The macro picture: demand is doubling, AI is tripling
The International Energy Agency's April 2026 update, Key Questions on Energy and AI, is not what I'm citing here. I'm relying on the IEA's own text instead. placeholder
Correction on sourcing: the IEA's report says placeholder
I'll state the IEA findings plainly, drawn from the IEA's published executive summary and press release:
- Data centre electricity use is projected to roughly double, from 485 TWh in 2025 to 950 TWh in 2030. That is around 3% of global electricity demand by 2030.
- Electricity use by AI-focused data centres is projected to triple over the same period.
- Data centre demand grew 17% in 2025, while AI-focused data centres grew about 50%.
- Bottlenecks in transformers, gas turbines, chips, grid connections and approvals are limiting near-term growth. The IEA also flags a shortage of high-bandwidth memory that it expects to persist through at least the end of 2027.
- The IEA says per-task AI energy use is falling fast, but agentic, reasoning and video workloads can use hundreds or thousands of times more energy per query than a simple text prompt.
These are global figures for data centres. They are not a sum of what Amazon, Microsoft, Google, Meta, Oracle and CoreWeave will use.
Why a "who needs the most" ranking doesn't hold up
Three mismatches make a clean ranking misleading:
- Power versus energy. Gigawatts measure a rate of draw. Terawatt-hours measure energy over time. A campus announced at 5 GW is not using 5 GW today.
- AI versus everything else. Company-wide electricity figures include search, video, ads, storage and ordinary cloud workloads.
- Different boundaries. Reporting years, fiscal calendars and what counts as a data centre differ from company to company.
The source report itself concedes that no company publishes a comparable AI-only forecast through 2030. A qualitative table of "very high" and "high growth" labels is an opinion, not a measurement. Treat it that way.
What the companies actually disclose
Google
Alphabet's 2025 Environmental Report covers calendar year 2024. Per the report schedule summarized in the research record, it lists about 30.8 TWh of data-centre electricity and about 32.2 TWh of total electricity. The 30.8 TWh figure is data-centre use, not AI-only use. Anyone who writes "Google used 30.8 TWh on AI" is wrong.
Microsoft
Microsoft's datacenter sustainability page reports efficiency, not a total. placeholder Microsoft's published figures are:
- Global power usage effectiveness (PUE) of 1.17 for fiscal 2025 (July 1, 2024 to June 30, 2025), versus 1.16 in fiscal 2024.
- Those metrics cover only facilities Microsoft fully owns and controls that had been operational for 12 months.
- A pledge to match 100% of electricity consumption with zero-carbon purchases by 2030, and 40 GW of new renewable supply contracted across 26 countries.
PUE says how much overhead a facility burns, not how much power the company draws. That is why Microsoft can look efficient and still be among the biggest consumers.
Meta
Meta's Hyperion campus in Richland Parish, Louisiana, is the best-documented gigawatt project in the comparison. Per CNBC and other coverage of Meta's July 13, 2026 announcement, Meta said the campus will scale to 5 GW at a cost of more than $50 billion. That is up from roughly $27 billion and an original 2 GW design. A Meta spokesperson said the project should reach 2 GW by 2030, though no timeline exists for the full 5 GW buildout.
So "5 GW by 2030" would be wrong. The first phase is about 2 GW by 2030.
IEEE Spectrum reports that Meta secured power for Hyperion by negotiating with Entergy for construction of three new gas-turbine power plants. Reports on the final generation count vary widely. Different outlets cite different numbers of plants and capacity, so I'd treat those figures with caution.
On Meta's reported consumption, I'd note one caution. The 18.1 TWh figure for 2024 in the source article is described there as data-centre electricity. The available primary-source excerpt only confirms a 2023 data-centre figure of about 15.0 TWh. Check the boundary before comparing it with Google's data-centre number.
