Introduction: When Investment Becomes a National Security Question
On October 7, 2026, the U.S. Department of the Treasury disclosed an enforcement action that, on first reading, seemed almost too small to matter and, on second reading, revealed how profoundly the relationship between American capital and foreign artificial intelligence development had changed. In April 2025, a Chinese investment fund controlled by Amidi, LLC—the parent of the organization that operates as Plug and Play Tech Center, one of Silicon Valley’s best-known accelerator networks—had invested approximately $92,478 in Shanghai Qiongche Intelligent Technology Company Limited, a private company that does business as Noematrix and that develops artificial intelligence, robotics, and what the industry now calls embodied intelligence. The sum was trivial by the standards of a venture ecosystem in which a single frontier-model financing round can exceed one hundred billion dollars. Yet Treasury imposed a $200,000 civil penalty, more than twice the value of the investment itself, because Amidi had failed to submit a notification that the Outbound Investment Security Program required of it. The government did not say the investment was prohibited; it said the investment was reportable and had not been reported. [1]
The details that emerged in the weeks afterward made the case more instructive rather than less. Treasury had identified the transaction not through a voluntary disclosure but through its own compliance and market-monitoring efforts, and it had actually assessed the penalty in July 2026 before announcing it publicly in October. [1] Press reporting traced a longer relationship between Plug and Play’s China operation and Noematrix, including participation in financing rounds in September 2024, December 2024, and April 2025, and noted that Noematrix’s co-founder is a professor and vice dean at Shanghai Jiao Tong University’s School of Artificial Intelligence, with subsequent rounds in 2026 drawing in a university-linked fund and a state-owned data enterprise. [16] Law firms advising private fund sponsors immediately observed that Treasury had disclosed very little about what diligence had been performed, what aggravating or mitigating factors were weighed, or how the penalty was calculated, which left investors to infer the government’s expectations from the bare fact that a sub-$100,000 cheque written by a controlled foreign subsidiary had produced the first civil penalty in the program’s history. [14] Secretary of the Treasury Scott Bessent framed the matter in unmistakably strategic language, describing the action as one that
“underscores Treasury’s commitment to safeguarding U.S. national security” — Scott Bessent, Secretary of the Treasury [1]
Fifteen days later, on October 22, 2026, Treasury issued a Notice of Proposed Rulemaking to implement Subtitle C of the Comprehensive Outbound Investment National Security Act of 2025, the COINS Act, which Congress had enacted on December 18, 2025, as Title LXXXV of the Fiscal Year 2026 National Defense Authorization Act and had grafted onto the Defense Production Act of 1950, the same statutory home that houses the Committee on Foreign Investment in the United States. [2][29][31] The proposed rule would expand the outbound investment framework well beyond its original China-centered focus on semiconductors, quantum information technologies, and artificial intelligence. It would add high-performance computing and supercomputing, hypersonic systems, and—at Treasury’s own discretion, since Congress had not named the sector—biotechnology; it would extend “country of concern” status to Cuba, Iran, North Korea, and Russia in addition to China, Hong Kong, and Macau; and it would convert the development of what the rule now calls “super intelligence systems” from a largely notifiable category into a prohibited one, on the reasoning that training compute is no longer a reliable proxy for capability. [2][6][7][9] The proposal was published in the Federal Register on October 26, 2026; comments are due December 7, 2026; the statute requires a final rule by roughly mid-March 2027; and until that final rule takes effect, the regulations in force since January 2, 2025 continue to govern. [3][5][10]
I have chosen the title Capital End-Use because the next phase of American technological competition cannot be understood through semiconductor export controls alone. For the better part of a decade, Washington’s instinct has been to regulate the movement of physical and intangible technology: the GPUs, the lithography tools, the electronic design automation software, the technical data. Export controls ask, in essence, where a thing is going and what it will be used for. But financial investment supports technological development through channels that never require a single accelerator to cross a border. Venture capital, growth equity, private equity buyouts, strategic minority stakes, joint ventures, and convertible financing deliver not merely money but managerial expertise, recruiting networks, board seats, commercial introductions, and the imprimatur of credible institutional backing. The term Capital End-Use therefore poses a question that is the financial analogue of the export-control question: where does investment ultimately flow, and what technological and strategic capabilities does it help to create?
The title is especially apt for the Five-Layer AI Economy that organizes this paper’s analytical structure. Capital moves across Energy, Chips, Datacenters, Models, and Applications and Agents, but the strategic consequences of investment differ profoundly among those layers. Financing a gas-fired power plant that happens to serve a computing campus is not equivalent to financing a high-bandwidth-memory fabrication line, an autonomous-weapons research program, or a frontier model developer. Their ownership structures, technological functions, and national-security implications must be examined separately, and the regulatory perimeter that is now being drawn around them must be understood as layer-specific rather than uniform. The argument of this paper is that, by 2027–2030, the intersection of technological capability, financial participation, and geopolitical jurisdiction could become a defining feature of American industrial policy and of international investment governance more broadly, with consequences that extend from Sand Hill Road to Abu Dhabi, from Hsinchu to Shanghai, and from the Federal Register to the limited-partnership agreements that govern trillions of dollars of pooled capital.
Central Thesis
Semiconductor export controls seek to regulate access to sensitive technology. Outbound investment controls seek to regulate particular financial transactions that could support the development of sensitive technological capabilities by foreign adversaries. The two instruments address different but interconnected dimensions of national security, and the second is now being built on the scaffolding of the first. A country may restrict advanced GPUs from reaching a foreign AI developer while still needing to consider whether its investors are helping to finance that developer’s models, engineering talent, commercial expansion, or related infrastructure. I propose Capital End-Use as an economic and geopolitical framework for evaluating not simply the destination of investment, but the ultimate technological capabilities, institutional advantages, and strategic consequences it may generate.
