AI Energy Consumption Forecast 2027
By Axis Intelligence Research
Co-authors: Sarah Mitchell and Aidan Jad | Last updated: October 28, 2026 | License: CC BY 4.0
Quick Answer:
Axis Intelligence Research projects global data center electricity use of about 740 TWh in 2027 in its base case, up from 485 TWh in 2025 (IEA). That is above Gartner’s 702 TWh, but only because the IEA’s 2025 starting point is 38 TWh higher; Axis expects slightly slower growth. AI-focused facilities reach 48% of the total.
Key Findings
- Axis Intelligence Research projects global data center electricity consumption of about 740 TWh in 2027 in its base case, with an upside of 775 TWh and a downside of 700 TWh (as of October 28, 2026).
- Axis Intelligence Research projects AI-focused data centers will consume about 357 TWh in 2027, or 48% of all data center electricity, up from roughly 30% in 2025.
- Axis Intelligence Research finds that about 51% of the AI electricity added in 2027 comes from hardware shipped in 2026, so most of next year’s increase was bought before it began.
- Axis Intelligence Research finds that the entire 38 TWh gap between its 2027 base case and Gartner’s 702 TWh comes from the 2025 baseline (485 TWh at the IEA, 447 TWh at Gartner); on growth, Axis is lower (+52% over two years versus +57%).
- Two forecasters cut their near-term numbers in 2026: the EIA lowered 2027 US commercial-sector power growth, which includes data centers, from 5.3% to 2.8% between its May and October outlooks, and Gartner’s 2027 figure for AI-optimized servers fell from 500 TWh (November 2024) to 258 TWh (June 2026).
How Much Electricity Will AI Use in 2027?
Data centers worldwide will probably use between 700 and 775 TWh in 2027, and AI-focused sites will draw just under half of it. For scale, the IEA put the whole sector at 485 TWh in 2025, after 17% growth that year, with AI-focused facilities up 50%.
Axis Base Case and scenarios, 2025–2027
| Metric (world, TWh) | 2025 | 2026 (Axis) | 2027 Downside | 2027 Base | 2027 Upside | Source |
|---|---|---|---|---|---|---|
| AI-focused data centers | 148 | 239 | 348 | 357 | 389 | IEA (2025, chart reading); Axis Intelligence Research (2026–27) |
| Non-AI data centers | 337 | 358 | 353 | 380 | 388 | IEA (2025); Axis Intelligence Research |
| All data centers | 485 | 595 | 700 | 740 | 775 | IEA (2025); Axis Intelligence Research |
Totals for 2026 and 2027 are rounded to the nearest 5 TWh, so rows don’t always add exactly. The IEA publishes the 2025 AI-focused figure only as a bar in Figure 1.5 of Key Questions on Energy and AI; the 148 TWh is an Axis measurement of that chart (±3 TWh), consistent with the IEA’s statement that AI-focused use “more than triples” to around 465 TWh by 2030. Rows for 2026–27 are projections, not reported data.
A note on definitions, because two Axis pages carry two different “AI shares” for 2027. Our AI data center energy consumption statistics page publishes an AI-attributable estimate of 395.4 TWh (56.33%) for 2027, built on Gartner’s server segments with cooling allocated to AI servers. This page uses the IEA’s definition, which counts whole facilities built around accelerators. The first answers “how much power do AI servers pull, wherever they sit”; the second answers “how much power do AI data centers pull.” They are different metrics, and neither replaces the other.
Here’s what the table can’t tell you. The 2026 column doesn’t move between scenarios, because 2026 is mostly in the books: Nvidia has reported two quarters of fiscal 2027 and guided the third. Even the downside adds load in 2027, because the racks delivered in late 2026 get energized in the first half of 2027. Every interconnection engineer knows the pattern: the purchase order clears, the substation lags, and the load shows up a year later.
How Does the Axis Base Case Get to 740 TWh?
Most published forecasts apply one growth rate to a market total. This one splits consumption into two loads and builds the AI side from what was actually bought.
The formula, in plain text:
AI-focused consumption (year t) = consumption (t−1) + ½ × spend (t−1) × k (t−1) + ½ × spend (t) × k (t)
Non-AI consumption (year t) = 337 TWh × (1 + g)^(t − 2025)
Spend is Nvidia’s data center revenue (company-reported, used as a proxy for all AI hardware bought each year), k is the power intensity of that spend, and the two halves reflect hardware arriving, on average, mid-year.
