📎 Run a site or newsletter? Use the Cite or Embed buttons just above to link to this tool or embed the live version on your own page, free, no signup, just keep the credit.
How the math works
Three plain cost models, computed live above:
- Buy, the hardware price upfront, plus electricity: TDP × a 0.7 typical-load factor × your active hours × your $/kWh. Hardware prices come from our tracked data layer, with used gear labeled as such.
- Rent, a cloud GPU by the hour (RunPod, Vast.ai and similar; the default rate is a typical RTX-4090-class community price, check current rates, they move weekly).
- API, pay per token. Enter a blended (input+output) price per million tokens for the model class you'd use.
The break-even months are simply: hardware price ÷ (what you'd pay monthly the other way − your electricity). Dots on the chart mark them.
What the calculator deliberately ignores
For the record: resale value (used hardware holds value well, which favors buying more than these curves show), cloud spot-price swings, API price cuts (historically frequent, long-horizon API totals are probably overestimates), and most importantly everything that isn't money: privacy, working offline, latency, no rate limits, and the simple joy of owning the machine. Plenty of people run local AI knowing it costs more than an API. That's a feature, not an error.
Use it with the fit calculator
This page answers "should I buy at all?" Its sibling, Can I run it?, answers "what exactly can the hardware run?", model sizes, quantization, context length, and which machines fit. Between the two you can go from curiosity to a defensible purchase decision without leaving the site.
Disclosure: hardware links and the cloud-GPU links on this page are affiliate/referral links (details). They never change the math.
Worked example: a used RTX 3090 against renting and an API
The calculator's defaults describe a common setup: a used RTX 3090 at $1,100 (350W), used three hours a day for 300,000 tokens a day, electricity at $0.16 per kWh, a rented 3090-class GPU at $0.59 an hour, and an API at $1 per million tokens. Here is what those inputs produce.
| Option | Monthly cost | How it is computed |
|---|---|---|
| Buy | $3.53 in electricity, after $1,100 upfront | 350W × 0.7 load × 3 hours × 30 days × $0.16 |
| Rent | $53.10 | $0.59 × 3 hours × 30 days |
| API | $9.00 | 9 million tokens × $1 per million |
| Break-even | Result | How it is computed |
|---|---|---|
| Buying beats renting after | 22 months | $1,100 ÷ ($53.10 minus $3.53) |
| Buying beats the API after | about 17 years | $1,100 ÷ ($9.00 minus $3.53) |
| Local cost once paid off | $0.39 per million tokens | $3.53 ÷ 9 million tokens |
| Electricity per year | 268 kWh, about $43 | 245W × 3 hours × 365 days |
Read that plainly. At 300,000 tokens a day, buying beats renting inside two years but never catches the API on cost alone. What flips it is volume: at 3 million tokens a day the API bill is $90 a month and the card pays for itself in about 13 months, after which every million tokens costs 39 cents of electricity. Price is only one of the three reasons people buy, though, and the calculator says so.
Is it cheaper to buy a GPU or rent one for local AI?
Buying wins on hours. The break-even is the hardware price divided by what you save each month against renting. At three hours a day on the defaults it is 22 months; at eight hours a day the rental bill rises to $141.60 a month against $9.41 in electricity, and the same card pays off in about eight months. Renting wins for bursts and for trying hardware you may never buy. The used-market side of this is tracked monthly in our used GPU price index.
How much does electricity cost to run a local LLM?
Watts, times a load factor, times hours, times your rate. The 350W card above costs $3.53 a month at three hours a day. A 128GB Mac Studio M4 Max, at 160W, costs about $1.61 a month for the same hours, which is part of why unified-memory machines suit always-on use. The calculator uses a 0.7 load factor because generation does not hold a card at its full rating; it does not model idle draw, which matters for a box that runs all day (see the section above on what it ignores).
When does an API beat both?
At low or bursty volume. Below roughly a million tokens a day, an API at a dollar per million costs less per month than the electricity-plus-amortization of owning anything, and it needs no setup. The crossover moves with the API price: at $3 per million, the 300,000-token-a-day case pays off a $1,100 card in about four years instead of seventeen. Enter your real token count and your real API rate; the defaults are only a starting point.
What the price does not capture
Three things push people to buy even when the arithmetic says rent: data that cannot leave the building, models no API serves (most new open-weight releases for their first weeks), and the freedom to run a model as long as you like with no meter running. The trade is that you carry the hardware risk, and a used card has no warranty. Which machine to buy, once you have decided to buy, is the subject of the cheapest way to run a 70B and which Mac for local LLMs.
Where the numbers come from
Hardware prices come from our tracked catalogue, with used prices dated to the month they were observed; power figures are the vendor's rated TDP. The rental and API rates are editable defaults, not claims about any provider, because both move often. The formulas are printed under the calculator on purpose, so you can check every figure, and the same math is exposed through the public API. Our methodology page lists the sources we rely on.