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

AI power demand will exceed what can be built in the next few years, but financial and supply-chain constraints will cap growth far below the 100- to 1,000-fold increase envisaged by Nvidia’s Jensen Huang, while the current buildout is already creating future oversupply.

• The section was researched, modelled and written with Anthropic’s Claude, using research agents to examine 27 questions, the firm’s own research and models, and external sources; the team checked the claims, sources and logic.
• The authors estimate that the work consumed hundreds of millions of tokens and several tens of kWh of energy, and conclude that AI should support rather than replace analysts because experienced researchers can often reach conclusions faster.
• Vertiv, viewed as the listed company closest to actual data-centre construction, increased its forecast for annual new data-centre power from 13–20 GW in 2024 to 20–35 GW in 2026, approximately 75% higher in eighteen months.
• Clean Energy Transition’s base case is roughly 1,000 TWh of data-centre electricity consumption by 2030, nearly twice its consensus estimate from two years earlier, with all major forecasters also raising their projections.
• The firm says its earlier view that AI would create lasting electricity demand and benefit electrical, cooling and power-equipment makers has proved correct, but argues that money and supply-chain capacity—not megawatts—will ultimately constrain deployment.

Takeaway: Investors should focus on the equipment makers benefiting from near-term AI infrastructure demand while recognising that the current rush to build is already sowing the conditions for a later oversupply; the feasibility of extreme long-term demand remains unknowable.