Sovereign AI

America innovates, Europe regulates, China imitates, Australia procrastinates

American Tech colonialism comes for friend and foe alike. Don’t get captured.

Generated with Google Gemini

On Friday 12 June, Anthropic received a letter from the US government. Three days earlier Fable 5, their most capable model, had shipped. Now the government wanted it inaccessible to every foreign national on earth, including the ones sitting inside Anthropic’s own offices.

A nationality rule is simply absurd, and impossible to enforce. So Anthropic did the only thing open to them and switched off Fable 5 for the public and Mythos 5 internally. A restriction aimed at foreigners resulted in a global blackout, and the engineers who built the thing, a fair few of them foreign nationals themselves, were locked out of their own work by close of business.

The government pulled up the ladder. And if you live anywhere that isnt the US, it should have made one thing very plain — the most powerful tools of the coming decade can be taken away from you, at a stroke, by a government that was never yours.

American tech is the path of least resistance. Convenient, neatly packaged, integrated to within an inch of its life — the Ikea flatpack of the digital world, except someone in Washington keeps the allen key. And lately the political climate’s bolted a sting onto the tail, and it’s getting harder to swallow. The bloke who sold you the furniture can let himself in and walk off with the wardrobe, and theres bugger all you can do about it. No wonder Europe’s edging for the door.

“America innovates, Europe regulates, China imitates, Australia procrastinates.” It is a good line I have used myself. It flatters everyone who repeats it. It lets America cast itself as building the future, lets Europe feel principled, and lets the rest of us shrug and call inaction prudence. It’s wrong though.

The real divide isnt innovation against regulation. It’s sovereignty against dependence. Who controls the model, compute, data (a can of worms I daren’t get into) and the off switch, and who is renting all of them from someone else and praying the landlord stays friendly.

“If you’re not at the table, you’re on the menu.” — Mark Carney, World Economic Forum, Davos, January 2026

Let me be clear about one thing first. I am not here to wave the CCP flag. The treatment of their own people, the surveillance, Hong Kong, the Uyghurs — the list is long and grim. But I am not American either. From where I sit, the interesting question was never which superpower is the kinder custodian of frontier AI. It’s why on earth you would hand either of them the keys.

Here is the part the cliché cannot explain. China gives its best work away, for now at least. DeepSeek, Qwen, models now credible for local and offline coding work, and near-frontier models in their full form, free to download. America keeps its technological secrets locked up. Yes, Google open-sourced Gemma and Meta shipped Llama, so “the US ships nothing open” is too strong. But the frontier flagship stays shut, and now, as of June, can be switched off by decree.

Why? Because they were never built for you. They were built for America. Even when it required the world’s brightest minds. Much like the Manhattan Project, which was made and then the technology kept for America.

You might say shutting down Fable is spite, or Anthropic getting what its hype machine asked for. Back in February, Anthropic, the one that bangs on most about safety, refused two lines in a Pentagon contract. No fully autonomous weapons, and no mass surveillance of American citizens. For declining to let its model pick targets and watch civilians, the response was swift. Federal agencies were ordered to stop using Anthropic. The company was branded a “supply chain risk”, a label normally reserved for firms with direct ties to foreign adversaries, and the government dangled a Korean-War-era law to compel it. Hours later, OpenAI and the ever scrupulous Sam Altman signed the deal.

Put another way, the firm that asked for guardrails got branded a threat, and the firm that signed got a contract. Now, months later, the same government turned the rival’s flagship off globally.

American frontier AI answers to the American state first, and to the rest of us a long way second. If at all. Anyone selling neutral tools and a rising tide for all humanity is selling you ethical vapourware. And before you write this off as one administration’s nonsense, it’ll be true of the next lot too, just more insidious. Theres a reason the Five Eyes program (FVEY) works, and that the sticking point was “no mass surveillance of American citizens” and not “no mass surveillance”. The carve-out was never for you.

Heres where the rented tools stop being abstract. The system running America’s targeting in Iran is the Maven Smart System — built by Palantir on a roughly $1.3bn Pentagon contract — with Anthropic’s Claude embedded to rank targets by strategic importance and draft an automated legal justification for each strike. Yes, you heard me right. The model writes the paperwork that says the killing was legal.

In the first 24 hours of Operation <insert childish name here> it generated hundreds of strike coordinates. Over a thousand targets fell in a day, a tempo CENTCOM openly credits to AI. On day one a Tomahawk hit an elementary school in Minab. At least 168 people were killed, more than a hundred of them children under twelve. It wasnt a rogue algorithm. The school sat under 100 yards from an IRGC installation, and the failure was the targeting, not an AI malfunction. Which is worse, if anything. The machine worked exactly as designed.