The paper’s central question is therefore the following. As artificial intelligence becomes the organizing technology of the global economy, can the United States distinguish productive international capital formation from investments that materially strengthen the sensitive technological capabilities of strategic adversaries—without undermining its own innovation ecosystem, its financial competitiveness, and the allied partnerships on which its technological leadership depends?
The paper proceeds in six sections. Section 1 traces the evolution of American technology-security policy from hardware controls to investment governance and explains the respective roles of the Commerce and Treasury Departments. Section 2 examines the financial architecture of frontier AI—venture capital, private equity, convertible instruments, and strategic acquisitions—and shows why different instruments create different national-security exposures. Section 3 applies the Five-Layer AI Economy to distinguish the strategic consequences of investment across energy, chips, datacenters, models, and agents. Section 4 evaluates the international consequences for China, the financial hubs of Hong Kong and Singapore, the allied semiconductor economies of Taiwan, Japan, and South Korea, and the sovereign wealth funds of the Gulf. Section 5 proposes a practical Capital End-Use Assessment for corporations and policymakers and sets out three scenarios through 2030. Section 6 distills the argument into seven pillars, and the conclusion returns to the question of what it means for the destination of money to matter less than the capabilities that money creates.
Section 1: From Controlling Hardware to Scrutinizing Investment — The Expansion of American Technology Security Policy
The modern American system for protecting sensitive technology was built, layer upon layer, around the physical object. The Export Administration Regulations classify items by Export Control Classification Number; the Entity List names the parties who may not receive them; the foreign direct product rule reaches goods made abroad with American tools. When the Biden Administration launched its sweeping October 2022 controls on advanced computing and semiconductor manufacturing equipment, the architecture was still fundamentally one of chokepoints: identify the irreplaceable tool, the irreplaceable chip, the irreplaceable software, and deny it. Chris Miller, the Tufts historian whose book Chip War was published days before those controls, has made the point repeatedly that the strategy worked because even China’s most advanced fabrication facilities depend on tools imported from the United States, Japan, and the Netherlands before the controls took hold, and that this dependence remains the single most important bottleneck in Beijing’s ambitions. [45] Speaking in Singapore on October 22, 2026—the same day Treasury released its outbound proposal—Miller estimated that China may need five years or longer to catch up in advanced chips precisely because access to cutting-edge chipmaking equipment remains constrained. [48]
Yet the same scholarship that explains why hardware controls matter also explains their limits. Miller himself has described even well-designed export controls as instruments that can only
“throw sand in the gears” — Chris Miller, Fletcher School, Tufts University [46]
of a determined competitor’s technological program. The gears still turn. And what keeps them turning, in a world where the physical chokepoints are increasingly well guarded, is capital: the money that funds the domestic substitute, hires the engineer, builds the alternative supply chain, and sustains a loss-making national champion until it reaches scale. Washington’s turn toward investment governance is the recognition that denying a tool while financing the firm that is trying to build its replacement is an incoherent strategy. The purpose of this section is to explain how that recognition hardened into statute, which agencies are responsible for which pieces of the resulting system, and why the newest piece—the COINS Act and its implementing rule—represents something qualitatively different from what came before.
1.1 The Evolution of American Technology Restrictions: From Export Controls to Outbound Investment Governance
The intellectual case for outbound investment screening predates its legal form by several years. In February 2023, Georgetown’s Center for Security and Emerging Technology published a detailed empirical study documenting that U.S. investors had participated in transactions worth $40.2 billion into 251 Chinese AI companies over the period studied, representing roughly 37 percent of the $110 billion raised by all Chinese AI companies, and that the majority of those transactions occurred at venture stages such as angel, seed, and pre-seed—precisely the stages at which capital carries the greatest intangible benefit in validation, network, and expertise. [28] The CSET authors did not argue for a blanket prohibition; they argued for clearly scoped objectives and improved disclosure, and they posed a question that still animates the debate: whether the United States was worried about China gaining a first-mover advantage in particular fields, or about something broader, and how the answer should shape the design of any review regime. [28]
The legal form arrived in three stages. Executive Order 14105 of August 9, 2023 declared a national emergency with respect to the advancement by countries of concern of sensitive technologies critical to military, intelligence, surveillance, and cyber-enabled capabilities, and directed Treasury to write regulations. Treasury’s final rule, published October 28, 2024 and effective January 2, 2025, created the Outbound Investment Security Program at 31 C.F.R. Part 850, prohibiting certain investments and requiring notification of others in Chinese entities engaged in advanced semiconductors and microelectronics, quantum information technologies, and specified AI systems, with AI notifications triggered either by end-use or by a technical threshold of training compute exceeding 10^23 operations. [32] The third stage was congressional: after several years in which bills such as the Senate’s FIGHT China Act competed for position, the House and Senate agreed on the COINS Act as part of the FY2026 NDAA, and President Trump signed it on December 18, 2025. [29][31] The statute codified the executive program, expanded it, instructed Treasury to issue or update regulations within 450 days, authorized $150 million annually for fiscal years 2026 and 2027 to build the program’s staff, and included a seven-year sunset. [30][31]
The October 22, 2026 proposal is the fourth stage, and it is the stage at which the regime acquires the full shape Congress intended. The Congressional Research Service’s March 2026 summary captured the political consensus behind it: broad support for the program’s tailored approach combined with persistent worry that foreign firms would simply fill any market share American investors vacated. [27] The proposed rule attempts to answer the first concern with precision and the second with reach. It would replace the current regulations with a new Part 850, carry forward most existing exceptions, add the exceptions Congress enacted, introduce the six technology sectors the statute contemplates, and extend the geographic perimeter. [5][8][10] It would also, in two important respects, go beyond what the statute required: by adding biotechnology as a prohibited sector on Treasury’s own authority to designate technologies that enable military, intelligence, surveillance, or cyber capabilities, and by treating the development of any super intelligence system as prohibited rather than merely notifiable. [2][6]
One distinction deserves emphasis because it has already produced confusion in commentary: the proposal was announced by Treasury on October 22, 2026 and published in the Federal Register on October 26, 2026, and the comment deadline of December 7 runs from the latter. [9][10] The announcement date governs the public narrative; the publication date governs the administrative record.
