If your AI provider changed the deal tomorrow, what would break?

Every business that has started using AI is now renting its intelligence from a handful of companies. That is not automatically a problem. You rent your electricity too. It becomes a problem when the thing you are renting sits close to the thing that makes you money, and you have no practical way to move.

The argument that put this on the agenda

Speaking on CNBC in July 2026, Palantir's chief executive Alex Karp made the case loudly. Technical customers, he said, want to know they “own the means of production” and that it is “not being transferred to someone else”. His worry is that a business hands its proprietary data to a frontier AI lab, gets a modest amount of value back, and quietly transfers its edge in the process.

Two things are worth saying before you take that at face value. Karp sells the alternative, so it is a sales argument as much as an observation. And on the narrow question of whether the big labs train on your business data, the published policies do not support him. The broader worry is sound though, and the specific version of it is far more useful to you than the headline.

What your provider can actually see

Start with the part most people get wrong, because the answer is better than the fear. On the paid, commercial tiers that a business actually uses, the major labs state plainly that they do not train on your data. Anthropic's policy, updated in March 2026, is that by default it will not use inputs or outputs from its commercial products, including its API, to train its models. OpenAI's developer documentation says data sent to the API is not used to train or improve its models unless you explicitly opt in. Google draws the same line for its paid Gemini and Vertex AI usage.

Now the part people get wrong in the other direction. Those promises attach to the paid tier. The free and consumer tiers are governed by different terms. Google's free tier may use what you send to improve its products, with human review. Anthropic's consumer plans train on chats by default unless you opt out. Same company, same underlying model, completely different deal.

The risk that is real, and it is not training

Depend on one provider and you inherit their roadmap. During 2026 OpenAI published shutdown dates for more than twenty model versions, after which the model simply stops being reachable. Fine tuned models retire alongside the base model they were built on, so a business that had invested in tuning a model to its own data either redid that work or lost it. None of this is unusual. It is ordinary platform behaviour, and it is also how a working automation becomes a broken one on a date you did not choose.

Prices move too, and so do rate limits and terms. That is not malice. It is what building on someone else's product has always meant.

And a platform partner can become a competitor

In February 2026 Figma and Anthropic shipped a joint feature bringing Claude generated code into Figma. Anthropic's chief product officer sat on Figma's board at the time. In April he left that board, and days later Anthropic released Claude Design, a tool that turns a written prompt into a working prototype. Figma's shares fell around seven percent on the day, according to reporting at the time.

Nothing there was improper, and both companies were entitled to do exactly what they did. That is the lesson rather than a complaint. A frontier lab is a company with its own ambitions, and the surface you have built on can become the surface it decides it wants. If your business is a thin layer over one lab's product, you have quietly bet that the lab will never want your customers.

The control plane, in plain terms

The answer is not to avoid the frontier models. They are extraordinary and you should use them. The answer is to stop calling them directly from the middle of your business.

Put a thin layer of your own code between your business and whichever model answers. Everything that is yours lives in that layer. Your prompts, your business context, your rules about what the AI may do, your log of what it did, and the decision about which model gets which job. People call this a control plane. It is not exotic. For most businesses it is one file that every AI call passes through.

• Your prompts and business context live in your code, not in a vendor's prompt library, so they move when you move.
• The provider is chosen by a setting, not hardcoded in twenty places. Swapping becomes an afternoon rather than a rewrite.
• Every call is logged with the model that answered it, so when a model is retired you know precisely what breaks.
• A second provider is wired up and tested before you need it, not on the morning you need it.
• Sensitive work is routed by rule. The client contract does not go to the same place as the blog draft.

Our guide on the harness argues that the wrapper around the model is where your compounding advantage lives. The control plane is that same idea pointed at a different risk. The harness is about the quality of what comes out. The control plane is about who you depend on.

The cheapness is real, and it is not where people think

Two things are true at once. The price of intelligence has collapsed, and buying your own hardware is usually not how a small business gets hold of that.

Andreessen Horowitz put a number on the collapse in 2024, estimating that for a model of a given standard the cost falls by roughly ten times every year. Stanford's AI Index found that a query at the quality of GPT-3.5 fell from around twenty dollars per million tokens in late 2022 to around seven cents two years later, a drop of more than two hundred fold. Nobody sensible expects that to run forever. It has been remarkably steady so far.

