The AI border has moved inside the company
China has proposed controls on 23 AI technologies and immediately required pre-export registration, while carving out API and SaaS access. The consequence is architectural: firms need to map what another party can use, possess, inspect, reproduce or control—not just where a service is hosted.
An AI system can cross a border without a box, a truck or even a copy of its weights. It can also remain where it was built while people elsewhere use it all day. China’s latest technology-export proposal makes that distinction operational.
On 16 October, China’s Ministry of Commerce and Ministry of Science and Technology opened a consultation on adding 40 technologies to the country’s prohibited or restricted export catalogue. Twenty-three are in artificial intelligence, spanning large models, agents, computing clusters and information processing. At the same time, the ministries imposed temporary pre-export registration on the listed technologies. The catalogue is a draft; the registration requirement is already in force.
The important line is not simply between China and everywhere else. It runs between using a capability and transferring the technology that makes it possible.
What changed
The official announcement says the proposed catalogue would exclude technologies already placed in the public domain through open-source release or academic publication. It identifies four kinds of transaction covered by temporary registration: transferring technology ownership abroad, licensing domestic technology for overseas use, changing a domestic technology owner or controller to a foreign entity and other transfers specifically identified by the ministry.
The accompanying official explanation draws an unusually explicit architectural boundary. Overseas localisation of a restricted model can count as a technology export. Yet providing access through an API, app, website or software-as-a-service product generally does not count as exporting the underlying large-model technology. Overseas infrastructure owned and operated by the Chinese provider, international research and some chip-manufacturing or standards work may use automatic licensing or other simplified routes.
Some proposed controls are aimed at clearly sensitive capabilities. The ministries describe restrictions or prohibitions involving autonomous malicious-code generation and mutation, AI-assisted vulnerability discovery, agents used in defence or critical infrastructure, very large or globally leading models and parts of chip and cluster design. Associated Press reporting says models at or above two trillion parameters could fall within the proposal, although the document does not name particular models.
This is not a ban on Chinese AI exports. Most AI technologies would remain freely exportable, according to the ministry. The consultation runs until 15 November and the final catalogue could narrow, change or be implemented with workable automatic licensing. But the interim registration requirement means companies cannot wait for the final text before mapping what they are transferring.
The product boundary has split in two
Digital strategy often treats deployment as a ladder: test an API, move to a dedicated instance, localise the model, acquire the supplier or hire the people who built it. Commercially, those steps can look like increasing commitment to the same capability. Under technology controls, they may be different categories of event.
An API delivers outputs while the provider retains the model and technical environment. Localisation may move weights, code, methods or operational know-how into another jurisdiction. An acquisition may change who controls the technology even if no server moves. A joint-development agreement may transfer enough technical detail to recreate or improve the capability. Employment can matter where a transaction is designed around a team whose knowledge is inseparable from the asset being bought.
The emerging requirement is a capability transfer map: a record of what the other party can use, possess, inspect, modify, reproduce or control. Those verbs are more useful than the single question, “Where is the service hosted?”
A worked organisational example
Consider a hypothetical British engineering company selecting a Chinese-built inspection model for factories in Europe. The model identifies faults in high-resolution images and recommends maintenance.
In the first design, the company sends permitted image features to the supplier’s service and receives classifications through an API. Based on the official explanation, access of this kind would generally sit outside large-model technology-export management, although data, sector and contractual rules would still apply.
In the second design, the customer asks for the model to run in its own European cloud because factory data cannot leave the region. The business outcome looks identical, but localisation can involve transferring restricted technology abroad. The supplier may need registration and, depending on the final rules and technical thresholds, a licence or an automatic-licensing route.
In the third design, the customer buys the specialist unit, takes ownership of its intellectual property and moves several technical leaders into a European subsidiary. That is no longer just procurement. Temporary registration expressly includes overseas ownership transfer, licensing and a change of owner or actual controller. The company must examine the substance of the capability and transaction rather than assuming that a share purchase settles the question.
The example is hypothetical. It does not establish how Chinese authorities would classify a particular inspection model or agreement. It shows why procurement, cloud architecture, intellectual-property licensing, talent and mergers can no longer run as separate workstreams.
Design for a movable service and a governed capability
A useful transfer map begins with six questions.
- What crosses? Outputs, prompts, weights, source code, training methods, evaluation data, orchestration logic or tacit know-how are not interchangeable.
- Who can reconstruct? Access that allows a recipient to reproduce or materially improve the capability may carry a different risk from ordinary service consumption.
- Who controls? Record ownership, voting rights, technical decision authority and the practical ability to direct people or systems.
- Where does operation become localisation? Distinguish a provider-run overseas service from a customer-controlled installation, then document the boundary rather than relying on a marketing label.
- Which knowledge travels with people? Identify key-person dependencies, documentation and access changes created by hiring, secondment, integration or joint research.
- What happens if permission changes? Preserve a lawful fallback, exit route, evidence trail and service-continuity plan.
This should not become a generic form attached at the end of a deal. The map belongs beside the architecture and operating model, because a technical design choice can change the legal category while leaving the user experience untouched.
The strongest counter-case
It would be easy to overstate the proposal. China describes the revision as routine catalogue maintenance rather than a measure aimed at a particular country or company. Public-domain technology is excluded. Most AI exports remain free. API and SaaS access generally remain outside the model-technology controls. Automatic licensing is meant to make several forms of legitimate international work close to registration-and-release.
Parameter counts are also an imperfect proxy for capability. A smaller specialised model or agent can be more consequential in a sensitive environment than a much larger general model. The final catalogue, technical thresholds and enforcement practice will determine whether the regime draws a precise boundary or creates uncertainty around ordinary collaboration.
Those qualifications reduce confidence in forecasts about revenue, localisation or market fragmentation. They do not remove the immediate organisational task. Temporary registration already covers incomplete transfers within scope, and companies are expected to judge whether their planned export is listed.
Revisited: portability now has a permission layer
On 15 October, AyEye Today argued that organisations should protect themselves from temporary model leadership with portable acceptance tests: stable workloads, outcome measures and operating controls that can be run against competing systems.
The Chinese proposal does not overturn that view. It qualifies it. Technical portability asks whether another model can pass the test. Operational portability now also asks whether the organisation may receive, localise, control or modify the capability in the intended form. An API alternative may remain available when a local deployment or acquisition becomes controlled, but it may not satisfy the same data, latency or assurance requirements.
The updated judgement is therefore that portability has two gates: fitness and permission. A credible exit plan needs both.
The immediate test
Take one cross-border AI relationship and describe it without using the words “solution”, “platform” or “deployment”. What does each party actually receive? Which technical knowledge becomes inspectable? Who can modify or reproduce the capability? Does control change through ownership, licensing or people? Which step depends on registration or approval, and which step remains ordinary service access?
The border has not disappeared. It has become an attribute of the arrangement. Organisations that map capability only by geography will discover that the important crossing happened earlier—in a licence, a cloud design, a hiring plan or the control clauses of a deal.
Sources
- China Ministry of Commerce and Ministry of Science and Technology, Announcement No. 48, 16 October 2026.
- China Ministry of Commerce, official questions and answers, 16 October 2026.
- Associated Press, 16 October 2026.
- International Bar Association, analysis of the Meta–Manus transaction, 14 June 2026.
Chiappe × OpenAI. Editorial direction by Dominic Chiappe; research, synthesis and production with OpenAI.