1.2 Two Federal Agencies, Two Strategic Responsibilities: Commerce and Treasury
The American system now has two regulators of technology security whose jurisdictions touch but do not overlap, and the difference between them is the difference between an object and a transaction.
The Department of Commerce, through the Bureau of Industry and Security, regulates exports, reexports, and in-country transfers of controlled items: advanced computing chips, semiconductor manufacturing equipment, software, and technology in the sense of technical data. Its tools are the licensing requirement, the presumption of denial, the Entity List, and the civil and criminal penalties available under the Export Control Reform Act. Its central question is whether a particular item may go to a particular party for a particular use. The Department of the Treasury, through the Office of Investment Security that also staffs CFIUS, regulates covered national security transactions by U.S. persons: equity acquisitions, certain debt that converts or confers rights, greenfield investments, joint ventures, and limited-partner commitments. Its tools are prohibition, mandatory notification, and civil penalties, and its central question is whether a U.S. person may participate financially in an entity engaged in a covered activity. Pillsbury partner Matthew Rabinowitz described the COINS regime to Military.com as effectively the inverse of CFIUS: where CFIUS reviews foreign capital entering American companies, the outbound program reviews American capital entering foreign ones. [68]
The two regimes are complementary precisely because they fail in different ways. Export controls cannot reach a Chinese robotics startup that uses no controlled American components; investment controls can. Investment controls cannot stop a shipment of H200 accelerators through a Malaysian intermediary; export controls can, or at least can try. The January 2026 episode illustrates how fluid the first regime has become and, by contrast, how durable the second may prove. On January 13, 2026, BIS announced a final rule—published and effective January 15—revising its license review policy for exports to China and Macau of NVIDIA H200, AMD MI325X, and comparable chips with total processing performance below 21,000 and DRAM bandwidth below 6,500 gigabytes per second, from a presumption of denial to case-by-case review, subject to third-party testing in the United States, enhanced know-your-customer certifications, and, under a parallel Section 232 action, a 25 percent duty. [20][21] Under Secretary for Industry and Security Jeffrey Kessler explained the logic in a single sentence:
“Export controls should evolve with changes in technology, while protecting national security.” — Jeffrey Kessler, Under Secretary of Commerce for Industry and Security [20]
The sequel was instructive. Chinese customs reportedly halted H200 imports and domestic firms were warned against purchasing them unless necessary, so that an American license did not translate into a sale; NVIDIA’s second-quarter fiscal 2027 results reported that shipments of Hopper products to China were less than one percent of data-center revenue, and the company’s third-quarter outlook assumed no data-center compute revenue from China at all. [22][33][34] Meanwhile, on May 31, 2026, BIS issued guidance clarifying that a license continues to be required for advanced computing items destined for any entity headquartered in Country Group D:5 or Macau, or whose ultimate parent is headquartered there, even when the receiving entity sits in Singapore, the United Arab Emirates, or Malaysia—a headquarters-based test that reaches through geography to ownership in much the way the outbound rules reach through intermediaries to the covered foreign person. [23][24] The export regime, in other words, has been oscillating between loosening at the product level and tightening at the entity level, which is exactly the environment in which a stable, statutory investment regime becomes attractive to policymakers who want a rule that does not depend on the next licensing decision.
1.3 The COINS Act and the Emergence of Financial National Security
The COINS Act did three things that an executive order could not. It gave the program a statutory basis that will survive a change of administration; it fixed the geographic and technological perimeter in law while leaving Treasury discretion to expand it; and it created compliance tools—a confidential non-binding feedback process and the authority to publish a non-exhaustive database of covered foreign persons—that transform the program from a pure prohibition into something closer to an advisory regime. [9][30][31]
The structure of the statute matters. Subtitle C, which Treasury’s October 2026 proposal implements, is the outbound investment subtitle proper; a separate Subtitle B creates parallel authorities, including sanctions, directed at covered foreign persons engaged in defense or surveillance sectors of a country of concern. [29][73] Within Subtitle C, the statute distinguishes notifiable technologies, in which a transaction is permitted but must be reported, from prohibited technologies, in which the transaction is barred outright, and it defines “covered foreign person” as a foreign person that is a person of a country of concern engaged in a covered activity, or an entity that holds specified interests in or derives specified revenue from such a person. [31][76] The proposed rule retains the existing concept of “knowingly directing,” under which a U.S. person who directs a transaction by a non-U.S. entity that would be prohibited if undertaken directly is itself in violation, and the statute extends that concept to notifiable transactions as well. [77]
The technological expansion is the heart of the matter. The statute names semiconductors and microelectronics, artificial intelligence systems, quantum information technologies, high-performance computing and supercomputing, and hypersonic systems; the proposed rule adds biotechnology, including certain pharmaceutical compounds, which multiple law firms immediately flagged as a discretionary expansion that Congress had not required. [2][6][8] The geographic expansion is nearly as consequential in principle and far less so in practice: because comprehensive sanctions already bar most new investment in Cuba, Iran, North Korea, and Russia, the addition of those jurisdictions changes little on the ground, and Treasury’s proposal notes that Venezuela—which Congress had included “under the regime of Nicolás Maduro”—is omitted because that regime no longer exists. [9][30] China, Hong Kong, and Macau remain the program’s gravitational center.
