Anthropic disclosed in an investor presentation that it successfully ported its flagship model, Claude, to run on AMD GPUs in just three days with minimal human intervention. The company used its frontier model, Mythos, to write and port nearly 100 percent of the required software kernels. Historically, optimizing AI workloads for non-Nvidia hardware required months of specialized engineering effort due to Nvidia’s proprietary CUDA ecosystem.

According to Chief Compute Officer Tom Brown, Anthropic trains and runs Claude across four major chip families as part of a long-term multi-chip strategy. By treating compute as an interchangeable pool, the company aims to optimize its price-to-performance ratio across vendors while bypassing hardware bottlenecks. The lab is also establishing “Anthropic Direct” to lease and operate custom data centers directly.

If Anthropic’s claims hold up under external verification, automated code translation could significantly lower the software barrier protecting Nvidia’s compute monopoly and 70+ percent gross margins. However, the reported data originates directly from Anthropic’s investor materials and has not been independently verified.

Why it matters

  • Demonstrates AI-driven kernel translation can bypass Nvidia CUDA lock-in, enabling hardware switching.

  • Allows AI startups to negotiate better compute pricing by diversifying across AMD and other chip providers.

  • Highlights Anthropic’s strategy to reduce infrastructure TCO via direct data center leasing.

Source: the-decoder.com