Access to the same Western expert datasets is a major reason Chinese open models are catching up.

High-quality post-training data plausibly helps, but compute, architecture, distillation, domestic experts and engineering execution are competing explanations. No public attribution supports 'a big reason' with confidence.

Restricting U.S. vendors would preserve the advantage rather than shift purchases to Chinese or third-country suppliers.

The answer depends on how scarce the experts, task designs and verification pipelines really are. Jason asserted scarcity but did not demonstrate it.

Chinese labs can reproduce most expert datasets at acceptable cost and quality using domestic talent.

China has a large technical workforce, but dataset design, tacit knowledge, language, verification and access to frontier failure modes may still create nontrivial scarcity.

A new data restriction could trigger retaliation disproportionate to its security benefit.

Retaliation is a credible policy cost, but it should be weighed against a demonstrated security gain rather than used as a veto on controls.

Expert-data exports are tolerable while U.S. labs lead but become strategically problematic if China catches up.

The better test is the input's marginal capability effect, scarcity and end use. Relative rank affects urgency, not whether a control works.

If a non-frontier model is good enough for most of one sophisticated user's tasks, it will be good enough for most enterprise workloads.

The experience is useful evidence, not a representative workload study. High-stakes reasoning, regulated workflows, latency, security and integration can change the choice.

The frontier lead will remain scarce enough that Anthropic and OpenAI can preserve durable model-layer pricing power.

Current revenue supports scarcity, but model rankings, open-weight releases and Google's integrated stack can change quickly. A present premium does not prove a stable duopoly.

Enterprises can reliably route workloads among open-weight, frontier and specialized models without orchestration costs erasing the savings.

The architecture is already practical and open-weight models support self-hosting, but governance, evaluation and maintenance remain real costs that Friedberg only briefly acknowledged.

The quality gap at sophisticated tasks and the total cost of self-hosting will remain large enough to steer economic share toward frontier labs.

This fits today's market, but both inference efficiency and open-model capability are moving too quickly for a confident long-run rating.

A successful one-month internal build is representative of the cost and durability of replacing packaged enterprise software.

Initial build cost omits long-term maintenance, access control, auditability, integrations and vendor accountability. Those costs vary sharply by application.

Compliance, integrations and data gravity will keep systems of record materially stickier than no-code tools.

These are concrete switching barriers, though they protect incumbents from rapid replacement rather than guaranteeing perpetual pricing power.

No-code and generic application vendors face substantially more AI substitution risk than data platforms and systems of record.

Their core value is closer to code generation and interface assembly, while data gravity, compliance and transaction history create stronger moats elsewhere.