What is new, and what justifies the phrase “financial national security,” is the attention to the structure of capital itself. The proposed rule would continue to treat a pooled investment fund’s investment adviser as the fund’s parent, so that a non-U.S. fund advised from New York is a controlled foreign entity for whose transactions the adviser is responsible. [10] Its preamble states that a U.S. person’s commitment to a venture, private equity, or fund-of-funds vehicle that already holds a covered foreign person may constitute an indirect investment in that person regardless of whether the limited-partner commitment is itself covered—an observation that Ropes & Gray noted appears in the preamble rather than the rule text, which is precisely the sort of ambiguity that will dominate the comment period. [6] And it adopts a knowledge standard under which a U.S. person is treated as knowing a fact if it has actual knowledge, awareness of a high probability, or reason to know, assessed against whether the person conducted a reasonable and diligent inquiry, including whether it considered information Treasury had provided through its new feedback process. [4][9]
1.4 The White House’s Strategic Investment Doctrine
The administration’s investment doctrine was set out in the America First Investment Policy, a National Security Presidential Memorandum of February 21, 2025. The memorandum encouraged inbound investment from allies through a proposed fast track, directed CFIUS to tighten scrutiny of investments with Chinese touchpoints, and announced that the administration would consider new or expanded restrictions on U.S. outbound investment to China in semiconductors, artificial intelligence, quantum, biotechnology, hypersonics, aerospace, advanced manufacturing, directed energy, and other areas that support China’s military-civil fusion strategy, including the possible application of restrictions to publicly traded securities and the use of sanctions under the International Emergency Economic Powers Act. [25][26] The memorandum’s governing maxim—that economic security is national security—was not new, but its explicit linkage of outbound restriction to the sectors later named in the COINS Act and in Treasury’s proposal makes the October 2026 rule the administrative fulfillment of a stated presidential priority. [10][26] Treasury’s October 22 announcement was careful to pair the restrictive message with a liberal one, asserting that implementation of the COINS Act is consistent with
“maintaining a strong, open investment environment” — U.S. Department of the Treasury, Press Release, October 22, 2026 [2]
The doctrine sits alongside two other strands of policy. The first is industrial: the CHIPS and Science Act incentives, the 25 percent investment tax credit for qualified semiconductor capital expenditures, and the export-promotion logic behind the January 2026 H200 decision, which the administration framed as a way to strengthen the American technology ecosystem by keeping foreign customers on American platforms. [20][67] The second is terminological, and it would be a mistake to treat it as trivial. On September 29, 2026, President Trump signed Executive Order 14434, “Inaugurating the Era of Super Intelligence,” which directs executive departments and agencies, to the maximum extent permitted by law, to use the terms “Super Intelligence” and “SI” in place of “Artificial Intelligence” and “AI” in official correspondence, websites, reports, and other non-statutory materials, and which requires the Assistant to the President for Science and Technology to propose legislative language for a federal definition within sixty days. [17][19] The order changes no statute, imposes no obligation on private parties, and carries no penalty; it currently defines super intelligence by reference to the existing statutory definition of artificial intelligence at 15 U.S.C. § 9401. [18] But because the Treasury proposal was drafted after the order, it speaks throughout of “SI systems,” and the Fact Sheet accompanying the NPRM opens with a footnote explaining that usage. [4] Readers of the rule should understand that “super intelligence system,” in this regulatory context, does not mean a system possessing superhuman intelligence; it means what the regulation defines it to mean, which is a broad category of AI systems whose development Treasury proposes to treat as prohibited because no single benchmark reliably distinguishes advanced models. [6][7]
1.5 Investment as a Vehicle for Technology and Institutional Capability
The final conceptual step is to explain why money, as such, is a national-security concern. The answer is that investment is never only money. A Series A led by a credible firm signals to subsequent investors, recruits, and customers that a company is worth backing; a board seat brings governance experience and introductions; a corporate venture investment often arrives bundled with technology partnerships and distribution; a limited-partner commitment from a prestigious American endowment can anchor a fund that then deploys capital across dozens of portfolio companies the endowment will never diligence. Holland & Knight’s January 2025 summary of the original rule described its target as investments that contribute “capital as well as intangible benefits” to Chinese companies engaged in covered activities, and the intangible benefits are the point. [32] The Amidi case makes this concrete: the Chinese fund’s $92,478 was never going to decide Noematrix’s fate, but Plug and Play’s repeated participation across multiple rounds, alongside Sequoia China and later a Shanghai Jiao Tong University fund and a state-owned data group, placed an internationally recognized Silicon Valley brand on the capitalization table of a company whose work on embodied intelligence has obvious dual-use implications. [16]
This leads to the proposition that organizes the rest of the paper. The economic impact of an investment should be assessed through both its financial value and the capabilities it helps to establish, and the second measure will often exceed the first by orders of magnitude. A $10 million investment that accompanies a technology license, a hiring pipeline, and a path to international markets can be more strategically consequential than a $1 billion passive position in a listed index. Treasury’s exceptions for publicly traded securities and for certain limited-partner interests reflect an implicit acceptance of this view: the regime targets the transactions through which capability is transferred, not the transactions through which returns are merely earned. [8][31]
Section 1’s central argument can be stated plainly. The emerging regulatory perimeter extends beyond controlling the technologies that cross borders to examining certain investments that can accelerate sensitive technological development abroad. Export controls and investment controls are now two halves of a single strategy, and the second half is being written in the Federal Register at this moment.
