AI Data Centers and the Power Grid: A Stock-and-Flow Analysis of Surging Demand

Last updated: 29 October 2026.

AI data center power demand is growing faster than the electricity system can be expanded, and the gap is a systems problem before it is an engineering one. Demand for computing can be added in months. Generation, transmission and grid connections take years. When a fast-moving stock meets a slow-moving one, the outcome depends on the delay between them, and the delay is where the stress shows up. This article uses a stock-and-flow lens to explain what the best available numbers say, why the mismatch persists, and what the structure implies for the next decade.

What the numbers say

Three sources frame the discussion. All figures below are the ones their publishers report; projections are scenarios, not forecasts of certainty.

• Level today. The IEA’s Energy and AI report estimates data centres used about 415 TWh of electricity in 2024, roughly 1.5% of global electricity consumption.
• Trajectory. In its base case the IEA expects consumption to reach about 945 TWh by 2030, a little more than double the 2024 level and roughly the electricity use of Japan today. By our arithmetic this is compound growth of about 15% a year.
• Where it concentrates. The United States and China together account for nearly 80% of the projected growth to 2030, according to the IEA’s analysis as reported in the press.
• Uncertainty. The IEA estimates AI-specific servers were around 15% of data centre energy demand in 2024. Independent researchers, including Alex de Vries, argue the AI share is probably understated, and the IEA itself notes that data centre operators disclose energy use only in limited ways.

Converting energy to power helps the intuition. 415 TWh a year is an average draw of roughly 47 GW (415 TWh divided by 8,760 hours). 945 TWh is roughly 108 GW. The increase of about 60 GW of continuous load is the equivalent of dozens of large power plants running around the clock, concentrated in a small number of regions.

A stock-and-flow model of the problem

Systems dynamics models describe the world in stocks (things that accumulate) and flows (rates that fill or drain them). Two stocks matter here.

1. Demand stock: the electrical load that data centres want to connect, driven by investment in AI computing. The inflow is new data centre capacity planned or contracted. It responds to commercial incentives and can move in a year or two.
2. Supply stock: the generation and grid capacity actually available to serve that load. The inflow is capacity commissioned, which must pass through study, permitting, procurement, construction and connection. The outflow is retirement.

The two stocks are linked by a balancing loop: when demand exceeds supply, prices rise and political pressure builds, which in time brings new supply. But the loop contains a long delay. In systems terms, a balancing loop with a long delay tends to overshoot or oscillate, because decisions are made on a picture of the system that is already years out of date. We discuss that mechanism more generally in delays in systems thinking.

Figure 1. A conceptual stock-and-flow picture, not data. Demand can ramp quickly; supply follows in steps after a long delay, leaving a growing gap.

Why the delay is so long

Lawrence Berkeley National Laboratory’s Queued Up: 2026 Edition tracks power plants seeking transmission interconnection in the United States as of the end of 2025. Its findings explain why the supply side lags:

• About 1,312 GW of generation and roughly 749 GW of storage were waiting in interconnection queues.
• For projects that were built in 2025, the median time from interconnection request to commercial operation was more than five years in the regions with available data.
• Of projects that requested interconnection between 2000 and 2020, only 13% had reached commercial operation by the end of 2025, while 75% had withdrawn.
• Solar (about 773 GW) was the largest generation type in the queue, but gas (about 253 GW) grew 86% over the prior year.

Two lessons follow. First, a large queue is not the same as future supply: most of what enters a queue never gets built. Second, the queue itself is a stock with its own dynamics. Studies are sequential, withdrawn projects force re-studies of others, and uncertainty makes developers submit speculative requests, which lengthens the queue for everyone. That is a reinforcing loop inside the balancing one.

The IEA adds another signal. It reports that around 20% of planned data centres could face delays from grid connection constraints. This is a demand-side stock waiting on a supply-side flow.

What systems thinking predicts

Workarounds that move the problem

When the legitimate path is slow, participants find faster ones. For data centres that means siting in regions with spare capacity, building generation on site, signing power deals with existing plants, or delaying retirements of older plants. Each fix works locally and shifts pressure elsewhere. This is the pattern known as shifting the burden, and several of these choices lock in fossil capacity for decades, which connects to the dynamics of carbon lock-in.

Efficiency gains that do not reduce total demand

More efficient chips and cooling lower the energy per unit of computing. But if cheaper computing leads to far more computing, total consumption can still rise. The mechanism, known as the rebound effect or Jevons paradox, is covered in the rebound effect in systems thinking. The IEA’s own scenarios reflect this uncertainty: the spread between its lower and higher cases is large because nobody knows how fast demand for AI services will grow.

A system that grows into its limits

A fast-growing stock that depends on a slower-growing resource eventually meets the constraint. The classic structure is described in limits to growth. In the data centre case the limit may not be total energy but local grid capacity, water for cooling, equipment such as transformers, or skilled labour. The first limit hit becomes the new bottleneck.

Flexibility as a lever

Not all computing loads must run at the same time. Training workloads and some batch inference can be scheduled to match times when the grid has slack, and facilities can reduce draw in short peaks. Flexible load changes the structure: it turns a rigid demand stock into one that responds to the supply signal, shortening the effective feedback delay. It is among the few options that act within months rather than years.

Reading claims about AI energy critically

Because the topic is politically and commercially charged, check any figure you meet against four questions:

1. Energy or power? TWh (energy over a year) and GW (instantaneous capacity) are different. Peak capacity requirements matter more for the grid than annual totals.
2. Scope. Does the number include only AI servers, all data centres, or the wider digital economy such as networks?
3. Scenario or measurement? Projections for 2030 are model outputs with assumptions; 2024 numbers are estimates, since operators disclose little.
4. Geography. Global averages hide local stress. The US and China dominate the growth, and within countries a handful of regions carry most of it.

Frequently asked questions

How much electricity do data centres use?

The IEA estimates about 415 TWh in 2024, around 1.5% of global electricity demand, and projects about 945 TWh in 2030 in its base case.

Is AI the main driver?

AI is a major driver of growth but not the whole of the load. The IEA estimates AI servers were roughly 15% of data centre energy demand in 2024, and some independent analysts think that is an underestimate. Conventional cloud and storage still make up most of today’s total.

Why can’t the grid just build more capacity quickly?

Because interconnection, permitting, equipment lead times and construction form a chain of delays. Berkeley Lab reports a median of more than five years from request to operation for projects built in 2025.

Will efficiency improvements solve the problem?

They reduce energy per task, but rebound effects can offset the savings if cheaper computing raises total use. The net result depends on how demand responds to falling costs.

Conclusion

AI data center power demand is best understood as two stocks moving at very different speeds, joined by a long delay. That structure predicts shortfalls, workarounds that lock in other problems, and a premium on anything that shortens the feedback loop, such as flexible load and faster interconnection. Watching the queue and the connection times will tell you more about the next few years than any single demand forecast.

Sources and further reading

• International Energy Agency, Energy and AI (2025).
• Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition.
• Scientific American, “Data Centers Will Use Twice as Much Energy by 2030, Driven by AI”, for commentary on the IEA estimate.

Related concept: AI data center power demand in our glossary.