Will AI Energy Demand Taper Off?

Anyone following the AI boom will have heard two arguments pulling in opposite directions. The first says hardware and algorithms are improving so fast — chips getting quicker per watt, models shrinking through quantisation and distillation, software stacks tightening at every layer — that energy use should eventually level off. The second says the opposite: tech giants are signing multi-gigawatt power deals, bringing retired nuclear plants back online, and forecasting demand curves that only go up. I wanted to work out which story actually holds, so I went through the industry data myself. My conclusion is that the second camp has it right. Efficiency keeps improving, but it is not enough on its own. Total demand is going to keep accelerating for the rest of the decade.

Why Efficiency Doesn't Mean Less Energy

The reason the efficiency argument falls short has a name: Jevons' Paradox. Back in 1865, the economist William Stanley Jevons noticed something counterintuitive about James Watt's improved steam engine. Watt's design used coal far more efficiently than earlier engines, so logic would suggest England's coal consumption should have fallen. Instead, it took off. Cheaper, more practical power meant coal got used in far more places for far more purposes, and the efficiency gain was swallowed many times over by growth in total use.

I think we are watching the same mechanism play out with AI, just compressed into a much shorter timeframe. The cost of running a single inference task, or generating one token at a fixed level of quality, has been falling far faster than most people realise — Epoch AI puts the rate anywhere from 9x to 900x per year depending on the benchmark, with a typical figure closer to 40x — as chips, precision formats, and software stacks all improve together. That falling cost does not sit still; it unlocks work that was not economical before: background agents running continuously, multi-step reasoning chains, real-time video generation, assistants embedded in everything. Each is a new source of demand that did not exist when running a model cost ten times as much. Multiply the growing number of tasks by the number of queries each one generates, and the resulting demand outpaces the efficiency curve by a wide margin. Cheaper compute does not reduce total power draw; it expands the market for compute, and the expanded market wins.

What the Forecasts Actually Say

It is one thing to reason about a mechanism in the abstract, and another to see whether the people who model energy grids for a living agree. When you look at the major forecasts, the pattern holds up consistently across the board.

The International Energy Agency modelled global data centre electricity demand under several scenarios, including ones assuming aggressive efficiency gains across chips and cooling. Even in those optimistic cases, demand does not plateau. The IEA's central projection has worldwide data centre power roughly doubling from about 415 terawatt-hours (TWh) in 2024 to more than 940 TWh by 2030 — accounting for around 3% of all global electricity generation.

McKinsey looked specifically at the United States, where the shift is even starker. In 2023, American data centres drew about 147 TWh, or roughly 3.7% of the national grid. By 2030, McKinsey projects that figure will reach 606 TWh — nearly 12% of total US electricity demand. That is the data centres' share of national power more than tripling in just seven years.

Gartner's research tracks where that power goes. Its most recent forecast shows AI-optimised servers overtaking all conventional server workloads combined by 2027, drawing roughly 258 TWh that yeargartner-2026. An earlier Gartner report warned that the physical grid cannot keep pace either way: power availability is projected to operationally constrain 40% of existing AI data centres by 2027gartner-2024, a concern echoed by the OECD's own analysis of AI's grid impactoecd-power.

Why This Isn't a Repeat of the Cloud Story

There is an obvious objection to all this: didn't we already go through an efficiency-versus-growth story once before, and come out fine? During the 2010s, global data centre compute capacity grew by more than 500%, yet total electricity consumption barely movedmasanet-2020. That split was real. The industry was moving from inefficient, on-premises server rooms scattered through every company basement to centralised hyperscale facilities with power usage effectiveness ratings around 1.1 to 1.2 — about as efficient as a data centre gets.

The AI build-out cannot repeat that trick, because the underlying physics have changed. A typical hyperscale cloud rack drew around 5 to 15 kW, cooled by ordinary perimeter air conditioning. A modern AI rack, built around NVIDIA's H100, H200, and Blackwell chips — each drawing 700 to more than 1,000 watts — routinely pulls 30 to more than 100 kW and needs direct-to-chip liquid cooling or immersion to control the heat. The workload has changed too, from CPU-driven web services and databases to GPU- and ASIC-driven parallel computation. The binding constraint has shifted from land and real estate to power generation and grid interconnection. Hyperscale cloud already captured the easy efficiency wins. What is left is raw electrical load, concentrated into a much smaller physical footprint than anything the industry has built before.

The Real Constraint Is the Grid, Not the Chip

If data centre growth does stall out over the next few years, I do not think it will be because anyone ran out of demand, or because some efficiency breakthrough made further scaling pointless. It will be because of what sits between the power station and the rack.

Three separate bottlenecks are compounding each other. First, securing grid approval and transmission interconnects for large campuses now takes three to seven-plus years in key hubs like Northern Virginia, Texas, Frankfurt, and Dublin. Second, the physical supply chain — high-voltage transformers, switchgear, and backup generators — is backlogged three to four years. Third, regional grids are running near peak reserve margins, prompting utilities and regulators to scrutinise or halt major new connection requests.

Put those three together and it is not surprising that around 40% of planned AI data centre projects are facing energisation delays. That pressure is exactly why operators are increasingly turning to co-located generation, on-site microgrids, and long-term power purchase agreements with nuclear and geothermal providers — not because those are the cheapest options, but because they are the only way to get guaranteed power on a timeline that matches the build.

Where I've Landed

Pulling this together, here is the argument as I now see it. Efficiency really is improving, and that is not a mirage — per-token costs are falling fast, and that is genuine progress. But Jevons' Paradox means falling cost is exactly what drives total consumption up, because it makes an entire tier of workloads viable that could not have existed at the old price point. Layer the forecasts on top of that mechanism — global demand roughly doubling by 2030, data centres claiming double-digit shares of major grids, and AI hardware claiming the lion's share of that load — and the trajectory is clearly accelerating, not flattening.

The genuine limiting factor from here is not silicon. It is gigawatts — specifically, how quickly civil infrastructure can deliver high-voltage power to the rack. That is a civil engineering and regulatory challenge as much as a technological one, and it is the one that will determine how far AI can scale.

Sources and Further Reading

• International Energy Agency (IEA): Energy and AI — Electricity forecasts
• McKinsey & Company: AI's Power Binge — charting US data centre demand to 2030
• Gartner Research: Data centre electricity demand to grow 26% in 2026, AI servers to overtake conventional by 2027, and power shortages will restrict 40% of AI data centres by 2027
• OECD Economics Department: Wired for power — the energy behind the AI revolution
• Masanet et al. (2020), Science: Recalibrating global data center energy-use estimates
• Epoch AI: LLM inference prices have fallen rapidly but unequally across tasks