Amazon, Oracle, CoreWeave, xAI, Crusoe, Nebius, OpenAI
The research record found no comparable company-specific electricity totals or verified 2030 AI-only estimates for these. Constellation's recent newsroom listing does show a September 30, 2026 20-year power purchase agreement with Amazon adding 190 MW of nuclear capacity at Calvert Cliffs. That is a procurement deal, not a consumption figure.
One useful rule from the source: electricity used by Oracle-hosted OpenAI workloads shouldn't be counted twice, once for Oracle and once for OpenAI, when estimating industry totals.
Power deals are not load
Two nuclear deals show how easily procurement gets confused with demand.
Microsoft and Constellation (September 20, 2024). Constellation announced a 20-year power purchase agreement tied to restarting Three Mile Island Unit 1 as the Crane Clean Energy Center. It would add approximately 835 MW of carbon-free generation, and the agreement supports Microsoft's goal of matching data-centre power use in PJM with carbon-free energy. Constellation's release says restarting requires Nuclear Regulatory Commission approval and state and local permits. It says the plant is expected online in 2028.
Google and Constellation (October 6, 2026). The companies announced a 20-year agreement for 890 MW of new nuclear capacity on the PJM grid, through upgrades at 11 Constellation-owned units in Illinois, Pennsylvania and New Jersey. The first uprate is expected by 2028. Separately, they signed a 15-year energy supply agreement for 2,700 MW from Constellation's existing PJM fleet.
Don't add 890 and 2,700 together and call it new generation. Constellation's release says the 890 MW is entirely new capacity. The 2,700 MW deal is a financial structure meant to keep existing plants economically viable. Neither figure measures Google's AI load.
The release also describes a demand-flexibility component, with load-shaping and demand-response capabilities to curtail non-critical consumption during high-stress grid events. That is a notable detail. It suggests big AI loads are being asked to behave like grid participants, not just customers.
Scale arithmetic: what a gigawatt means
If a facility ran at a steady load all year, the conversion is simple: 1 GW times 8,760 hours is 8.76 TWh. At an assumed $0.08 per kWh, that is about $701 million a year.
| Continuous load | Annual energy | Cost at $0.08/kWh |
|---|---|---|
| 100 MW | 0.876 TWh | about $70 million |
| 500 MW | 4.38 TWh | about $350 million |
| 1 GW | 8.76 TWh | about $701 million |
| 2 GW | 17.52 TWh | about $1.40 billion |
| 5 GW | 43.8 TWh | about $3.50 billion |
These are illustrations, not forecasts. Real consumption depends on how much capacity is energized, utilization, outages, contract structure, demand charges and transmission costs. The 5 GW row is not Hyperion's expected consumption.
The IEA also says AI server power density rose 11-fold between 2020 and 2025, with a further fourfold increase expected by 2027. By then a single advanced rack could have peak demand equal to about 65 households. That is the physical pressure behind the transformer and cooling bottlenecks.
What this means for Windows and Microsoft admins
- Capacity can be a supply-chain problem, not a software one. The IEA's bottleneck list explains why new regions and GPU capacity can arrive late, regardless of announcements.
- Efficiency per task is improving, but total demand isn't falling. Agentic features like Copilot workflows are exactly the kind of energy-intensive use the IEA flags.
- Costs may flow downstream. This is general industry reasoning, not a sourced prediction. Power and grid constraints are a plausible pressure on AI service pricing and availability over time.
- Read sustainability claims carefully. Matching electricity with clean purchases is an accounting claim, not proof that facilities draw clean power every hour.
The bottom line
Amazon, Microsoft, Google and Meta are the likeliest to have the largest absolute AI-related demand, given their existing scale. That is reasoned inference, not a measured ranking. Oracle, CoreWeave, xAI, Crusoe and Nebius may grow faster in percentage terms from smaller bases. The IEA warns that the pace of growth is sensitive to financing and market sentiment as well as to power.
The reliable lesson is that the AI race has moved from chips to connections. The winners will be whoever turns announced gigawatts into energized, connected, utilized capacity.