Here is the part that catches people out. That collapse reaches you through hosted services, and that includes hosted open weight models. It is a reason to rent, not a reason to buy.

Three ways to run a model, and what each one is for

Rent a frontier model

The most capable models, reached through an API and priced per token. Right for the hard, high value work where a wrong answer costs you real money. You depend on the provider, which is exactly what the control plane is there to manage.

Rent an open weight model

Open weight models are the ones whose weights are published, so anybody can host them. Families such as Llama, Qwen, Gemma and Mistral publish theirs. You can run them yourself, and you can also rent them from hosting companies for very little. Together AI, one such host, publishes a price of five cents per million input tokens for OpenAI's smaller open weight model, and around a dollar per million for Meta's Llama 3.3 70B.

For most businesses this is where the cheapness actually lives. It is also a control move, because the same open model is offered by several hosts. If one puts its price up or falls over, you point your control plane at another and the model itself does not change underneath you.

Own the hardware

Running the model on a machine you own is the strongest form of independence, and the hardware has finally made it plausible. Nvidia's DGX Spark, a desk sized machine with 128 gigabytes of unified memory, launched at 3,999 dollars in October 2025 and rose to 4,699 dollars in February 2026 on memory supply constraints. A single RTX 5090, the first consumer graphics card with enough memory to hold a compressed 70 billion parameter model, lists at 1,799 pounds in the UK and often sells for a good deal more. A well specified Mac Studio does the same job through its unified memory.

The honest arithmetic on owning the box

Hardware is the smallest part of the bill. Business electricity in the UK sits somewhere around twenty two to twenty eight pence per kilowatt hour, so a workstation drawing six hundred watts around the clock costs roughly a hundred and ten pounds a month simply to keep switched on. Then there is your time, because updates, drivers and security patches become your job rather than your provider's.

Now put a real workload against it. A busy internal assistant handling a thousand conversations a day generates a few tens of millions of tokens a month. Rented from a host on a small open weight model, that is a few pounds. The electricity alone on the machine you bought to avoid that bill is many times higher, before you have paid for the machine.

So why run locally at all

Because the good reasons were never about money. The work never leaves your building, which for a firm holding files under legal privilege, or medical records, or payroll, can be the whole argument on its own. There is no network in the way, so it is fast. It keeps working when the line goes down. And it puts a floor under the business, so a price rise, a retired model, or a lab shipping a product that competes with you becomes an inconvenience rather than an emergency.

There is a quieter benefit too. A business that has proved it can run a model on its own hardware negotiates differently, because switching has stopped being theoretical.

Be honest about the trade. Stanford's AI Index found the gap between the best open weight models and the best closed ones narrowing sharply, from around eight percent to under two percent on its benchmark suite in a single year. That is a real convergence, and it is measured on hard reasoning problems rather than the routine work most businesses actually need. For classification, extraction, summarising and first drafts, a small open model is usually more than adequate. For the genuinely difficult judgement, the frontier is still the frontier.

One caution, because the direction of travel is easy to overstate. Menlo Ventures surveyed nearly five hundred enterprise decision makers in late 2025 and found open weight models accounted for about eleven percent of the models those companies called, down from nineteen percent the year before. That is large companies rather than small ones, and it measures which model they chose rather than where it ran. It is still a useful corrective to the idea that the whole market is quietly moving to open models to save money.

A sensible shape for most businesses

1. 1.Reach the frontier models through your own control plane, and spend them on the hard, high value work.
2. 2.Move the high volume, lower stakes work to a rented open weight model, which is where the price collapse actually shows up on your bill. Our guide on cutting your AI bill covers the routing.
3. 3.Run one small model on your own hardware for anything that should never leave the building, or simply prove that you can, so switching is a decision rather than a crisis.
4. 4.Write down what would break tomorrow if your main provider disappeared. If you cannot answer that on one page, that is the work.

What Karp got right

Strip out the salesmanship and a real point survives. The compounding value in your business was never the model. It is your data, your processes, and the judgement encoded in how you use them. Rent the intelligence. Own the context, the rules, the logs, and the route out.

We run our own company this way, with the model swappable behind a seam and the rules and the logs kept firmly on our side of it. If you want to work out what would actually break in your business if your AI provider changed the deal next month, that is a short conversation and a worthwhile one.