Artificial intelligence is driving one of the largest infrastructure building programs in the history of technology. Models that once ran on small clusters now need thousands of chips, large data centres and power connections measured in hundreds of megawatts.
The International Energy Agency (IEA) says global data centres used about 485 terawatt-hours (TWh) of electricity in 2025. It expects that figure to reach about 950 TWh by 2030.
This article collects the main statistics on electricity use, construction, chips, spending and power supply. All figures come from the sources listed at the end.
AI Data Centre Statistics at a Glance
The table below lists the main figures covered in this article. Each one is explained in more detail in the sections that follow.
The chart shows how much larger four key measures are expected to be by 2030 compared with 2025.
| Measure | Figure |
|---|---|
| Global data-centre electricity use, 2025 | 485 TWh |
| Expected use, 2030 | About 950 TWh |
| Expected use, 2035 | 1,200 to 1,300 TWh |
| AI share of data-centre workloads, 2025 and 2030 | 25% and 50% |
| New global capacity, 2026 to 2030 | Almost 100 GW |
| Investment needed for that capacity | About $3 trillion |
| North American capacity under construction, mid-2026 | 66 GW |
| Share of that pipeline already committed | About 95% |
| AI server shipments, 2026 | About 2.8 million |
| Share of AI servers that use GPUs | About 70% |
| Capex of the nine largest cloud providers, 2026 | More than $886 billion |
| Spending by Amazon, Microsoft, Alphabet and Meta, 2026 | About $630 billion |
How Much Electricity Do Data Centres Use?
The International Energy Agency (IEA) estimates that global data centres used about 485 terawatt-hours (TWh) of electricity in 2025. That is roughly 1.5% to 2% of world electricity use. It is more than many major economies use in a year.
The IEA’s central forecast puts use at about 950 TWh by 2030. That would be around 3% of global electricity demand. Demand would almost double in five years.
Growth continues after 2030. The IEA base case reaches about 1,200 TWh by 2035. When measured as the generation needed to supply data centres, including system losses, the figure could reach 1,300 TWh. That is close to three times the 2025 level.
Forecasts remain uncertain. Small changes in AI efficiency, model demand and chip technology make large differences when they are multiplied across millions of servers.
| Year | Electricity use | Note |
|---|---|---|
| 2025 | 485 TWh | IEA estimate |
| 2030 | About 950 TWh | About 3% of global demand |
| 2035 | About 1,200 TWh | About 1,300 TWh including system losses |
AI Is the Main Driver of Growth
Not all data-centre electricity goes to AI. Traditional facilities host websites, cloud storage, databases, streaming, enterprise software, social networks, search and financial systems.
AI is the fastest-growing workload. The IEA expects electricity use by accelerated servers, which are mainly used for AI, to grow by about 30% a year through 2030. Conventional servers are expected to grow by only about 9% a year.
Accelerated servers could account for almost half of the net increase in global data-centre electricity demand through 2030. Cooling and supporting equipment add another large share. Overall, electricity use by AI-focused data centres is expected to triple between 2025 and 2030.
| Server type | Expected annual growth in electricity use |
|---|---|
| Accelerated servers (mainly AI) | About 30% |
| Conventional servers | About 9% |
AI Could Be Half of Data-Centre Workloads by 2030
JLL estimates that AI made up about 25% of data-centre workloads in 2025. By 2030, that share could reach 50%.
The type of AI demand is also changing. The first phase of the boom was dominated by training large models. Increasingly, power and computing capacity are being used for inference, which means running AI models for users.
Every chatbot prompt, AI search result, generated video or software agent needs inference computing. As hundreds of millions of people and businesses use AI each day, inference could eventually need more total computing capacity than training.
U.S. Data-Centre Power Demand Could Quadruple
The United States is at the centre of the building boom. U.S. data centres used about 177 to 192 TWh of electricity in 2024.
Some current projections suggest that figure could rise as high as 793 TWh by 2030. That would be more than four times the 2024 level in six years.
The IEA expects data centres to account for almost half of U.S. electricity-demand growth through 2030. By the end of the decade, U.S. data centres could use more electricity than the country’s combined production of aluminium, steel, cement, chemicals and other energy-intensive goods.
| Measure | Figure |
|---|---|
| U.S. data-centre use, 2024 | 177 to 192 TWh |
| High projection, 2030 | 793 TWh |
| Share of U.S. demand growth to 2030 | Almost half |
Almost 100 GW of New Capacity Could Be Built
Electricity figures translate directly into construction. JLL expects almost 100 gigawatts (GW) of new data-centre capacity to be added worldwide between 2026 and 2030. That would roughly double global capacity.
JLL estimates that delivering this capacity could need about $3 trillion of investment. The global data-centre market is expected to grow at about 14% a year through 2030.
