Energy

What it is about

Every AI answer is electricity turned into computation and then into heat. A data centre draws that electricity from the grid, converts it several times, feeds it to thousands of chips and moves the heat back out, often with the help of water. In 2025, data centres worldwide used about 485 TWh, around 1.5% of global electricity, according to the International Energy Agency (IEA). For 2030 the IEA expects about 950 TWh.

The energy layer decides where AI capacity can be built, how fast and at what price. The tightest limit today is the grid: connections, transformers and turbines take years to arrive. For a founder this layer reaches into everyday choices, from the model you pick to the cloud region you run in and the climate figures you report.

Power and energy: what GW and TWh mean

Two units run through every debate about AI and electricity, and they are easy to mix up. Power, measured in kilowatts (kW), megawatts (MW) or gigawatts (GW), says how much electricity flows at one moment. Think of the width of a pipe. Energy, measured in kilowatt hours (kWh) up to terawatt hours (TWh), says how much has flowed over a period, like the water that came through the pipe in a year.

Grid connections, power plants and the size of a data centre are planned in power. Electricity bills, emissions and national statistics are counted in energy. The link is simple arithmetic: energy equals power times hours times utilisation. One gigawatt running through all 8,760 hours of a year delivers 8.76 TWh. The US Lawrence Berkeley National Laboratory (LBNL) uses exactly this conversion. Its projection of 325 to 580 TWh for US data centres in 2028 corresponds to about 74 to 132 GW at an assumed average utilisation of 50%.

That is why two well-known figures cannot be compared directly. Epoch AI estimates about 30 GW of AI data centre capacity at the end of 2025, derived from the rated power of the chips sold plus an allowance for servers, network and cooling. The IEA counts about 485 TWh of electricity used by all data centres in 2025. The first figure is a capacity, the second a quantity of energy over a year.

Rule of thumb: when a company announces a 1 GW campus, it means capacity. Its annual use is at most 8.76 TWh and usually lower, because sites rarely run at full load all year. For scale, all data centres in Ireland used about 7.7 TWh in 2025.

How electricity reaches the chip

A data centre is a large electricity customer with a small power grid of its own inside. The order below is the standard textbook chain; the figures come from the sources named. Every conversion step loses a little energy as heat, and these losses later appear in the efficiency figure PUE.

1. 01Transmission grid. High-voltage lines bring electricity from power plants, often over long distances. In many regions they are already full where data centres cluster.
2. 02Grid connection and substation. A grid connection with its own substation steps the voltage down. Getting it is today the longest wait in the whole project, often 5 to 10 years (IEA, 2026).
3. 03Transformers and switchgear. Transformers lower the voltage further for the building; switchgear distributes and protects. Large power transformers currently take about 2 to 3 years to arrive (IEA).
4. 04UPS and batteries. An uninterruptible power supply (UPS) with batteries bridges seconds to minutes when the grid fails. LBNL assumes UPS efficiencies of 77 to 85% at small sites and 90 to 99% at hyperscale and AI facilities.
5. 05Backup generators. Diesel or gas generators start within seconds and carry longer outages. Northern Virginia alone has more than 4,000 diesel generators at data centres with over 11 GW, according to a columnist citing state permits (commentary).
6. 06Rack. Power distribution units feed each server rack. Power supplies convert the electricity for the GPUs and other chips, which turn almost all of it into heat.

Data centres build two or three levels of redundancy, so their backup capacity exceeds the actual load. Generators mostly run for tests. In Virginia, a late-2025 change that widened what counts as an emergency drew more than 500 public objections (Virginia Mercury, commentary).

Why AI racks need so much power

An AI chip performs trillions of multiplications a second, and almost every watt it draws ends up as heat. AI systems pack many such chips tightly together, because chips that sit close to each other exchange data faster. The result is a lot of power in a small space.

The numbers show the jump. The average rack drew almost 9 kW in 2025, according to the Uptime Institute's global operator survey, and more than 80% of sites have no rack above 30 kW. A rack with 72 of NVIDIA's Blackwell GPUs (GB200 NVL72) needs about 132 kW according to the server maker Supermicro, roughly fifteen times the average (vendor data). The IEA reports that the power density of AI servers rose elevenfold between 2020 and 2025 and expects another fourfold rise by 2027, when a single rack could draw as much as 65 households at peak.