This is the deployment Anthropic balked at. The rupture started when they began asking how Claude had been used in the January operation to capture Maduro in Venezuela. They asked a question about their own model, and got branded a threat to national security for the cheek of it.

To bring it home, because you dont have to be in Tehran for this to bite, picture your bank running fraud checks, your hospital triaging its lists, your states power grid balancing load — all on a model rented from an AWS, Microsoft, Google, hosted in your own country, metered by the token. Convenient. Right up until a letter lands.

Then what? Theres no graceful degradation. You dont fail over to a competitor, you fail back to 2019 — to the spreadsheet nobody maintains and the bloke who knew the manual process, who took redundancy in the first round of “AI transformation”. HAL with a P45. Im sorry Dave, Im afraid I cant do that — and also, Dave, you were let go in 3 months back.

This is the line item the rip-and-replace crowd never costs in. Every workflow you hand to a hosted frontier model is a dependency with a foreign off switch — and the switch isnt yours, isnt your governments, and answers to a politics you dont get a vote in. Resilience used to mean a generator in the basement. Now it means models on your own metal.

So whats the threat here? These are just selling tokens, aren’t they? Well, let’s check out America’s track record.

John Perkins wrote up the playbook in Confessions of an Economic Hitman. A foreign states compliance was never ensured through tanks first. It was loans a country could not repay, IMF conditions it could not refuse, and debt converted quietly into obedience. When the chequebook failed there were cruder tools (Iran in 53, Guatemala in 54, Chile in 73, Iraq in 03) but the elegant version was always financial. Make them dependent. Then make them grateful. Thats the threat tech dependence poses, especially AI. It’s a new tool in the same old arsenal.

You no longer need to bury a nation in dollars when you can host its data, its models and its workflows, and hold the keys to all three. And the keys are real, make no mistake. The US CLOUD Act lets American authorities compel an American company to hand over data it holds, wherever on earth that data physically sits. Jurisdiction follows corporate control, not data location. [1, 2] Your Microsoft and AWS data centres, your sovereign cloud, are sovereign right up until the US judges you inconvenient. No wonder Europe is pushing so hard to remove American tech from government.

Take DeepMind. About as British as a warm flat beer on a wet bank holiday, headquartered in London, and not remotely independent for all that. Google owns it, and Google’s compute is American down to the silicon. So one of the best AI labs on earth sits exactly as exposed as Anthropic did, one letter away from being throttled, repurposed or walled off from its own staff, and not a thing Westminster could do about it. The CLOUD Act doesnt care whats carved above the door. It follows the corporate parent home to California. Makes you wonder whether DeepMind, sat inside Europe rather than just off it, might have stood a better chance.

Palantir’s public-sector role “an unacceptable point of weakness.” — Committee report, June 2026

Enter Palantir. Seeded by In-Q-Tel, the CIA’s own venture arm, now wired into ICE deportations and sitting on a £330m contract to run the plumbing of the NHS. In Australia, it runs the data backbone of Coles — 840 stores, ten billion rows of supply chain, rostering and sales. Everyone panicked about the in-store cameras and missed the real point. It was never the cameras. It’s the ontology underneath the entire business, and once it’s in, you dont rip it out — it’s a years long project. Britain is already feeling the grip. In June a cross-party Commons committee told the government to terminate that contract, and MPs warned that leaning on Palantir makes it harder for the UK to ever publicly disagree with the US. Palantir’s response? It is suing the Mayor of London for blocking a separate police deal. Dependence is not to be refused.

And if data plus models sounds like a theoretical risk, take a glance at Cambridge Analytica. Robert Mercer funded it, funded the Brexit data operation, funded the 2016 Trump campaign — and Mercer was no bystander to the tech. He was an early AI researcher who helped pioneer the statistical language modelling today’s LLMs descend from. The man who weaponised data against democracies was a founding father of the maths now sitting in your pocket.

So what stops a superpower using insidious means to remove the inconvenient? Get between them and their money and you could find yourself quietly destabilised, eased out of office, in the background, by a hand you never see. The old playbook was debt. The new playbook could be, perhaps already is, tech dependence by design. Dare I call it compute colonialism? A bit rich coming from a British bloke living in Australia, granted. But I call it like I see it.

