Power & Pantry

Disclosure: Power and Pantry is reader-supported. If you buy through links on this page, we may earn a commission from Amazon or another program at no extra cost to you. See the About page for how this site is funded.

How Much Power Does a Local AI Use? (2026)

The short answer

There is no watt figure for a model, because neither Ollama nor LM Studio publishes one, and power draw belongs to the hardware, not to the model file. The same model runs on a fanless laptop and on a gaming tower and draws differently on each. The number to plan with is your machine's ceiling. For a laptop that is the wattage printed on its charger: current MacBook Airs ship with 30 W or 35 W adapters and support a 70 W one. For a desktop, the honest figures are the maker's own: Apple measures its Mac mini at up to 65 W at the wall and its Mac Studio at up to 145 W, and NVIDIA rates the graphics card in a mid-range gaming PC at up to 145 W for the card alone. At laptop ceilings, an outage evening of AI is a small load: 3 hours at 70 W is 210 watt-hours, about an eighth of a refrigerator's 1,800 watt-hour day, and the usable 850 watt-hours of a 1 kilowatt-hour station runs that 70 W load for about 12 hours. The expensive version of local AI is not asking questions. It is leaving the machine awake and working around the clock, which at a 70 W ceiling is 1,680 watt-hours a day, nearly a second refrigerator. A desktop built around a flagship graphics card is a different class again: NVIDIA's largest card carries a 575 W ceiling by itself, before the rest of the machine draws anything.

The number that does not exist

Search this page's title and the results are full of precise-looking answers, watts per model, watts per answer, watts per thousand words. Trace any of them back and the trail ends at a roundup that measured one machine on one day, or at no source at all. The two toolmakers whose software the setup uses publish nothing of the kind. Ollama's documentation and FAQ deal in memory and behavior: how much RAM a model class wants, how long a model stays loaded in memory (five minutes by default), how large a context window it holds. LM Studio's system requirements list platforms, RAM, and graphics memory. Neither document contains a power figure, for a model, for the app, or for an answer. That silence is not an oversight to route around. It reflects what power draw actually is: a property of a specific machine, its chip, its cooling, and its screen, working at a specific moment. A model file is data. It no more has a wattage than a photograph has a brightness; the display decides that, and here the machine decides.

What a model does decide is time, and time is the term the model controls in the energy equation. Energy is watts multiplied by hours. A larger model on the same hardware answers more slowly, so the hardware spends longer at load producing each answer, and the watt-hours per answer rise with it. That is this site's reasoning from the published figures, labeled as reasoning: no maker publishes the speed figures that would turn it into a measurement, as the setup page already noted when it declined to print speeds. Hold both halves at once. There is no per-model watt number, and model size still matters to your battery, through the clock.

What the hardware makers do publish

Three kinds of figure exist, and this page never blurs them. An adapter rating is a ceiling on what a charger can supply, printed on the charger and in the maker's tech specs; a machine at work draws at or below it. A measured maximum is the maker putting the machine on a meter under a defined heavy test; Apple's power tables state that their figures are measured from the wall, include power supply and system losses, define the maximum as a compute-intensive test application that maximizes processor usage, and define idle as the machine sitting with only the Finder open. A card ceiling, NVIDIA's total graphics power, is the most a graphics card is rated to draw, and it covers the card only, not the computer around it. With those labels on, the published figures are these.