AI loads also move fast. When a large training run pauses or resumes, its power draw can swing by more than half of the rated capacity within one second (IEA), and grid operators have to plan for that. Chip supply sets its own ceiling: global production of HBM memory supports at most about 25 GW of new AI servers a year until 2027, according to the IEA.

Cooling, PUE and water

All this heat has to leave the building. Two questions decide how much electricity and water that costs: how the heat is taken away from the chips, and how it finally leaves the site. The first three entries below describe the server room, the last two the way heat leaves the building (LBNL).

• Air cooling: fans blow cool air over the servers. Simple, proven and still the standard in most data centres, whose racks average under 10 kW.
• Direct liquid cooling: liquid cooling with cold plates sitting directly on the chips and coolant circulating in a closed loop. Dense AI racks such as the GB200 NVL72 are built for it.
• Immersion cooling: whole servers sit in a non-conductive fluid. It is being tested, for example at a tropical data centre testbed in Singapore; we found no reliable figures on how widely it is used.
• Cooling towers: evaporate water to carry heat away. They save electricity and consume water, which raises local concerns in dry regions.
• Dry coolers: release heat to the air through large radiators. They save water and need more electricity on hot days.

1. 01PUE. PUE, Power Usage Effectiveness, is the total electricity of a site divided by the electricity of its IT equipment; 1.0 would mean zero overhead. The global average has barely moved for six years and stood at 1.54 in 2025 (Uptime Institute). Sites built since 2020 reach 1.48, sites above 20 MW 1.44. The best hyperscale sites run at 1.1 to 1.2 (IEA), and Google reports 1.09 for its fleet (company figure).
2. 02WUE. WUE, Water Usage Effectiveness, gives the litres of water used on site per kWh of IT electricity. Only 47% of operators collect water data at all (Uptime Institute), and no global average exists.
3. 03Direct and indirect water. Much of the water sits in the power plants that generate the electricity. For the US in 2023, LBNL counts about 64 billion litres of direct cooling water and about 800 billion litres of indirect water through electricity generation, roughly twelve times as much, as researchers at the University of Wisconsin-Milwaukee summarise. An earlier summary of the same report gave 66 billion litres; the gap is probably rounding or a different boundary.

Liquid cooling inside the rack and the water use of the building are two separate questions. Liquid-cooled racks can run warmer, which allows more free cooling with outside air and can save both water and electricity. Whether a site consumes water depends on how it releases its heat (LBNL). In hot, humid climates cooling can take up to 40% of a data centre's electricity, according to researchers at the National University of Singapore.

Training, inference and the energy of a single request

Training a model is a large, one-off burst of computing that can run for weeks. Inference, every answer a model gives, is small per request and repeats billions of times. At scale, use outweighs training. Sasha Luccioni and colleagues calculated that for a mid-sized open model (BLOOMz-7B), the energy spent on answering equals the energy of training after about 593 million requests.

How much energy one request uses depends on what is counted. Google published the first production measurement in August 2025: a median text prompt to its Gemini app used 0.24 Wh in May 2025, including idle machines, processors, memory and cooling. Counting only the active AI chips gives 0.10 Wh. OpenAI states 0.34 Wh for an average ChatGPT query and Microsoft researchers a median of 0.34 Wh for large models, both cited by the IEA. All three are company figures without independent audit. From the outside, Epoch AI estimated about 0.3 Wh for a typical GPT-4o query.

Older figures of 3 to 4 Wh per query came from estimates made in 2022 and 2023 with older hardware, low utilisation and other assumptions. Google reports that its energy per prompt fell 33-fold within one year. The type of task matters even more: an image uses about ten times the energy of a text answer, a video hundreds to thousands of times (IEA). Long inputs push the number up as well; Epoch estimates about 40 Wh for an input of 100,000 tokens. Reasoning models and agents that work through many steps use far more than a simple answer.