Calibrating k: what a dollar of AI hardware draws
k is an Axis Intelligence Research calculation, fitted to 2025:
- AI-focused consumption rose from about 99 TWh in 2024 to 148 TWh in 2025 (148 ÷ 1.5, using the IEA’s 50% growth figure), an increase of 49.3 TWh.
- Nvidia reported data center revenue of $115.2 billion for fiscal 2025 and $193.7 billion for fiscal 2026.
- k = 49.3 ÷ (0.5 × 115.2 + 0.5 × 193.7) = 31.9 TWh of annual consumption per $100 billion of Nvidia data center revenue.
That ratio is empirical, not physical. It folds in Google’s TPUs, AMD parts, Chinese domestic accelerators, the CPUs and storage in AI halls, and cooling, all scaled to the one spending series that is reported quarterly and audited. Track that series on our Nvidia statistics page.
A buyer-side check lands in the same place. Nvidia’s CFO told analysts in August 2025 that cloud providers and enterprises were on track to spend about $600 billion on data center infrastructure and compute in calendar 2025, and CEO Jensen Huang put a gigawatt AI factory at $50 billion to $60 billion (Nvidia Q2 fiscal 2026 earnings call). That buys 10 to 12 GW. At the IEA’s own 2025 data center capacity factor of 48%, it works out to roughly 42 to 50 TWh a year, bracketing the 49.3 TWh the IEA observed. Gartner’s figures point the same way: 447 TWh on 104 GW of demand in 2025 implies a 49% load factor.
Driver inputs
| Driver | Value | Basis | Source |
|---|---|---|---|
| AI-focused consumption, 2025 | 148 TWh | Chart reading, ±3 TWh | IEA Figure 1.5, April 2026 |
| Non-AI consumption, 2025 | 337 TWh | 485 − 148 | IEA, April 2026 |
| Power intensity, k | 31.9 TWh per $100bn | Calibrated on 2025 | Axis Intelligence Research |
| Nvidia DC revenue, fiscal 2027 | ≈ $376bn | Q1 $75.2bn + Q2 $89.0bn reported; Q3 from $108.0bn total guide × 92.5% data center share; Q4 assumes the same 12.2% sequential step | Nvidia; Axis Intelligence Research |
| Nvidia DC revenue, fiscal 2028 (base) | ≈ $430bn | $1 trillion Blackwell and Rubin order outlook through 2027, minus fiscal 2026 and 2027 revenue | Nvidia GTC keynote, March 16, 2026; Axis Intelligence Research |
| k for 2027 shipments (base) | −15% | More dollars per gigawatt as Rubin systems and memory costs rise | Axis Intelligence Research assumption |
| Non-AI growth (base) | +6.2% a year | Observed 2025 rate: 337 ÷ 317.3 | IEA; Axis Intelligence Research |
What separates the three scenarios
- Base (740 TWh): spending follows Nvidia’s order outlook; each 2027 dollar draws 15% less power; half of each year’s shipments draw power in-year.
- Upside (775 TWh): 2027 spending runs 50% above fiscal 2027 and power intensity holds; non-AI load grows 7.3% a year, the pace implied by the IEA’s 2030 Base Case (945 TWh total, 465 TWh AI-focused).
- Downside (700 TWh): 2027 spending stays flat, power intensity drops 30%, and grid congestion means only 30% of 2027 shipments are energized in-year; non-AI load grows 2.4%, Gartner’s 2027 rate for conventional servers.
Sensitivity: which assumption matters
If power intensity is off by 20% in either direction, the 2027 total moves by about 36 TWh. If 2027 chip spending is off by 20%, it moves by about 12 TWh. The number to watch isn’t next year’s capex headline. It’s how many watts each dollar buys, so system and memory pricing in Nvidia’s quarterly commentary matters more to this forecast than a capex beat.
Why Do 2027 Data Center Electricity Forecasts Disagree?
Mostly the starting point and what gets counted, not optimism.
