Buying hardware to run models locally pays off only in some situations. A fair comparison is simple arithmetic, and it is worth doing before you buy. This guide gives the formula, not the prices, which change often. Current hardware prices are on the products page.
The break-even formula
months to break even = hardware cost ÷ (monthly hosted spend − monthly electricity)Fill in the numbers
- Hardware cost: the extra you would spend. If you already own a capable machine it is zero, and the answer is simply “use it”.
- Monthly hosted spend: what you would otherwise pay for the same usage, at the hosted model quality you need. This is the number people most often guess wrong; use real usage.
- Electricity: watts under load ÷ 1,000 × hours of use per month × your price per kWh. For example, a 350 W system running two hours a day at 15 cents per kWh is about 0.35 × 60 × 0.15, a little over $3 a month. Idle draw adds a little.
When the hardware pays for itself
Electricity per month$3.15
Saved per month$46.85
Break-evenabout 32 months (2.7 years)
Local: hardware plus electricityHosted
Example: $1,500 of hardware against $50 a month of hosted spend, with a 350 W system used 2 hours a day at $0.15 per kWh.
Costs that people forget
- Quality gap. A local model that is a grade lower may cost you time on hard tasks. Count that time.
- Upgrades and resale. Hardware loses value, and models keep growing.
- Setup and upkeep. Drivers, updates and troubleshooting are your job (checklist).
When local wins
- Heavy, steady use, where per-token fees add up.
- Private or regulated data that you would rather not send out (privacy).
- Work offline, or with a hard requirement for predictable cost.
- Tinkering and learning, which have value of their own.
When hosted wins
- Occasional use, where the hardware would sit idle.
- Tasks that need the largest models, which no consumer machine runs well.
The two combine well: local for routine and private work, hosted for the occasional hard problem. To see what your current machine already does, use My hardware; to compare candidates, use Compare.