• Boundary: only the chips, or the whole system with idle servers and cooling. In Google's data this choice alone makes a factor of about 2.4.
• Utilisation: a chip that serves many users at once uses less energy per answer. Lab tests with one request at a time miss this effect of production batching (Google).
• Model and length: model size and the number of tokens in and out drive the computing effort. Two measurements of the same open model (Llama 3.1 70B) differed by about a factor of six.
• Who measures: companies measure in production and publish without audit. Outsiders estimate with assumptions; Epoch states a possible error of up to a factor of two.

The grid is the bottleneck

In many places the scarce resource is the connection to the grid and the equipment needed to build it. A new data centre needs a grid connection, and the grid operator must check whether lines and substations can carry the extra load. In many jurisdictions this takes 5 to 10 years, according to the IEA. In Europe's established hubs Frankfurt, London, Amsterdam, Paris and Dublin, the think tank Ember reports average waits of 7 to 10 years, in some cases up to 13.

Queues are long partly because they are inflated. Texas grid operator ERCOT recorded more than 230 GW of connection requests from large loads in January 2026, against a record peak demand of 85 GW; some requests are speculative (IEA). Jonathan Koomey and co-authors warn regulators that requests are often duplicated and that IT energy forecasts have a long history of overshooting. On the supply side, LBNL counts 2,061 GW of power plants and storage waiting for a US grid connection at the end of 2025. Of the capacity requested between 2000 and 2020, only 13% was built.

Then there is the hardware. Large power transformers take about 2 to 3 years (IEA). Wood Mackenzie measured 128 weeks in the US in mid-2025, a supply deficit of about 30% and prices 77% above 2019 (market research). An electrical contractor interviewed by POWER magazine disputes the shortage and quotes 12 to 14 months for standard units. Gas turbines take about five years, and orders rose 70% in 2025 (IEA). BloombergNEF expects data centres to need about 400,000 tonnes of copper a year over the next decade.

Scarcity shows up in prices. In PJM, the grid of the eastern US, the auction for 2027/28 capacity cleared at the price cap, a record price for the third auction in a row, and the rise in forecast demand came almost entirely from data centres. Flexibility can help: a Duke University study estimates that the largest US grids could absorb 76 GW of new load if that load pauses for 0.25% of the year, about 85 hours. This is a modelled upper bound that leaves out transmission limits.

Own power plants: what runs and what is announced

Because the grid is slow, operators build power plants on site. The IEA expects 15 to 27 GW of gas generation at data centres by 2030, mostly in the US. Only about a fifth of the tracked US projects had started land clearing or construction. An IEA chart from June 2026 shows a US pipeline of 47 GW; that figure includes proposals and early-stage plans. To run reliably without the grid, such plants need 30 to 70% more capacity than the load they serve. The IEA also expects 20 to 25 GW of batteries at data centres by 2030.

Nuclear power for AI consists mostly of contracts for future electricity. The list below separates plants that deliver today from those that are announced.

• Crane (formerly Three Mile Island 1): Microsoft holds a 20-year contract for about 835 MW. The plant is still offline; operator Constellation targets a restart in 2027 (company statement, September 2025).
• Susquehanna: Amazon Web Services buys up to 1,920 MW from an existing, operating plant owned by Talen (contract from 2025, IEA).
• Meta: contracts with Oklo, TerraPower and Vistra for more than 6 GW. Vistra's plants exist; the reactors of Oklo and TerraPower are yet to be built (IEA).
• Small modular reactors: conditional purchase commitments grew from about 25 GW at the end of 2024 to 45 GW at the end of 2025. The first projects are expected to deliver around 2030 (IEA).

Read every energy deal with three questions: does the plant exist, is it connected, and when was the status last confirmed? Announcements often stay on lists long after plans have changed.

Clean power contracts and how emissions are counted

Large AI companies buy clean electricity mainly through power purchase agreements (PPAs): long-term contracts with a wind, solar or nuclear plant. The tech sector signed about 40% of all corporate renewable PPAs worldwide in 2025 (IEA). BloombergNEF counts 55.9 GW of corporate clean energy contracts in 2025, the first decline after eight years, with Amazon, Meta, Google and Microsoft holding about 49%. The two shares use different definitions: the IEA counts renewables only, BloombergNEF includes nuclear.

