AMD has signed a strategic partnership with Anthropic and will make an equity investment of up to $5 billion (about 6.9 trillion won). With AMD adding Anthropic to its roster of major customers after OpenAI, Meta, Oracle and Microsoft, cracks are appearing in the AI accelerator market long dominated by Nvidia as the sole powerhouse.
◇ Large-scale supply of "Helios" AI accelerators to Anthropic
On the 22nd (local time), on the first day of "AMD Advancing AI 2026," which opened at Moscone West in San Francisco, AMD announced the partnership. The core of the deal is supplying Anthropic with the "Helios" AI accelerator rack solution with a capacity of up to 2 gigawatts (GW).
Under the agreement, AMD will begin building the first 1 GW of facilities in the first half of 2027. One GW is roughly equivalent to the output of a single nuclear power plant, indicating that Anthropic is moving to secure AI infrastructure on a massive scale. The remaining 1 GW will be added in phases, bringing the total contract scale to the power output of two nuclear plants.
Anthropic plans to build data centers by bundling AMD's latest AI Semiconductor graphics processing unit (GPU) "MI455X," the EPYC server central processing unit (CPU) "Venice," Pensando networking for data transfer, and operating software. In particular, Anthropic's large expansion of its collaboration with AMD underscores confidence in AMD's AI accelerators, which is significant.
AMD also plans to pursue a strategic equity investment of up to $5 billion in Anthropic. Because it is uncommon for a chipmaker to supply a customer with a large volume of GPUs and also invest equity, some say the two have formed a tight relationship that goes beyond a simple "supplier-customer" arrangement and extends into capital ties.
◇ From Claude to chip development… stronger software ties
Separately from the hardware supply deal, the two companies also announced a plan to use Anthropic's AI model Claude directly in AMD chip development. AMD aims to use Claude as a development assist tool in efforts to improve its chips and software, raising the completeness of "ROCm," a software platform designed to extract maximum GPU performance on real workloads.
AMD plans to go a step further by rolling out Claude broadly across its engineering and product development organizations. In the chip industry, strengthening ties with customers from the software development stage rather than stopping at hardware sales is widely seen as a response to Nvidia's developer lock-in strategy built around the CUDA ecosystem.
The announcement took on greater significance alongside previously released news of a partnership with Microsoft's Azure. That is because AMD has effectively secured all major AI corporations—OpenAI, Meta, Oracle, Microsoft and Anthropic—as customers. In a briefing the same day, AMD said eight of the world's top 10 AI corporations are already running workloads on its Instinct GPUs.
An AMD official said, "Until now, the AI accelerator market was effectively controlled by Nvidia alone," and added, "As AMD secures large AI corporations one after another as customers, it is increasingly establishing itself as a practical alternative supplier."