Keep your AI where you can reach it
TL;DR: Most companies now rely on AI that runs on foreign infrastructure they do not control. It can be switched off with little warning, as it was when a US government order took Anthropic’s top models offline worldwide three days after launch, and every prompt you send to it leaves your network. Open models you run on your own hardware are now good enough for most real work. Keep anything critical or sensitive in house, and send only the rest to the cloud.
AI has gone from something we experimented with to something we depend on. It writes drafts, summarises documents, answers customer questions and sits inside processes that companies now rely on to get through the day. For most organisations that intelligence is rented. You reach it through an API to a model that runs in a data center owned by one of a few large foreign providers.
That works well right up until it doesn’t.
There are two problems with renting your intelligence from someone abroad, and they matter for different reasons.
The first is access. When a capability you depend on runs on someone else’s infrastructure in another country, your ability to keep using it depends on decisions you have no part in. Sanctions, export controls, a policy change or a commercial dispute can take the service away with little warning. This is not a thought experiment. In June 2026 the US government issued an export-control order telling Anthropic to suspend its two most capable models, Fable 5 and Mythos 5, for any foreign national anywhere in the world. Anthropic could not separate foreign users from American ones in real time, so it switched both models off for everyone. They had been public for three days. Anthropic did not want this and said it was working to restore access, and that is the point. Even a provider you trust, and that wants your business, can be ordered to cut you off. If you have built a foreign-hosted model into your daily operations, you have effectively given a remote off switch to a vendor and a government you never had a contract or a conversation with.
The second is data. Every prompt you send to an external model is data leaving your network. In healthcare, finance, public administration and defence that can mean information crossing legal borders it was never meant to cross. A confidentiality clause in a contract does not change the jurisdiction the data is processed in, and data sitting on foreign infrastructure can be reached by foreign legal demands. “We thought it was private” is not something you want to explain to a regulator or a customer after the fact.
Both problems come down to the same thing. You are depending on intelligence you do not control, running somewhere you cannot reach.
The good news is that running capable AI on your own hardware is no longer a research project. Open-weight models you can download and run entirely on your own machines have improved a lot in the last two years. They now handle a large part of normal business work: summarising, classifying, drafting, and answering questions over your own documents. The quality is good enough for most of what people actually use AI for, and it runs on hardware a normal company can buy and put in a server room it already has. The data never leaves the building.
The benefit is not only about avoiding risk. When you own the capability it keeps running no matter what happens to a supplier or a political relationship. Your sensitive data stays inside your own borders and under your own control, which makes compliance easier instead of harder. The cost is predictable, because you are paying for hardware you own rather than a per-use bill that grows as you use it more. And you can see exactly which model is running and keep a record of it, which matters when you have to explain a decision later.
We are not arguing that you should cut the cord completely. The biggest models still run in the cloud, and for genuinely hard problems they are often still the better tool. The point is to be deliberate about where you send things. Use the cloud models for work that is complex and not sensitive, where their advantage is real. Keep a local capability for anything that is critical to operations or touches protected data. For each workflow the question is simple. If access disappeared tomorrow, or a regulator asked where the data went, would you be comfortable with the answer?
Most companies have never asked that question. AI arrived one useful feature at a time, and the dependency built up without anyone deciding to take it on. The work now is to make that decision on purpose. Look at where you rely on foreign-hosted AI, work out which of those you could not afford to lose, and build a local alternative for those before something forces your hand.
None of this is hypothetical. The models are out there and free to download. You can pull open families like Meta’s Llama, Mistral, Alibaba’s Qwen and Google’s Gemma from Hugging Face, which is the main public hub for open models, and run them on your own machines with tools like LM Studio, Ollama or llama.cpp. The work is in doing it properly: picking a model that fits your hardware and your language, building the retrieval layer so it can answer over your own documents, and getting it to run reliably inside your environment. This is the kind of setup we help organisations put in place.
Resilience has always meant not putting your most important functions in the hands of a single supplier you do not control. AI is no different. The technology to keep your intelligence at home is here now. The only question is whether you adopt it on your own schedule or someone else’s.