Most claims of 100% renewable energy rest on annual matching. Over a year, the company buys as much green electricity or as many certificates as it consumes, while some hours still run on gas or coal. Certificates carry no time stamp. The stricter standard is 24/7 carbon-free energy, which matches consumption in every hour on the same grid and needs storage, flexible demand or clean power that runs around the clock to close the last gaps.

This choice decides which emissions a company reports. Scope 2 emissions, those caused by purchased electricity, can be counted in two ways. The location-based method uses the average emissions of the local grid; the market-based method uses what the company has contracted. For 2024, Google reports 345 g CO2 equivalent per kWh location-based and 94 g market-based. A Guardian analysis from 2024 found the in-house data centre emissions of Apple, Google, Microsoft and Meta for 2020 to 2022 to be about 7.6 times higher location-based than officially reported. The GHG Protocol, the global accounting standard, is revising its rules to require hourly and local matching for clean energy claims; the full revised standard is planned for the end of 2027.

In total, the IEA expects data centre emissions to reach about 350 million tonnes of CO2 by 2035, about 2% of power sector emissions (forecast). One company shows how far the figures can diverge: in Google's 2025 reporting year, its electricity use rose 37% and its total emissions 18%, while its electricity-related emissions fell 3% (company report).

Rebound, waste heat and local costs

Each answer is getting more efficient, and total use keeps rising. Google's energy per prompt fell 33-fold in a year, yet data centre electricity worldwide grew about 17% in 2025 and the use of AI-focused data centres about 50% (IEA). Economists call this the rebound effect, also known as the Jevons paradox after a 19th-century observation about coal: when something becomes cheaper to use, people use more of it. Sasha Luccioni, Emma Strubell and Kate Crawford conclude that efficiency gains in AI give no guarantee of a smaller footprint.

The heat itself can be useful. In Odense, Denmark, Meta's data centre feeds a district heating network: ammonia heat pumps lift the waste heat to 70 to 75 °C, enough for more than 12,000 homes in 2023, according to the engineering firm Ramboll, which worked on the project. Such projects need a heating network close by, and the heat pumps use electricity themselves. Liquid-cooled AI racks deliver warmer coolant, which can make reuse easier.

Other costs land locally. Shaolei Ren and colleagues estimate that air pollution from US data centres and the power plants that feed them could cause public health costs of more than 20 billion dollars a year by 2028 (projection). In drought-hit Uruguay and Chile, residents fought Google data centre plans over water use; in both cases Google moved to air cooling.

What this means for your decisions

You pay for electricity with every API call, through the price per token and your cost per task. The biggest lever is the model. Luccioni's measurements show that large general-purpose models use far more energy for simple jobs than small task-specific ones, with no clear gain in accuracy. Our reading of the sources: a small language model, model routing that sends easy tasks to small models, shorter prompts, caching and batch processing lower energy and cost together.

The second lever is the region. The same workload has a very different carbon footprint on a coal-heavy grid than on one rich in hydro or wind, and in water-stressed regions the cooling design matters. Ask your cloud or model provider for PUE, WUE, the hourly clean energy share of the region and both location-based and market-based emissions. Treat per-query figures as company figures until someone audits them. Batch jobs such as evaluations or embeddings can run at hours or in regions with cleaner power.

For your own reporting, state both emission methods and the boundary of any per-query figure. In the EU, data centres with at least 500 kW of IT load will carry an A to G sustainability label based on PUE and WUE, with the first labels expected in 2027. That will make provider claims easier to compare and part of responsible AI in practice.

