Hello reader, Four years after ChatGPT's launch, the dust is settling across the hundreds of AI labs worldwide. The market is breaking into clear buckets. At one end are US frontier labs backed by Big Tech: OpenAI, Anthropic, and Google. At the other are Chinese players such as Z.ai, Moonshot, and MiniMax that have reportedly distilled frontier capabilities into open models that anyone can download and host. Then come the middle powers. Companies like India's Sarvam and France's Mistral set out to build frontier models, often with a sovereignty pitch. If the Chinese players weren't in the picture, they might have sold their less-capable models to enterprises that did not need the most advanced US systems but wanted a cheaper, locally hosted alternative. But Chinese models can be downloaded for free and run anywhere, leaving few takers for the middle powers' offerings. Today's featured story examines their workaround: hosting Chinese models for enterprises on GPUs originally bought to train and serve their own AI models. This can buy AI labs time by giving them early streams of revenue as they build toward a frontier model. But for many of them, it could signal a complete pivot from model development to an inference cloud business. Pranav Balakrishnan, AI reporter | 
