The Anthropic ban in the US is the best advertisement for Chinese AI, according to the SCMP

By South China Morning Post · June 24, 2026.

Columnist Alex Lo publishes in the South China Morning Post an opinion column analyzing the unintended consequences of U.S. export restrictions on Anthropic's artificial intelligence models. The central argument is paradoxical: by cutting off foreigners' access to the company's most advanced models, Washington would be involuntarily pushing the rest of the world toward Chinese open-source alternatives.

The immediate trigger is two concrete news items. Goldman Sachs, in April 2026, and JPMorgan Chase, the week before the article's publication (mid-June 2026), decided to disconnect their employees in Hong Kong from access to Anthropic's models. Both institutions made this decision based on a strict interpretation of the company's terms of use, which in turn reflect the restrictions Washington has imposed to limit China's access to the frontier AI models developed in the United States.

The hardest blow came with a U.S. Department of Commerce directive published in June 2026, which orders the denial to all foreigners—including Anthropic's own foreign employees—of access to the company's two most advanced models, identified in the article as Fable 5 and Mythos 5. According to Lo, the export-control directive left Anthropic barely 90 minutes to cut off access, citing national security reasons. The speed and forcefulness of the measure underscore the urgent and political character of the decision, which transcends commercial logic.

The impact on Hong Kong is the article's connecting thread. The columnist cites a Financial Times report warning that preventing access to the world's most advanced AI models poses a direct threat to Hong Kong's recovery as an international financial center, given that its adoption is massive in other parts of the world, especially for programming tasks. For the Goldman Sachs and JPMorgan bankers in the city, the disconnection is not a minor inconvenience: it represents a tangible competitive loss relative to their colleagues in other financial centers that do retain access to these tools.

Lo's thesis is that the most significant effect of these restrictions will not fall on mainland China—which already has access banned and has developed its own alternatives—but on the rest of the world, which until now freely used American models. By closing that access, the U.S. forces users in third countries, multinational companies and knowledge workers in centers like Hong Kong to seek substitutes. And the most obvious substitutes are Chinese models, which the article describes as good enough and available at a fraction of the cost of their American equivalents.

In general, and as sector context, the debate over AI export controls has been intensifying for months in Washington. The official logic is that frontier models, capable of accelerating research in sensitive areas such as synthetic biology, offensive cybersecurity or weapons design, should not be accessible to potentially adversarial actors. However, critics of these policies—including part of the American tech industry itself—argue that overly broad controls damage the competitive position of U.S. companies in global markets without actually preventing China from developing its own capabilities.

Lo's article illustrates precisely this tension. The Chinese AI industry has advanced rapidly in recent years, driven in part by the previous restrictions on chips and software that forced its companies to develop their own capabilities. Open-source or low-cost models from Chinese companies—although the article does not name them explicitly—already compete on performance with many Western commercial offerings for a wide spectrum of uses. The gap in standard use cases—writing, analysis, code, translation—has narrowed enough that the 'good enough' label is plausible for most corporate users.

From the perspective of agentic AI and enterprise systems, the forced disconnection of Anthropic in Hong Kong has relevant implications. Investment banks are heavy users of AI agents for tasks such as document analysis, report generation, due diligence assistance and automation of regulatory compliance workflows. Losing access to the most capable models of a trusted provider in the middle of a technology rollout generates operational friction, migration costs and service-interruption risks. The speed with which the disconnection was executed—90 minutes of margin—aggravates the perception of regulatory unpredictability.

For developers and companies building solutions on frontier model APIs, the episode reinforces the argument in favor of multi-provider architectures and the adoption of open-source models that can be deployed locally, without dependence on external political decisions. In this sense, the article points indirectly to a structural change in how enterprise technology teams design their AI stacks: sovereignty over the model—knowing that it cannot be disconnected with 90 minutes' notice—becomes a selection criterion as important as performance or price.

Lo does not address in depth the European regulatory perspective or the EU AI Act framework, but the situation is equally relevant for European companies with operations in Asia. If American frontier models become out of reach for subsidiaries in certain jurisdictions, the choice between Chinese, European providers or open-source models takes on a geopolitical dimension that goes beyond the technical comparison.

In prospective terms, the columnist suggests that the medium-term result of this escalation of restrictions could be the consolidation of an AI market fragmented along geopolitical blocs: an American ecosystem for allies and U.S. citizens, and a Chinese ecosystem for the rest—or at least for those who cannot or will not accept the conditions of access to American models. This bifurcation, if consolidated, would have profound consequences for the standardization of interfaces, the interoperability of agents and the global governance of AI.

The article is short—the outlet itself labels it as a 3-minute read—and its argument is deliberately provocative, as befits an opinion column. It offers no quantitative data on market share or cites comparative performance studies between American and Chinese models. Its value lies in pointing out a real political contradiction: the very measures designed to contain Chinese technological advancement may be accelerating the global adoption of its products.