Gavekal Technologies

Beijing is all-in on AI agents: its AI-Plus strategy calls for agents to be nearly ubiquitous in the economy by 2030. But widespread adoption still faces many hurdles: agent reliability is still spotty, training data is hard to access and to standardize, and general agent-builders such as Alibaba are short on the domain expertise needed to build industry-specific agents.

Conventional wisdom has held that while China’s AI buildout is constrained by lack of access to advanced chips, the US will be slowed by the scarcity and cost of electricity. Solving the power constraint is indeed a problem: AI data centers are suddenly adding a lot of new electricity demand on a system that has not grown in two decades. But aside from building new power plants and transmission lines—a long slog—there is a lot the US can do to squeeze more juice out of the current infrastructure.