There’s something faintly absurd about governments building walls around AI models just as some of them become small enough to run on a laptop.
On 16 October, China’s commerce and science ministries published a draft adding 40 technologies to the country’s export-control catalogue. Twenty-three of them are AI: large models, agents, chip designs and computing clusters. Models above two trillion parameters, or ranked in the world’s top ten on certain benchmarks, could be restricted. An interim registration rule took effect the same day, and comments close on 15 November.
Nobody is running a two-trillion-parameter model on a MacBook, admittedly. Those need a data centre and a sympathetic finance director. But plenty of smaller, very capable open-weight models run quite happily on one.
I understand why governments want this. AI is now tangled up with national security, industrial policy and a fair amount of national pride. Nobody should be surprised politicians have noticed.
But my first thought wasn’t that companies should stop using Chinese models. It was a more practical question. What happens when the thing you’re restricting is already sitting on somebody’s laptop?
The draft has quietly answered that itself. Technology already in the public domain through open source or published research is exempt. Serving models to overseas users through apps, websites and APIs isn’t treated as export at all. What’s in scope is weights, source code and core components leaving the country, including installing a covered model on servers abroad. Future open-source releases would need a filing, followed by automatic licensing.
So this isn’t a ban on Chinese AI, and anyone telling you it is hasn’t read it. The legal detail matters, particularly if you’re the one who has to comply. What it is, in effect, is a government accepting that it can’t recall what’s already out there and trying to control what comes next.
Which is exactly why I think businesses risk drawing the wrong conclusion.
Yes, I’ve used Chinese models
Of course I have. If you’re building with AI, they’re hard to ignore. They’re capable, often cheap, and open weights have given developers options that didn’t exist a few years ago.
For experimentation, the calculation is simple. Does it do the job, and is the cost reasonable? If so, try it.
Most of us understand there are risks in using technology from another jurisdiction. Those risks didn’t appear on 16 October because a ministry published a draft. The question has always been whether the reward justifies them. For a contained experiment, it often does. For a high-grade production system that depends entirely on one Chinese model, hosted service or distribution channel, I’d be much more nervous.
Here’s where I part company with the usual argument. I wouldn’t be comfortable building a critical business that depended entirely on one American provider either.
That isn’t hypothetical. At 5:21pm on Friday 12 June, Anthropic received a US government export-control directive requiring it to suspend access to its Fable 5 and Mythos 5 models for foreign nationals. The only way to comply was to switch them off for every customer. Access came back on 1 July, after the controls were lifted. If your product depended on one of those models, your contingency plan for nearly three weeks was refreshing a status page.
Nobody crossed a border. A letter arrived at the end of a working week.
We’ve managed to turn a basic architecture problem into a geopolitical loyalty test.
The trouble with having a favourite model
I see a lot of teams pick the frontier model they like and build everything around it. Planning, research, repetitive processing, tool use, drafting, review: one model, everywhere.
I understand how it happens. Something works, people get comfortable, and before long half the organisation’s workflows assume it’ll always be there.
Even without geopolitics, that’s fragile. Models change. Pricing changes. New releases behave differently. A model that was brilliant at a task can become oddly unreliable after an update, while one you dismissed six months ago turns out to be excellent.
I’ve seen enough of that to stop treating any model as a permanent member of the architecture. I’ve swapped models in my own work and haven’t yet found a use case where the alternative made the job impossible. That doesn’t mean the results were identical, or that I changed one configuration value and went for lunch.
It means I had options. In a market moving this fast, options are worth more than loyalty.
Model agnostic doesn’t mean model indifferent
There’s a fashionable way of talking about model independence as if every model were an interchangeable plug. It isn’t quite true.
Models reason differently, follow instructions with varying reliability and come with their own strengths and odd habits. In an agentic workflow, where a model calls tools, recovers from errors or makes a judgement that shapes the next step, those differences can matter enormously.
So design to swap models, yes. But test the replacement against the actual work. A successful API response is not the same thing as the business process working.
