Q2 2026 Special Section - AI 2.0: Agent Demand, Engine-Room Bubbles.

Jun 29
2 min read

First, a word on how this section was made

This Special Section was researched, modelled and written entirely by AI, using Anthropic's Claude in a Cowork project, with our team as a guide throughout. That is a real change from two years ago, when tools like ChatGPT mostly helped with spelling and grammar. Claude drew on our own research and models, the open web, and some external sources. Our initial approach was deliberately broad: Claude used research agents to look into 27 separate questions about AI and power, and we then helped it narrow the work to the story told here. Along the way we challenged its claims, checked its sources and its logic, and spent real time guarding against errors and over-confidence. We estimate the work used hundreds of millions of tokens and several tens of kWh of energy, comparable to running a household fridge for a month or two. The lesson is not that the machine replaces the analyst. A person still leads, questions and edits, much as a senior analyst runs a team. We learned a great deal about how to use these tools well, and also that in many cases an experienced analyst would have reached the answer far faster than by fighting with the AI model. When using AI, having robust processes and focusing on our insights are more important than ever.

This work began with a provocation. Nvidia's Jensen Huang argued that AI could ultimately need on the order of 100 to 1,000 times more energy, which raises an obvious question: is that plausible? Our answer, in short. The demand is real and, for the next few years, larger than what can be built. But money and supply-chain limits, not megawatts, are the ceiling; a 100 to 1,000-fold rise is nowhere near feasible on any horizon we can model, even if the very long-term end state is unknowable. In our view the rush to build is already sowing the next oversupply.

AI power demand is real, and the forecasts keep rising

Two years ago, we argued that AI would be a lasting new source of electricity demand, and that the makers of electrical, cooling and power equipment would be the main winners in our universe. That view has proved correct. The clearest gauge is Vertiv, the listed company closest to what is actually being built. It has lifted its view of new data centre power from 13 to 20 GW a year in 2024 to 20 to 35 GW a year in 2026, a rise of about 75% in eighteen months. Our own base case, roughly 1,000 TWh of data centre electricity by 2030, is close to double the consensus of two years ago. Every major forecaster has moved the same way.

Download the full Special section below.