What to buy
One user: a MacBook Pro with 36–64 GB of memory; the 128 GB configuration cost $8,100 on 29 September 2026. A small team: a Mac Studio M4 Max with 128 GB for $11,000 on a desk, drawing up to 145 W in the base M4 Max configuration, according to Apple. Models with 8–32 billion parameters: open Russian models and open models from around the world. Software: LM Studio or Ollama. By our calculation, the full cost is about $300–$400 a month with no engineer, because an employee who can update software handles maintenance.
What it solves
Drafts of emails and business proposals from your templates, answers to routine customer questions from a knowledge base, sorting incoming requests into a spreadsheet, checking contracts against company details, and meeting minutes from transcripts. A bank IT head’s personal agent on a Mac works with email, meeting transcripts and messengers at about 70 tokens per second.
When the cloud is cheaper
If your data holds no customer personal data, code or finances, or the volume is under 60 million tokens a month, the cloud wins: a mid-range cloud model costs $92–$250 a month, a cheap one $12–$48, and a Russian cloud with a data processing agreement $47–$900. By our calculation, a Mac Studio pays off in 40 months against a $350 monthly bill and in 71 months against $250, and against a bill under $94 it never pays off. For a small business, the local option is about customer data and independence from a provider, not about savings.
How to start in a week
Collect 50 real requests. Install LM Studio on any Mac with 16 GB and an 8-billion-parameter model, and run them. If the quality suits you, buy the memory the task needs, using the calculator. If not, try a 32-billion model on a rented GPU from about $0.50 an hour. Our pilot: one agent in 14 days from $1,499, on your Mac or on rented GPUs in Russia, with a metric in the contract.