Where should your AI run?
Where your AI runs is a decision, and most businesses make it by accident. They sign up for whatever tool came first, and from that moment every document they feed it travels to a vendor's servers and gets billed on the way. That can be the right call. It should be a call.
There are two paths. Rent the model through the vendor's pipe (the API, covered in Six words that explain all of AI) and let their computers do the work. Or run the model on hardware you own, in your office, so nothing leaves the building. This post lays out what each one costs, what each one trades away, and the five questions that decide which fits your business.
Why this became a question again
Twenty years ago everything moved to the cloud, and for ordinary software that was the right call. Software stores your data. AI reads it. Every file you hand a model gets processed, and depending on the account, it can be retained and used for training. The economics changed too: the compute that AI needs is expensive enough that the vendor's meter runs the whole time you use it. Those two facts, data exposure and recurring cost, are why the location of the machine matters in a way it stopped mattering for email a decade ago.
Path one: rent through the API
You pay a vendor to run the model on their hardware. Your request goes out, their answer comes back, and you are billed by the amount that travels through the pipe or by a flat monthly subscription.
What you get: a low upfront cost, a fast start, the newest models the day they ship, and nobody on your payroll maintaining a machine. What you give up: your data leaves the building, the cost recurs every month and scales with use, and you operate under the vendor's terms, which can change. The consumer-versus-business account distinction I covered in The three-tier data rule lives entirely on this path.
Path two: own the hardware
You buy a desktop-class machine with enough memory to hold a capable model, install an open-weight model on it, and run everything in the office. Your data never leaves. The monthly bill drops to electricity. You control the configuration and the upgrade schedule.
What you give up: money up front, a slower start, and the fact that someone has to maintain it. You also run models that trail the best cloud models. How far behind matters, so here is the measured answer.
Epoch AI, which tracks model capability across many benchmarks, found that since January 2026 the most capable open-weight models have lagged the frontier closed models by an average of about four months, a gap slightly wider than the three months it measured between 2023 and October 2025 (Epoch AI, May 29, 2026). Models small enough to run on a single consumer graphics card catch the frontier within roughly six to twelve months, with the caveat that benchmark scores can flatter smaller open models, so the lag in day-to-day usefulness is probably somewhat longer (Epoch AI).
Four months behind the frontier is nothing for sorting an inbox, drafting a first pass at a quote, summarizing a contract, or reading a stack of intake forms. It matters for the hardest reasoning and the longest multi-step work. Most of what a small business repeats every week is in the first category.
What each path costs, October 2026
Prices in this section are current as of this writing and will move. Treat them as the shape of the decision, not the quote.
Subscriptions for the chat tools. The major vendors' published consumer and business plans run from roughly $20 to $250 per person per month, with most teams landing near the low end. Twenty people on a business plan is a real line item by the end of the year.
Usage through the API. Billed by volume, often cents per task. This is how agents are billed, because an agent makes many calls to finish one job. The bill is small at first and grows with how much you automate, which is exactly the case where owning starts to make sense.
Your own hardware. A desktop machine with enough unified memory to run capable local models currently runs from roughly $2,500 to $10,000, depending on how much memory you buy; Apple's Mac Studio line, the common choice for this work, starts around $2,500 and climbs past $5,000 for the configurations that hold larger models. Add electricity and someone's time.
The break-even is arithmetic. If your AI bill is $500 a month, a $5,000 machine pays for itself in ten months. At $1,000 a month it pays back in five. Below a couple of hundred dollars a month, the machine never catches up and the cloud wins on cost alone.
So the honest rule: do not buy hardware to save money this year. Buy it for control of your data once you are already spending real money and a local model can do the work. If you want to see what a purpose-built local machine looks like, and what we put on it, that is what Maai Machines is for.
The five questions
These decide it. Answer them for your business, not for businesses in general.
- Data sensitivity. If you handle patient records, client financials, Social Security numbers, or anything a specific law governs, that work pushes toward your own hardware. Everything else can rent.
- Upfront capital. A few thousand dollars available now favors owning. Cash-tight favors renting, and there is no shame in that answer.
- Ongoing spend. If your AI bill is already a few hundred dollars a month and climbing, owning starts to pay back. If it is a single $20 subscription, it does not.
- Integration complexity. How many of your systems does the AI need to touch, and how messy are they? Cloud tools come with connectors built. A local setup means someone builds or configures them.
- Who maintains it. The cloud answer is nobody on your payroll. The hardware answer is that someone has to, whether that is you, a vendor, or a technical employee.
For most small businesses today, the honest answer is rent, on a business-tier account, for everything except the most sensitive work. For some, the accountants, the medical practices, the firms already spending real money, the honest answer is a machine of their own for that work and the cloud for the rest. Both are fine. The point is to choose and to know why.
A note on the hybrid
The two paths are not exclusive, and the setups I build are usually both. Confidential and regulated work runs on a local machine. Public-facing drafting and research runs through a business-tier cloud account, where the newest model is worth the pipe. The three data tiers decide which job goes where. That split gets you most of the data control of owning and most of the capability of renting, and it is the configuration I would start from if I were setting up a ten-person professional firm tomorrow.
If you want to work out which path fits, book a free fit call. Bring your current AI spend and the kind of data you handle; you leave with a recommendation and the reasoning behind it. The decision matrix above is in the newsletter welcome kit as a one-page card.
Sources: Epoch AI, Open models lag state-of-the-art closed models by 4 months (May 29, 2026); Epoch AI, What's the best model you can run on a single consumer GPU?. Hardware and subscription price ranges reflect published vendor pricing as of October 7, 2026.