AI Devours the World (Spring 2026) [pdf]

The 2026 AI infrastructure investment plans of Microsoft, Google, Amazon, Meta total $700bn, more than double the global telecom industry ($300bn). These companies have moved away from asset-light models, with capex-to-revenue ratios rising to 30-50%.

OpenAI's monthly revenue has grown explosively since December 2022, reaching roughly $4-5bn annualized as of mid-2026, and Anthropic is also growing rapidly. ChatGPT recorded over 900 million weekly active users, but the paid conversion rate is only 5%, and over 80% of users enter fewer than 1000 prompts annually, meaning it has not yet become an everyday essential tool.

Enterprise adoption is spreading rapidly but often remains stuck at the pilot stage. Coding automation is delivering results first, and according to a16z data, coding is the overwhelming #1 category in enterprise AI spending. Marketing, customer support, and back-office automation follow.

Model commoditization hypothesis: As benchmark performance across major frontier models converges and there are no network effects, it is suggested that LLMs themselves could become a low-margin infrastructure layer. The analysis suggests a path similar to 2010 mobile data networks — explosive demand growth → capacity shortage → pricing model restructuring → commoditization.

Supply bottlenecks: Multi-year backlogs are occurring in GPUs, memory, and power supply, and data center construction volume has surpassed office construction in the US. Semiconductor industry capex is also surging but still failing to keep pace with demand.

Benedict Evans offers a tentative hypothesis: "Chat interfaces are not good UX, and general-purpose use requires apps. Models will become commodities, and innovation will move up the stack."