Forecast reconciliation table
| Firm | Published | Metric | Base year | Horizon | Value | Scope note |
|---|---|---|---|---|---|---|
| Gartner | Jun 10, 2026 | All data centers | 2025 (447 TWh) | 2027 | 702 TWh | Servers + cooling and infrastructure |
| Gartner | Jun 10, 2026 | AI-optimized servers only | 2025 (95 TWh) | 2027 | 258 TWh | Excludes cooling |
| Gartner | Nov 12, 2024 | AI-optimized server power requirement | 2023 | 2027 | 500 TWh | Incremental AI servers; scope may differ from 2026 table |
| IDC | Sep 24, 2024 | AI datacenter energy | 2023 | 2027 | 146.2 TWh | AI workloads |
| IEA | Apr 2026 | All data centers | 2025 (485 TWh) | 2030 | 945 TWh | Base Case table; text rounds to about 950 |
| IEA | Apr 2026 | AI-focused data centers | 2025 (≈148 TWh) | 2030 | ≈465 TWh | Facility type, not workload |
| Axis Intelligence Research | Oct 28, 2026 | All data centers | 2025 (485 TWh) | 2027 | 740 TWh | Base case; range 700–775 |
| Axis Intelligence Research | Oct 28, 2026 | AI-focused data centers | 2025 (≈148 TWh) | 2027 | 357 TWh | Base case |
The IEA does not publish a 2027 value, and none is interpolated here.
Put the growth rates side by side and the disagreement reverses. Gartner’s path from 447 to 702 TWh is +57% over two years; the Axis base case from 485 to 740 TWh is +52%. Axis lands 38 TWh higher in 2027 for exactly one reason: it starts 38 TWh higher in 2025.
The 2025 baseline itself is contested. The Energy Institute’s Statistical Review of World Energy 2026 publishes a data center table for the first time, using S&P Global Energy data, and puts 2025 power demand at 787.8 TWh, up 19.7% on 2024. Its method differs at the root: an installed-base view of facility capacity in 47 countries, multiplied by assumed running rates and PUE, where the IEA builds up from server shipments. Neither is wrong; they measure different things, and any page quoting “data center electricity” without naming the method is quoting half a number.
Track record: the near-term number keeps shrinking
In November 2024, Gartner said the power needed to run AI-optimized servers would hit 500 TWh a year in 2027. Its June 2026 table puts AI-optimized server consumption in 2027 at 258 TWh, 48% lower (the two releases may not define the segment identically). Over the same period Gartner raised its long-range view, to more than 1,200 TWh for all data centers by 2030.
The EIA did something similar for the US. Its October 2026 outlook has commercial-sector electricity sales, the category that holds data centers, growing 2.8% in 2027; the May 2026 edition had 5.3%.
Short-term forecasts falling while long-term ones rise is what a supply-constrained market looks like. Demand isn’t softening. The interconnection queue is setting the pace.
Biggest AI Energy Trends for 2027
AI servers overtake conventional servers on the meter
Gartner’s June table has AI-optimized servers at 258 TWh in 2027 against 200 TWh for conventional servers, the first year the AI line is higher. Conventional servers grow 2.4% that year; AI servers grow 47.8%. The measured side of that split, year by year, lives on our AI data center energy consumption statistics page. What would prove it wrong: AI server shipments stalling in the first half of 2027, visible in Nvidia’s quarterly data center revenue.
Memory and the grid, not demand, set the ceiling
The IEA’s April 2026 report says bottlenecks across the value chain make its more aggressive near-term scenarios less likely. The hardest number in it: analysts estimate current high-bandwidth memory production can support at most around 25 GW of AI-ready servers a year through 2027, with the memory shortage lasting at least until late 2027. Gas turbine orders jumped 70% in 2025, which tells you where the queue for firm power sits. Gartner expects 40% of existing AI data centers to be operationally constrained by power availability by 2027. What would prove it wrong: the EIA restoring its 2027 commercial-sector growth toward the 5% range in its November or December outlooks.