What baffles me is that America seems not to give a single toss how it’s seen, so long as it’s the biggest dog in the yard. There will be a long tail to that. Contempt compounds. You humiliate your allies at Davos, switch off their tools on a whim, and reach into a British hospital’s records because a US statute says you can — people bloody notice. Sentiment hardens into policy. Policy hardens into procurement rules. Europe is already drawing up plans to rip US tech out of government, more than 229,000 Britons have signed petitions to bin Palantir, and the same export stunt that pulled Fable spooked the very investors Anthropic needs for the IPO it just filed for. Branding your own crown jewels a national-security risk is a poor look on a prospectus. None of this lands tomorrow. It lands over a decade — in lost contracts, in allies quietly building their own, in a valuation that stops assuming the world’s custom is a birthright. America is busy winning the argument that nobody should ever depend on America again. Own goal of the bloody century.

Heres the irony that ought to be keeping a few San Fran boardrooms up at night. China isnt giving its models away out of the goodness of its heart, thats commoditise-your-complement, oldest trick in the book. Make the expensive thing free and you torch your rival’s margins on the way past. Android did it to iOS. DeepSeek and Qwen are doing it to OpenAI. Follow it to the end and it gets properly funny. A frightening slice of America’s economy is propped on the promise that AI is a licence to print money. Make the models free, open, and good enough to run yourself, and you dont just dent OpenAI — you knacker the whole “AI will pay for itself” story the US market is leaning its full weight on. China could erode the financial case for American AI without firing a shot, and take a chunk of Wall Street down with it. That last bit is a plausible risk rather than a forecast, so dont quote me a date. And for a CFO it isnt even ideology. A model you run yourself has no meter ticking, no per-token toll, no landlord who can change the locks. Funniest of all, that would make China, not OpenAI, the one actually delivering OpenAI’s founding promise — AI for all of humanity, not one nation’s balance sheet. Irony thick enough to stand a spoon in.

“Vendor lock-in isn’t inevitable… the current position leaves us seriously exposed.” — Dame Chi Onwurah MP, Commons Science, Innovation and Technology Committee, June 2026

So what do you actually do about it, short of unplugging and fashioning a tin-foil hat? Start by killing the laziest part of the cliché. “Europe regulates” is meant as the insult. The continent that lawyered itself out of the race. Look again. Mistralis a roughly €20bn outfit shipping frontier and open-weight models, and it has turned European governance into the selling point. Data that stays in the EU, under EU law, for buyers who dont want Washington holding the keys. Regulation isnt the handbrake. Its becoming the product. Europe is even moving to loosen GDPR so its own labs can train harder, and the privacy hawks are furious, which tells you how seriously the sovereignty argument is now being taken.

Then theres “Australia procrastinates”, which stings because its accurate. And most infuriating of all, procrastination is a choice, and the choice it makes is dependence by default. We dont have to. Maincode in Melbourne is building Matilda on Australian infrastructure — and pointedly dropped the “sovereign” tag along the way, deciding it sounded too defensive. Make of that what you will. Tellingly, it still wont open-source the weights, not out of secrecy, but because it doesnt trust that anything it discloses wont be hoovered up by bigger players who operate on Australian shores unchecked. A small company guarding its IP is the whole national problem in miniature.

Heres the practical bit that matters. Theres a real difference between using the hosted DeepSeek app, which ships your data to servers in China, and downloading open weights to run on your own metal, where nothing leaves the building. One is a different landlord. The other is owning the house. Sovereignty isnt a flag — it’s where your data sleeps and who can switch it off.

If you handed me a magic wand, or just some clout and a spine, Id stop hedging and build the thing properly. Take the nations Carney was speaking to and drag them into an AI Manhattan Project. Pooled compute, frontier models, and (for once) a grown-up answer to where the data comes from and how the thing is governed. India, Europe, ASEAN minus China — thats well over two billion people and an economy in the same league as either hegemon. Not a fortress each behind its own little wall, but one shared stack with values baked in from the start. The talent’s there. The need’s there. Whats missing is the nerve.

And its not a fantasy. In January, at Davos of all places, Mark Carney told the room the old order had ruptured rather than shifted, and that on AI middle powers should band together with like-minded democracies rather than be forced to choose between, in his words, hegemons and hyper-scalers. Dwell on hyper-scalers for a second. A sitting G7 prime minister naming the cloud giants as a strategic threat, out loud, on that stage. He got a standing ovation. Washington’s reply, the next day, was to tell him Canada only survives because of the United States, and to threaten 100% tariffs, rather proving his point. As Carney put it, if youre not at the table, youre on the menu.

So the question was never whether you trust America more than China. Its simpler and harder than that. When the next letter lands at Anthropic’s door, or OpenAI’s, or more troubling still Googles, AWS’ or Microsoft’s — whose off switch is it, and will it be yours?

https://prompttoprod.substack.com/p/america-innovates-europe-regulates