JPMorgan Cuts Anthropic AI Access in Hong Kong

JPMorgan Chase has blocked its Hong Kong employees from accessing Anthropic’s Claude AI models, the Financial Times reported on June 18, 2026, citing three people familiar with the matter. The move strips staff at the world’s largest bank by market capitalisation of a tool that had been available through an internal drop-down list of approved large language models. It’s the second such restriction imposed by a Wall Street institution in less than two months — and the contractual mechanism behind both decisions reveals something deeper than a routine vendor dispute.

The timing is deliberate. JPMorgan’s decision follows Goldman Sachs, which in late April 2026 quietly removed Claude from its internal AI platform for Hong Kong-based bankers — a move Reuters confirmed on April 29 after the Financial Times broke the story. Together, these two restrictions represent a significant narrowing of Anthropic’s footprint inside global finance’s most important Asian hub.

The broader backdrop is a US-China technology rivalry that has moved, with remarkable speed, from semiconductors to software. In February 2026, Anthropic publicly accused three Chinese AI laboratories — DeepSeek, Moonshot AI, and MiniMax — of orchestrating what it called “industrial-scale campaigns” to extract Claude’s capabilities through approximately 24,000 fraudulent accounts and more than 16 million engineered exchanges, a technique known as model distillation. OpenAI had raised similar alarms eleven days earlier. Google’s Threat Intelligence Group confirmed its Gemini model had faced comparable incursions. The line between a legitimate AI vendor relationship and an intelligence risk was no longer theoretical.

Hong Kong sits precisely on that line.

The mechanism behind both the Goldman and JPMorgan restrictions is the same, and it deserves careful attention: neither bank was ordered to act by a regulator. Both arrived at their decisions by reading their own contracts with Anthropic and concluding that the terms did not permit use in Hong Kong.

That conclusion was not difficult to reach. Anthropic’s own supported-countries page does not list Hong Kong as a market where its commercial API or Claude.ai are officially accessible. A company spokesperson told the Financial Times that Claude models had “never been officially supported” in the territory, though the company declined further comment. In other words, both Goldman and JPMorgan had, for some period, been operating outside the literal scope of their vendor agreements — and only a formal compliance review brought that to light.

For Goldman, that review came after an internal consultation with Anthropic, described by the FT as a “strict interpretation” of its licensing arrangement. JPMorgan’s decision followed the same logic: the wording of Anthropic’s usage terms prompted the bank to delist Claude models from its internal tool selection interface. JPMorgan and Anthropic did not respond to Reuters’ requests for comment.

What makes this notable is the asymmetry. Other AI models remain available on Goldman’s internal platform: OpenAI’s ChatGPT and Google’s Gemini are still accessible to Hong Kong staff. This isn’t a blanket retreat from AI tools in the city. It’s specifically Anthropic — and the reason traces directly to where Hong Kong sits in Anthropic’s geographic framework, which groups the territory alongside mainland China in its access restrictions.

In September 2025, Anthropic tightened its terms of service further, prohibiting access from companies whose majority ownership is directly or indirectly attributable to entities headquartered in unsupported regions. The stated logic was that subsidiaries in supported jurisdictions had been used as pass-through access channels. The Goldman and JPMorgan moves are the downstream consequence of that policy meeting enterprise contracts.

Why JPMorgan Blocking Anthropic in Hong Kong Matters Beyond Finance

The standard reading of this story positions it as a compliance footnote — two banks tidying up their vendor agreements. That reading is too narrow.

What does JPMorgan blocking Anthropic in Hong Kong actually mean for AI access policy?

JPMorgan’s restriction signals that frontier AI vendors are now enforcing geographic access through private contract terms, effectively creating a parallel export-control regime that operates faster and with less procedural visibility than formal government regulation. Where traditional export controls require rulemaking, public comment, and enforcement by customs authorities, vendor-imposed regional restrictions are expressed in bilateral contracts, enforced by IP-level blocks, and largely invisible to outside observers until a major institution is affected.

The picture is more complicated than simple US-China tension. Hong Kong has historically operated outside the Great Firewall that blocks Claude, ChatGPT, and other Western AI models in mainland China. That status has eroded not because of anything the Hong Kong government has done, but because of decisions made in San Francisco boardrooms. Anthropic has been explicit that it maintains defence and intelligence relationships with the US government and has actively advocated for semiconductor export controls on China. Its regional access policy is, in part, a reflection of those commitments.

The National Security Law imposed on Hong Kong in 2020 fundamentally changed the territory’s legal architecture — creating obligations to share information with mainland authorities that US technology companies regard as incompatible with their own security commitments. Goldman Sachs CIO Marco Argenti had, as recently as February 2026, described an active partnership with Anthropic to develop autonomous AI agents for trade accounting and client onboarding. The Hong Kong restriction doesn’t end that relationship. It draws a geographic boundary around it.

