Українська

AI

An AI model on your own computer: how much memory it needs, how many months the hardware takes to pay off against a ten-seat subscription, and why pay-per-request adds up differently

The question usually goes like this: ten people pay for a chat assistant every month, so what if we buy one powerful computer and run a model on it? The answer depends on what you compare it with. Against a subscription the hardware pays off in a little over a year. Against pay-per-request on a cheap model it does not. And there is a third reason to run a model in-house that has nothing to do with money.

Updated 2026-11-11 8 min Figures carry a check date
Illustration generated with AI (OpenAI gpt-image-2.5-flare) and labelled with IPTC metadata.

How much memory a model needs

An open model runs on your own computer through a program such as Ollama: it loads the model file and answers requests the same way a chat does. The main limit is memory. The model file has to fit into the graphics card's video memory or into a Mac's unified memory, with room left for context (the text of the request and the answer) and for the system itself.

The file sizes below come from the Ollama library pages on 11 November 2026, in the standard compressed form the program downloads by default. We took general-purpose models from Google, Alibaba and OpenAI in three sizes; the library has others, including models for coding or document recognition.

ModelFile16 GB graphics card32 GB graphics cardMac, 64 GBMac, 128 GB
gemma4:12b7.7–8.0 GB✓✓✓✓
gpt-oss:20b14 GB≈ tight✓✓✓
qwen3.8:27b18 GB✗✓✓✓
gemma4:31b19–20 GB✗✓✓✓
gpt-oss:120b65 GB✗✗✗✓
qwen3.5:122b81 GB✗✗✗✓

The marks are our estimate: ✓ — the file is smaller than memory with room for context and the system, ≈ — almost no room, ✗ — the file is larger than or equal to memory. Check by running it on your machine: how much memory the context takes depends on the length of your documents.

The rows differ in more than size. Models of 8–20 GB are a level that handles draft letters, summaries and classifying requests. Models of 65–81 GB are closer to what a subscription offers but need an expensive machine. We did not compare quality with subscription models, so the numbers below are about money only.

Hardware and electricity

Prices come from the Rozetka catalogue on 11 November 2026, the lowest for new devices of the required configuration. Power draw comes from manufacturers' documentation: Apple publishes measured wall power at idle and at maximum load, NVIDIA the graphics card power.

OptionMemory for the modelPricePower draw
GeForce RTX 5060 Ti 16 GB graphics card (for an existing computer)16 GBfrom UAH 42,656—
Mac mini M4 Pro, 64 GB, 1 TB64 GB unifiedUAH 139,9995 W idle, 140 W maximum
Mac Studio M4 Max, 128 GB, 1 TB128 GB unifiedUAH 234,9996 W idle, 145 W maximum*
GeForce RTX 5090 32 GB graphics card (no computer)32 GBfrom UAH 368,700575 W for the card alone

* Apple gives the power draw for the Mac Studio M4 Max with 36 GB of memory; the page has no separate measurement for the 128 GB version.

What stands out in this table is the RTX 5090 price: the graphics card alone costs 1.57 times as much as a complete 128 GB Mac Studio and gives four times less memory for a model. For the task 'run a large model', a Mac with large unified memory is the cheaper option. For the task 'run an 8–14 GB model', a 16 GB graphics card in a computer you already have is enough.

We count electricity at the fixed price for household consumers — UAH 4.32 per kWh including VAT, in force from 1 June 2024 to 30 April 2027 (Cabinet of Ministers Resolution No. 483, Annex 3). It fits a sole trader with the computer at home. For an office the price is non-household and depends on the supply contract; use the one on your bill.

Two scenarios for the Mac mini: eight hours at maximum load on 22 working days, idle the rest of the time — 27.4 kWh, or UAH 118 a month. Round the clock at maximum load — 100.8 kWh, UAH 435. For the Mac Studio, UAH 124 and 451 respectively. Even at UAH 10 per kWh the first scenario for the Mac mini would cost UAH 274. Electricity decides nothing here.

Payback against a team subscription

Business plans of chat assistants on 11 November 2026: a standard seat on ChatGPT Business and Claude Team costs $20 a month billed annually and $25 billed monthly. For ten people that is $200 (UAH 9,100) or $250 (UAH 11,375) at the National Bank rate of UAH 45.5013 per dollar.

Payback = hardware price ÷ (monthly subscription − monthly electricity), using the eight-hours-a-day scenario.

Computer payback against a subscription and pay-per-request, months
Mac mini vs 10 seats, billed monthly12.4
Mac mini vs 10 seats, billed annually15.6
Mac Studio vs 10 seats, billed annually26.2
Mac mini vs 5 seats, billed annually31.6
Mac mini vs pay-per-request, gpt-6.1-sol40.2

Calculation: Rozetka price ÷ (seats × $20 or $25 × 45.5013 − electricity of UAH 118 for the Mac mini, UAH 124 for the Mac Studio). The last row is against 8,800 requests a month through the API, calculated in the next section. Prices on 11.11.2026.

