# Wall Street's Claude blackout and the export control that enforces itself

URL: https://www.thedeepfeed.ai/posts/2026-06-18-wall-street-claude-blackout-export-control/
Category: Policy
Published: 2026-06-18
Author: the-deep-feed
Tags: anthropic, export-controls, jpmorgan, goldman-sachs, us-china, ai-policy
Kind: deep

> JPMorgan cut Hong Kong staff off from Claude on June 18, weeks after Goldman did the same. The reason wasn't security — it was licensing fine print. That is how a vague export control the courts would never uphold gets enforced anyway: by private compliance teams.

## TL;DR

- On **June 18**, JPMorgan removed **Anthropic**'s Claude from the approved-tool dropdown for its **Hong Kong** staff, weeks after **Goldman Sachs** quietly did the same. Two of the largest banks on earth walled the city out of a frontier model.
- The stated reason was not model safety. It was **licensing fine print** — Anthropic's usage terms left Hong Kong in a gray zone after the June 12 export-control shutoff, and the banks' legal teams resolved the ambiguity by switching access off.
- This is the mechanism the June 14 analysis missed. The **export control doesn't have to survive a court** to work. It only has to make the terms scary enough that risk-averse compliance departments self-enforce — pushing the cost of legal uncertainty onto private firms.
- A Goldman partner called frontier AI *"state-supervised strategic infrastructure"* and the shutoff *"a kill switch of sorts."* The durable effect isn't capability denied to China. It's a **fragmenting enterprise market**, where access to a model now depends on the passport of the person typing the prompt.

On June 18, an internal dropdown menu at JPMorgan Chase got one entry shorter. Staff in Hong Kong who opened the bank's list of approved AI tools found Anthropic's Claude no longer on it. No memo about model danger, no incident, no breach. The bank's legal team had reviewed Anthropic's licensing terms, decided Hong Kong sat on the wrong side of a line, and removed the option.

Weeks earlier, [Goldman Sachs had done the same thing](https://www.ndtv.com/feature/after-goldman-sachs-jpmorgan-chase-blocks-anthropic-access-for-hong-kong-employees-11658054) to its own Hong Kong bankers. The Financial Times, which [first reported the JPMorgan move](https://finance.yahoo.com/technology/ai/articles/jpmorgan-chase-cuts-off-anthropic-041524918.html), noted the bank had simply removed Claude from the internal dropdown of approved models. The two moves rhyme too closely to be coincidence, and together they mark something the [June 12 export-control shutoff](/posts/2026-06-14-export-control-frontier-model-shutoff/) only hinted at: the order that switched off two **Anthropic** models worldwide is now reaching into the private sector and switching off access one institution at a time — not by force, but by fear of the fine print.

The framing on the ground was blunt. As one market-intelligence account put it the day the news broke:

> JPMorgan just cut its Hong Kong staff off from Anthropic's Claude — joining Goldman Sachs in locking the city out of top AI models. The reason? Not security. The fine print.
>
> — [@PrismMarketView](https://x.com/PrismMarketView/status/2067637874040750102), Jun 18, 2026

That distinction, security versus fine print, is the whole story.

## The fine print, not the firewall

The [June 14 piece](/posts/2026-06-14-export-control-frontier-model-shutoff/) argued that you cannot export-control software at all: the legal definitions don't fit, there's no chokepoint to control, and the open-weights ecosystem exists to destroy manufactured scarcity. Within 24 hours of the shutoff, Zhipu opened the weights of GLM-5.2 as a rebuttal, and the analysis held that the control would fail on all three layers.

It was right about the model. It was wrong about the enforcement surface. The control was never going to be tested in court, because it doesn't run through courts. It runs through every general counsel who reads Anthropic's usage terms and asks a simple question: *if I let my Hong Kong desk use this, and Commerce later decides that was a "deemed export" to a foreign national, am I the one explaining it to a regulator?*

For a bank, the answer to that question is always the same. You don't litigate the ambiguity. You remove the tool.

| Layer the June 14 piece tested | Predicted outcome | What actually happened |
|---|---|---|
| **Legal** — does the rule survive scrutiny? | Fails; software-as-speech precedent (Bernstein) | Untested — banks never challenged it, they complied |
| **Practical** — is there a chokepoint? | None; weights leak, APIs proxy | The chokepoint was built privately: corporate IT dropdowns |
| **Strategic** — does scarcity hold? | No; GLM-5.2 weights opened in 24h | Capability stayed available; *access by geography* fragmented |

The third row is where the surprise lives. The open-weights rebuttal was real — GLM-5.2 is out, and a Hong Kong developer who wants frontier capability has options. But a JPMorgan analyst in Hong Kong cannot use those options on the bank's systems either, because the bank's posture isn't "use the best model," it's "use only what legal has cleared." Capability proliferated. Sanctioned, auditable, enterprise access did the opposite.

![A bank approval dropdown rendered as a labeled control panel: a vertical menu listing four AI tools with one entry struck through in red, a legal-review stamp routing into the toggle, and a Hong Kong region tag flagged beside it.](/post-images/wall-street-claude-blackout-export-control/approval-dropdown.jpg)

## The kill switch found its operators

When the original directive landed, the open question was who would actually pull the trigger on a global API. The state issued the order; Anthropic flipped the switch for everyone because it had no surgical option. But the directive's reach into a Hong Kong trading floor was always going to depend on someone local choosing to comply.