Machine or partPublished figureWhat kind of figure it is
MacBook Air, M4 (2025)30 W or 35 W adapter in the box, depending on configuration; a 70 W adapter is supported for fast charging. Battery: 53.8 watt-hours (13-inch), 66.5 watt-hours (15-inch)Adapter ratings and battery capacity, from Apple's tech specs
MacBook Pro, current modelsApple's environmental reports test the 14-inch with a 70 W adapter and the 16-inch with a 140 W adapter. Idle with the display on measures about 4 to 5 W; sleep measures about 0.3 WAdapter ratings, and Apple-measured idle and sleep figures
Mac mini (2024), M4Idle 4 W, maximum 65 WApple-measured at the wall
Mac mini (2024), M4 ProIdle 5 W, maximum 140 WApple-measured at the wall
Mac Studio (2025), M4 MaxIdle 6 W, maximum 145 WApple-measured at the wall
Mac Studio, M3 Ultra configurationIdle 9 W, maximum 270 WApple-measured at the wall
GeForce RTX 5060 (desktop graphics card)145 W for the card. NVIDIA's required system power for a PC built around it: 550 WCard ceiling, plus NVIDIA's sizing guidance for the whole PC
GeForce RTX 5090 (desktop graphics card)575 W for the card. NVIDIA's required system power for a PC built around it: 1,000 WCard ceiling, plus NVIDIA's sizing guidance for the whole PC

Two gaps in that table are facts in themselves. Apple publishes measured maxima for its desktops and not for its laptops: the laptop documents give adapter ratings and idle figures, and no under-load wattage for a MacBook appears in them. And no maker publishes a whole-system measured draw for a Windows PC doing AI work. NVIDIA's two numbers bracket the truth without stating it: the card's ceiling is a floor for what the system can draw, and the required system power is a sizing recommendation for the power supply, not a draw. A Windows tower's real figure lives between its card ceiling and its supply rating, and the only instrument that reads it is a plug-in energy meter between the wall and the machine, read while the model is answering. That meter reading, taken on your own machine, beats every table on this page, including this one.

The planning method: budget the ceiling, count the hours

Every power page on this site plans at the conservative end, the way the heated blanket tables plan at full draw and let anything better be margin. For local AI the conservative figure you already own is the ceiling: your laptop's charger wattage, or your desktop maker's measured maximum. Multiply it by the hours you honestly expect to use the machine, and the product is the watt-hours to reserve. Three hours of questions at a 70 W ceiling is 70 × 3 = 210 watt-hours. The same three hours at a 35 W ceiling is 105 watt-hours. Real sessions draw less than the ceiling, because a machine spends part of every minute thinking at partial load, waiting for you to read, and idling between questions at figures like the 4 to 5 W Apple measures for an awake laptop. The ceiling is not a prediction. It is the budget line that cannot be exceeded, which is the line a plan should be built on. Enter your machine at its ceiling watts in this site's generator wattage calculator, alongside the fridge and the lights, and the total it returns is the outage plan with AI honestly inside it.

Runtime on a power station

The planning rule is the one every power page here uses: a station delivers about 85 percent of its rated capacity, so a 1 kilowatt-hour station offers about 850 usable watt-hours. Runtime is usable watt-hours divided by the draw. Four draws cover the published figures above: 35 W and 70 W for laptops at their adapter ceilings, and 145 W and 270 W for desktops at Apple-measured maxima. Read the tables as continuous full-draw work from a full charge, with nothing else plugged into the station.

Planning drawWhat it stands for300 Wh station (255 usable)500 Wh station (425 usable)1 kWh station (850 usable)2 kWh station (1,700 usable)3.6 kWh station (3,060 usable)
35 WA MacBook Air at its in-box adapter ceiling7.3 hours12.1 hours24.3 hours48.6 hours87.4 hours
70 WA laptop at a 70 W adapter ceiling3.6 hours6.1 hours12.1 hours24.3 hours43.7 hours
145 WA compact desktop at its maker-measured maximum1.8 hours2.9 hours5.9 hours11.7 hours21.1 hours
270 WA large desktop at its maker-measured maximum0.9 hours1.6 hours3.1 hours6.3 hours11.3 hours

A graphics-card desktop does not get a row, because its whole-system figure is the one that does not exist. The bounding arithmetic still says something useful. A flagship card at its 575 W ceiling, drawing nothing else at all, which never happens, would use a 1 kilowatt-hour station's 850 usable watt-hours in about 1.5 hours; the processor, drives, and fans shorten that further. Even a mid-range card at its 145 W ceiling puts the card alone in the third row's territory before the rest of the machine is counted. The practical conclusion is not that desktops are banned from outage plans. It is that a laptop is the outage AI machine, and a big desktop belongs on the grid side of the plan, or on the largest station class for short, deliberate sessions. The stations themselves are compared in this site's portable power station comparison.