Against ten subscription seats the Mac mini pays off in a little over a year. The Mac Studio takes over two years, and with five people four and a half (53.1 months). But this comparison assumes the local model fully replaces the subscription. More on that below.

Payback against pay-per-request

A subscription is not the only way to pay. Through an API you pay for tokens, and on cheap models that is very little. Take the same team: ten people, 40 requests a day each, 22 working days — 8,800 requests a month, 2,000 input tokens and 500 output each. Prices from the OpenAI API pricing page on 11 November 2026.

Model through the APIPrice per million tokens (input / output)8,800 requests a monthMac mini payback
gpt-6-luna$0.10 / $0.50$3.96 (UAH 180)over 2,000 months, i.e. never
gpt-6.1-terra$1 / $5$39.60 (UAH 1,802)83 months
gpt-6.1-sol$2 / $10$79.20 (UAH 3,604)40 months

This is where the calculation flips. If the team's tasks are drafts, summaries and classification, the things an 8–20 GB model handles, there are cheap models for such tasks through the API too. A UAH 139,999 computer against UAH 180 a month will not pay off within its service life. How many requests per person per day make the API cheaper than a subscription is calculated in the article on what AI costs for a team of ten.

What the calculation leaves out

  • Quality. A subscription gives access to the provider's strongest models. A 20 GB local model is a different class. Comparing prices is fair only for tasks the weaker model handles; check that on your own examples.
  • Concurrent requests. One computer handles ten people's requests in turn or shares memory and speed between them. We did not measure response speed; test it with several people at once.
  • Maintenance time. Someone has to install the program, update models, set up access from other computers and keep the machine running. In the calculation this time is zero, in reality it is not.
  • Interface. A subscription is a ready chat with projects, files and sharing. Locally, an interface has to be installed separately.
  • Residual value. After two years the computer can be sold or kept in use. We did not count it, and that favours the hardware.

When the reason is data, not price

There are tasks where price is not the point at all: the data cannot be handed to a third-party service. Patients' medical records, materials of a legal case, primary documents with personal data, internal financial reports the company is not ready to take outside its perimeter. According to OpenAI, the ChatGPT business plan does not train models on company data by default, but the data are still processed on the provider's servers.

A local model solves exactly this: the request does not leave the computer. For such tasks the payback calculation above becomes secondary, and the question changes: is the quality of a model that fits the available hardware good enough. Which data can go to a cloud assistant and which cannot is in the article on data for an AI assistant; industry examples are in the articles on AI for a clinic and for lawyers and accountants. Check each model's licence for commercial use on its page.

Buy the hardware or not

Should you buy a computer for your own AI model
  1. Can the data not go to a third-party service?Medical data, legal privilege, internal reports. Yes → a local model; then choose hardware by model size
  2. Is an 8–20 GB model enough for the tasks?Tested on your examples, not judged by a description. No → stay on the subscription; a large model needs a machine from UAH 234,999
  3. Do you already pay for a subscription for five or more people?And each uses it daily. No → pay-per-request through the API is cheaper than any hardware
  4. Is there someone to set up and maintain the machine?Installation, updates, network access. No → do not buy until there is
  5. Is the payback shorter than your planning horizon?Mac mini vs ten seats — 12–16 months. Yes → you can count the hardware as a saving

A first step for a week

  1. Install Ollama on a computer you already have with 16 GB of memory or more, and download an 8 GB model, for example gemma4:12b.
  2. Collect 20 real team tasks from last week: letters, meeting summaries, classifying requests. Remove personal data from them if you also test in the cloud.
  3. Give the same tasks to the local model and to the subscription model with the same instructions. Count how many answers could be sent without edits.
  4. If the local model gives at least two thirds of usable answers, price hardware for a bigger model. If not, the question is closed at no cost.

Our position

Your own model on a computer is rarely a way to save money. Against a ten-seat subscription it pays off in a little over a year, but only if the weaker model really replaces the subscription and someone looks after the machine. Against cheap models through the API the hardware never pays off. It is worth buying when data cannot go outside: then it is a payment for control, not a saving.

The position does not hold if the company already has a powerful machine sitting idle (then the hardware cost is close to zero), or if the request volume is tens of times larger than in our example, for instance automated processing of thousands of documents a day: there pay-per-token grows with volume while the hardware does not, and you need to calculate on your own figures.

Checked on 11 November 2026: model sizes on the ollama.com library pages; prices in the Rozetka catalogue; power draw on Apple support pages (Mac mini, Mac Studio) and the GeForce RTX 5090 specifications page on nvidia.com; plans on chatgpt.com/pricing, claude.com/pricing and the OpenAI API pricing page; electricity price in Cabinet of Ministers Resolution No. 483 on zakon.rada.gov.ua; exchange rate from the National Bank of Ukraine. The arithmetic was recalculated in code. Request and token counts are illustrative; model quality and speed were not measured.

Still not sure this applies to you?

Describe the situation. The first reply says whether this is our work and how long the diagnostic takes.

Request a diagnostic