For the banks, the calculus is straightforward: the productivity gains from AI access at a Hong Kong desk do not outweigh the compliance, legal, and reputational risk of operating outside a vendor’s supported terms in a territory that carries elevated intelligence-exposure risk.

The JPMorgan decision will accelerate a review that legal teams at every major multinational with Anthropic enterprise agreements should already have begun. The question is no longer whether Hong Kong is a compliant deployment location — the answer is clearly no. The question is how many other institutions have been quietly using Claude in the territory without having audited their contracts.

The Hong Kong Monetary Authority has already signalled its engagement. Reuters reported in late April that the HKMA had contacted a range of major banks to understand developments around Anthropic’s newer models and to remind them to update their risk assessments. That kind of regulatory outreach, even when framed as information-gathering, is a prompt to act.

The second-order effects extend to Hong Kong’s ambitions as an AI hub. The territory has spent the better part of a decade positioning itself as a bridge between Chinese capital and Western technology. That bridge is narrowing. If the two most prominent US AI safety companies — Anthropic and, to a degree, OpenAI, which restricted Chinese API traffic in 2024 — treat Hong Kong as within their China risk perimeter, the competitive disadvantage for Hong Kong-based financial technology firms compounds quickly.

There’s also a signal effect for Anthropic’s enterprise strategy. The company has made substantial inroads in global banking: its Claude for Financial Services product launched in July 2025, expanding in October with real-time market-data connectors and pre-built agent skills for tasks such as discounted cash flow models and coverage reports. Goldman’s six-month embedded-engineer partnership was the most visible sign of that momentum. Yet the Hong Kong restrictions reveal a structural tension in Anthropic’s commercial model: the same security posture that makes the company attractive to US defence agencies makes it unavailable in one of the world’s most important financial centres.

For institutions choosing AI infrastructure in 2026, this imposes a new due-diligence requirement. Geography is now a compliance variable in AI vendor contracts, in the same way it is for data residency, cross-border transfer restrictions, and sanctions screening. Legal teams that built their enterprise AI agreements before this framework solidified are carrying risk they may not have priced.

It’s worth taking seriously the argument that these restrictions are disproportionate, or at minimum, poorly calibrated.

Hong Kong is not mainland China. Its legal system retains common-law foundations and its financial regulatory architecture — overseen by the HKMA — remains broadly aligned with international standards. Anthropic itself has acknowledged that distillation is “a widely used and legitimate training method” in the AI industry; its objection to the Chinese lab campaigns was not to the technique but to the alleged use of fraudulent accounts to circumvent access restrictions.

Critics argue, with some justification, that corporate AI access policy is doing the work of geopolitics without the accountability of formal regulation. As one analyst framed it in a widely circulated April 2026 essay: “The block, in other words, was self-imposed by the customer, on the basis of contract terms set by the vendor, not by any government regulator on either side.” There’s no public comment period, no appeals process, no parliamentary scrutiny.

There’s also a competitive-dynamics dimension that Anthropic’s critics are quick to identify. OpenAI’s ChatGPT and Google’s Gemini remain available on Goldman’s Hong Kong platform. If Anthropic’s stricter geographic policy results in it losing ground to rivals in major Asia-Pacific markets, the company’s commercial sustainability — and, by extension, its ability to fund frontier safety research — takes a hit. The irony is that the safety-focused company may be hardening its access policy in ways that cede the field to competitors less exercised about where their models end up.

Professors and ethics consultants studying the distillation controversy have raised a further point. Lia Raquel Neves, founder of the ethics consultancy EITIC, noted that since Anthropic itself recognises distillation as a legitimate practice, “the central point of controversy lies not only in the technique itself, but in the alleged fraudulent access and possible violation of contractual terms.” The legal and moral weight of Anthropic’s position rests on contract enforcement, not on a principled objection to knowledge transfer. That distinction matters when assessing whether geographic blanket restrictions are a proportionate response.

What JPMorgan’s decision confirms — coming seven weeks after Goldman’s and on the same contractual basis — is that AI access is now a geopolitical asset class. The question of which staff in which offices can use which model is no longer a procurement decision. It’s a foreign-policy adjacency.

Hong Kong built its financial primacy on the promise that it could offer something mainland China could not: legal predictability, open information flows, and access to the world’s best tools and capital. That promise is being tested from multiple directions simultaneously — by Beijing’s legal architecture, by Washington’s export-control instincts, and now by the private contract decisions of AI companies headquartered in San Francisco.

JPMorgan’s move won’t be the last. Every major institution with an Anthropic enterprise agreement and a Hong Kong desk is, today, looking at its contract language with fresh eyes. The banks that haven’t yet acted are simply the ones that haven’t finished reading.