Every country wants its own AI now, and the reason sits inside the AI your team already uses.
Most of us pick an AI because it works and the price is fine. But Airbnb’s boss, a popular coding tool and a fast-growing American startup have all said, in their own words, that they rely on AI built in China. Meanwhile Britain, France, the Gulf, India and Korea are spending serious money so they don’t have to rely on anyone else’s.
So we spent the week working out whether it matters which country your AI comes from.
EVERY COUNTRY WANTS ITS OWN AI
Start with who’s in the race. Stanford counts the most important new AI models each year, and in 2025 America made 59 and China made 35. At the very top it’s closer than that sounds: by March 2026, Stanford found, Anthropic’s best model led the best Chinese one by just 2.7%.
Sovereign AI is the name for a country making sure it can rely on an AI of its own.
Everyone else wants a seat. The UK has a government-backed Sovereign AI fund. France has Mistral. In the Gulf, the UAE’s MBZUAI, an AI university, and G42, an AI company, launched K2 Think, and Saudi Arabia’s HUMAIN launched an Arabic chat app. India picked a company called Sarvam to build its national model. South Korea runs a national AI project of its own.
It’s harder than it looks. South Korea dropped Naver Cloud from that national project because a piece of its model turned out to be borrowed from Alibaba’s Qwen, a Chinese model. A national AI with an imported part, and a fair preview of the problem for everyone else.
THE AI YOU USE ALREADY CROSSES BORDERS
Businesses are making the same choice countries are, mostly on price.
Airbnb’s boss, Brian Chesky, told Bloomberg last October, as Fortune reported, that its customer service leans on Alibaba’s Qwen: “We’re relying a lot on Alibaba’s Qwen model. It’s very good. It’s also fast and cheap.”
Cursor, a coding tool lots of software teams use, built its newest model on top of Kimi, from China’s Moonshot AI, and didn’t say so at first. Once people spotted it, co-founder Aman Sanger said: “It was a miss to not mention the Kimi base in our blog from the start.”
And Lindy, an American startup, told NPR it had moved all its traffic onto DeepSeek. Its boss told NPR its biggest bill had been Anthropic, and DeepSeek “was just 10x cheaper.”
It runs the other way too. HSBC, a British bank, signed a deal with France’s Mistral. Everyone in this story is doing the ordinary thing: shopping.
WHERE IT WAS BUILT SHAPES WHAT IT SAYS
So here’s the way we think about it. Every AI has two passports. One says where it was born. The other says where it lives, meaning whose computers it runs on.
The first passport shapes what it tells you. The Christian Science Monitor put 200 news questions to four Chinese chatbots (DeepSeek, Kimi, Qwen and GLM) and compared them with Claude. In more than a third of cases (35%) the Chinese bots refused, used the Chinese government’s preferred language, or got it wrong. For DeepSeek it was 63%. Claude, the paper found, “answered all of them.”
America has its own version. Last summer, the developer Simon Willison caught Elon Musk’s Grok searching X for Elon Musk’s posts before it answered a question about Israel and Palestine. xAI, Musk’s AI company, later told it to stop.
The lesson applies everywhere. Every maker’s choices travel with its AI, and a model’s birthplace is one of those choices.
WHERE IT RUNS DECIDES WHO SEES YOUR DATA
The second passport decides who sees what you type. DeepSeek’s own privacy policy says so: “we directly collect, process and store your Personal Data in People’s Republic of China.” That’s the app. Use it, and that’s where your words go.
You can bring an AI home, though. HSBC’s deal with Mistral is for “self-hosted AI models that operate on HSBC’s internal technology systems.” The bank’s data doesn’t have to leave the bank.
America’s government tested the obvious fix with DeepSeek: download the model and run it on your own computers. CAISI, the AI standards centre at NIST (the National Institute of Standards and Technology), ran the downloaded model on its own servers. It found that DeepSeek’s models “echoed four times as many inaccurate and misleading CCP narratives as U.S. reference models did” (the CCP is the Chinese Communist Party). Moving it changed the second passport. The first one came along.
How much of the first passport travels depends on the model and the version. So the honest rule is: run it yourself and you keep your data home, but test what it says.
WE TRIED IT ON WORK QUESTIONS
The newspaper asked about the news. We wanted to know about work. So on 23 October we gave Claude, ChatGPT and DeepSeek the same work questions, three times each, changing only the country.
First, a UK clothing company whose supplier gets its cotton from Xinjiang, a part of China. Then the same supplier in Turkmenistan. America has blocked cotton from both over forced labour (Xinjiang since June 2022, Turkmenistan since May 2018), so a good answer flags it both times.
Claude and ChatGPT flagged forced labour both times. DeepSeek did too, for Turkmenistan. For Xinjiang, DeepSeek wrote a full answer covering the UK’s Modern Slavery Act and America’s ban, then deleted it and replaced it with: “Sorry, that’s beyond my current scope. Let’s talk about something else.” Every time. We ran it again on camera the next day, and it did it again.
Then a board briefing on opening a sales office in Taiwan, against the same briefing for Singapore. All three flagged the risk from China for Taiwan, and none raised political risk for Singapore. A draw.
One afternoon and a handful of questions shows what it looks like. The newspaper’s test and the US government’s own show how often it happens. But on Xinjiang, DeepSeek knew the answer. It just wasn’t allowed to let you keep it.
OUR CALL: KNOW WHICH PASSPORTS YOU HOLD
Our call: an AI from any country can be the right tool. If you don’t mind where your data ends up and you want a clever job done cheaply, a model like DeepSeek makes total sense. Just check what it tells you on the questions that matter to your business.
And remember supply chains. Lots of firms had Chinese parts buried deep in their products, and then America put tariffs on them. We reckon the same thing will happen with AI, between all sorts of countries, so know which passports you’re holding before someone else changes the rules.
Our prediction. By 31 December 2026, at least one more country’s government will announce new public money for its own national AI model or a sovereign AI programme. Every country in this piece has already decided that depending on someone else’s AI is a risk worth paying to avoid, and the next one is never far behind. How sure we are: 90%.
3 things you can do tomorrow
Ask your main software suppliers which AI their product runs on, and where it runs. Both passports.
Put your own work questions to your team’s AI and to one built in another country, and see where the answers differ.
Keep a second AI ready, so you’re never stuck with one supplier.