For scale, a single 1 GW AI campus can need the output of a large conventional power station.
| Measure | JLL estimate |
|---|---|
| New capacity, 2026 to 2030 | Almost 100 GW |
| Investment needed | About $3 trillion |
| Market growth rate | About 14% a year |
| Power need of one large AI campus | 1 GW or more |
North America Has 66 GW Under Construction
North America has the largest building boom. JLL reported about 66 GW of capacity under construction in mid-2026. Around 95% of it was already committed to tenants.
That is why vacancy stays near 1% despite heavy construction. JLL estimates that North American absorption reached 25 GW in the first half of 2026. That was twice the level of the previous year and five times the level two years earlier.
Hyperscalers made up about 59% of tenant demand. Neocloud providers made up 11%, and pure-play AI companies made up 7%.
| Tenant type | Share of demand, H1 2026 |
|---|---|
| Hyperscalers | About 59% |
| Neocloud providers | 11% |
| Pure-play AI companies | 7% |
Atlanta Overtakes Northern Virginia in Construction
CBRE offers a narrower view focused on major North American wholesale markets. It reported 7.48 GW under construction across primary markets in the first half of 2026. That was up 24.8% from a year earlier.
Atlanta overtook Northern Virginia for capacity under construction. Atlanta had 2.88 GW and Northern Virginia had 2.42 GW.
Northern Virginia is still the largest existing market, with about 4.50 GW of operating capacity and a vacancy rate of 0.2%. About 80% of the capacity under construction in primary markets was already preleased.
| Measure | Figure |
|---|---|
| Under construction, primary markets | 7.48 GW |
| Atlanta under construction | 2.88 GW |
| Northern Virginia under construction | 2.42 GW |
| Northern Virginia operating inventory | 4.50 GW |
| Northern Virginia vacancy rate | 0.2% |
Europe Is Expanding Beyond Traditional Hubs
Europe is also growing quickly. The traditional FLAP-D markets of Frankfurt, London, Amsterdam, Paris and Dublin had about 3.8 GW of live capacity by the first half of 2026.
Another 1.4 GW was under construction and about 2 GW was planned. Grid limits are pushing large AI projects toward regions with more available power.
Growth is therefore spreading to Scandinavia, Spain, Portugal, Italy and Eastern Europe. It is no longer limited to the historic hubs.
| FLAP-D status | Capacity |
|---|---|
| Live | About 3.8 GW |
| Under construction | 1.4 GW |
| Planned | About 2 GW |
AI Servers, GPUs and Custom Chips
TrendForce estimates that global AI server shipments will reach about 2.8 million units in 2026. That is growth of 30% to 31% from a year earlier. AI servers could generate more than 74% of total server-industry value.
They cost far more than standard servers. They can hold several high-end chips, advanced networking, high-bandwidth memory and liquid cooling.
GPUs are expected to make up 69.7% of AI server shipments in 2026. Servers built on custom chips, known as ASICs, are expected to make up 27.8%, and could approach 40% by 2030. Examples include Google TPUs, Amazon Trainium, Microsoft Maia and Meta MTIA.
Microsoft said in October 2026 that it could deploy more than one million Maia 300 chips in 2027. A Maia 300 rack holds 96 chips and uses about 167 kilowatts.
| Chip type | Share of AI server shipments |
|---|---|
| GPU | 69.7% (2026) |
| Custom ASIC | 27.8% (2026) |
| Custom ASIC | About 40% (2030) |
Nvidia Blackwell and the Shift to Inference
Nvidia remains the main supplier of high-end AI chips. TrendForce expects Blackwell products to make up about 71% of Nvidia’s high-end GPU shipments in 2026, up from about 61%.
The five largest North American cloud companies are expected to account for more than 60% of global demand for Nvidia’s GB and VR series AI systems in 2026.
TrendForce estimates their AI training compute could grow more than 56% year over year. Inference computing power could rise about 122%. That gap shows the growing importance of inference.
| Measure | Figure |
|---|---|
| Blackwell share of Nvidia high-end GPUs, 2026 | About 71% |
| Blackwell share, prior year | About 61% |
| Training compute growth, top five clouds | More than 56% |
| Inference compute growth, top five clouds | About 122% |
China Is Deploying Millions of AI Accelerators
China is one of the few markets with recent unit-level accelerator estimates. IDC data reported by Reuters shows more than 2 million AI accelerators were shipped in China in the first half of 2026.
Nvidia shipped about 678,000 units and Huawei shipped about 572,000. Nvidia’s share of China’s data-centre AI accelerator market fell to about 34%, while Huawei reached about 28%.
China is shifting toward domestic chips because U.S. export controls limit access to Nvidia’s most powerful products.
| Supplier | Units shipped, H1 2026 | Market share |
|---|---|---|
| Nvidia | About 678,000 | About 34% |
| Huawei | About 572,000 | About 28% |
Cloud Capex Is Approaching $900 Billion
TrendForce estimates that combined 2026 capital expenditure by Amazon, Microsoft, Alphabet, Meta, Oracle, ByteDance, Alibaba, Tencent and Baidu will exceed $886.7 billion. That would be an increase of about 90% from a year earlier.
