AI's real limit isn't the model. It's the power, water and chips behind it.

Your AI plans almost certainly assume the compute will simply be there: cheap, abundant, on tap, like electricity from a socket. It is a reasonable-feeling assumption, and it is quietly becoming the riskiest part of the plan. The models get the headlines and the excitement, while the power stations, water and chips they depend on get the constraints, and it is those constraints, not the cleverness of the model, that will decide whether your AI roadmap is actually deliverable.

This is the argument of a 2026 World Economic Forum report, Building Resilient and Scalable AI Value Chains. It makes the case that AI's growth is bounded by a physical value chain most strategies never look at: energy, which it calls "a highly visible choke point"; water, "the overlooked AI growth limiter"; and behind them chips, data centres, raw materials and the capital to build all of it. Its recommendation is to plan AI around those resource limits rather than treating the software as the whole story.

How real is the physical limit, in numbers?

Real enough that the energy sector is now planning around it. The International Energy Agency estimates that data centres consumed roughly 415 terawatt-hours of electricity in 2024, about 1.5% of global supply, and that this has been growing at around 12% a year. It projects demand will more than double to about 945 TWh by 2030, slightly more than the entire electricity consumption of Japan today, with AI the single biggest driver. In the United States specifically, the IEA expects AI processing to require more electricity than all heavy industry, steel, cement and chemicals, combined. Water, used in vast quantities to cool these facilities, is the quieter version of the same story.

Put plainly, the thing you are treating as an infinite utility is a finite, contested and fast-growing draw on real resources. Compute is not free of physics, and as demand doubles, the assumption that it stays cheap and available everywhere gets harder to bank on.

Why should a physical constraint be on my strategy agenda?

Because it arrives in your plan as cost, availability and risk, the three things strategy is supposed to manage. Power and water near data centres are finite and increasingly contested; grid connections face real queues; chips and capacity sit with a handful of suppliers and regions. As demand doubles, access gets harder and prices get less predictable, so the AI you budgeted as cheap may not stay that way. None of this is exotic. It is ordinary supply-chain risk, familiar to any serious leader, simply applied to your newest and least-examined dependency. It is the same blind spot behind why most organisations fail at AI adoption: the exciting part gets the attention while the enabling conditions are assumed.

How should this change the way I plan AI?

By treating compute as a supply chain to be managed, not a tap that is always open.

1. Treat compute as a supply chain, not a utility. Know where your AI capacity actually comes from, who controls it, and how exposed you are to shocks in price or availability.
2. Budget for cost volatility. Do not assume today's price per unit of compute holds. Model scenarios in which it gets scarcer or dearer as global demand doubles.
3. Match ambition to what is deliverable. Sweeping AI plans that assume unlimited cheap compute can stall on grid queues and capacity limits. Scope to what the value chain can realistically supply.
4. Make efficiency a strategy, not an afterthought. The cheapest and most resilient megawatt is the one you never needed. Efficient models and disciplined use cut both cost and exposure at once.
5. Ask the resilience questions. Concentration among a few suppliers and regions, energy and water availability, and continuity if a link is disrupted, these belong on the board agenda, not just in procurement.

What does this change for me as a leader?

It adds a dimension to AI strategy that most plans leave out entirely: the physical one. The organisations that win with AI will not only have the best use cases. They will also have asked whether the power, water, chips and capital to run those use cases at scale are actually, reliably there, and what happens to their costs and plans if they are not. That is not pessimism. It is the ordinary discipline of not betting your strategy on a resource you have never checked the supply of.

It is the same clear-eyed posture running through the end of business as usual: the tools are extraordinary, and they still run on a finite, physical world. Treat AI as resting on a real value chain with real limits, and you plan with your eyes open, rather than discovering the constraint the hard way when the grid, the water or the budget says no. These are exactly the kind of dependency questions a board should be asking before it commits to an AI-heavy future.

About the author

Thomas Green

British technology futurist, AI keynote speaker and advisor. Thirty years across enterprise technology and AI strategy, helping leaders navigate the future of work. The futurist who died.