Rack density quadruples, and cooling follows
AI server power density rose 11-fold between 2020 and 2025 and is set to rise another fourfold by 2027, according to the IEA, which describes a single rack approaching the peak draw of 65 households. The IEA’s tables show global PUE (power usage effectiveness, total facility power divided by IT power) improving from 1.38 in 2025 to 1.29 in 2030; Berkeley Lab sees the US average falling from 1.45 in 2024 to 1.36 in 2030, as more load sits in liquid-cooled AI halls. Denser racks lower the overhead per kilowatt-hour; they don’t lower the kilowatt-hours. The water side of that trade-off is on our AI data center water usage statistics page. What would prove it wrong: a shift back toward air-cooled inference fleets, which neither model assumes.
Behind-the-meter generation becomes a normal line item
Slow interconnection is pushing US developers toward onsite gas. The IEA’s satellite tracking found around one-fifth of those projects had started land clearing or construction, estimates 15 to 27 GW of onsite gas could power data centers by 2030, and notes that covering critical AI load reliably requires overbuilding onsite plant by 30% to 70%. The pipeline of conditional small modular reactor offtake deals grew from 25 GW at the end of 2024 to 45 GW by April 2026 (IEA). None of that nuclear capacity produces a kilowatt-hour in 2027; the gas can. The reactor pipeline is on our small modular reactor statistics page. What would prove it wrong: turbine delivery slots slipping past 2028, which would push onsite projects off the 2027 map.
Cheaper queries, heavier workloads
Energy per AI task has fallen by at least an order of magnitude a year recently, and the IEA calculates that answering every conventional web search with a simple AI text query would use under 4 TWh a year. Reasoning, video and agentic tasks can use hundreds to thousands of times more energy per request. Efficiency per token and total load are moving in opposite directions, and the total is winning. What would prove it wrong: hyperscalers reporting falling inference capex as a share of total spend.
Power prices become a data center story
The EIA expects PJM wholesale prices in 2026 to run 41% above 2025, the largest increase of any hub it tracks, and sees the US residential price at 18.7 cents per kWh in 2027, up from 17.3 cents in 2025. The IEA’s view is that load growth only raises prices where supply is already tight; PJM, which covers the Northern Virginia cluster, is that case. Track it on our AI data center electricity prices page.
What Does 2027 Look Like in the United States?
Three credible sources disagree on where the US even stands, before anyone forecasts anything.
| 2025 US data center electricity | Value | Method | Source |
|---|---|---|---|
| IEA | 224 TWh | Server shipments, Base Case table | IEA, April 2026 |
| Berkeley Lab (Axis calculation) | ≈234 TWh | 192 TWh (2024) × 1.22 reference-case growth | Berkeley Lab, June 2026 |
| Energy Institute / S&P Global | 312.6 TWh | Installed-base capacity view | Energy Institute, 2026 |
Berkeley Lab’s June 2026 update put US data center use at 192 TWh in 2024, 4.7% of US electricity, and projects 22% growth in 2025 and 29% in 2026. Applying those rates gives about 302 TWh for 2026 (an Axis Intelligence Research calculation). Its next published points are 464 TWh in 2028 and 649 TWh in 2030 (range 521–843 TWh); it does not publish 2027, and neither do we.
The EIA’s 2.8% growth for the whole commercial sector in 2027 looks small next to that, but the commercial category also holds offices, stores and every other non-industrial building, so the two aren’t contradictory. They differ in what they assume about the wires: Berkeley Lab models current trends and leaves grid availability out of scope, while the EIA’s cut reflects connections that are actually slowing. For load planning in a single US market, the EIA’s regional tables are the closer read. Compare state exposure on our data center electricity by state page.
What Could Change This Forecast?
Signposts to watch
| Date | Signal | Confirms | Where to watch |
|---|---|---|---|
| Late Oct 2026 | Hyperscaler Q3 results and capex guidance | Base if 2026 guidance holds; upside if 2027 is guided sharply higher | Company investor pages |
| Nov 10, 2026 | EIA Short-Term Energy Outlook | Downside if 2027 commercial growth is cut below 2.8% | EIA |
| Mid-Nov 2026 (date not yet set) | Nvidia fiscal Q3 2027 results and Q4 guide | Base if data center revenue is near $100bn and Q4 guide implies about $112bn | Nvidia investor relations |
| Nov 16–Dec 10, 2026 | Gartner infrastructure conferences (London, Tokyo, Las Vegas) | Any revision to the 702 TWh 2027 figure | Gartner newsroom |
| Late Jan–Feb 2027 | Hyperscaler Q4 results with full-year 2027 capex | Upside if 2027 spending is guided more than 50% above 2026 | Company investor pages |
| Throughout 2027 | Memory and system pricing commentary | Downside if dollars per gigawatt rise faster than 15% | Nvidia, memory makers |
Follow the deals behind these signals on our AI data center power deals tracker and the spending series on our China AI capex statistics page.
