Anthropic's IPO roadshow video reveals details of its multi-chip compute strategy, confirming that the company trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs. Co-founder Tom Brown described an early bet on hardware diversity, noting that each architecture requires a tailored software and inference stack. Anthropic has developed internal tooling, referred to as Mythos, to automate kernel porting; this system enabled Claude to run on AMD GPUs for the first time in just three days with minimal human intervention, writing or porting nearly 100% of the required kernels. The company frames its advantage around "total effective compute," defined as the product of physical capacity and discovered compute multipliers. A 4x multiplier, for example, reduces the physical compute needed to train a given model by 75% or allows a significantly more capable model within the same budget. Anthropic states these multipliers compound and stack, creating a self-reinforcing advantage that is difficult to replicate. The firm anticipates this cycle will accelerate toward recursive self-improvement, where Claude autonomously discovers further multipliers, increasing the rate of discovery. Previous evidence of this multi-chip approach appeared in a September 2026 postmortem detailing bugs isolated to TPU deployments.