Compute, capital and climate this week. Anthropic reported its first-ever positive operating income on $11.5bn of quarterly revenue. A record-low Danube forced Romania to shut down its only nuclear power plant. And fusion start-ups collectively passed $7.1bn in funding, from investors who, by their own admission, are betting on AI’s growing power demand as much as they are on fusion itself.

The AI industry’s growth curve and the energy system’s stress curve are now colliding in public, in real time. For organisations building AI products, that means energy can no longer be treated as somebody else’s problem.

The pressure is productive.

Microsoft Research’s analysis of 13.5 million GitHub Copilot sessions found that 87% of LLM calls now originate from the agent itself, rather than directly from the user. A single prompt can trigger a cascade of autonomous tool calls, with agents asking models dozens of questions behind the scenes to complete one task.

That changes the equation. Agentic AI isn’t simply increasing the number of prompts we send; it is fundamentally changing how compute is consumed. More tasks are becoming multi-step, autonomous and continuously active, turning what might once have been a single model interaction into an entire chain of inference.

This isn’t a temporary spike in demand. It’s a structural shift.

At the same time, the cost of running that compute is falling fast. Google’s Gemini 3.7 Flash launched this week at roughly half the price of its predecessor, while OpenAI previewed a GPT-5.6 Sol mode running up to 14× faster on Cerebras hardware.

For anyone building AI products, that’s excellent news. Cheaper, faster inference makes more products viable and puts AI within reach. But it also creates a classic Jevons paradox: when something becomes cheaper and more efficient, we don’t necessarily use less of it. We find more reasons to use it.