EN ▾
Get API key

uncensoredlocalllm.comGuide

Local LLM Electricity Cost vs Paying per Token

Running a model on your own GPU is not free: the card draws real power while it generates and the PC keeps drawing some when it idles. Here is how to estimate it and when a per-token API costs less.

Updated

Isometric GPU connected to a power meter and coins, illustrating electricity cost of local LLMs
Isometric illustration of a local GPU connected to a hosted cloud, representing local versus API coding models

Key points

  • A single-user setup at about 30 tokens per second spends roughly $0.63 of electricity per 1M output tokens, before hardware.
  • At light use the API is cheaper. At heavy daily use a local rig can beat it on power alone.
  • Idle draw matters more than most people expect when the PC stays on all day.
  • Hardware purchase, not electricity, usually decides the real break-even.

The short answer

The local llm electricity cost depends on three numbers: how many watts your machine pulls while generating, how many hours a day it generates, and what your utility charges per kilowatt-hour. For a typical single-GPU desktop the monthly figure lands between a few dollars and a few tens of dollars. That is small next to the price of the hardware, but it is not zero, and it is large enough to matter when you compare it with paying per token.

The calculation below uses assumptions you can swap for your own. The API side uses the real prices of this service: $0.15 per 1M input tokens and $0.50 per 1M output tokens.

The formula

Electricity is energy times price. Energy in kilowatt-hours is power in kilowatts times hours.

monthly_kwh  = (load_kw * load_hours_per_day + idle_kw * idle_hours_per_day) * 30
monthly_cost = monthly_kwh * price_per_kwh

Use a plug-in power meter if you have one. It measures the whole machine at the wall, which is what you pay for. Reading the GPU's rated board power from its spec sheet gives only part of the picture, because the CPU, memory, drives, fans and power supply losses add to it.

For the examples I assume:

  • 450 W at the wall while the model is generating (a card rated near 350 W plus the rest of the system).
  • 60 W at the wall when the PC is on but idle.
  • $0.15 per kWh. This is an assumption. Rates vary a lot by country and by time of day, so replace it with the number on your bill.
  • 30 tokens per second of generation for a single user. Your speed depends on the model, its quantization and the card.

Three usage levels, worked out

Take a light, a moderate and a heavy user. The machine stays on all day in each case, so idle hours are counted.

UsageGenerating per dayEnergy per monthElectricity per monthOutput tokens per month
Light1 hour54.9 kWh$8.243.24M
Moderate4 hours90.0 kWh$13.5012.96M
Heavy12 hours183.6 kWh$27.5438.88M

Here is the moderate row step by step. Generating: 4 hours times 0.45 kW is 1.8 kWh a day. Idle: 20 hours times 0.06 kW is 1.2 kWh a day. Total 3.0 kWh a day, so 90 kWh over 30 days, which at $0.15 is $13.50. Tokens: 4 hours is 14,400 seconds, times 30 tokens per second is 432,000 tokens a day, or 12.96M a month.

Note how the light user pays $8.24 for only 3.24M tokens. Idle power is most of that bill: 23 idle hours at 60 W is 1.38 kWh a day, more than double the 0.45 kWh spent generating.

What the same tokens cost through an API

Chat use sends more tokens in than it receives, because each request carries the system prompt and the conversation so far. Assume three input tokens for every output token, which is a plausible ratio for chat with history. Then each month's bill at $0.15 and $0.50 per 1M is:

UsageOutput costInput cost (3x output tokens)API totalLocal electricity
Light3.24M x $0.50 = $1.629.72M x $0.15 = $1.46$3.08$8.24
Moderate12.96M x $0.50 = $6.4838.88M x $0.15 = $5.83$12.31$13.50
Heavy38.88M x $0.50 = $19.44116.64M x $0.15 = $17.50$36.94$27.54

Two things stand out. First, below about four hours of generation a day, the API is cheaper than the electricity alone. Second, at twelve hours a day the local rig wins on power, because the idle share shrinks and you are using the silicon you already paid for. The API figure also includes input processing, which your GPU does for free in cash terms but still spends energy on. Prefill is fast, so I left it out of the local side; that slightly favors local.

