The AI model can be Chinese. The servers Australian. The data local. The business built on top entirely yours.
That changes how we should think about competition, sovereignty and value creation.
A discussion with Kevin on Big Stakes pod brought me back to a question I think more investors, founders and policymakers should ask:
When we restrict foreign AI, which part of the stack are we actually protecting?
A full-stack capitalist perspective looks beyond the model developer. It follows ownership, control and margins through chips, energy, cloud infrastructure, models, applications, distribution and customer relationships.
Those layers do not have to belong to the same company, or the same country.
An open-weight model makes this particularly interesting. Downloadable model parameters can allow a local operator to run the model without sending prompts to its original developer, retain a particular version and build products around it, subject to the licence.
The developer’s nationality still matters to the assessment. But it does not tell us who receives the data, controls the deployment or captures the commercial value.
Operational control can transfer without complete knowledge transferring.
Local hosting does not remove the need to examine software dependencies, external connections, model behaviour and updates. Sensitive government applications deserve demanding standards.
But those standards should identify the risk that remains after safeguards are applied.
If a deployment is genuinely isolated, a claim that information reaches foreign servers needs an actual transmission pathway. If the concern is unreliable behaviour, test it. If the concern is strategic dependence, examine switching costs, update control and available alternatives.
Each problem calls for a different response.
There is also an economic question we should stop treating as secondary.
A restriction can protect domestic model suppliers while raising costs for domestic businesses building on top of models. Investors in one layer may benefit while entrepreneurs in another lose a useful input.
Cheaper, capable models could shift more value toward applications, proprietary workflows, distribution and customer trust. That is an opportunity for businesses that will never train a frontier model themselves.
Chinese developers can benefit too. Widespread adoption can build an ecosystem of skills, integrations and derivative products around their technology, even when they never receive the end user’s prompts.
Both sides can gain. Dependence can still emerge. The task is to understand where.
My test for a restriction is straightforward:
What specific risk are we addressing? What evidence would change the decision? Why are narrower safeguards insufficient? And what capability, competition and business formation are we giving up?
Governments can have legitimate security interests, and their restrictions can still impose substantial competitive costs.
A full-stack capitalist should count both.
Sovereignty includes the capacity to inspect, operate, adapt and replace the technology your economy depends on. Domestic ownership matters. So does the freedom of domestic businesses to build.