What the layer is made of

Key figures, with source and date

• 485 TWhused by data centres worldwide in 2025, about 1.5% of global electricity. In 2024 it was about 415 TWh.IEA, Key Questions on Energy and AI · 16 Apr 2026 · neutral
• +17%growth in data centre electricity use in 2025. AI-focused data centres grew about 50%.IEA · 16 Apr 2026 · neutral
• about 950 TWhexpected for data centres in 2030, around 3% of global electricity. AI-focused sites would use about 465 TWh, roughly half (our calculation from IEA figures).IEA, Key Questions on Energy and AI · 16 Apr 2026 · neutral, forecast
• 0.24 to 0.34 Whper typical text request, as reported by Google (Gemini) and OpenAI (ChatGPT). Images, video, long inputs and agents use far more.Google; OpenAI via IEA · 21 Aug 2025 · vendor
• 1.54global average PUE in 2025, almost unchanged for six years. The best hyperscale sites reach 1.1 to 1.2.Uptime Institute; IEA · 29 Jul 2025 · industry body
• about 12xas much water was used indirectly by US data centres through power generation (about 800 bn litres) as directly for cooling (about 64 bn litres) in 2023.LBNL, summarised by UW-Milwaukee researchers · 25 Aug 2025 · research
• 5 to 10 yearsis the wait for a grid connection in many jurisdictions. Large transformers take 2 to 3 years, gas turbines about 5.IEA · 16 Apr 2026 · neutral
• 23%of Ireland's metered electricity went to data centres in 2025, more than all urban households. In 2015 it was 5%.Central Statistics Office Ireland via Irish Examiner · 7 Jul 2026 · government
• 15 to 27 GWof gas power plants expected directly at data centres by 2030, mostly in the US. Only about a fifth of tracked projects are under way.IEA · 16 Apr 2026 · neutral, forecast

Regions compared

Three patterns stand out. Mature hubs such as Virginia, Dublin and Amsterdam are hitting grid limits. New hubs such as Johor, the Nordics, India, Brazil and the Gulf are racing to approve capacity. In other places a single AI campus would claim a large part of the national grid, as in Kenya. Approved, requested and queued capacity is much larger than what gets built, so the table labels each figure.

Sources: IEA (16 Apr 2026), EPRI (26 Feb 2026), Reuters (14 Jul 2026), Carbon Brief (16 Apr 2025), Ember (19 Jun 2025), Linklaters (14 Oct 2026), CSO via Irish Examiner (7 Jul 2026), Statnett via Argus (26 Mar 2026), AMRO (May 2026), The Star (20 Oct 2026), Morgan Lewis (3 Mar 2026), CEEW (19 Jan 2026), WEF (23 Jul 2026), Semafor (6 May 2026), Daily Nation (11 Dec 2025), Canal Solar (1 Jun 2026), Logic(s) (1 Nov 2025). Several figures come from secondary reports of the original. National statistics use different definitions and are not directly comparable.

Examples

• Microsoft and Constellation plan to restart the Crane nuclear plant, formerly Three Mile Island 1, in 2027, a year earlier than first planned (company statement, September 2025).
• In Texas, connection requests from large loads such as data centres passed 230 GW in January 2026, against a record peak demand of 85 GW (IEA).
• In Odense, Denmark, waste heat from Meta's data centre warms more than 12,000 homes through district heating (Ramboll).
• In Kenya, a planned 1 billion dollar data centre by Microsoft and G42 would need about a third of the country's power capacity and is unlikely to proceed (Semafor, May 2026).
• Since December 2025, new data centres in Ireland must bring their own generation or storage and cover 80% of their demand with new renewables within six years (CRU).

Why it matters

The energy layer explains why AI capacity is scarce in some places and abundant in others, why data centres move to the Nordics, Johor or Texas, and why electricity prices around AI hubs are rising. In many regions the grid now sets the pace of expansion. Every delay there travels up the stack into the price and availability of the models you use.

It also decides what you can honestly claim. Customers, investors and regulators ask about energy, water and emissions, and the EU label and the revised GHG Protocol will make the answers comparable. Knowing the difference between capacity and consumption, annual and hourly matching, location-based and market-based figures lets you read provider claims with a critical eye.

Common misconceptions

• One AI query uses about 3 Wh.

  For a short text prompt, current company figures and estimates are about 0.24 to 0.34 Wh and about 0.26 ml of on-site water (Google, OpenAI, Epoch). Long inputs, images, video and agents use far more.

• GW and TWh are interchangeable.

  GW measures power at one moment, TWh energy over time. One gigawatt running for a full year delivers at most 8.76 TWh.

• A low PUE means a green data centre.

  PUE measures the overhead for cooling and power conversion. It says nothing about carbon, water or idle chips, and the global average has stayed around 1.54 for years (Uptime Institute).

• Liquid cooling uses a lot of water.

  Liquid cooling in the rack can save water. Water use depends on how the building releases its heat: cooling towers consume water, dry coolers use more electricity (LBNL).