The more useful idea is to stop assuming one model should do every part of the job. In Ascendancy IQ I’ve come to think of intelligence as something you allocate to the work. A frontier model for hard decisions and orchestration. A cheaper model for repetitive, bounded tasks. Something in between for the daily grind. There’s no reason they need to come from the same company, or the same country.
That gives you a more practical kind of independence. Instead of one preferred model and a slightly dusty backup, you’re already running a system that knows how to use several. If one provider becomes unavailable or uncompetitive, there’s still somewhere to route the work. You just haven’t built the business on the assumption that one model will be there forever.
Where the regulatory argument gets awkward
Export controls aren’t meaningless. Governments have plenty of levers: companies, cloud services, chip supply, procurement, licensing and distribution. Those create real legal obligations, whether or not somebody can find a technical workaround. Chips are the clearest case. US restrictions have already limited China’s access to advanced AI chips and the equipment to make them, and that has genuinely bitten.
But controlling a commercial market is different from making a distributed technology disappear. Once weights are published, people run them locally. Files get copied. Developers move between tools and providers. China’s own draft concedes as much by exempting what’s already public.
I’m not suggesting anyone ignore the law or treat a VPN as a compliance strategy. That would be spectacularly poor advice. I’m saying the power to restrict legitimate commercial activity is not the same as the power to make an already-distributed model vanish. The more capable local models become, the more obvious that gets.
Why Dubai may think differently
Where you build will shape how you respond.
Dubai’s commercial position has long been about connecting markets rather than picking a camp, and I’d expect businesses here to be particularly keen to keep their options open.
At state level it’s less tidy. When Microsoft put $1.5 billion into Abu Dhabi’s G42 in 2024, part of the arrangement was G42 removing Huawei equipment and stepping back from Chinese hardware. On chips and infrastructure, the UAE has largely chosen the American side, which is where the GPUs are.
But that rather proves the point. Governments pick sides on infrastructure. Businesses still have to ship products, and the sensible ones keep their architecture flexible even when the politics isn’t.
The objective is a dependable product, not an ideological statement every time you choose a model. If Mistral works best for one part of a workflow, use it. If an American frontier model is right for difficult reasoning, use that. If a Chinese open-weight model can do a repetitive job at a fraction of the cost, evaluate it properly, including the security, licensing and regulatory implications.
The model should earn its place in the architecture.
What I’d tell a founder building today
I wouldn’t hand them a list of approved countries.
I’d ask them to show me the parts of the business that stop working if their preferred model disappeared tomorrow. Then I’d ask whether they’d actually tested an alternative. Not sent a few prompts to another API. Tested the real workflow, with the same constraints, the same quality bar and the same unpleasant edge cases that turn up once customers are involved.
Some systems will move easily. Others will have accumulated far more provider-specific behaviour than anyone realised. That’s normal. The point is to find out before you’re forced to.
I wouldn’t run three expensive live providers for every trivial task. That would be ridiculous. But I would avoid dependencies nobody understands and nobody has tested.
The best model today may not be the best one in six months. It may not even be the best one for the next step in the same workflow. That’s an extraordinary amount of choice, and it’s worth designing for.
Geopolitics will complicate access, pricing and compliance. Some restrictions will matter a great deal. Others will prove harder to enforce than the press releases suggest.
You can’t control which walls governments build, or when the letter arrives.
You can control how much of your business falls over when it does.
Sources
- AP via PBS NewsHour: China says it will curb tech exports, including AI, citing national security risks (16 October 2026)
- The Next Web: China proposes export controls on its advanced AI models and chip designs (October 2026)
- DEV Community: China’s AI export draft spares API calls and public weights. Overseas deploys aren’t (October 2026)
- Anthropic: Statement on the US government directive to suspend access to Fable 5 and Mythos 5 (12 June 2026)
- Fortune: the White House on Microsoft’s G42 investment and Huawei’s exit (June 2024)
