Short answer: Artificial intelligence is often described as a software revolution. In practice, it is also becoming an energy and infrastructure race. AI models do not run in an abstract “cloud”. They run in buildings full of servers that need electricity, cooling, grid connections, chips, land, water and long-term access to reliable power.
The important caveat is that there is no single global meter for “AI electricity”. ChatGPT, Gemini, Claude, open-source models, enterprise AI and Chinese AI systems are not reported as one official electricity category. The most defensible proxy is therefore data-centre electricity demand, with AI/GPU servers as the fastest-growing part of that demand.
According to the International Energy Agency (IEA), data centres consumed about 415 TWh of electricity in 2024, roughly 1.5% of global electricity demand. In the IEA base case, this rises to about 945 TWh by 2030, just under 3% of global electricity demand.
That is roughly the electricity consumption of Japan today.
The core global numbers
| Measure | Year | Value | How to read it |
|---|---|---|---|
| Global data-centre electricity demand | 2024 | ~415 TWh | Around 1.5% of global electricity use |
| Global data-centre electricity demand | 2030 | ~945 TWh | IEA base case, close to Japan-scale demand |
| Global data-centre electricity demand | 2035 | ~1,200 TWh | IEA base case |
| High-growth 2035 scenario | 2035 | >1,700 TWh | IEA “Lift-Off” case |
| AI/GPU server electricity demand | 2024–2030 | ~30% annual growth | One of the main drivers of new demand |
This does not mean AI will “use all the electricity”. That would be too simplistic. The real issue is more practical: the growth is fast, locally concentrated and arriving faster than grids, transformers, substations and new firm power can be built in many places.
AI is not the same as all data centres
Data centres are not only AI. They also run cloud services, banks, streaming platforms, e-commerce, telecoms, storage, business software and the ordinary internet.
But AI is the new growth engine. The IEA says electricity demand from accelerated servers — the servers used for AI workloads — is expected to grow by around 30% per year and account for almost half of the increase in data-centre electricity demand by 2030.
Country-scale comparison: how large is 945 TWh?
Country comparisons are imperfect because national electricity use changes every year. They are still useful for scale.
| Comparison | Approximate annual electricity scale | What it means |
|---|---|---|
| Global data centres in 2024 | ~415 TWh | Already similar to a large industrial country’s electricity use |
| Global data centres in 2030 | ~945 TWh | Roughly Japan-scale electricity demand |
| UK total electricity demand | ~280–300 TWh/year | 2030 global data centres would be more than 3× current UK electricity demand |
| Germany total electricity demand | ~500 TWh/year | 2030 global data centres would be close to about 2× Germany’s electricity use |
| Current Bitcoin mining | ~125–170 TWh/year, with 2024 Cambridge estimate at 138 TWh | Today’s data centres already use several times more electricity than Bitcoin mining |
The useful conclusion is not that AI is uniquely “bad”. It is that the digital economy is becoming a major electricity customer in its own right.
The United States: one number is no longer enough
For the United States, a single point estimate is misleading. The DOE/Lawrence Berkeley National Laboratory report says:
- US data centres consumed about 176 TWh in 2023, around 4.4% of US electricity use.
- By 2028, demand could rise to 325–580 TWh, or 6.7–12% of US electricity use.
A sensible 2030 range is therefore scenario-based:
| US 2030 scenario | Annual data-centre electricity demand | How to read it |
|---|---|---|
| Conservative / base-low | ~425–500 TWh | IEA-style baseline plus lower continuation of LBNL growth |
| Main realistic range | ~500–700 TWh | If a large share of hyperscaler buildout is actually connected |
| Stress / upside case | ~800–1,000+ TWh | If high utility-queue and gigawatt project assumptions materialise; risk of overstatement |
This is why Microsoft, Amazon, Google, Meta, Oracle and OpenAI are no longer only buying cloud capacity. They are trying to secure gigawatts of reliable power.
