How much do power, heat and noise really cost when you run AI at home?

"No monthly bill" is one of the usual selling points of local AI, and it is close to true. But a machine doing heavy work draws power, turns all of it into heat, and may need fans to get rid of that heat. Before you put one in your office or bedroom, it helps to know the actual numbers.

All of the electricity figures below use one formula you can redo with your own rate:

def monthly_cost(watts, hours_per_day, cents_per_kwh, days=30):
    kwh = watts / 1000 * hours_per_day * days
    return kwh * cents_per_kwh / 100

print(monthly_cost(200, 4, 11.142))   # about 2.67

Power by machine

Apple's figures are maximums for the whole computer, in specific configurations, so an AI job will often use less. NVIDIA's 575 W is the card alone, and it does not include the processor, memory and fans. To be cautious, the examples below assume 800 W for a complete RTX 5090 PC under load. That figure is our assumption, not a measurement.

What it costs in Québec

Hydro-Québec's residential Rate D, in effect since April 1, 2026, charges 7.065¢ per kWh for the first 40 kWh a day on average over the billing period, and 11.142¢ above that. Which rate applies to your AI machine depends on your other use. In a home with electric heating in January, the extra power probably lands in the higher tier. Other provinces charge differently, so use the rate on your own bill.

Here is the monthly cost for four hours a day of heavy use, every day:

Idle matters if the machine stays on all day as a server. A DGX Spark at 35 W runs up 25 kWh a month, or $1.78 to $2.81, before it does any work. A Mac Studio M5 Max at 7 W costs $0.36 to $0.56. Tom's Hardware notes a typical enthusiast desktop idles at 75 to 100 W, which is $3.80 to $8 a month.

Heat and noise

Every watt becomes heat in the room. Apple rates the M5 Max Mac Studio at 682 BTU per hour at maximum, and the M5 Ultra at 1,314. An 800 W PC gives off about what a small space heater does on its low setting. In a Québec winter that offsets some of your heating, while in July it works against your air conditioning.

Published noise measurements are hard to compare, because reviewers use different distances, rooms and workloads:

Those two numbers do not mean the 5090 is quieter: the distance is more than twice as far, and a card in a closed case behaves differently. Apple's Mac Studio specification page gives no noise figure. The safest test is your own ears, so listen to the exact model under a long load before you put it next to your desk.

Does running AI at home ever cost less than a hosted model?

On electricity alone, roughly break-even on small models. An RTX 5090 running gpt-oss-20b wrote 205 tokens per second in LMSYS's test, or about 738,000 tokens an hour. At an assumed 800 W and 11.142¢, that hour costs about 8.9¢, or about 12¢ Canadian per million tokens. The DGX Spark, at 49.7 tokens per second and about 160 W, comes to about 10¢ per million. At the lower Québec rate, these fall to about 8¢ and 6¢.

Meanwhile, OpenRouter listed the same model from about US$0.066 per million output tokens on November 2, 2026, the same order of cost, with input cheaper still. These are rough figures drawn from different tests.

The purchase then decides the matter. Spread a US$6,950 DGX Spark over an assumed three years and it comes to about US$193 a month, before any electricity. At US$0.066 per million tokens, that monthly sum buys nearly three billion output tokens from a hosted service, far more than one person will use.

So for most people, local AI does not pay for itself in dollars. It pays in other ways: your data stays with you, it works offline, nobody can change or retire the model on you, and you learn how these systems work. Larger hosted models cost far more than this cheap example, so the gap narrows when a local model can really replace an expensive one for your task.