• 100% renewable means green power every hour.

  Usually it means annual matching with contracts and certificates. Hourly matching on the same grid (24/7) is the stricter standard, and the GHG Protocol is revising its rules in that direction.

• Nuclear power already runs today's AI.

  Most deals are contracts for future supply. The Crane restart targets 2027, the first small modular reactors around 2030 (Constellation, IEA). Existing plants such as Susquehanna deliver through contracts.

• More efficient AI will lower total energy use.

  Google's energy per prompt fell 33-fold within a year, while data centre electricity worldwide grew 17% in 2025. Efficiency can raise total use, a rebound effect (Luccioni, Strubell and Crawford).

Perspectives from leading experts

Ten perspectives from an energy agency, research, critical scholarship and the regions where data centres are being built. Each summary paraphrases a published work, linked below it. Interests are marked.

• Fatih Birol

  Executive Director of the International Energy Agency, Paris · neutral

  Data centre electricity rose 17% in 2025, AI-focused sites grew much faster, and consumption is set to double by 2030. He presents AI both as a large new electricity consumer and as a tool to save energy elsewhere, and stresses that secure power supply is the precondition for AI.

  Data centre electricity use surged in 2025 (IEA news release) · 16 Apr 2026

• Sasha Luccioni and Emma Strubell

  Researcher at Hugging Face, Montreal, Canada, and researcher at Carnegie Mellon University, USA · research

  Measuring 88 models on ten tasks, they found that generating images used more than 1,400 times the energy of classifying text. General generative models used far more energy than small task-specific models for simple jobs, without clearly better accuracy. At scale, the energy of answering overtakes the energy of training.

  Power Hungry Processing: Watts Driving the Cost of AI Deployment? (ACM FAccT 2024) · Jun 2024

• Shaolei Ren and co-authors

  Researchers at UC Riverside and partners, work on the water and health costs of AI, USA · academic

  Data centres and the power plants that feed them emit air pollutants, a cost largely missing from sustainability reports. The team projects the annual US public health burden above 20 billion dollars in 2028, with households in the hardest-hit counties carrying up to about seven times the national average. They propose scheduling and siting computing with health costs in mind.

  Health-Informed Computing, first published as The Unpaid Toll (arXiv) · 23 Oct 2025

• Alex de Vries-Gao

  Data scientist at VU Amsterdam, Netherlands · research

  Working from chip supply, he estimates that the AI accelerators shipped in 2023 and 2024 draw up to 9.4 GW with full systems and cooling, comparable to the electricity use of the Netherlands. He argues that AI's energy use is underestimated because companies stopped disclosing it and the EU AI Act requires energy reporting only for training.

  Commentary in Joule, VU Amsterdam news release · 22 May 2025

• Jonathan Koomey

  President of Koomey Analytics, long-time researcher on IT electricity use, USA · independent analysis

  Future data centre demand is deeply uncertain because both computing demand and efficiency gains are uncertain. He asks regulators to test the assumptions behind forecasts of explosive growth, since grid requests are often inflated or duplicated and IT energy forecasts have a history of overestimation.

  Separating fact from fiction in data center electricity forecasts: a guide for regulators (with GridLab) · 23 Mar 2026

• Kate Crawford

  Professor at USC and researcher at Microsoft Research, born in Australia · critical research

  Companies keep the energy and water data of AI largely secret, and betting on future breakthroughs such as fusion distracts from what is possible now, such as smaller models and better data centre design. She calls for mandatory measurement and public reporting, independent audits and legal benchmarks, because voluntary measures rarely create accountability.

  Generative AI's environmental costs are soaring and mostly secret (Nature) · 20 Feb 2024

• Shabrina Nadhila

  Energy analyst for Asia at the think tank Ember, Jakarta, Indonesia · policy research

  Solar and wind could cover up to about 30% of Southeast Asia's data centre electricity by 2030 without batteries. Without action, the boom risks derailing the region's energy transition; in Malaysia, data centre demand could rise from 9 TWh in 2024 to 68 TWh in 2030, about 30% of national consumption.

