Open vs closed AI models: what a small business actually needs to know

The short answer: open AI models are ones you can download and run on your own computers. Closed models are ones you rent over the internet, the way most businesses use ChatGPT or Claude today. For nearly every small business, the honest position is that you'll rent the thinking either way, and the part worth worrying about is what gets built on top of the model, because that's the part you own. The rest of this page explains both sides properly, including the fight that broke out between the biggest companies in tech in July 2026.

Start with the van question

Own a van and you can respray it, fit it out with racking, drive it wherever you like. You also pay when the clutch goes, and it sits on your insurance. Hire one and it just works. Somebody else services it, and you pay every month you keep it.

That's open versus closed AI, near enough.

An open-weight model is the van you own. You download the model itself, run it on your own machines, change how it behaves, and keep your data inside your own building. The upkeep is yours too: the hardware, the updates, the person who gets the call when it stops working at four o'clock on a Friday.

A closed model is the hire van. You use somebody else's AI over the internet and pay as you go. Less control, far less hassle. If it breaks, that's their problem, and the newest, most capable engines tend to arrive on the hire fleet first.

Neither is better. They're different deals, the same way owning and hiring are different deals. What matters is knowing which deal you're actually signing up for, and most businesses have never been shown the difference in plain English.

"Open weights" is not quite "open source"

One wrinkle worth thirty seconds of your time, because it changes what you're allowed to do.

Most AI that gets called open source is really open weights. You get the finished thing, the dish on the plate, free to take away. What you don't get is the recipe: the training data, the method, the kitchen it was cooked in. True open source would hand you all of it.

For almost every business this distinction is academic, because you were never going to rebuild the model from scratch anyway. But the paperwork that comes with an open-weight model still decides what you may do with it commercially. Some licences are generous. Some have conditions buried in them. If a supplier ever proposes building on an open model, the licence is a question for them to answer before anyone spends a pound.

What happened in July 2026, and why the giants suddenly care

This argument stopped being a technical one on 24 July 2026.

That Friday, Jensen Huang, the chief executive of NVIDIA, posted on X for the first time in his life. He used it to share an open letter called "Open Weights and American AI Leadership", signed by 25 companies including Microsoft, Meta, Dell and IBM, urging Washington not to put early restrictions on models anyone can download and run. Within a day the list had roughly doubled, and OpenAI and Google had added their names. NVIDIA followed up by launching a group called the Open Secure AI Alliance to carry the argument.

Anthropic, the company behind Claude, didn't sign. Three days later it published its own position: it says it has never asked for a ban on open-weight models, and what it wants is safety testing on the most powerful models, open and closed alike, along with tighter controls on advanced chips reaching authoritarian states.

So within one week, nearly every big name in AI declared where it stands on the van question.

Why now? Because the answer decides where the money settles. NVIDIA sells the machinery open models run on, so a world full of open models is a world that buys more of its hardware. The big labs sell access to closed models, so their margins live on the other side of the fence. The Moonshots podcast panel, Peter Diamandis's show, made the sharpest observation of the week on this: the safety arguments and the commercial interests overlap so heavily on every side that separating them is close to impossible.

None of that makes either camp wrong. It does mean every headline you read about open versus closed AI was written about companies arguing their own book. Keep that in mind and the coverage gets a lot easier to read.

The three ways a business can "own" AI

When an owner asks me "can we have our own AI?", the question splits three ways, and only one of them is usually worth their money.

1. Build a model from scratch

A research team, a building full of computers, and a bill with more zeros on it than any small business will ever see. Nobody reading this needs it, and nobody selling it to you at small-business prices is telling the truth.

2. Run an open-weight model yourself

This one is real, and July's news is partly why it keeps coming up. You take a model somebody else released, run it on your own machines, maybe train it further on your own material. Total control of where your data lives.

Understand what you're buying, though. You've just become a small IT department. Hardware to buy or rent, updates to apply, security to watch, and somebody on call when it goes quiet mid-quote. Owners who ask for this are usually picturing an asset. What turns up is a second business to run. It earns its keep in a few specific situations, which we'll get to, but it's rarely the right first move.

3. Rent the thinking, own everything else

This is the third way, and it's what most owners actually mean by "our own AI" even when they can't name it yet.

The model, the thinking part, is rented from one of the big providers. Everything that makes it yours is not rented. Your AI knows your prices, your suppliers, the way you word a quote, which customers get chased and which get left alone, and what your team is allowed to send out without you seeing it first. None of that comes out of any model, open or closed. It gets built, it belongs to you, and nobody else has a copy.

Here's the part that July's fight actually proves. The model market can lurch in a single week, and a business set up this way swaps the engine underneath in an afternoon. The work that took months, the layer that knows your business, carries straight over. That layer is the asset. The engine was always replaceable.

The four questions that actually decide it

Almost nobody asks the open-or-closed question because they care about model weights. They ask it because of four worries, and every one of them deserves a straight answer before you sign anything with anyone.

• Is our data training somebody else's AI? Get the answer in writing, in the contract, not from a salesperson. The main providers offer business tiers with commitments on this. If your supplier can't point to the clause, walk.
• What happens if we fall out with you? If the answer is anything other than "you keep the lot and it carries on working", walk away. This applies to AI suppliers exactly as it applies to accountants.
• Are we tied to one provider forever? Ask what it would cost to move, and watch how long the pause is. A well-built system treats the model as a replaceable part.
• What is this going to cost us? Ask whose account the bill lands on and whether you can read it line by line. You should be able to see every pound.

Notice that not one of those four is answered by choosing open over closed. They're answered by the contract, the architecture, and the person you're buying from.

When an open model genuinely fits

Fair's fair: there are cases where running your own open-weight model is the right call.

• Hard data-residency rules. If regulation or a client contract says certain data never leaves your premises, an open model on your own hardware answers that cleanly.
• Very high, very predictable usage. Past a certain volume, owning the van beats hiring it. Most small businesses are nowhere near that line, but some are.
• No internet where the work happens. Rare, but real for some field and manufacturing settings.

If none of those describes you, the hire fleet plus a well-built layer of your own is almost certainly the better deal, and it leaves your capital free for work that actually wins you customers.

The bit nobody in Washington is fighting about

Here's the thing I keep coming back to with the businesses I work with. The giants are fighting about the engine because engines are their business. Your business runs on something else, and the ceiling on what can be built there is far higher than most owners have been shown.

The small end is the familiar stuff: follow-up that never slips, quotes out the same day, information entered once instead of three times. Worth having, and it's usually where the first win comes from. But it's the entry point, not the destination.

The bigger end is a whole operating layer for the business. Your quoting, your job records, your finances and your customer history connected and talking to each other instead of sitting in separate tools. Every person on the team with AI that knows their own role, so the office manager, the estimator and the person on site each get help with their actual job. A business that keeps a working memory of everything it has ever quoted, bought, promised and delivered, and can answer questions about itself in seconds. Owners who read their whole month in one page instead of digging through five systems.

None of that is science fiction. It's what gets built, layer by layer, on top of whichever model you rent. And none of it is an open-versus-closed decision. It's a mapping decision: knowing which piece pays for itself first and building in that order. Get that right and the model underneath becomes what it should have been all along, a part you can swap.

If you want the wider picture of what AI is actually doing inside UK small firms right now, our guide to how small UK trades use AI in 2026 covers the four jobs it does well, and the AI automation pillar guide goes deeper on the whole stack.