| Company / project | Power scale | Approximate annual electricity at high utilisation | Note |
|---|---|---|---|
| Microsoft + Three Mile Island / Crane Clean Energy Center | 835 MW | ~6.6 TWh/year | Nuclear restart target around 2028 |
| Amazon AWS + Susquehanna | up to 1,920 MW | ~15 TWh/year | Main full-scale impact more likely around 2032 |
| Google + Kairos Power | up to 500 MW | ~4 TWh/year | First unit targeted around 2030, larger buildout after 2030 |
| Meta + Clinton nuclear | 1,121 MW | ~8.8 TWh/year | Long-term nuclear power-purchase agreement |
| OpenAI / Oracle / SoftBank Stargate | 7–10 GW | ~43–79 TWh/year | Only if capacity is built and highly utilised |
The brake is equally important: announced gigawatts are not guaranteed terawatt-hours. Some projects will be delayed, some will not run at full load and some capacity can be counted more than once in market commentary.
Regional overview
| Region | Latest / current evidence | 2030 indication | Confidence |
|---|---|---|---|
| World | ~415 TWh in 2024 | ~945 TWh IEA base case | High for order of magnitude |
| United States | 176 TWh in 2023 | ~425–700 TWh main range; stress case higher | High uncertainty but strong evidence of rapid growth |
| China | roughly 100+ TWh implied today | around ~279 TWh from IEA regional implication | Lower transparency |
| Europe / EU | about 62–70 TWh today | roughly 107–115 TWh by 2030 | Medium to high |
| Great Britain / UK | around 7.5 TWh in 2024 in NESO data | around 13–25 TWh by 2030 depending on scenario | Good for scenarios |
| Middle East / GCC | capacity about 1 GW in 2025 | around 3.3 GW capacity over the next five years; UAE >6 TWh by 2030 | Weaker TWh data |
For the UK reader, the immediate issue is not that AI data centres will dominate national electricity demand by 2030. The issue is where new large loads connect, whether grid reinforcement arrives on time, and whether new demand competes with household electrification, heat pumps, EVs and industrial demand.
What powers data centres?
The IEA estimates the physical electricity mix used to supply data centres globally is roughly:
| Source | Approximate share |
|---|---|
| Coal | ~30% |
| Renewables | ~27% |
| Natural gas | ~26% |
| Nuclear | ~15% |
This is different from a marketing claim such as “100% renewable” based on certificates. The table describes the physical electricity systems where data centres are located.
In the United States, natural gas is currently the largest source of data-centre power. That is one reason hyperscalers are signing nuclear, renewable and “firm clean power” contracts: AI data centres need power that is available every hour, not only when the wind blows or the sun shines.
Why AI is pulling technology companies into energy
AI changes the relationship between technology and energy because the bottleneck moves from software alone to physical infrastructure.
A leading AI company can have the model, the talent and the capital, but still be constrained by:
- grid connection delays,
- transformer availability,
- local planning rules,
- cooling and water constraints,
- electricity prices,
- power-purchase contracts,
- access to firm low-carbon electricity.
This is why nuclear power is suddenly part of the AI conversation. Nuclear plants are expensive and slow, but they produce large amounts of steady electricity. That quality is valuable for high-load data centres.
Comparison with Bitcoin: the same debate, larger and more complex
For years, Bitcoin was the best-known example of an energy-intensive digital technology. Bitcoin mining became a symbol of the question: should a purely digital activity be allowed to consume as much electricity as a smaller country?
AI moves the same debate into a larger and more complex arena. Bitcoin is one global network with relatively measurable mining activity. AI is a broader infrastructure layer: models, cloud platforms, inference, training, business applications, chips, data centres and network capacity. The comparison is not perfect, but it is useful for scale.
The Cambridge Digital Mining Industry Report 2025 estimates Bitcoin mining electricity consumption at about 138 TWh in 2024. Live annualised Cambridge CBECI estimates vary by methodology and assumptions, but a reasonable current order-of-magnitude range is about 125–170 TWh per year.
| Comparison | Annual electricity demand |
|---|---|
| Bitcoin mining, 2024 Cambridge estimate | ~138 TWh |
| Bitcoin mining, current annualised estimate range | ~125–170 TWh |
| Global data centres, 2024 | ~415 TWh |
| Global data centres, 2030 | ~945 TWh |
| US data centres, 2030 main range | ~425–700 TWh |
Today’s global data centres already use about three times as much electricity as Bitcoin mining. By 2030, global data-centre demand could be six to seven times today’s Bitcoin mining electricity use, if Bitcoin stayed near today’s range.