Where the model enters the math: time at load

The setup page sized models to RAM: an 8 GB machine plans on small models in the 1B to 3B range, a 16 GB machine on the 7B to 8B class, a 32 GB machine up to the low 30s. Those classes differ in memory, and through time they differ in energy. On one machine, a 33B model produces each answer more slowly than a 3B model does, so the machine sits at load longer per answer, and the watt-hours follow the minutes. Two documented behaviors stretch the time further. A model that does not fit in the graphics hardware's memory falls back to ordinary RAM and the processor, wholly or in part, and answers slow down; Ollama's documentation describes that fallback, and the setup page flagged it as a comfort issue. It is also an energy issue, for the same reason: more minutes at load for the same answer. And a model stays loaded in memory for five minutes after each answer by Ollama's default, which keeps the machine in its working state between back-to-back questions rather than letting it drop back toward idle. None of this comes with a number attached, and this page attaches none. The planning translation is simple: on a battery budget, the smaller model that answers your question is also the cheaper model, and choosing it is an energy decision as much as a quality one.

The daily budget: cheap by the session, real by the day

Put the session figures next to the loads this site already plans around. A refrigerator uses about 1,800 watt-hours a day, as the refrigerator runtime article works through, and the essentials load used across the power pages, fridge, lights, router, and device charging, totals 2,535 watt-hours a day. A 3-hour AI session at a 70 W ceiling costs 210 watt-hours: about 12 percent of the refrigerator's day, and about 8 percent added to the essentials load. At a 35 W ceiling the same session costs 105 watt-hours, about 6 percent of the fridge's day. A desktop session at a 145 W measured maximum costs 435 watt-hours for the same 3 hours, about a quarter of the fridge's day, and half the usable capacity of a 1 kilowatt-hour station. Those are the numbers that earn local AI its place in the load plan: small enough to include without restructuring the plan, large enough that leaving it out of the arithmetic is how plans quietly fail.

The failure mode has a specific shape, and it is not a long conversation. It is the machine that never sleeps. Left awake around the clock at its 70 W ceiling, a laptop would draw 1,680 watt-hours a day, 93 percent of a second refrigerator. Left awake but idle, at the 4 to 5 W Apple measures for a display-on laptop, the same day costs about 100 watt-hours, and asleep, at about 0.3 W, it costs about 7. The spread between those three days is the whole lesson: the load is not the questions, it is the hours at load and the hours awake. A plan that gives AI a session, two or three hours at a stated ceiling, and then lets the machine sleep has priced it. A plan that leaves a model loaded and a laptop open on the counter has priced it too, at the highest number on this page, whether the plan says so or not.

One more buffer belongs in the picture. A laptop carries its own battery: 53.8 watt-hours in the 13-inch MacBook Air, 66.5 in the 15-inch. At its 35 W ceiling that battery is about an hour and a half of full-draw work, which is a buffer for carrying the machine between rooms, not a power source for the plan. The energy still comes from the station when the laptop recharges; the battery only decides when the station pays. Plan the station for the session, and let the battery be the slack it is good at being.

On the grid, the same work costs pennies

Off the battery, the question changes from hours to money, and the money is small. At the US average residential price of 18.16 cents per kilowatt-hour and Hawaii's 47.18 cents, the highest in the country (US Energy Information Administration, August 2026), the 210 watt-hour session costs 0.21 × 18.16, about 4 cents, in the average state, and 0.21 × 47.18, about 10 cents, in Hawaii. The 435 watt-hour desktop session costs about 8 cents and about 21 cents. A month of daily 3-hour laptop sessions at the 70 W ceiling uses 6.3 kilowatt-hours, about $1.14 at the US average and about $2.97 in Hawaii. Local AI is not a grid-bill problem at household scale. It is a battery-share problem during an outage, which is why this page spends its length on watt-hours instead of dollars.