Section 2: The Foreign Investment Architecture of Frontier Artificial Intelligence — Venture Capital, Private Equity, and Strategic Ownership
If Section 1 explained why Washington has come to regard investment as a national-security instrument, Section 2 explains what, precisely, is being regulated. The distinction is important because the public conversation about “American money in Chinese AI” tends to collapse an extraordinarily varied set of financial relationships into a single image of a wire transfer. The reality is a layered architecture in which a university endowment in New England commits to a fund-of-funds in Delaware, which commits to a growth fund domiciled in the Cayman Islands and advised from Hong Kong, which purchases preferred shares in a Shanghai company through a variable-interest-entity structure, which in turn holds a convertible note in a Shenzhen robotics subsidiary. Each link in that chain is a different legal instrument with different rights, different disclosure obligations, and different national-security implications. Treasury’s proposed rule attempts to regulate the chain without strangling it, and whether it succeeds will depend on how well policymakers and investors understand the instruments themselves.
The scale of what is at stake has grown enormously since the 2023 CSET study. Stanford’s 2026 AI Index reports that U.S. private AI investment reached $285.9 billion in 2025, more than twenty-three times China’s $12.4 billion, within a global corporate AI investment total of $581.7 billion that more than doubled year over year—while cautioning that the private figures
“likely understate” — Stanford Institute for Human-Centered Artificial Intelligence, 2026 AI Index Report [38]
China’s actual spending because government guidance funds and state-initiated vehicles sit outside private-capital databases. [38][39] The OECD, measuring the outbound direction of flows, found that U.S. venture investors accounted for roughly 56 percent, or about $124 billion, of the worldwide value of outgoing venture investment in AI in 2025, followed by investors in the United Kingdom at 9 percent, China at 8 percent, and the EU27 at 7 percent. [40] In the first half of 2026, two companies alone—OpenAI and Anthropic—raised a combined $217 billion, which was 43 percent of all global startup investment in that period. [64] These figures describe a market in which American capital is simultaneously dominant and concentrated, and in which the marginal cross-border transaction that reaches a covered foreign person is a tiny fraction of total activity. The regulatory design problem is to find that fraction without imposing costs on the whole.
2.1 Venture Capital and the Financing of Foreign AI Startups
Venture capital is the instrument most directly implicated by the outbound rules, and for good reason: it is the stage at which money is most tightly bundled with capability. A seed or Series A investor typically acquires preferred equity with information rights, pro-rata rights, protective provisions, and frequently a board seat or observer seat. Beyond the term sheet, early-stage investors provide what the industry calls platform services: recruiting, customer introductions, follow-on fundraising, and access to a network of portfolio companies and corporate partners. The 2023 CSET finding that most U.S. participation in Chinese AI came at angel, seed, and pre-seed stages is therefore more worrying from a capability-transfer standpoint than a comparable dollar figure at a late growth stage would be. [28] The categories of Chinese company most likely to attract such investment—semiconductor design, frontier and open-weight model developers, robotics and embodied intelligence, autonomous systems, and advanced computing—map almost exactly onto the covered activities in the existing and proposed rules.
Three features of venture investing complicate the rule’s application. The first is corporate venture capital. Many of the largest American technology companies operate venture arms that invest in foreign startups for strategic rather than purely financial reasons, often alongside commercial agreements; under the proposed rule, the investing subsidiary is a U.S. person and the commercial agreement may itself bear on whether the target is a covered foreign person. The second is the accelerator model, of which Plug and Play is the archetype: an organization that takes small stakes in very large numbers of companies, frequently through locally domiciled funds, and whose value proposition is precisely the network and the brand rather than the cheque. The Amidi penalty demonstrates that Treasury regards such structures as fully within scope and that the smallness of any individual position is no defense. [1][13] The third is the geographic question the paper’s outline poses: whether restricting selected foreign investments changes the geographic distribution of American venture capital. The early evidence suggests that it does, though less through legal compulsion than through the compliance economics discussed in Section 2.5—limited partners increasingly demand “China exposure” reporting from their general partners, and general partners increasingly conclude that the marginal Chinese deal is not worth the diligence burden, the reputational risk, and the possibility of a Treasury inquiry. [50]
2.2 Private Equity, Limited Partnerships, and the Complexity of Indirect Investment
Private equity and the broader world of pooled vehicles present the hardest tracing problems in the regime. The structure is familiar: limited partners commit capital to a fund managed by a general partner; the general partner’s affiliated adviser makes investment decisions; the fund may invest directly in operating companies or in other funds; and each layer may be domiciled in a different jurisdiction for tax, regulatory, or commercial reasons. A U.S. pension plan that commits to a fund-of-funds that commits to an Asia growth fund may have no practical knowledge of, and no contractual ability to prevent, an investment by that growth fund in a Shenzhen semiconductor firm.
