AI data centres used 50% more electricity in 2025.

all data centres combined? up 17%. and the IEA expects the total to roughly double by 2030 (yes, double).

so AI energy demand isn't a side story for engineering and sustainability students anymore.

it's pretty much the job market walking towards you.

What's actually happening with AI energy demand

The IEA's Key Questions on Energy and AI came out in April and the numbers are pretty blunt:

→ data centres used 485 TWh in 2025 → the IEA projects around 950 TWh by 2030, roughly 3% of global electricity demand → AI-focused data centres alone are expected to triple over the same period

and the money behind it is the bit that made us look twice.

Eco-Business's write-up of the report notes Big Tech capex passed $400 billion in 2025 and is expected to rise about 75% in 2026.

it also notes orders for gas turbines for data centres rose 70% in 2025.

so a lot of this new load isn't waiting for wind farms.

the grid is now the bottleneck, and September showed both responses:

  1. the flexibility pitch. Google, Nvidia and Emerald AI launched the AI Energy Management Alliance, with members including Anthropic and utilities. The idea: data centres dial their power down when the grid is stressed, in exchange for faster connections.
  2. the pushback. In Virginia, Loudoun County's board voted to start a process that could pause final decisions on data centre and substation applications for up to 12 months.

Emerald AI's founder argues in Fortune that moderate flexibility could unlock 100GW on the existing US grid (citing a Duke study).

interesting claim. also written by someone whose company sells that flexibility (we'll come back to this).

How AI is used inside the energy sector

this is the half most headlines skip.

AI isn't just a new load, it's also a tool for running the energy system.

the IEA lists the real uses:

→ monitoring transformers and grid equipment to cut unexpected failures → squeezing more out of existing grid capacity instead of building new lines straight away → optimising production in energy-heavy industry, which could cut energy costs by 3 to 10 percentage points

well-documented use cases could save over 13 EJ by 2035, about 3% of global final energy consumption.

BUT the IEA says the energy sector "is not yet taking full advantage" of AI.

and the single biggest barrier in its survey? a lack of digital skills.

honestly, as a student that's the best news in the whole report. the gap is people who can do both energy and data.

What uni won't teach you about AI in energy

most energy and engineering degrees are solid on the physics. where they fall short:

  1. demand is treated as a number you're given. in power systems modules the load is a forecast. in real life the biggest new load is a customer that negotiates, shifts and lobbies.
  2. data skills are a side module. a term of MATLAB isn't enough when the IEA names digital skills as the main thing holding the sector back.
  3. messy data is the norm. only 10% of global electricity consumption is covered by open electricity data policies, so you won't be handed clean datasets at work.
  4. nobody teaches you to read a deal. power purchase agreements, flexibility contracts and "carbon-free" pledges are where your engineering meets someone's marketing.

The critical-thinking move: 5 questions for any AI energy claim

next time you read a headline about AI and power, run it through this:

  1. energy or capacity? TWh is energy used over a year. GW is power at a moment. "100GW unlocked" is about peak hours, not total supply. mixing them up is the most common mistake in this debate.
  2. who's saying it, and what do they sell? an IEA projection and a founder's op-ed can both be useful. they're not the same kind of evidence.
  3. measured or modelled? "demand grew 17% in 2025" is measured. "could save 13 EJ by 2035 if barriers are overcome" is modelled. both matter, only one already happened.
  4. where does the electron actually come from, and when? a "100% clean" pledge over a year can still mean gas at 7pm in January. ask about the hour, not just the annual total.
  5. who pays, and who can say no? local councils, grid operators and bill payers all have a veto of some kind. Loudoun just showed that.

What to do this month

  1. read the IEA executive summary (about 20 minutes) and turn it into a one-page cheat sheet of numbers you can quote in an interview.
  2. plot real demand data. download a week of demand data from your country's grid operator, plot it in Python and mark the peak hours. now you understand what "flexibility" is actually selling.
  3. audit one deal. pick any data centre power deal in the news, run the 5 questions above and write 200 words. that's a portfolio piece, not homework.
  4. learn one forecasting model. a simple time-series model on that same demand data. it's exactly the digital skill the IEA says the sector is short of.

so TLDR: AI is the fastest growing new load on the grid AND one of the best tools for running it. the students who can talk about both will be the ones every utility wants.

if a data centre offered to cut its power during your grid's worst hours, would you let it skip the connection queue?

Sources

Get the next issue by email

theVANTAGEd is written by students to bring critical thinking to AI across real industries. One issue every two weeks.