◦DeepSeek training·~2k H800 GPUs for V3 · a few MW · orders of magnitude lessDeepSeek
Energy Sourcing — The Nuclear Rush
committed ~10 GW+ nuclear PPAs
online today 0 GW (all 2027+)
Reactor restarts & SMR deals
☢Meta·Vistra (existing plants + uprates), TerraPower Natrium and Oklo Aurora — up to 6.6 GW, announced Jan 2026Meta
☢Microsoft·Crane Clean Energy Center — restarting Three Mile Island Unit 1, 835 MW, targeting 2027Microsoft
☢Amazon·X-energy Xe-100 SMRs, up to 960 MW at Richland WA (320 MW phase 1) + Talen Susquehanna nuclear PPAAWS
☢Google·Kairos Power KP-FHR fluoride-salt-cooled SMRs, 500 MW fleet, first unit targeted 2030Google
None of these reactors are online yet — every deal above targets 2027 or later. They exist because 24/7 carbon-free baseload is the only thing that matches a datacenter's flat, always-on demand curve at gigawatt scale, and interconnection queues for new gas or transmission capacity can run 5+ years. Nuclear is a bet that AI demand keeps compounding long enough for these reactors to matter.
Water Footprint
▸ cooling doesn't run on electricity alone
US direct + indirect (2023)
17.4B + 211B gal/yr
Google, 2025
10.9B gal (+34% YoY)
Per query, cooling only
~0.3 mL
Per query, full accounting
5–25 mL
Microsoft WUE (2025)
0.27 L/kWh
Amazon WUE (2025)
0.12 L/kWh
Evaporative cooling towers are the default for hyperscale AI datacenters because they're cheap and efficient — but they consume potable or reclaimed water directly. "Full accounting" figures also include water evaporated upstream to generate the electricity itself. An estimated 92.6% of US datacenter water use still goes unreported in corporate sustainability disclosures.
Methodology
▸ how we estimated
TRAINING GW is peak power draw of the cluster used to train the flagship model, averaged over the training run. Where the lab doesn't disclose, we back-compute from GPU count × TDP × utilization (typ. 0.4–0.6).
INFERENCE GW is steady-state serving power for the model's primary public product. This is harder to pin down — most figures are estimates derived from disclosed query-volume × per-query Wh.
PROJECTIONS use publicly-announced datacenter capacity (Stargate, Hyperion, Colossus 2, etc.) scaled to when the next flagship model is expected. Anything after 2027 is noted as speculative.
Sources: Epoch AI, IEA Electricity 2024, OpenAI's Stargate announcements, Meta capex calls, xAI public statements, MIT Technology Review, Data Center Frontier, IEEE Spectrum, SemiAnalysis. Error bars are wide — treat numbers as order-of-magnitude.
▸ power-draw --summaryexit 0
jasper@grid:~/ai-gigawatts$summary
✓ top 10 AI models draw ~7.7 GW in 2026 (estimated, operational)
→ projected ~22 GW by end-2027 if announcements land on time
⚠ inference now dominates: ~80–90% of all AI electricity use
⚠ xAI Colossus ran on unpermitted gas turbines (EPA violation, 2025)
# DeepSeek trained V3 on ~2k GPUs — proves efficiency gaps are possible
✗ most labs do not publish training energy — all numbers are estimates
Guide — read the AI power race in five steps
Estimated electricity draw of the ten biggest AI model families, in gigawatts.
1Start with the summary
Combined draw today, the 2027 projection and the biggest single campus.
2Compare the models
Bars show current versus projected GW per model family, with its cluster and GPU count.
3Split training and inference
See how much of each lab’s power goes to building models versus answering prompts.
4Put it in context
Compare AI campuses with a reactor, the Hoover Dam, Belgium and all US data centres.
5Check the sourcing
Nuclear deals and water use per company, with the caveat that every figure is an estimate.
Training and running large AI models now takes data-centre campuses measured in gigawatts — the scale of power stations. This page estimates how much electricity the ten largest AI model families draw today and what the announced build-outs would add by 2027, and puts those numbers next to things you know.
What’s the difference between gigawatts (GW) and gigawatt-hours (GWh)?
GW is power — the rate at which electricity is drawn at a given moment. GWh is energy — power over time. A 1 GW data centre running all year uses about 8,760 GWh. This page uses GW because that is what a grid operator has to provision for, and what data-centre announcements quote.
Who uses the most power: OpenAI or xAI?
It depends on what you count. In announced capacity, OpenAI leads: its Stargate programme targets about 10 GW across seven US sites by 2029. In running capacity in 2026, xAI’s Colossus sites in Memphis are ahead after a very fast build-out, while Stargate’s first campus in Abilene, Texas had about 0.3 GW running in April 2026 and aims for 1.2 GW by late 2026.
Training or inference — which uses more?
Inference, increasingly. Training a frontier model is a huge but one-off burst; answering billions of prompts every day adds up faster. Industry estimates put inference at roughly 80–90% of AI computing today. The Training vs Inference panel shows the estimated split per model. Google is one of the few to publish a per-prompt figure: about 0.24 Wh for a median Gemini text prompt (2025).
Why are DeepSeek and Mistral so much lower?
DeepSeek reported training V3 on about 2,000 Nvidia H800 GPUs for roughly two months — a few megawatts including cooling, orders of magnitude below the frontier US clusters. Both DeepSeek and Mistral use mixture-of-experts models that activate only part of the network per token (DeepSeek V3: 37B of 671B parameters), and both serve far fewer users than ChatGPT or Gemini. Mistral also published a life-cycle analysis putting a one-page answer at about 1 g CO₂e.
How reliable are these numbers?
Treat them as estimates with ±30–50% error bars. Most labs publish neither training energy nor inference power. The figures combine data-centre announcements, GPU counts × power per chip, utility filings and third-party analyses (Epoch AI, IEA, LBNL). The ranking is more robust than any single number.
Are AI data centres switching to nuclear?
Every hyperscaler has signed nuclear deals, but most new reactors arrive in the late 2020s or 2030s. Meta announced up to 6.6 GW in January 2026 across Vistra (existing plants plus upgrades), TerraPower and Oklo. Microsoft has a 20-year deal to restart Three Mile Island Unit 1 (835 MW, now the Crane Clean Energy Center), targeted for 2027. Amazon backs X-energy small modular reactors (320 MW first phase, up to 960 MW) and buys from Talen’s Susquehanna plant. Google ordered a 500 MW fleet of Kairos Power reactors, first unit targeted for 2030.
How much water do AI data centres use?
A lot, mostly for evaporative cooling. Google reported 10.9 billion gallons of water consumption in 2025, up 34% on 2024. Efficiency varies widely: Microsoft reports a water usage effectiveness of 0.27 L/kWh and Amazon 0.12 L/kWh for 2025. Per prompt, estimates range from a fraction of a millilitre for on-site cooling to tens of millilitres once the water used to generate the electricity is included.
How big is a gigawatt, really?
One gigawatt is roughly one large nuclear reactor, or the average electricity use of about 800,000 US homes (at ~1.2 kW each). Belgium uses about 9 GW on average. All US data centres together drew about 20 GW on average in 2023 (176 TWh, Lawrence Berkeley National Laboratory) — and AI is the main reason that number is expected to double or triple by 2028.