When closing the door pushes the customer to the shop across the street: the paradox of AI controls
🕒 Published on Zendoric: June 24, 2026 · 09:00
Goldman Sachs and JPMorgan are cutting off their Hong Kong teams from Anthropic's models due to a directive from Washington. Alex Lo's column in the SCMP makes an uncomfortable case: the ban doesn't slow China down, it hands market share to its open alternatives. Beyond the geopolitics, the episode leaves an architecture lesson for any company.
There are political decisions whose real effect surfaces in the least expected place. As columnist Alex Lo recounts in the South China Morning Post, a June 2026 directive from the U.S. Department of Commerce forced the denial of access to all foreigners —including Anthropic's own foreign employees— to its two most advanced models, cited in the article as Fable 5 and Mythos 5, with barely 90 minutes to carry out the cutoff. Earlier, Goldman Sachs (April) and JPMorgan (mid-June) had already cut off their teams in Hong Kong, citing a strict reading of the terms of use.
Lo's thesis is deliberately paradoxical: the blow does not fall on mainland China, which had already been barred from access and has developed its own capabilities, but on the rest of the world that until now used U.S. models normally. By closing that door to them, the author argues, multinationals and knowledge workers are pushed toward Chinese open-source models, described as 'good enough' and available at a fraction of the cost. It is worth taking this argument for what it is —an opinion column— and not as a settled prediction, but the dynamic it describes is recognizable.
Beyond the geopolitical standoff, there is an operational lesson here that goes beyond Anthropic and Washington. What is truly instructive for any technical team is the word 'unpredictability': a trusted provider can become unreachable with 90 minutes' notice because of a decision unrelated to the contract. That turns sovereignty over the model —knowing it cannot be switched off on you by surprise— into a selection criterion as legitimate as performance or price.
The sensible response is not to pick a side, but to design for resilience: multi-provider architectures, abstraction layers that allow switching models without rewriting the application, and the option to deploy open weights on one's own infrastructure when the use case justifies it. The implicit accusations about the intentions of one party or another belong to the realm of political debate; the engineering lesson, by contrast, is clear and applicable today. Anyone building on a single frontier model, whatever its origin, is accepting a concentration risk that should be measured and mitigated in advance.