That someone turned out to be the compliance function at the world's largest banks. A Goldman Sachs partner, Bobby Molavi, framed what the episode revealed about the new shape of the market:

> The US administration's move to block foreign nationals from accessing Anthropic's most advanced models over security fears was an unprecedented intervention in the affairs of a leading AI lab. Frontier AI is becoming "state-supervised strategic infrastructure" — an example of a kill switch of sorts being installed.
>
> — Bloomberg, reporting Goldman Sachs partner Bobby Molavi, [via @neilsethinew](https://x.com/neilsethinew/status/2067917810567704660), Jun 19, 2026

"State-supervised strategic infrastructure" is the phrase a bank uses when it has decided to treat a software vendor the way it treats a sanctioned counterparty. Once a tool carries that label internally, the calculus stops being about productivity. It becomes about exposure. And the cheapest way to zero out exposure is to remove the tool from the menu for the jurisdiction in question.

This is why the order didn't need to be coherent to be effective. The [June 14 argument](/posts/2026-06-14-export-control-frontier-model-shutoff/) showed that "remote access over an API" had been ruled *not* an export in three prior Commerce advisory opinions — meaning the legal theory behind the Hong Kong restriction is shaky at best. It doesn't matter. The banks aren't betting on the legal theory being sound. They're betting that being wrong is cheaper than being the test case.

## Why Hong Kong, specifically

As [Semafor framed it](https://www.semafor.com/article/06/18/2026/jpmorgan-restricts-anthropic-in-hong-kong), JPMorgan faced pressure from both the White House and Anthropic's own usage terms. The geography is not incidental. Hong Kong is where US-China financial plumbing physically overlaps — American banks running large operations inside a jurisdiction that Washington increasingly treats as Chinese territory for export-control purposes. That makes it the first place where the abstract category "foreign national" collides with a concrete payroll.

Anthropic's own terms reportedly flag Hong Kong as a region that is not officially supported for its most advanced models, which gave the banks' lawyers the hook they needed. The shutoff didn't have to name Hong Kong. It only had to create enough doubt that a conservative reading of the license excluded it — and conservative readings are the only kind a bank legal department produces.

The result is a precedent that travels. If licensing ambiguity plus regulatory pressure is enough to strip Claude from Hong Kong, the same logic applies anywhere a US export interpretation might later reach: Singapore desks staffed by Chinese nationals, mainland-facing teams, joint ventures. Each institution will draw its own line, and each line will be drawn on the safe side.

## The market that's actually fragmenting

Step back from the banks and the shape becomes clear. The export control's framers wanted to deny capability to a rival state. What they built instead is a system where the same model is available or forbidden depending on who is asking and from where — adjudicated not by Commerce but by thousands of private compliance reviews, each optimizing for its own liability.

Anthropic, notably, is thriving through all of it. The company surpassed OpenAI in enterprise spending share in May and raised at a reported $965 billion valuation, with revenue growth that reads less like a SaaS curve than a phase change in how expertise gets priced. The shutoff and the bank bans haven't dented the business. But they have quietly redrawn its map: there is now a tier of the global economy (large, regulated, cross-border institutions operating in contested jurisdictions) where the safest move is to not use the best model at all.

That's the durable export. Not a capability gap with China; GLM-5.2's open weights closed that within a day. The lasting product is a norm: frontier AI access is now a function of geopolitics, enforced by the people who can least afford to guess wrong. The state wrote the rule. Wall Street is enforcing it, because the alternative is becoming the case that finds out whether the rule was ever enforceable in the first place.

No bank wants that job. So the dropdown gets one entry shorter, and the kill switch the [June 14 piece](/posts/2026-06-14-export-control-frontier-model-shutoff/) described as legally hollow turns out to work perfectly — precisely because no one is willing to test whether it should.

## Sources

- [Financial Times — JPMorgan blocks Anthropic AI access for Hong Kong staff (via Reuters)](https://finance.yahoo.com/technology/ai/articles/jpmorgan-chase-cuts-off-anthropic-041524918.html)
- [Semafor — JPMorgan restricts Anthropic in Hong Kong](https://www.semafor.com/article/06/18/2026/jpmorgan-restricts-anthropic-in-hong-kong)
- [American Banker — JPMorganChase blocks Hong Kong staff from using Claude](https://www.americanbanker.com/news/jpmorganchase-blocks-hong-kong-staff-from-using-claude)
- [The Next Web — JPMorgan cuts off Anthropic access for Hong Kong staff](https://thenextweb.com/news/jpmorgan-cuts-off-anthropic-access-for-hong-kong-staff)
- [GovInfoSecurity — JPMorgan Pulls Anthropic Claude Access in Hong Kong](https://www.govinfosecurity.com/jpmorgan-pulls-anthropic-claude-access-in-hong-kong-a-32013)
- [Storyboard18 — JPMorgan, Goldman Sachs restrict Anthropic AI use in Hong Kong amid US crackdown](https://www.storyboard18.com/digital/jpmorgan-goldman-sachs-restrict-anthropic-ai-use-in-hong-kong-amid-us-crackdown-101657.htm)
- [NDTV — After Goldman Sachs, JPMorgan Chase Blocks Anthropic Access For Hong Kong Employees](https://www.ndtv.com/feature/after-goldman-sachs-jpmorgan-chase-blocks-anthropic-access-for-hong-kong-employees-11658054)
- [The Banker — JPMorgan blocks Anthropic access for Hong Kong staff](https://www.thebanker.com/content/25c65f6e-7d60-4974-ba5f-78eeea847107)
- [The Economic Times — US banking giants restrict Anthropic AI access abroad](https://m.economictimes.com/news/new-updates/do-not-use-us-banking-giants-jpmorgan-and-goldman-sachs-restrict-anthropic-ai-access-abroad-heres-why/amp_articleshow/131836251.cms)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-06-18-wall-street-claude-blackout-export-control/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt