Short answer: It was plausible, but it was never a directly measured or universally verified fact. Alex de Vries-Gao’s 2025 analysis estimated that AI hardware could require as much as 23 gigawatts (GW) by the end of 2025. If that were an average load running continuously, it would equal about 201 terawatt-hours (TWh) per year—potentially comparable with or higher than modeled Bitcoin-mining consumption. But the estimate was a forecast, and no authoritative global dataset currently isolates AI electricity use from other data-center workloads well enough to declare an uncontested winner.
What the headline actually means
“Power” and “electricity consumption” are different measurements. Power is an instantaneous rate, expressed in watts, megawatts or gigawatts. Electricity consumption is energy used over time, expressed in kilowatt-hours or terawatt-hours.
The original claim is therefore best written as: AI could use more electricity than Bitcoin mining by the end of 2025. A 23-GW estimate cannot be compared directly with an annual Bitcoin TWh estimate without converting the units.
- 1 GW running continuously for one year is approximately 8.76 TWh.
- 23 GW × 8,760 hours equals approximately 201.5 TWh per year.
That 201.5-TWh figure is a calculation, not a reported measurement. It applies only if 23 GW represents average continuous demand. If it represents installed capacity, nameplate demand or a peak, actual annual consumption would be lower.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstall#1 Best Overall
Where the 23-GW forecast came from
The estimate came from Alex de Vries-Gao, a researcher associated with Vrije Universiteit Amsterdam’s Institute for Environmental Studies and founder of Digiconomist. His analysis used indirect indicators such as AI-chip production, shipment information, hardware power ratings, deployment and utilization assumptions rather than a meter covering every AI data center.
The published analysis projected AI-related power demand of up to 23 GW in 2025 and discussed possible AI electricity use of roughly 85–134 TWh in 2027 under its assumptions. It also suggested that AI could approach half of total data-center electricity use in a high-growth scenario. Read the underlying study at ScienceDirect; contemporary reporting is available from Wired.
Those figures describe a model-based scenario, not an audited global account. The result depends on how many accelerators are deployed, how hard they run, how long they remain in service, and whether facility overhead such as cooling is included.
Rank #2
How Bitcoin’s electricity use is estimated
Bitcoin mining has no global utility meter either. The Cambridge Bitcoin Electricity Consumption Index (CBECI) models network-wide demand and publishes a theoretical lower bound, a best estimate and an upper bound. Its annualized figures use a seven-day moving average to smooth short-term changes.
Cambridge’s methodology makes assumptions about profitable hardware, machine efficiency, electricity prices, hardware availability and facility overhead. Miners are decentralized, so the true total cannot be observed directly. The live index and its scenarios are at CBECI, with assumptions explained in the methodology.
As a result, the AI-versus-Bitcoin comparison changes depending on which Bitcoin scenario, date and boundary are selected. A current live estimate also cannot automatically prove or disprove a forecast made for the end of 2025.
AI versus Bitcoin: a like-for-like comparison
| Question | AI estimate | Bitcoin mining estimate |
|---|---|---|
| What is being measured? | Projected AI hardware or related power demand; exact boundary varies | Modeled network-wide mining electricity demand |
| Typical unit in the headline | 23 GW, an instantaneous-rate or capacity-style figure | Annualized TWh scenarios |
| How is it produced? | Chip shipments, power ratings, deployment and utilization assumptions | Hardware-efficiency and mining-economics model |
| Direct global meter available? | No | No |
| Main uncertainty | Utilization, workload allocation, hardware life and facility overhead | Hardware mix, electricity costs, miner behavior and overhead |
A fair test must align geography, period, metric, system boundary and operational assumptions. Comparing AI’s 23 GW directly with Bitcoin’s annual TWh is a category error; converting the AI figure to an annualized TWh is necessary, but still leaves the continuous-operation assumption.
Do data centers already use more electricity than Bitcoin?
