Anthropic is treating graphics processing units as fungible units of compute, a shift that could reduce the importance of the NVIDIA CUDA ecosystem, according to IPO roadshow material from the company. The company's Claude Mythos Preview model reportedly required only three days to determine how to run another model on AMD MI355X GPUs. This process typically takes between six and 12 months when preparing a new compute platform for inference-related operations.
The ability of highly capable models to port other models to new platforms with relative ease suggests a significantly reduced barrier to switching between hardware providers. Some observers suggest this development could signal the end of NVIDIA's CUDA moat for all practical intents and purposes. However, the demise of the ecosystem may be premature given that it consists of a massive collection of libraries, tools, and years of progressive optimization, alongside a large number of developers who continue to build on the platform.
NVIDIA has implemented several strategies to ensure the resilience of its CUDA moat. These measures include providing free GPUs and training to university researchers and coupling the NVIDIA Collective Communications Library (NCCL) with proprietary hardware such as InfiniBand and NVLink. The NCCL is a tool and library used for connecting multiple GPUs.
