Benedict Evans' Spring 2026 deck: AI is past hype, now grinding through deployment

Benedict Evans returns with his annual macro slide deck, this time framing AI as having cleared the novelty phase and entered the messier work of integration. The argument: model capability is no longer the bottleneck — distribution, workflow fit, and economics are. Chatbot usage has plateaued in shape even as raw numbers climb, and the interesting question is which product surfaces actually convert curiosity into recurring use.

Capex remains the loudest signal. Hyperscaler spending has compounded past the point where conventional payback math works, which forces a bet that inference demand scales into the buildout rather than the other way around. Evans flags the gap between training cost curves and unit economics at the application layer, where margins are thin and switching costs are weak. Enterprise adoption is real but slower and narrower than the keynote slides suggest, concentrated in code, support, and document workflows.

The through-line is that AI is following the diffusion pattern of every prior platform — slower than the believers want, faster than the skeptics admit, and ultimately decided by whoever figures out the boring distribution problems. Open weights, agent frameworks, and the GPU supply chain all get treated as second-order issues compared to the question of where durable demand actually lives.