Methodology and Limitations
Included: forecasts and estimates from the IEA, Gartner, IDC, Lawrence Berkeley National Laboratory, the EIA and the Energy Institute, each taken from the producer’s own publication (retrieved October 28, 2026). Excluded: forecasts that could not be traced to the organization that produced them.
Scope: “AI-focused data centers” follows the IEA’s facility-level definition, which differs from Gartner’s server-level “AI-optimized server” category. The Base Case is built on IEA data and inherits its scope; the Energy Institute series is shown for comparison, not blended in.
Limitations. Nvidia revenue is a proxy for all AI hardware; if custom chips from Google, Amazon or Meta take share faster than in 2025, the calibrated k understates load. k is fitted on a single year. Fiscal years are treated as calendar years (11 of 12 months overlap). The 2025 AI-focused value is measured from an IEA chart (±3 TWh); a 3 TWh error moves the 2027 total by only about 2 TWh, because the non-AI share shifts the other way. The model has no regional detail and cannot say where the load lands.
Grade-back: we will compare these projections with reported 2027 figures when the IEA and Berkeley Lab publish them, and the next edition of this page will open with that comparison, right or wrong.
Frequently Asked Questions
How much electricity will AI use in 2027?
About 357 TWh for AI-focused data centers in the Axis Intelligence Research base case, inside a data center total of about 740 TWh. That’s 48% of all data center consumption, up from roughly 30% in 2025.
Will AI servers use more power than regular servers in 2027?
Yes, according to Gartner’s June 2026 forecast: 258 TWh for AI-optimized servers versus 200 TWh for conventional servers.
Is AI electricity demand slowing down?
No, but near-term forecasts are being cut. The EIA lowered 2027 US commercial-sector growth from 5.3% to 2.8% between May and October 2026, while Gartner raised its 2030 view and the IEA held its own. The constraint is memory supply and grid connections, not appetite.
Does one ChatGPT-style query use a lot of energy?
Not a simple one. The IEA calculates that if every conventional web search became a simple AI text query, the total would be under 4 TWh a year, less than 1% of data center use. Agentic and video tasks can cost hundreds to thousands of times more per request.
Will AI data centers raise my electricity bill?
It depends on your grid. The EIA expects 2026 PJM wholesale prices 41% above 2025, while most other hubs stay close to 2025 levels. Where supply is tight, new load pushes prices up; where it isn’t, the IEA finds it can spread fixed costs and lower them.
Why do data center electricity estimates for 2025 range from 447 to 788 TWh?
Method. Gartner (447 TWh) and the IEA (485 TWh) build up from equipment; the Energy Institute and S&P Global (787.8 TWh) start from installed facility capacity and assumed running rates. Always check which one a number comes from before comparing it.
About This Dataset
The file ai-energy-consumption-forecast.csv contains 257 rows: every forecast in the reconciliation tables, every driver input with its source URL and retrieval date, and the Axis Base Case for 2026–2027 in three scenarios. License: CC BY 4.0. Updated when a signpost above resolves or a new IEA, Gartner or Berkeley Lab release appears.
Cite this page:
- APA: Axis Intelligence Research, Jad, A., & Mitchell, S. (2026). AI energy consumption forecast 2027: Trends, scenarios and the Axis base case. Axis Intelligence. https://axis-intelligence.com/ai-energy-consumption-forecast/
- MLA: Axis Intelligence Research, Aidan Jad, and Sarah Mitchell. “AI Energy Consumption Forecast 2027: Trends, Scenarios and the Axis Base Case.” Axis Intelligence, 28 Oct. 2026.
- Chicago: Axis Intelligence Research, Aidan Jad, and Sarah Mitchell. 2026. “AI Energy Consumption Forecast 2027: Trends, Scenarios and the Axis Base Case.” Axis Intelligence, October 28, 2026.
Data: Axis Intelligence Research, AI Energy Consumption Forecast 2027, 2026.