The proposed rule addresses this architecture in several ways. It retains the existing exception for passive limited-partner investments below a specified size or percentage threshold, provided the limited partner does not exercise control or direction; it carries forward the exception for publicly traded securities; and it continues to treat the investment adviser as the parent of the funds it advises, which has the effect of making a U.S.-based adviser responsible for the covered transactions of the non-U.S. funds it manages. [8][10][31] The preamble’s observation that a limited-partner commitment to a fund with pre-existing covered holdings may constitute an indirect investment is the proposal’s most contested statement, because it implies a diligence obligation extending to a fund’s existing portfolio and not merely to its future deployments. [6] The concept of “knowingly directing,” which the statute extends to notifiable transactions, reaches the American executive who sits on the investment committee of a foreign fund and votes to approve a covered deal. [77]
The practical difficulty is one of proportion. Treating every diversified international allocation as equally sensitive would convert the program into a de facto prohibition on Asian private markets, which neither Congress nor Treasury intends. Distinguishing the passive allocation from the controlling or directing one requires reliable documentation of beneficial ownership, control, and investment direction at each layer—documentation that many foreign managers have historically been unwilling to provide and that the proposed knowledge standard now makes commercially essential. The compliance literature that emerged within days of the Amidi announcement emphasized that a list-based screen alone is insufficient, because Noematrix appeared on no sanctions list; it was a covered foreign person because of what it did, not because of who had designated it. [15]
2.3 Convertible Securities and Financial Instruments That Change Ownership
A recurring theme in the design of both the existing and the proposed rules is that the legal form of an instrument can matter as much as its economic value. Ordinary commercial lending—a term loan at market rates with no conversion feature and no governance rights—is excepted, because it transfers money without transferring the attributes of ownership. But convertible debt, contingent equity, warrants, structured notes with equity kickers, and financing arrangements that confer board rights, veto rights, or rights to intellectual property or technical information are covered, because they either are equity in substance or will become equity when conditions are met. [31][32] The proposed rule largely preserves these categories while clarifying that temporary ownership of an issuer’s shares in the ordinary course of underwriting does not by itself make the underwriting a covered transaction. [7]
The economic logic is sound. A venture lender that advances $20 million against warrants for 5 percent of a Chinese robotics company has acquired a contingent ownership interest and, in many structures, information and consent rights that exceed those of a passive equity holder. A SAFE or a convertible note in a seed-stage company is, for all practical purposes, an equity investment with a deferred price. The regime’s attention to these forms reflects a judgment that the national-security relevance of a transaction lies in the relationship it creates between investor and investee, and that relationships can be created through debt as easily as through stock. The corresponding compliance lesson is that treasury and legal functions at U.S. investors must evaluate the rights conferred by an instrument, and the circumstances under which those rights crystallize, rather than relying on its label.
2.4 Strategic Acquisitions, Joint Ventures, and Corporate Expansion
Strategic acquisitions and joint ventures sit at the opposite end of the spectrum from the passive limited-partner commitment: they involve control, operational integration, and the direct transfer of technical and managerial knowledge. An American industrial company that acquires a majority of a Chinese machine-vision firm, or forms a joint venture with a Chinese partner to develop autonomous warehouse robots, is not merely financing a covered activity; it is participating in it. The outbound rules reach acquisitions of equity, the establishment of joint ventures, and greenfield investments that result in a covered foreign person, and they interact with export controls, sanctions, antitrust review in multiple jurisdictions, and—increasingly—the target country’s own investment-security rules. [31]
That last interaction has become one of the most important developments of 2026. In April 2026, China’s National Development and Reform Commission blocked the acquisition of Manus, a Chinese-founded AI-agent startup that had re-registered in Singapore, by a major American technology company, in what Morgan Lewis described as the first security-review block of a cross-border AI transaction under China’s principles-based standard for “important information technology that may affect national security.” [50] On June 1, 2026, China’s State Council published a new Regulation on Overseas Investment that, as Stefanie Kam of the S. Rajaratnam School of International Studies has analyzed, elevates outbound-investment oversight from ministerial rules to administrative regulation, adds a national-security review, and restricts indirect technology and data transfer so as to block offshore transfers of sensitive Chinese technology. [51] The Foundation for Defense of Democracies noted that the new rules also allow Beijing to restrict engineers from leaving the country, a tool reportedly deployed when Manus’s co-founders were barred from departing. [52] The strategic implication, which Morgan Lewis drew out, is that an offshore restructuring—the traditional method by which a Chinese-origin company exits the perimeter of the U.S. outbound rules—may now be read by American regulators as evidence that the company has not in fact left the Chinese regulatory perimeter, since Beijing itself treats it as Chinese. [50] Capital End-Use, in other words, is now being asserted by both governments simultaneously, and a transaction can be caught in a pincer in which each side claims the capability for itself.
The outline for this paper called for hypothetical acquisition cases, and four will suffice to show how the analysis changes with the layer involved. In the first, an American logistics-automation company proposes to acquire a controlling stake in a Chinese robotics firm whose manipulators are used in both e-commerce fulfillment and defense-adjacent manufacturing; the transaction is a covered acquisition of a covered foreign person engaged in an AI end-use, and the acquirer’s export-control exposure for technical data flowing into the joint engineering team compounds the outbound-investment prohibition. In the second, an American memory-technology licensor proposes a joint venture with a Chinese packaging house to co-develop high-bandwidth-memory stacking; the joint venture is itself a covered transaction in semiconductors and microelectronics, and the licensor’s contribution of process know-how may independently require an export license. In the third, an American enterprise-software company proposes a minority investment in a Singapore-domiciled model developer whose research team and ultimate parent are in Hangzhou; under the proposed rule’s definition of covered foreign person and BIS’s headquarters-based test, the Singapore domicile is unlikely to be dispositive. [23][24] In the fourth, an American industrial conglomerate proposes to acquire a European factory-automation company that has a Chinese subsidiary developing autonomous inspection agents; the European acquisition is outside scope, but the acquirer inherits a controlled foreign entity whose future covered transactions become the acquirer’s responsibility to notify or prevent. [1]
2.5 Understanding the Boundaries of Regulation
The regime is best understood as three concentric rings. The innermost ring is prohibition: transactions involving prohibited technologies—advanced semiconductors above specified performance thresholds, supercomputers, quantum computers and certain quantum sensing and networking systems, AI systems designed for military, intelligence, or mass-surveillance end uses, and, under the proposal, any super intelligence system development, hypersonic systems, and specified biotechnology—may not be undertaken. The middle ring is notification: transactions involving notifiable technologies may proceed but must be reported to Treasury, with the information Congress and Treasury specify. The outer ring is exception: publicly traded securities, qualifying limited-partner interests, ordinary bank lending, specified financial services, transactions secondary or ancillary to a covered transaction, binding uncalled capital commitments made before the effective date, and a de minimis carve-out for transactions that do not exceed $50,000 in aggregate over the preceding twelve months. [5][7][8]
The de minimis provision deserves particular care, because it is easy to misread. It applies to certain notification obligations and does not operate as an unrestricted exemption from the prohibitions; a $40,000 investment in a prohibited-technology developer remains prohibited. The Amidi transaction, at $92,478, would have exceeded the proposed threshold in any event, and the penalty was assessed under the current rule, which has no comparable aggregate carve-out. [1][8] Investors should therefore read the de minimis exception as administrative relief for small notifiable positions, not as a safe harbor for small prohibited ones.
