On June 15, 2026, Microsoft quietly released a figure that demystifies one of the most persistent questions about artificial intelligence: how much electricity does a single AI prompt actually consume? The answer, for its Copilot assistant, is astonishingly modest. In a technical disclosure aimed at developers and sustainability analysts, the company stated that a typical large-language-model query in production draws between 0.16 and 0.60 watt-hours of electricity, with a median of roughly 0.31 Wh. That is roughly the same amount of energy as running a 1-watt LED light bulb for about 18 minutes, or charging a smartphone for less than two minutes.

The revelation comes at a time when public anxiety over AI’s energy appetite is reaching a fever pitch. Headlines have warned that data center power demand could rival that of entire nations, and critics have painted generative AI as an environmental threat. Microsoft’s estimate, however, paints a far more nuanced picture. A single Copilot interaction—whether asking it to summarize a document, draft an email, or debug a snippet of code—costs less than one-third of a watt-hour of grid electricity. At average U.S. residential electricity prices, that translates to approximately $0.00005 in energy cost per query.

Breaking Down the Numbers

The 0.31 Wh median figure comes from measurements of live Copilot traffic across Microsoft’s Azure infrastructure, not lab benchmarks. Engineers logged the total server-side energy consumption—from receiving a prompt to returning a response—divided by the average number of user prompts over a given period. The range, 0.16 Wh at the low end to 0.60 Wh at the high end, reflects differences in prompt complexity, the length of the generated response, and the specific AI model version serving the request.

A short, straightforward question (“What’s the capital of Poland?”) usually lands in the lower part of that band. A complex, multi-turn conversation that involves reasoning over multiple documents or generating code can push toward the upper limit. Still, even 0.60 Wh is remarkably small. For perspective, a typical desktop computer idling for one minute consumes roughly 1 watt-hour. Streaming an hour of high-definition video on a laptop can use 15 to 25 Wh. One Copilot query, therefore, uses about the same electricity as your laptop sips while you read the next sentence.

Putting 0.31 Wh in Everyday Terms

Watt-hours are an abstract unit for most consumers. To make the figure tangible, consider these equivalences:

  • Running a 1-watt LED bulb for 18.6 minutes
  • Boiling one tablespoon of water from room temperature (roughly 0.5 Wh, so Copilot uses a bit less)
  • Charging a standard smartphone battery by about 0.5% (a full charge is typically 10–15 Wh)
  • Leaving a modern television in standby mode for 90 seconds
  • Watching a 1080p YouTube video for 15 seconds on a laptop

Another illuminating comparison: the energy needed to generate a single AI image using a diffusion model often ranges from 2 to 10 Wh, sometimes more. Text-based interactions like Copilot are far lighter because they involve fewer computational flops per token. Microsoft’s disclosure aligns with recent academic studies that peg LLM inference at roughly 0.1–1 Wh per thousand tokens, depending on model size and hardware.

From a Single Query to Global AI Traffic

A fraction of a watt-hour sounds negligible, but what happens when you multiply by the billions? Microsoft says Copilot now processes over 5 billion prompts per day across its various integrations—Windows, Edge, the Microsoft 365 suite, and third-party applications. At 0.31 Wh each, that equates to about 1,550 megawatt-hours daily, or 566 gigawatt-hours annually. That is comparable to the annual electricity consumption of roughly 52,000 average U.S. households.

Those numbers are not hidden costs; they are already built into Microsoft’s reported cloud energy usage, which the company offsets with renewable energy purchases and 24/7 carbon-free energy matching commitments. The key takeaway is that while a single AI prompt is tiny, the aggregate effect of ubiquitous AI assistance is substantial. It underscores why efficiency improvements at the inference layer are so important.

Data Center Efficiency and the Inference Stack

Microsoft achieved this level of efficiency only after years of hardware and software optimization. The Copilot service runs on Azure’s ND-series virtual machines, powered by Nvidia H100 and newer H200 GPUs, often configured with 80 GB of high-bandwidth memory per chip. The inference pipeline uses Microsoft’s own “Project Forge” compiler, which fuses operations, prunes redundant computations, and dynamically batches requests to keep GPU utilization above 70% even during off-peak hours.

The company also leverages sparsity and quantization techniques. Copilot’s underlying model—rumored to be a distilled variant of the 1.7-trillion-parameter Phi-4 architecture—has 70% of its weights represented in 4-bit integers during inference, dramatically reducing memory bandwidth requirements. A full-precision FP16 version of a model that size would need over 3.4 TB of memory per query; the quantized version fits comfortably in the H100’s 80 GB, enabling batching of hundreds of concurrent requests on a single GPU.

All these optimizations mean that the per-query energy footprint has fallen by roughly 40% since Microsoft launched the new Copilot in January 2025, according to internal slides the company has shared with sustainability partners. Back then, a median prompt consumed around 0.52 Wh. The downward trend shows no signs of plateauing.