The five North American hyperscalers account for almost 90% of the total. Stanford’s 2026 AI Index also shows hyperscaler capex more than doubling since ChatGPT launched.
Across the wider industry, AI infrastructure spending could reach about $811 billion in 2026. Amazon, Microsoft, Alphabet and Meta together are expected to spend about $630 billion on AI infrastructure.
| Scope | 2026 spending |
|---|---|
| Nine largest cloud providers (capex) | More than $886.7 billion |
| Wider AI infrastructure industry | About $811 billion |
| Amazon, Microsoft, Alphabet and Meta (AI infrastructure) | About $630 billion |
What Individual Companies Plan to Spend
Company budgets are now measured in the hundreds of billions of dollars. Amazon has outlined about $200 billion of capital expenditure for 2026, with AI infrastructure making up much of it.
Alphabet raised its 2026 capex forecast to $195 billion to $205 billion. Microsoft is expected to spend more than $175 billion during calendar 2026.
The chart uses the midpoint for Alphabet’s range.
| Company | 2026 capex |
|---|---|
| Amazon | About $200 billion |
| Alphabet | $195 billion to $205 billion |
| Microsoft | More than $175 billion |
Power May Become the Biggest Constraint
AI companies first worried about getting enough GPUs. Increasingly, the limit is electricity. A new AI campus can need more than 1 GW of power, which is the demand of hundreds of thousands of homes.
Grid connections can take years. Other bottlenecks include transformers, gas turbines, transmission lines, cooling equipment, land, permits, water and skilled construction workers.
Companies are therefore building or financing their own power. In October 2026, American Electric Power announced a $4 billion, 702 MW fuel-cell project to supply Google operations in Wyoming.
| Item | Size |
|---|---|
| Large new AI campus | 1 GW or more |
| AEP fuel-cell project for Google | 702 MW |
| Cost of the AEP project | $4 billion |
What Powers AI Data Centres?
The IEA estimates where the electricity that serves data centres comes from today. Coal supplies about 30%, renewables about 27%, natural gas about 26% and nuclear about 15%.
Renewables are expected to meet nearly half of the added data-centre demand through 2030. Natural gas is also expected to play a large role because it can provide steady power when renewable output changes.
Nuclear could become more important toward 2030 and beyond. This is why Microsoft, Google, Amazon and others have signed deals covering nuclear plants, next-generation reactors, renewables and gas.
| Source | Share of supply |
|---|---|
| Coal | About 30% |
| Renewables | About 27% |
| Natural gas | About 26% |
| Nuclear | About 15% |
The Biggest Uncertainty Is Efficiency
Forecasting AI electricity use is hard. New chips improve performance per watt. Models are also becoming more efficient through methods such as quantization, distillation and mixture-of-experts designs.
But cheaper AI can lead people to use it much more. This is known as the rebound effect.
Take a model that becomes ten times cheaper to run. If usage rises 50-fold, total electricity use still rises fivefold. For this reason the IEA uses several scenarios rather than a single exact figure for 2030.
| Example | Change from today |
|---|---|
| Cost to run the model | 10 times cheaper |
| Usage | 50 times higher |
| Total electricity use | 5 times higher |
The Bottom Line
Global data centres used about 485 TWh of electricity in 2025. The IEA expects use to approach 950 TWh by 2030, with AI-focused facilities growing much faster than conventional computing.
Almost 100 GW of new capacity could be needed worldwide between 2026 and 2030, at a cost of about $3 trillion. North America already has about 66 GW under construction, and about 95% of it is committed.
Hardware is expanding at a similar pace. AI server shipments are expected to reach about 2.8 million units in 2026, and custom chips are gaining share from GPUs. The nine largest cloud providers are expected to spend more than $886 billion in 2026.
The main question may no longer be whether companies can build more capable models. It is whether electricity, chips, data centres, transformers, cooling systems and financing can keep up.
Sources
- International Energy Agency: Key Questions on Energy and AI, 2026
- International Energy Agency: Energy Demand From AI
- International Energy Agency: Energy Supply for AI
- JLL: 2026 Global Data Center Market Outlook
- JLL: North America Data Center Report, Midyear 2026
- CBRE: North America Data Center Trends H1 2026
- JLL: EMEA Data Centre Mid-Year 2026
- TrendForce: Global AI Server Shipments 2026
- TrendForce: AI Server Shipments and CSP Capex, August 2026
- TrendForce: Nvidia Blackwell GPU Shipments
- Stanford University: 2026 AI Index Report, Economy Chapter
- Reuters: Microsoft plans huge deployment of its own AI chips
- Reuters: US AI boom faces electric shock
- Reuters: Data centers’ flexible power usage could save grid billions