Matched offers are coming. This site is new, and the affiliate links for this corner are still being set up. When they are live, this box will hold current offers for the power stations this page's tables size, with prices re-checked at that time. The disclosure at the top of this page explains how the site earns.

What we could not verify

Questions

How many watts does an 8B model use?

There is no such figure, from the model's maker or from either toolmaker. The model is a file; the watts belong to your machine. On a 16 GB laptop, the class the setup page pairs with 8B models, plan at the machine's charger wattage, 35 W to 70 W on current Apple laptops, for the hours you use it. That ceiling budget overstates a real session, deliberately, the way every planning table on this site overstates: the surplus is margin, not error.

Will a 1 kWh power station run my laptop AI all day?

At a 70 W ceiling, the station's 850 usable watt-hours are about 12.1 hours of continuous full-draw work, and a real day of mixed use, questions between other tasks, the machine idling at a few watts in between, runs longer. Whether that is "all day" depends on what else the station owes. The same 850 watt-hours is less than half a refrigerator's daily 1,800, so a station carrying the fridge has no all-day AI budget at any ceiling. Run both loads through the generator wattage calculator before the outage, not during it.

Does a bigger model use more electricity?

Per answer, yes, by the only route available: time. A bigger model on the same hardware takes longer to produce each answer, and longer at load is more watt-hours, at whatever draw the machine runs. This page labels that as reasoning, not measurement, because no maker publishes the speeds or the draws that would settle it as a number. The effect is real enough to plan by: on a tight battery, pick the smallest model that does the job.

Is it cheaper to run the AI from the laptop's own battery?

It is the same energy, drawn at a different time. A 13-inch MacBook Air battery holds 53.8 watt-hours, about an hour and a half of full-draw work at the machine's 35 W ceiling, and recharging it afterward takes that energy, plus conversion losses, from the station or the wall. The battery is for mobility and for riding through the moments when the station is busy elsewhere. The plan still has to contain the session's watt-hours somewhere.

What about a desktop with a big graphics card?

Treat it as a space-heater-class load until a meter tells you otherwise. NVIDIA rates its mid-range RTX 5060 at a 145 W ceiling for the card alone and its flagship RTX 5090 at 575 W for the card alone, with the rest of the machine drawing on top of either. At the flagship's ceiling, a 1 kilowatt-hour station's usable capacity is gone in about 1.5 hours even if nothing else drew a watt. A desktop AI session in an outage is a short, deliberate event on a large station, or it is the first load cut from the plan.

How can I get my own machine's real number?

With a plug-in energy meter, the inexpensive kind that sits between the wall outlet and a plug and shows live watts. Put it on your charger or your desktop's power cord on an ordinary day, ask the model a long question, and read the meter while it answers. That reading is your machine, your model, and your room, and it replaces the ceiling in every table on this page with something better. Do it before an outage, the same day you do the Wi-Fi-off check from the setup page, and write the number into your plan.

A note on responsibility

The math on this page plans a load; it does not certify a machine or a setup. A computer working at load for hours makes heat, and it needs its vents clear and a room it can breathe in; Apple's own power tables note that hotter rooms push fan speeds and power draw up, so a machine in a hot, closed room costs more and ages faster than the same machine in a cool one. The standing rule from the setup page also stands here: a local model is a language tool, not an authority, and nothing it says about wiring, loads, or safety outranks your equipment's manuals. This page's data basis is stated at the top: the hardware makers' published figures, checked in November 2026, arithmetic shown where it is used, and the gaps named where a reader would otherwise assume a number exists.