The character of the load is also different. Bitcoin mining can be relatively mobile: miners can move towards cheap electricity, curtailed renewables, stranded energy or off-grid gas. AI data centres are heavier infrastructure. They need stable grid connections, latency, land, cooling, water, transformers and guaranteed power. Once built, they become long-term local energy-system actors.
Bitcoin’s energy mix is also more nuanced than old headlines suggest
Cambridge 2025 reports this estimated Bitcoin mining energy mix:
| Source | Share |
|---|---|
| Renewables | 42.6% |
| Nuclear | 9.8% |
| Non-fossil total | 52.4% |
| Natural gas | 38.2% |
| Coal | 8.9% |
| Oil | 0.5% |
| Fossil total | 47.6% |
Some Bitcoin mining can use otherwise curtailed, stranded or off-grid energy, including curtailed renewables or gas that might otherwise be flared. That matters, but it should not be used as a blanket defence of all mining. Grid-connected mining can still add load and affect local electricity systems.
What this means for the UK energy system
For the UK, the most realistic near-term questions are practical:
- Grid connection timing: can large AI/data-centre loads connect without delaying housing, heat-pump, EV and industrial electrification plans?
- Firm power: can the UK add enough reliable low-carbon power rather than relying mainly on gas at peak times?
- Location: will new data centres cluster in areas where the grid is already constrained?
- Waste heat and flexibility: can data centres become part of local energy planning instead of only a new load on the system?
- Household bills: who pays for the network upgrades if demand grows faster than planned?
None of this means households should panic. It does mean AI is now part of the same energy-planning discussion as EVs, heat pumps, industry, nuclear, renewables and grid expansion.
Bottom line: the next digital industry will be built on electricity
The AI debate is usually about model capability: better chatbots, faster coding, automated routine work and competition between technology companies. But underneath that software layer, a very physical industry is emerging.
AI infrastructure needs servers, chips, cooling systems, land, water, transformers and above all electricity. Not occasional electricity, but reliable power available every hour of the year. That is why technology companies are signing nuclear contracts, supporting new reactors, reserving gigawatts of capacity and pushing for faster grid connections.
AI will not “eat all the electricity”. But it is becoming one of the fastest-growing new electricity customers in the world. By 2030, global data centres could be using Japan-scale electricity. In the United States, data-centre demand could grow fast enough to influence where new power plants are built, how long older plants keep running and how much grid reinforcement costs.
The biggest change is therefore not only technological. It is strategic. Countries and companies that can combine computing power with cheap, reliable and cleaner electricity will have an advantage in the AI race. Those that cannot may discover that even the best algorithm eventually meets a simple physical limit: there is nowhere to plug it in.
Sources used for this article
The sources below were used to verify the main figures and context. Forecasts are scenarios, not guarantees; country comparisons are approximate scale comparisons.
- International Energy Agency (IEA) — Energy and AI: Energy demand from AI.
- International Energy Agency (IEA) — Energy supply for AI.
- DOE / Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report.
- Congressional Research Service — Data Centers and Their Energy Consumption: Frequently Asked Questions.
- European Commission — Data centres: an energy-hungry challenge.
- NESO — Future Energy Scenarios 2025 data workbook and data-centre demand scenarios.
- GOV.UK — UK Compute Roadmap.
- Cambridge Centre for Alternative Finance — Cambridge Bitcoin Electricity Consumption Index and Digital Mining Industry Report 2025.
- U.S. Energy Information Administration (EIA) — crypto-mining electricity analysis and international electricity data.
- Constellation / Microsoft — Crane Clean Energy Center / Three Mile Island restart announcement.
- Talen Energy / Amazon Web Services — Susquehanna nuclear power agreement.
- Google / Kairos Power — advanced nuclear agreement.
- Meta — nuclear energy procurement announcements.
- OpenAI / Oracle / SoftBank — Stargate infrastructure announcements.
