In-House Chips: Cost Dilemmas and Strategic Divergences Among Large Model Companies

In principle, it's no coincidence that large model companies are collectively "making chips." Let's first understand a simple concept: the computational power consumption for large model inference is hundreds of times higher than during the training phase, and it happens continuously 24/7. When Claude demonstrates astonishing long-text understanding capabilities, the underlying computational demand is already approaching the economic limits of traditional CPU-GPU architectures. Estimates suggest that maintaining inference for tens of millions of daily active users for frontier models like Claude could result in chip depreciation costs alone reaching hundreds of thousands of dollars per day. This financially bleeding model forces Anthropic to break through upstream in the supply chain.