The Coming Intelligence Import Bill

On 30 March 2026, a 28-year assumption about digital trade expired. Since 1998, members of the World Trade Organization had repeatedly agreed not to impose customs duties on electronic transmissions. At the Fourteenth Ministerial Conference in Yaoundé, they failed to renew that moratorium. No customs officer began inspecting software downloads or AI prompts the next morning, but the assumption that digital trade stays duty-free is now exactly that, an assumption.

When a company sends data to a model operating abroad and receives software, analysis, product designs or operational decisions in return, nothing physical crosses a port. Productive capability has crossed a border all the same. Today the payment appears as an API charge, an enterprise licence or a cloud bill. Tomorrow the intelligence import bill may contain more lines.

In July, Demis Hassabis proposed a US-led standards body for frontier AI, modelled partly on FINRA, the industry-funded organisation that oversees American securities brokerages. Laboratories would submit their most capable models for testing against dangerous cyber, biological and deceptive capabilities before release. The proposal deserves to be taken seriously, and it raises a question it does not answer. Who will build the models being tested? Standards can make imported intelligence safer. They cannot guarantee it remains affordable, available or under the importing country’s control. A country with excellent regulators and no capability of its own does not have sovereignty. It has well-managed dependency.

Countries have depended on foreign technology for decades, so it is fair to ask why AI should be treated differently. Three reasons. Advanced AI is becoming an input into decisions rather than a tool executing a specified process. The dependency compounds, because organisations redesign their data, workflows and security practices around a particular model until replacing it means rebuilding the institution rather than installing a substitute. And every model embodies choices about acceptable behaviour and tolerable risk that were made under another jurisdiction’s laws and political pressures.

The coming friction will arrive from both directions. Importing governments may tax intelligence to protect their revenue bases, which is part of what the dispute in Yaoundé was about. Producing countries may condition access through export licences, restrictions on model weights, or priority for domestic customers when capacity is scarce. Not every condition will be unreasonable. The point is that a supposedly borderless market is acquiring borders, and dependency does not always appear as a tariff. Sometimes it appears as a capability that is unavailable precisely when it is most valuable.

That would matter less if frontier capability were widely distributed. It is not. The Stanford AI Index reports that industry produced more than 90 per cent of notable models in 2025, with American organisations producing 50 and Chinese organisations 30. Private AI investment in the United States reached $285.9 billion, more than twenty times the recorded figure for China. None of this is evidence of wrongdoing. Concentration creates leverage regardless of how fairly it arose.

The strongest counterargument is open-weight AI. The UK’s AI Security Institute found that the most capable open model it tested performed similarly on its cyber evaluations to closed systems released only four to seven months earlier. Open weights weaken pricing power and blunt the threat of sudden withdrawal, and countries should use them. But operating today’s generation is not the same as being able to build tomorrow’s, and there is no guarantee the most capable future systems will be released openly. Open weights are a route towards capability, not an excuse to avoid building it.

Not every country can reproduce Silicon Valley, and pretending otherwise turns sovereignty into branding. Britain’s £500 million Sovereign AI programme and £1.1 billion hardware plan illustrate what a serious beginning looks like without approaching parity with the largest American laboratories. For most countries, minimum viable sovereignty will mean domestic expertise, competitive open weights, national or pooled allied computing, and dependable partners. Every country does not need to build everything. Every country needs a strategy for what it must be able to do without asking another country’s permission.

This is also why standards bodies and capability building belong together. Countries that bring models, computing and evaluation expertise to a standards regime participate as technical peers. Countries that bring only demand participate as customers.

There may never be a single universal tariff called an AI duty. Some of the coming cost will be importer taxation, some producer conditioning, some ordinary pricing power in a concentrated market. The name of the mechanism matters less than the position of the country facing it. Countries that can train, adapt and evaluate competitive systems will keep using foreign models from a position of choice. The next customs frontier may have no ports or cargo ships. It may be an API gateway, and the time to build an alternative is before that gateway acquires a toll.