Electricity per million tokens

A useful way to compare is energy cost per 1M output tokens during generation only. At 450 W and 30 tokens per second, one million tokens take about 33,333 seconds, or 9.26 hours. That is 4.17 kWh, and at $0.15 per kWh it is about $0.63. The API charges $0.50 per 1M output tokens, with input billed separately at $0.15.

So even before you count the GPU, a single user generating at 30 tokens per second spends more on power per output token than the API price. The gap closes only if your hardware is faster or your power is cheaper. At 60 tokens per second the same math gives about $0.31. At $0.08 per kWh and 30 tokens per second it gives about $0.33. Either change moves the answer, so run your own numbers.

Batching is the exception. A server that handles several requests at once spends much less energy per token, which is why hosted providers can price low. A desktop used by one person rarely benefits from it.

What the formula leaves out

  • Hardware. A GPU that runs a useful model costs far more than a year of the electricity above. Spread over its life, it usually dominates the total. The page on how much an LLM costs goes through the purchase side.
  • Cooling and noise. In a warm climate the heat from 450 W also loads your air conditioning. In winter it is partly useful heating.
  • Bigger models. A model that needs two cards roughly doubles the load figure.
  • Your time. Updating drivers, picking quantizations and fixing out-of-memory errors is work an API removes.

On the other side, a local model costs nothing per extra request and works offline. If you run it for many hours daily, that matters.

Ways to lower the local bill

If you stay local, a few habits cut the figure without touching the model.

  • Sleep the machine. Going from 60 W idle all day to sleep for 16 of the 20 idle hours saves about 29 kWh a month in the moderate case, roughly $4.32 at $0.15.
  • Cap the GPU power limit. Many cards lose little generation speed when the power limit is lowered by 20 to 30 percent. Test your own card, because results vary by model and driver.
  • Use a smaller quantization or model for routine tasks. Fewer bytes per weight means less memory traffic and often faster generation, which shortens the hours the card is busy.
  • Run on a time-of-use tariff. If your utility charges less at night, schedule batch jobs for those hours.

Check the sleep saving: 16 hours times 0.06 kW is 0.96 kWh a day, times 30 is 28.8 kWh, times $0.15 is $4.32. None of this changes the order of magnitude, though. Electricity stays a small part of the total next to the hardware itself.

Trying the API side with real numbers

You can check the API estimate against your own workload cheaply. New users can claim $0.50 of trial credit with one click on the API key page after sign-up, no card needed, valid for 7 days. Send a few realistic requests, then read the X-Balance-USD response header to see what they cost.

curl -i https://api.uncensoredlocalllm.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"uncensored","messages":[{"role":"user","content":"Summarize the pros and cons of running an LLM at home."}]}'

Compare the balance before and after, scale by your expected request count, and set it against your electricity estimate. Current rates and top-up bonuses are listed on the pricing page, and request details are in the docs.

Questions and answers

How much electricity does a local LLM use?

A single-GPU desktop typically draws a few hundred watts while generating and well under 100 W idle. Measure at the wall with a plug-in meter, then multiply by hours and your price per kWh.

Does the model sit idle most of the time?

For most personal setups, yes. Idle draw can be larger than generation energy, so putting the PC to sleep between sessions cuts the bill.

Is a local LLM cheaper than an API?

At light or moderate use, usually not once power is counted, and not at all once hardware is included. At many hours of daily generation it can come out ahead on power alone.

Can I test the API cost without paying?

Yes. The one-click trial credit of $0.50 needs no card, and the balance header shows what each request cost.

uncensoredlocalllm.com

Your key is one form away

Create an account, copy the key, change the base URL. That is the whole setup.