[DigitalToday reporter Chi-gyu Hwang (황치규)] Anthropic has signed a multibillion-dollar deal with AMD to secure up to 2 gigawatts of GPU capacity.

SiliconANGLE reported on Tuesday that AMD plans to invest up to $5 billion in Anthropic and broadly introduce the Claude AI model into its software development.

Anthropic will adopt AMD’s MI450 series GPUs and install them in systems based on the Helios rack design. The two companies plan to deploy this hardware from the first half of 2027.

A single Helios system contains 72 GPUs and delivers up to 1.4 exaflops based on FP8 data processing. It also includes 31 terabytes of HBM4 memory. This capacity is 50 percent more than Nvidia’s flagship Vera Rubin system. Helios also supports scale-out bandwidth of 43 terabits per second.

Helios will incorporate central processing units and Pensando data processing units in addition to GPUs. The Pensando chips track hardware reliability indicators such as latency and handle more than 12 infrastructure management tasks, including load balancing and traffic encryption.

The two companies will also work over several years to improve AMD’s software development efficiency. They will focus in particular on accelerating development of ROCm, the software platform provided with Helios. ROCm includes a compiler that converts customers’ AI models into AMD-compatible formats and tools that improve training and inference performance.

The agreement extends AMD’s large AI hardware deals won so far this year. Microsoft said earlier this week it is deploying Helios in its cloud data centres. Meta disclosed in February plans to buy up to 6 gigawatts of GPUs from AMD. Meta’s initial order is a customised GPU based on the MI450 line adopted by Anthropic.