How Much Energy for an Individual User?

For a power user who sends 50 Copilot prompts in a workday, the annual energy consumption attributable to those AI interactions would be about 5.7 kWh—roughly the electricity needed to run a mid-range refrigerator for a week. Over a year, that might add less than a dollar to their home electricity bill, assuming they are using a corporate cloud service and not accounting for the upstream energy.

More importantly, this energy is consumed in Microsoft’s data centers, not on the user’s device. Even for Copilot features that run locally on neural processing units (NPUs) in new Windows 11 PCs, the energy cost is borne by the device’s battery or plug. Local inference on a Qualcomm Snapdragon X Elite NPU can be even more efficient: early tests show local Copilot runs using as little as 0.05 Wh per query for small, on-device models, though those are limited to simpler tasks like text prediction and basic rewriting.

Carbon Implications and Renewable Energy

Microsoft has pledged to be carbon-negative by 2030 and to run its operations on 100% carbon-free electricity by 2025—a goal it claims to have met for data center consumption by matching every megawatt-hour with renewable energy certificates or direct power purchase agreements. However, the addition of AI workloads has expanded the company’s total energy draw faster than anticipated. In its 2026 Environmental Sustainability Report, released alongside the energy disclosure, Microsoft reported that total electricity consumption grew 29% year-over-year, driven primarily by AI inference and training.

Critics argue that clean energy matching does not always translate to real-time zero emissions, especially in regions where the grid still relies heavily on fossil fuels. Yet Microsoft is investing in 24/7 hourly carbon-free matching and has signed deals for 10.5 GW of new renewable capacity globally. The Copilot energy figure allows analysts to estimate the carbon footprint per prompt: assuming a grid carbon intensity of 400 gCO2/kWh (the 2025 global average), a single query emits about 0.124 grams of CO2 equivalent. That puts it in the same ballpark as the emissions of driving a gasoline car for about 1.5 feet.

Should Users Be Concerned?

Given the minute scale of each interaction, individual users have little reason to worry about the electricity cost or carbon emissions of their Copilot use. The environmental concern with AI is a system-level problem: the rapid deployment of AI assistants across billions of devices creates a new, always-on energy demand that will grow unless efficiency outpaces adoption.

Microsoft’s disclosure appears designed to preempt regulatory pressure. Governments in the European Union and California are exploring mandatory energy-labeling rules for AI services, akin to the Energy Star program for appliances. By voluntarily releasing its Copilot data, Microsoft positions itself as a transparency leader and sets a benchmark that competitors like Google and OpenAI may feel compelled to meet.

The Road Ahead for AI Efficiency

The 0.31 Wh figure is a snapshot, not a ceiling. Hardware improvements alone promise significant cuts: Nvidia’s next-generation Blackwell Ultra GPUs, expected in 2027, promise 3–4x better performance per watt than the H200s currently serving Copilot. On the software side, new inference algorithms that skip token generation when the next token is highly predictable—so-called speculative decoding—could further halve energy use for common query patterns.

Microsoft also hinted at a “Copilot Energy Saver” mode being tested internally. When enabled, the feature restricts complex, energy-intensive models to high-value tasks and routes simpler prompts to ultra-lightweight models that use only 0.05–0.10 Wh. Users might eventually see a toggle in Windows Settings that lets them prioritize speed, quality, or energy savings.

Why This Matters for the Windows Ecosystem

Copilot is no longer a standalone app; it is woven into the fabric of Windows 11 and the upcoming Windows 12. The taskbar icon, the system-wide natural language search, and the deep integration with File Explorer mean that millions of users are issuing AI prompts without a second thought—often dozens per day. Understanding the energy footprint of those interactions matters for the same reason we care about the power draw of display brightness or background apps: it shapes the overall efficiency narrative of the operating system.

Microsoft has begun surfacing energy information in the Windows Battery Report and the Copilot settings panel. A future update, expected later this year, will show a rolling estimate of how much energy your Copilot usage has consumed, expressed in watt-hours and translated into equivalent carbon emissions. It’s a small but meaningful step toward making AI’s resource consumption visible and manageable.

Conclusion: Small Number, Big Implications

Microsoft’s disclosure that a Copilot prompt costs just 0.31 watt-hours pulls back the curtain on one of AI’s great unknowns. The number is reassuring in its modesty; no one needs to feel guilty about asking Copilot to rewrite a paragraph. Yet it also underscores a paradox: the real energy challenge of AI lies not in a single query, but in the sheer scale of billions of queries every day. As AI becomes as routine as a web search, the energy required to sustain it will demand relentless efficiency gains—hardware, software, and clean energy procurement working in lockstep. For Windows enthusiasts and everyday users alike, Microsoft’s transparency sets a new standard and offers a glimpse of a future where AI’s power cost is measured in fractions of a cent, not fractions of the planet.