The proposal’s most important structural feature is the allocation of responsibility. By adopting a knowledge standard that includes “reason to know” and by defining reasonable and diligent inquiry with reference to Treasury’s own guidance and feedback, the rule places the burden of investigation squarely on the investor. [4][9] The new confidential feedback process is both a tool and a trap: it offers investors a way to reduce uncertainty before closing, but it also means that an investor who proceeds without seeking feedback on a borderline transaction may find that omission weighed against it in an enforcement proceeding. [9] The contemplated public database of covered foreign persons would help, but Treasury has been explicit that it would be non-exhaustive, and the Noematrix case shows that absence from any list is not evidence of absence from scope. [6][15]
The Amidi matter, finally, should be read as a case study rather than an anomaly. It involved a controlled foreign entity rather than a direct U.S. investment, which confirms that the obligation to notify and to take reasonable steps to prevent reaches through corporate structure. It involved a notification failure rather than a prohibited transaction, which confirms that Treasury will enforce the procedural ring and not only the substantive one. It was identified through market monitoring, which confirms that the program has investigative capacity. And it produced a penalty more than double the investment, which signals that the economics of non-compliance are unfavorable even for small positions. [1][11][12][13] The NYU Program on Corporate Compliance and Enforcement observed that Treasury’s announcement, unlike many sanctions enforcement releases, did not highlight specific compliance expectations for fund sponsors, leaving the industry to infer them. [14] The inference the industry has drawn is that an outbound-investment compliance program is now a cost of doing business for any U.S. investor with Asian exposure.
Section 2’s central argument follows directly. Investment security cannot be evaluated solely by the amount of money transferred. Ownership, intermediary structures, contractual rights, technological activities, and the investor’s knowledge determine how a transaction should be understood, and a regime that focuses on any one of those variables to the exclusion of the others will either miss the transactions that matter or burden the transactions that do not.
Section 3: Capital End-Use and the Five-Layer AI Economy — How Financial Restrictions Influence Energy, Chips, Datacenters, Models, and Agents
The regulatory analysis in Sections 1 and 2 was organized around instruments and institutions. This section reorganizes it around technology, because the strategic consequences of an investment depend on what the investment builds. The Five-Layer AI Economy is a way of seeing the artificial-intelligence industry as a vertically integrated stack in which electricity feeds chips, chips populate datacenters, datacenters train and serve models, and models animate applications and autonomous agents. Each layer has its own capital intensity, its own ownership patterns, its own geography, and its own relationship to the covered activities in the outbound rules. A policymaker who treats “AI investment” as a single category will either prohibit too much—sweeping in the gas turbine and the office-building lease—or too little, missing the subsidiary whose only product is a training run.
The scale of the stack in 2026 is without precedent in industrial history. The four largest American hyperscalers—Amazon, Alphabet, Microsoft, and Meta—reported second-quarter 2026 results in late July that pushed their combined calendar-2026 capital-expenditure guidance to roughly $730–760 billion, nearly double the approximately $410 billion they spent in 2025: Amazon raised its outlook to about $220 billion, Alphabet to $195–205 billion, Microsoft to roughly $175 billion on a calendar basis, and Meta to $130–145 billion, a range Meta raised again to $135–145 billion with its third-quarter results on October 21. [35][36][37] Alphabet disclosed that roughly 60 percent of its second-quarter spend went to servers and 40 percent to datacenters and networking, and Microsoft that about two-thirds went to short-lived assets such as CPUs and GPUs. [36] On the supply side, NVIDIA reported second-quarter fiscal 2027 revenue of $96.2 billion, up 106 percent from a year earlier, with datacenter revenue of $89.0 billion, up 117 percent, and guided to $108.0 billion for the third quarter. [33] Jensen Huang’s summary of the moment was four words long:
“Now, compute is revenue.” — Jensen Huang, Founder and Chief Executive Officer, NVIDIA [33]
Not every economist agrees that the revenue will follow the compute. Daron Acemoglu of MIT, who has argued for two years that the current generation of AI will materially affect only a small fraction of tasks over the coming decade, told Bloomberg bluntly that
“A lot of money is going to get wasted” — Daron Acemoglu, Institute Professor, Massachusetts Institute of Technology [43]
and has continued to warn, including in a February 2026 BBC interview, of the possibility of an investment bubble. [44] The International Monetary Fund’s October 2026 World Economic Outlook held global growth at 3.1 percent for 2026 while modeling a market-correction scenario in which the unwinding of AI-related investment reduces global GDP by as much as 0.4 percent and investment by as much as 1.3 percent by 2028, and its companion Global Financial Stability Report added an analysis of rising leverage in the AI ecosystem. [41] The same outlook observed that Taiwan and Korea source 62 percent of net AI computing-hardware trade and that China and the United States account for 57 percent of newly built datacenters since 2024. [41] The point for present purposes is not whether the boom will end well but that the capital flows it has unleashed are so large that even a small fraction misdirected to a covered foreign person would be strategically significant, and that the geography of the stack is already concentrated in exactly the jurisdictions the outbound regime is designed to police or to protect.
