Energy, chips and sovereignty: the physical economics that will decide where AI lands
Power, water, compute access, data residency and the EU's regulatory identity — as inputs to strategy
The question: Models improve on the fast clock; power grids, fabs and regulation move on the slow one. Which constraint will actually decide what your organisation can do with AI in 2028?
The binding constraint on AI is moving from algorithms to atoms: electricity, cooling water, chips and grid connections with multi-year lead times, plus a fragmenting regulatory map. Leaders who plan only on model capability plan on the wrong clock (Lesson 30).
What the lesson covers
Energy. The IEA's 2025 "Energy and AI" report put data-centre electricity demand on a path to roughly double by 2030, with AI the main driver — concentrated in a few regions, bottlenecked on grid connections and generation that take years to build. Water for cooling and local opposition add to it. For a firm this is not abstract: inference at volume is an energy bill and a location decision, and the cheapest tokens will increasingly be where the power is.
Chips and compute. Advanced accelerators come from a handful of firms and fabs; export controls and national strategies make access a geopolitical variable. "Sovereign AI" — national compute, national models, national data rules — is now policy in the EU, the Gulf, India and elsewhere. The practical consequence: where your data may live, which models you may run, and at what price, can change with a regulation or a trade decision, which is Lesson 8's switching-door argument with a flag on it.
The EU's regulatory identity. The AI Act (Lesson 18) plus the 2025 General-Purpose AI Code of Practice, data-protection and data-residency rules shape what can be deployed where; the EU is betting on trust and rules as its competitive position. For a European firm this is both constraint and asset: compliance is a cost, and "compliant by design" is a thing a regulated customer will pay for.
Efficiency as strategy. The same constraints make small and routed models (Lessons 8, 29), on-premise or regional hosting, and energy-aware workloads into competitive choices rather than engineering niceties. The firm that can do the job with a tenth of the compute is the firm that is not hostage to the grid connection.
Plan on two clocks. The leader's job is to keep a watchlist with triggers (Lesson 30) for the physical constraints as well as for the model releases: power price and availability in your regions, compute access, residency rules, the Code of Practice's obligations — each paired with the decision it would change. Optionality (Lesson 20) is the posture: avoid irreversible dependence on one region, one vendor or one rulebook.
Key points
- Data-centre electricity demand is on a path to roughly double by 2030, bottlenecked on grids and generation (IEA 2025).
- Chips and compute are geopolitical variables; sovereign-AI policy shapes where data may live and which models may run.
- The EU's rules are constraint and asset: compliance costs, and "compliant by design" sells to regulated buyers.
- Efficiency — small and routed models, regional hosting — is strategy, not engineering taste.
- Plan on two clocks: a watchlist with triggers for power, compute, residency and rules, and optionality against single dependence.
Framework — Power → Compute access → Residency & rules → Efficiency choices → Triggers
Five physical and regulatory constraints, each mapped to a decision it could change and a trigger that would change it.
The lab
Write the physical-constraints brief a board has not asked for yet.
Open this lesson, its lab and its quiz
Sources and further reading
- Energy and AI — IEA, 2025
- General-Purpose AI Code of Practice — European Commission, 2025
- Lesson 30 — the slow clock and the bottleneck watchlist — Applied AI Academy