  Solar & wind could power up to a third of ASEAN's data centres in 2030 (Ember) · 27 May 2025

• Lee Poh Seng

  Associate Professor at the National University of Singapore, leads the Sustainable Tropical Data Centre Testbed · academic

  Cooling can take up to 40% of a data centre's electricity, most of all in hot and humid climates. His testbed aims to cut energy use by about 40% and water use by 30 to 40% compared with typical sites, testing immersion cooling, direct-to-chip liquid cooling and warmer supply air.

  Taking a byte out of data centre emissions (NUS) · Apr 2024

• Tania Rodríguez

  Member of the citizens' movement MOSACAT, on TIME's AI 100 list 2024, Santiago, Chile · critical research

  Google's original plan in Cerrillos would have drawn 169 litres of water per second in a region marked by years of drought and went through a lighter environmental review. Residents' resistance forced Google to drop the design; she criticises that the air-cooled redesign was approved without a new assessment.

  Interview with MOSACAT members (Down To Earth) · 26 Jun 2026

• Karikari Achireko

  Director for Corporate Strategy and Global Partnerships, Africa Sustainable Energy Centre · policy research

  Africa's data centre demand is about 0.4 GW today and could reach 1.5 to 2.2 GW by 2030. Because grids are still unstrained by such loads, countries can plan power, fibre and water together in designated zones and set fair rules on who pays for grid upgrades, learning from Ireland and Amsterdam.

  Are Africa's power grids ready for the AI data centre boom? (World Economic Forum) · 23 Jul 2026

Terms for this layer

• Data CentreA building full of servers, storage and network equipment, with the power and cooling that AI runs on.
• PUEThe standard efficiency figure for data centres: total electricity of the site divided by the electricity that reaches the IT equipment.
• WUEHow much water a data centre uses on site, measured in litres per kWh of IT electricity.
• Liquid CoolingCooling chips with a liquid that flows through cold plates on the chip or around the whole server, needed for dense AI racks.
• Grid ConnectionThe physical link between a data centre and the power grid, and the approval to draw a given amount of power; today the longest wait in many AI projects.
• Power Purchase AgreementA long-term contract to buy electricity, often from a specific wind, solar or nuclear plant; the main way AI companies buy clean power.
• Scope 2 EmissionsThe greenhouse gas emissions caused by the electricity a company buys, counted either by the local grid (location-based) or by its contracts (market-based).
• Rebound EffectWhen efficiency makes something cheaper to use, people use more of it, so total consumption can rise.
• GPUA graphics processor: a chip that runs thousands of simple calculations at once, which made it the workhorse of AI.
• AI ChipA processor built for the calculations AI needs, such as a GPU or a custom chip, trading flexibility for speed and efficiency.
• HBMVery fast memory stacked right next to an AI chip, so its processors never wait for data. In 2025, it was a key bottleneck.
• TrainingThe one-time, very expensive process that sets a model's parameters by showing it vast amounts of data.
• InferenceRunning a finished model to get an answer. What you pay for every time you send a message.
• TokenThe small chunk of text a model reads and writes. Price, speed and memory limits are all counted in tokens.
• Reasoning ModelA model that generates intermediate working steps before its final answer, trading time and tokens for better results on hard problems.
• AI AgentA model connected to tools and given a goal, so it can take steps on its own instead of only answering.
• Cost per TaskWhat one finished piece of work really costs with AI: tokens, tools, retries and the human time to check it.
• Small Language ModelA compact language model that runs fast and cheaply, often on a laptop or phone, at the cost of some capability.
• Model RoutingSending each request to the model that fits it best, small and cheap for easy jobs, large for hard ones.
• Cloud ComputingRenting computing power, storage and software over the internet, on demand, instead of running your own machines.
• EmbeddingsText turned into lists of numbers that capture meaning. The basis of semantic search and RAG.
• Responsible AIBuilding and using AI in ways that are fair, safe, transparent and accountable, and to show it in daily practice.

