or difficult migration work.
AMD has improved its AI software ecosystem, but large scale production use will test it more heavily than smaller research deployments.
Anthropic will also need to optimise Claude for the platform. Efficient training and inference depend on software that can take full advantage of the available memory, compute, and networking resources.
The partnership could increase competition in AI hardware
NVIDIA remains the leading supplier of AI accelerators, but large AMD contracts could create stronger competition.
More competition may improve pricing, supply availability, software support, and product development. It could also give cloud providers and AI companies more negotiating power when purchasing hardware.
For AMD, the Anthropic agreement is a chance to show that Instinct accelerators and Helios racks can operate at the scale required by frontier AI companies.
For Anthropic, the partnership secures a large amount of future computing capacity and reduces reliance on a narrow set of suppliers.
The first gigawatt deployment in 2027 will be an important test. If AMD and Anthropic can bring the infrastructure online as planned, the deal could become one of the strongest demonstrations yet of AMD’s ability to compete in large scale AI computing.
Discussion (0)
Be the first to comment.