Local AI vs ChatGPT and Claude: what it really costs

Two ways to pay for AI

Renting it. A subscription per person (ChatGPT, Claude, Copilot) or pay-as-you-go API charges. Nothing to buy, the newest models, and the bill grows with every person and every request.

Owning it. A machine in your office runs open models: AI models anyone can download and run, such as Qwen, gpt-oss or Mistral. You pay up front, then mostly electricity. Your documents never leave the building.

What cloud costs

Business plans are priced per person, per month. As an example, Anthropic's Claude Team plan is US$25 a person on monthly billing, or US$20 billed annually, before tax (November 2026). If you use AI through an API as well, those charges come on top and grow with use.

What owning it costs

• The machine. A Mac Studio starts at £2,499 from Apple UK; more on choosing one in what hardware you need to run AI locally.
• Setup. Ours is a fixed £6,000: hardware advice, installation, fitting it to your jobs and showing your team how to use it.
• Electricity. A machine drawing 100 watts for 10 hours a day at 25p a unit costs about £91 a year.
• Upkeep. Models and software need updating every few months, by you or by someone you pay.

A worked example

Ten people on £25-a-month subscriptions spend £9,000 over three years.

A £2,499 Mac Studio, our £6,000 setup and three years of electricity at the figures above come to about £8,800.

That's close to even, which is the honest answer for most small teams. With five people, cloud is clearly cheaper. With twenty, or with API bills on top, owning it pulls well ahead. Try your own numbers in the local vs cloud AI cost calculator.

When cloud is the better choice

• A few people, light use, and nothing sensitive going in
• You need the very best model for hard reasoning, or the newest features as soon as they launch
• Nobody can look after a machine, and you don't want to pay someone to

When owning it is worth it

• Client records, contracts or health information you can't, or would rather not, send to an AI company
• A team big enough that per-person fees add up
• Heavy, repetitive work such as summarising or sorting thousands of documents, where API charges would mount up
• You want a fixed cost you know in advance

What you give up

Open models are good and getting better quickly, but the best cloud models are still ahead, by months rather than years. For summarising, drafting and searching your own documents, open models are more than good enough. For the hardest problems, a cloud model still does better, and many businesses end up using both: their own machine for sensitive work, a cloud tool for the rest.

The bigger picture

Cost is the main reason companies are moving to open models, according to Jeffrey Morgan, chief executive of Ollama, one of the most widely used tools for running them, speaking on Y Combinator's Lightcone podcast in September 2026.

If you'd like a straight answer for your business, see Private AI or book a free call.