Anthropic has quietly shut the door on capital from mainland China and Hong Kong ahead of its highly anticipated stock market debut, a move that underscores the deepening fracture between the world’s two largest economies over the control of frontier artificial intelligence. The San Francisco-based startup, valued at over $18 billion in its last private round, has informed prospective backers that entities with beneficial ownership ties to the People’s Republic will not be allocated shares in the pre-IPO financing window, according to people familiar with the matter.
The decision, while not entirely unexpected given the current regulatory climate, marks a significant escalation in the self-policing of critical technology infrastructure. It signals that the most coveted AI companies are now treating access to their cap tables as a matter of national security compliance, not merely commercial preference. For Wall Street, it raises a thorny question: in the race to monetise generative intelligence, is the cost of capital about to bifurcate along geopolitical lines?
The Silent Screening Process
Sources close to the transaction describe a rigorous “know your customer” protocol that goes well beyond standard Office of Foreign Assets Control checks. Anthropic’s legal and compliance teams, working in concert with lead underwriters, have reportedly constructed a waterfall structure that effectively filters out limited partners whose ultimate beneficial owners reside in, or are controlled by entities in, restricted jurisdictions. The screen applies not only to direct investments but also to feeder funds and special purpose vehicles designed to obscure provenance.
One banker involved in the syndication described the mandate as “unusually prescriptive.” Typically, issuers rely on broad representations and warranties in subscription agreements. Here, the company has demanded granular transparency on LP composition down to the natural-person level for any vehicle with exposure to Greater China. Several prominent Asia-focused funds, including at least two with significant Hong Kong-domiciled limited partners, were turned away during the book-building phase despite expressing willingness to accept side letters restricting information rights.
The message is unambiguous: Anthropic is prioritising regulatory certainty with Washington over the deep liquidity pools of Shenzhen and Singapore. It is a calculation that reflects the new reality for hard-tech unicorns. The Committee on Foreign Investment in the United States (CFIUS) has expanded its jurisdiction aggressively under the latest executive orders, and the spectre of forced divestment — or worse, a blocked offering — is a risk no underwriter wants to price.
A Precedent Set by the Chips Act
The architecture for this exclusion was laid down long before Anthropic filed its confidential S-1. The CHIPS and Science Act, coupled with the October 2022 export controls on advanced semiconductors, established a doctrine of “small yard, high fence” that now governs the flow of both silicon and capital into American AI labs. While those rules technically target the transfer of technology, the enforcement mechanism has increasingly focused on the ownership layer.

Anthropic’s primary rival, OpenAI, has maintained a similarly restrictive cap table since its “capped-profit” restructuring, though it has never pursued a traditional IPO pathway. Databricks, another likely candidate for a 2025 listing, has quietly rebuffed secondary interest from sovereign-linked vehicles in the Gulf and East Asia for over a year. The pattern suggests an emerging industry standard: if you build models that the Pentagon classifies as dual-use, your shareholder register becomes a compliance artefact.
This is not merely performative. Anthropic’s Claude models are already integrated into workflows at several US defence contractors and intelligence agencies. The company holds active contracts under the Department of Defence’s Tradewinds procurement vehicle. In that context, a Chinese limited partner on the cap table — even a passive one — creates a vector for foreign influence that CFIUS is structurally designed to interrogate.
The Secondary Market Workaround
Predictably, capital finds a way. Within days of the allocation notices, a grey market for Anthropic exposure has begun to form in Singapore and Geneva. Structured products offering synthetic beta to the company’s valuation — total return swaps, funded participation notes, and collateralised forward contracts — are being marketed to family offices and wealth managers who cannot clear the primary screen. Pricing on these instruments implies a 15 to 20 per cent illiquidity premium over the last primary round, a steep but not prohibitive hurdle for investors desperate for pure-play AI exposure.
This shadow market introduces its own risks. Counterparty creditworthiness becomes paramount when the underlying asset is unlisted and the derivative is bespoke. Moreover, the SEC has signalled increased scrutiny of “acting in concert” arrangements that effectively circumvent registration requirements. A enforcement action against a major broker-dealer for facilitating such structures would chill the secondary trade instantly.
For the company, the workaround is a double-edged sword. It preserves valuation momentum — crucial for employee retention and acquisition currency — but it dilutes the signalling value of a clean, blue-chip shareholder base. When the roadshow begins in earnest, institutional investors will ask pointed questions about the composition of the pre-IPO register. A cap table riddled with opaque synthetic holders could complicate the narrative of governance maturity that public markets demand.
The Long Game: Sovereign AI and the Fragmentation of Compute
Zooming out, Anthropic’s move is a single data point in a broader fragmentation of the global AI stack. Nations are increasingly treating foundation models as strategic infrastructure akin to telecommunications or energy grids. The UAE has poured billions into the Falcon series via the Technology Innovation Institute. Saudi Arabia’s Public Investment Fund backs a sovereign LLM initiative through the Saudi Data and AI Authority. China, cut off from H100s and now from equity in the leading Western labs, is doubling down on domestic alternatives — Zhipu, Moonshot, Baichuan — funded by state-guided capital and local cloud giants.

This splintering has profound implications for the economics of model training. The scaling laws that drive performance demand massive, contiguous compute clusters and the capital to sustain them. If the capital markets bifurcate, the compute supply chain follows. Nvidia’s allocation logic already weighs “strategic alignment” alongside creditworthiness when doling out Blackwell slots. A world where Chinese capital cannot fund Anthropic, and American capital cannot fund Zhipu, is a world where the marginal dollar of AI investment yields less collective intelligence than it would in a frictionless market.
Wall Street, ever the arbiter of efficiency, will eventually price this friction. The question is whether the premium for “geopolitically clean” AI equity becomes a permanent feature of the asset class or a transient anomaly of the current tensions. History suggests the former. The semiconductor industry never fully re-integrated after the Cold War export controls; the AI industry may prove no different.
Why it Matters
Anthropic’s exclusion of Chinese capital is not just a compliance checkbox — it is a declaration that the commanding heights of artificial intelligence are now contested terrain where ownership equals allegiance. By choosing regulatory certainty over the deepest pools of global liquidity, the company has effectively priced in a future where access to Western AI infrastructure is a privilege granted by Washington, not a product bought on open markets. For investors, the message is clear: the alpha in AI will increasingly accrue to those who can navigate the new geography of permission, not merely those who can pick the best models.
