Advanced Micro Devices has secured a potentially transformative AI infrastructure commitment from Anthropic, with the developer of the Claude model family planning to deploy up to two gigawatts of AMD Instinct MI450-series accelerators. The agreement, announced July 22, positions AMD’s next-generation Helios rack-scale architecture for one of the largest disclosed deployments by a leading artificial-intelligence laboratory.

The companies said deployment of the first gigawatt is expected to begin in the first half of 2027. Anthropic will use Helios systems featuring the Instinct MI455X, the flagship accelerator within AMD’s MI450 series, alongside EPYC processors code-named Venice, Pensando networking components and the company’s ROCm software platform. The announcement did not specify the number of accelerators to be purchased, the locations of the planned installations or the total value of the computing contract.

A gigawatt measures electrical capacity rather than computing performance or chip quantity. Its use as the central unit of the agreement reflects the scale at which frontier AI infrastructure is now being planned. The ultimate number of systems will depend on rack configurations, utilization assumptions, cooling requirements and the power allocated to supporting equipment such as networking, storage and data-center operations.

The phrase “up to two gigawatts” also makes the agreement a maximum commitment rather than a guarantee that the full capacity will be installed on a fixed timetable. The first-gigawatt milestone provides the clearest near-term target, while the second phase will depend on execution, infrastructure availability, Anthropic’s demand trajectory and the performance of AMD’s hardware and software in production environments.

Even with those qualifications, the agreement is a major commercial validation for AMD. The market for high-end AI accelerators has been shaped by customers’ need for tightly integrated systems capable of training increasingly complex models and serving large volumes of inference requests. Winning a large-scale commitment from Anthropic gives AMD an opportunity to demonstrate that its products can support both workloads at frontier-model scale.

The deployment will build on Anthropic’s existing use of AMD Instinct MI355X GPUs. That detail is significant because it indicates that the relationship is not beginning with an untested platform. Experience operating Claude workloads on the MI355X can provide engineering data, migration knowledge and performance baselines as the companies prepare for the more ambitious MI455X rollout.

AMD’s Helios design integrates 72 MI455X accelerators within a rack-scale system. According to AMD’s published specifications, a full configuration is designed to deliver up to 2.9 exaFLOPS of theoretical FP4 computing performance and 1.4 exaFLOPS at FP8, with 31 terabytes of high-bandwidth memory. Such lower-precision formats are increasingly important for AI training and inference because they can increase throughput and improve energy efficiency when models and software are designed to use them effectively.

The MI455X is based on AMD’s next-generation CDNA architecture and incorporates HBM4 memory. High memory capacity and bandwidth are central requirements for frontier AI systems, particularly when models, active data and intermediate calculations must be distributed across large numbers of accelerators. Improvements in memory and interconnect performance can reduce the time processors spend waiting for data, although realized performance depends heavily on software optimization and workload characteristics.

Helios also represents AMD’s effort to compete at the level of the complete rack rather than the standalone GPU. The design combines accelerators, CPUs, networking, cooling, power management and software into a reference architecture that manufacturers and infrastructure operators can adapt. This approach addresses the operational reality that large AI customers increasingly procure coordinated systems instead of assembling individual components independently.

The architecture uses AMD EPYC Venice processors as hosts and Pensando technology for networking and data movement. AMD has emphasized open networking standards and compatibility with the Open Compute Project’s Open Rack Wide format. That strategy is intended to give cloud providers and data-center operators more flexibility when integrating systems, while differentiating Helios from more proprietary infrastructure designs.

Software will be an equally important test of the partnership. Nvidia’s position in AI computing has been supported not only by accelerator performance but by an extensive software ecosystem that allows developers to build, optimize and deploy models across multiple generations of hardware. AMD has expanded ROCm’s compatibility with widely used frameworks and libraries, but large production deployments remain essential for strengthening developer confidence and identifying performance gaps.

AMD and Anthropic therefore plan a multi-year engineering collaboration centered on Claude and ROCm. Their teams will work to optimize training and inference workloads for Instinct accelerators, while Claude will be used to accelerate AMD software development. AMD said it will also adopt Claude broadly across its engineering and product-development teams, making Anthropic both an infrastructure customer and an enterprise AI supplier to the chipmaker.

The arrangement creates a feedback loop that could benefit both sides. Anthropic can work directly with AMD engineers to tune kernels, compilers, communication libraries and scheduling systems for Claude workloads. AMD, in turn, can receive detailed information about the requirements of a major frontier-model developer and apply those findings to ROCm, future accelerators and rack-scale products offered to other customers.