Very likely—but that is a broader claim than saying AI alone exceeded Bitcoin. The International Energy Agency estimates that all data centers consumed about 415 TWh in 2024, around 1.5% of global electricity use. It projects nearly 945 TWh by 2030. These totals include AI and non-AI services such as cloud storage, web hosting, video delivery, enterprise software, social platforms and conventional computing. See the IEA’s Energy and AI executive summary.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Consequently, the IEA total cannot establish that AI itself overtook Bitcoin in 2025. It does show that the larger data-center sector is already operating at a scale where AI growth matters to power systems.
Why AI demand is rising
- Training larger models and repeating training or fine-tuning runs.
- Serving inference requests to millions of users.
- Embedding AI in search, office software, coding tools, advertising, customer service and recommendations.
- Using more compute-intensive reasoning and multimodal models.
- Deploying dense GPU and accelerator servers that require substantial cooling and networking.
- Building dedicated AI campuses rather than adding occasional workloads to ordinary facilities.
The IEA notes that AI workloads can create rapid, large power swings. Its overview says traditional data centers commonly use about 10–25 MW, while hyperscale AI facilities can exceed 100 MW. Those are facility-scale examples, not a global AI total. See the IEA’s artificial-intelligence topic page.
Reasons the 2025 forecast could be too high
- Chip shipments are not electricity meters. Hardware can be delayed, underused, reassigned or retired early.
- Nameplate ratings exceed actual draw when servers are not fully loaded.
- Training clusters may run intensely for some periods but not continuously.
- New chips can deliver more computation per watt, while software optimization can reduce energy per task.
- Demand may shift toward smaller models, efficient inference or edge devices.
- Cooling, power conversion and backup overhead differ substantially among facilities.
Reasons it could be too low
- Inference could grow faster than training as consumer and business use expands.
- Reasoning models can require many more computations per answer.
- New applications may create demand not reflected in early infrastructure plans.
- AI hardware can be operated continuously to provide commercial services.
- Grid interconnection, transformer and transmission shortages can delay actual consumption even when underlying demand is strong.
The IEA warns that grid constraints may delay a significant share of planned data-center projects unless infrastructure expands; see its executive summary.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What happened by the end of 2025?
As of August 18, 2026, the public evidence supports a cautious three-part assessment:
Best Value
- The forecast was reasonable. AI-related infrastructure and electricity demand grew rapidly enough that a Bitcoin-scale crossover was credible.
- A precise crossover is not confirmed. Operators generally report combined data-center or corporate electricity use, not a global AI-only total. AI chips may run mixed workloads, and utilization is rarely disclosed.
- The direction was supported, but the winner is not auditable. The claim may have become true under particular assumptions, yet no single authoritative dataset proves it across all regions and workloads.
The most accurate classification is therefore: plausible forecast, unverified end-of-2025 statistic. It is neither responsible to state that AI definitely surpassed Bitcoin nor to call the forecast debunked.
Why similar electricity totals do not mean identical impacts
| Issue | AI workloads | Bitcoin mining |
|---|---|---|
| Main computation | Model training and inference | Proof-of-work hashing |
| Hardware | GPUs, AI accelerators, CPUs and networking | Specialized ASIC miners |
| Output | Models, predictions and digital services | Network security and block validation |
| Flexibility | Some training can be scheduled; inference depends on service requirements | Mining can often move by location or timing as profitability changes |
| Measurement confidence | No independent global AI-electricity index | Independent modeled range, not direct metering |
Electricity is also not the same as climate impact. Emissions depend on the electricity mix, equipment manufacturing, facility construction, cooling and hardware lifetimes. A sector can use more electricity yet produce fewer emissions per unit of energy if it operates on a cleaner grid. Regional effects can be significant even when a sector’s global share appears modest.
Final verdict
“AI could consume more power than Bitcoin by the end of 2025” was a credible high-end forecast, not a settled statistic. The 23-GW estimate can imply roughly 201 TWh per year only under continuous-operation assumptions. Bitcoin’s own electricity use is modeled rather than directly measured, while AI lacks a comparable public accounting system that separates its workloads from the rest of the data-center industry. The evidence supports a plausible crossover, but not a clean, audited global declaration that AI had definitely overtaken Bitcoin in 2025.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problems