Sources

1. Key Questions on Energy and AI (full report), International Energy Agency, 16 Apr 2026
2. Key Questions on Energy and AI: executive summary, International Energy Agency, 16 Apr 2026
3. Data centre electricity use surged in 2025, International Energy Agency, 16 Apr 2026
4. Powering the Digital Economy: key questions on energy and AI (presentation), IEA, at the Asia Clean Energy Forum, 11 Jun 2026
5. 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, Dec 2024
6. Backlog of power plants seeking grid connection eased somewhat in 2025 (Queued Up), Lawrence Berkeley National Laboratory, 1 Jul 2026
7. Global Data Center Survey 2025, Uptime Institute, 29 Jul 2025
8. GB200 NVL72 rack, Supermicro (vendor), Oct 2026
9. Measuring the environmental impact of delivering AI at Google scale, Elsworth et al., Google (vendor), 21 Aug 2025
10. How much energy does ChatGPT use?, Epoch AI, 7 Feb 2025
11. AI data centre power capacity, Epoch AI, 16 Jan 2026
12. From Efficiency Gains to Rebound Effects: The Problem of Jevons' Paradox in AI, Luccioni, Strubell, Crawford (FAccT 2025), 13 Jun 2025
13. Data centers consume massive amounts of water, companies rarely say exactly how much, McCauley and Scanlan, The Conversation, republished by Civil Beat, 25 Aug 2025
14. Data centers' diesel generators are a top concern for Virginians, Ivy Main, Virginia Mercury (commentary), 14 Jan 2026
15. Grids for data centres: ambitious grid planning can win Europe's AI race, Ember, 19 Jun 2025
16. Transformers in 2026: shortage, scramble or self-inflicted crisis?, POWER magazine, citing Wood Mackenzie, 2 Jan 2026
17. Data centre copper demand, MINING.COM, citing BloombergNEF, 12 Aug 2025
18. PJM capacity auction for 2027/28, Utility Dive, 18 Dec 2025
19. US grid headroom for flexible loads (Duke University study), Utility Dive, 11 Feb 2025
20. Powering Intelligence 2026, EPRI, via Energy Digital, 26 Feb 2026
21. Crane Clean Energy Center ahead of schedule, Constellation (vendor), 23 Sep 2025
22. Amazon, Meta, Google, Microsoft account for half of global clean energy purchases in 2025, ESG Today, citing BloombergNEF, 23 Feb 2026
23. What does 24/7 carbon-free energy mean?, David Roberts, Canary Media, 12 Nov 2021
24. Scope 2 standard advances to consultation, GHG Protocol, 1 Aug 2025
25. Data centre emissions over 660% more than Meta, Microsoft, Google and Apple report, TechInformed, citing a Guardian analysis, 16 Sep 2024
26. Google's AI boom drives up emissions and power use, Axios, 30 Jun 2026
27. Meta surplus heat to district heating, Odense, Ramboll (engineering firm on the project), 2023
28. Exposing Google's seizure of water in drought-impacted Uruguay, Logic(s) magazine, interview with Daniel Pena and Enol Nieto Jiménez, 1 Nov 2025
29. EU Commission adopts data centre sustainability rating scheme, Linklaters, 14 Oct 2026
30. Factbox: where authorities are restricting data centres amid AI boom, Reuters, via WHBL, 14 Jul 2026
31. Data centres used 23% of Ireland's electricity in 2025, Irish Examiner, citing the Central Statistics Office, 7 Jul 2026
32. Explainer: how China is managing the rising energy demand from data centres, Yuan Ye, Carbon Brief, 16 Apr 2025
33. Nordic data centres to support power demand growth, Argus, citing Statnett, 26 Mar 2026
34. Data centre boom in Malaysia, ASEAN+3 Macroeconomic Research Office (AMRO), May 2026
35. 12.3 GW energy capacity approved for data centres, The Star, parliamentary reply, 20 Oct 2026
36. Singapore announces data centre capacity allocation call, Morgan Lewis, 3 Mar 2026
37. Why India must plan for AI data centres' energy and water footprint, Council on Energy, Environment and Water (CEEW), 19 Jan 2026
38. Energy shortfall scuppers Kenya's 1 billion dollar Microsoft data centre, Semafor, 6 May 2026
39. Brazil receives 38 GW of grid access requests for data centres, Canal Solar, 1 Jun 2026
40. AI and the environment: Kenya pushes global accountability for data centre emissions, Kimani Mwangi, Daily Nation, 11 Dec 2025