Such optimization is particularly important because headline accelerator specifications do not automatically translate into equivalent application performance. Frontier models require thousands of processors to operate as a coordinated system. Bottlenecks can emerge in collective communications, memory management, network congestion, compiler behavior, fault recovery and orchestration. A deep engineering partnership can address those issues before and during deployment rather than leaving the customer to manage them after installation.

The commercial relationship is reinforced by AMD’s commitment to make a strategic equity investment of up to $5 billion in Anthropic. The companies characterized the investment as a future commitment, but did not disclose its timing, valuation, governance provisions or whether funding will be released according to deployment milestones. The investment nevertheless aligns AMD’s financial interests with Anthropic’s expansion and underscores the strategic value chipmakers now place on securing major AI-model customers.

The structure also demonstrates how capital, hardware and computing access are becoming increasingly interconnected. Frontier AI developers require enormous infrastructure commitments before the associated revenue is fully realized. Semiconductor companies, cloud providers and other technology partners may therefore use investments, financing support or long-term commercial arrangements to help customers secure capacity while establishing demand for their own platforms.

For Anthropic, AMD becomes an additional component of a deliberately diversified computing portfolio. The company has previously described the use of Amazon Trainium accelerators, Google tensor processing units and Nvidia GPUs, allowing workloads to be assigned according to cost, availability and technical suitability. In April, Anthropic announced a commitment for up to five gigawatts of new Amazon capacity over time, as well as a separate expansion involving multiple gigawatts of future Google and Broadcom technology.

The AMD deployment does not displace those relationships. Instead, it broadens the range of hardware available to train and serve Claude. Anthropic has said that hardware diversity can improve resilience and enable the company to match individual workloads with the most appropriate chips. A broader supplier base may also give the company more flexibility when negotiating capacity, planning model releases and responding to regional demand.

That strategy carries complexity as well as potential benefits. Supporting several accelerator architectures requires engineering teams to maintain different software toolchains, performance libraries and operational procedures. Models must be tested across platforms, and differences in numerical formats, memory architecture and networking can affect output consistency and efficiency. Anthropic’s direct engineering collaboration with AMD is intended to reduce those costs by optimizing the stack jointly.

The deal gives AMD a prominent opportunity to expand its share of the AI infrastructure market as customers seek alternatives and supplements to incumbent systems. Large model developers and cloud operators increasingly want multiple accelerator suppliers to improve supply security, preserve negotiating leverage and ensure that capacity expansion is not constrained by a single product roadmap.

AMD must still execute against demanding production requirements. The MI455X and Helios platform will need to be manufactured, integrated and deployed at high volume, while the broader data-center ecosystem must supply sufficient power, liquid cooling, networking and construction capacity. Any delay involving chips, memory, facilities or software could affect the first-half 2027 timetable.

The use of gigawatt-scale commitments also highlights the energy implications of continued AI growth. New clusters require access to electrical generation, transmission infrastructure, substations and cooling systems, often on timelines longer than semiconductor development cycles. The announcement did not identify power suppliers, data-center operators or geographic markets for the Anthropic installations, leaving important infrastructure details unresolved.

Those details will determine how rapidly the commitment becomes operational. Accelerator availability alone is insufficient when a project also requires land, interconnection approvals, equipment procurement and specialized construction. Infrastructure partners may deploy Helios systems through cloud facilities, dedicated campuses or a combination of operating models, but AMD and Anthropic have not publicly outlined that structure.

For enterprise customers using Claude, the practical significance will depend on whether expanded capacity produces greater availability, faster responses and more predictable pricing. More computing infrastructure can support larger training runs, stronger models and higher inference volumes, but the economics will also depend on utilization rates and operating efficiency. Anthropic will need to balance investments in future model capability with the cost of reliably serving existing customers.

The first-gigawatt deployment will consequently serve as a major benchmark for AMD’s AI ambitions. Successful implementation would demonstrate that Helios and ROCm can sustain demanding frontier workloads at industrial scale, potentially supporting further orders from Anthropic and other model developers. It would also give AMD a reference customer capable of influencing infrastructure decisions across cloud computing, enterprise AI and the wider semiconductor market.

For Anthropic, the agreement secures another path to the computing resources required to develop Claude while preserving a multi-platform strategy. For AMD, it combines a large prospective hardware deployment, a software-development partnership, an enterprise adoption commitment and a substantial equity investment. The result is a relationship spanning nearly every layer of the AI technology stack, from capital and data-center capacity to silicon, networking, software and production model workloads.