The framework below is illustrative. It identifies the primary Capital End-Use question at each layer; it does not assert that every investment in these industries is covered by Treasury’s regulations, and most are not.
| Layer | Primary capital end-use question | Typical investors | Regulatory touchpoints |
| 1. Energy | What facilities and computing capabilities will the financed power supply support? | Infrastructure funds, utilities, sovereign wealth, project finance | Generally outside outbound scope unless tied to a covered entity; CFIUS for inbound |
| 2. Chips | Does the investment enable sensitive semiconductor design, fabrication, packaging, or memory capability? | Strategic corporates, state funds, late-stage VC, PE | Prohibited above performance thresholds; notifiable below; export controls on tools and IP |
| 3. Datacenters | What advanced-computing activities and entities will the infrastructure support? | PE, REITs, hyperscalers, sovereign funds, infrastructure partnerships | Supercomputer and HPC prohibitions; BIS headquarters test for chip deliveries |
| 4. Models | What model-development capability will the investment enable? | Frontier VC, corporate strategic, sovereign funds | Proposed prohibition on SI-system development; military/surveillance end-use prohibitions |
| 5. Apps & Agents | Will the investment support sensitive autonomous, surveillance, military, or other covered applications? | Early-stage VC, accelerators, corporate VC | Notifiable AI end-uses; embodied-AI and robotics scrutiny (Amidi/Noematrix) |
3.1 Layer 1 — Energy: The Financial Foundations of Computational Expansion
Electricity is the most capital-intensive and the least regulated layer of the stack, and the combination is becoming uncomfortable. Hyperscale campuses now draw gigawatts; the Stargate UAE campus is planned at five gigawatts, and the IMF’s observation that two countries account for most new datacenter construction is in substance an observation about where generation and transmission are being built. [41][59] Investment in generation, transmission, nuclear restarts and small modular reactors, natural-gas peakers, and renewable portfolios is overwhelmingly undertaken by infrastructure funds, utilities, and sovereign investors whose interest in the computing load is purely as an anchor tenant.
The outbound rules do not reach ordinary energy investment, and they should not. A U.S. infrastructure fund that finances a combined-cycle plant in a third country is not engaged in a covered activity even if a downstream customer operates GPUs. But two qualifications are necessary. The first is that energy financing can indirectly enable covered development when it is structured as part of a bundled campus transaction in which the power, the facility, and the computing tenant are financed together, and when the tenant is a covered foreign person; the proposal’s treatment of transactions secondary or ancillary to a covered transaction is meant to exclude the incidental utility contract, not the integrated project. [7] The second is that the strategic importance of energy is rising faster than the regulatory attention paid to it. As compute demand scales, the ability to finance power-intensive industrial ecosystems becomes a competitive variable in its own right, which is why the Gulf’s hydrocarbon abundance has made it the preferred location for the largest campuses outside the United States and why state-level energy availability has become decisive in domestic site selection, a point developed in Section 3.6. [60][61]
3.2 Layer 2 — Chips: The Strategic Financing of Semiconductor Innovation
The semiconductor layer is where the outbound regime and the export-control regime are most tightly interwoven, and where the global competitive landscape is most legible. NVIDIA, AMD, and Intel design; TSMC, Samsung, and Intel fabricate at the leading edge; SK hynix, Samsung, and Micron supply the high-bandwidth memory without which an accelerator is inert; ASML, Applied Materials, Lam Research, and Tokyo Electron supply the tools. On the Chinese side, Huawei’s Ascend ecosystem, SMIC, CXMT, and a widening set of design houses are being financed by the third phase of the National Integrated Circuit Industry Investment Fund—the “Big Fund,” registered at 344 billion yuan—and by the 60-billion-yuan National AI Industry Investment Fund established in January 2025 within days of a tightening of American export controls. [53][54][78]
The outbound rules prohibit U.S. investment in Chinese entities developing or producing advanced logic, memory, and packaging above specified thresholds, and require notification below them; the proposed rule carries those categories forward and adds high-performance computing and supercomputing as a distinct sector. [8][31] The complementarity with export controls is direct: BIS denies the tool, Treasury denies the money that would fund the domestic replacement. The distinction that matters analytically is between supporting existing production and financing new capability. A passive position in a listed Chinese foundry is excepted; a growth investment in a startup developing EUV-adjacent patterning or HBM stacking is prohibited; and the difference lies not in the dollar amount but in whether the capital is building something that does not yet exist.
The question of whether access to international capital influences the timing, cost, and viability of domestic Chinese alternatives is contested among the most serious observers. Miller has argued, somewhat provocatively, that if one believed AI mattered and chips were its key ingredient, then
“China’s been underspending for the last four years on AI” — Chris Miller, Fletcher School, Tufts University [47]
and that Beijing is pushing firms toward Huawei’s domestic ecosystem for reasons of data security and distrust of foreign hardware as much as of ambition. [47] If that reading is correct, then the marginal American venture dollar was never the binding constraint on Chinese semiconductor progress; state capital and policy direction were. The outbound rules would then operate less as a brake on Chinese capability than as a prophylactic against the reputational and strategic embarrassment of American institutions financing the competitor, and as a signal to allies. If the alternative reading is correct—that early-stage Chinese design firms were materially dependent on the validation and network that American investors supplied—then the rules bite harder. The truth probably differs by sub-segment: state funds dominate fabrication, where capital requirements are enormous and returns are distant, while private and foreign capital has mattered more at the design and IP stages, which is exactly where the notification and prohibition categories are concentrated.
3.3 Layer 3 — Datacenters: When Infrastructure Financing Meets National Security
Datacenters are where the physical and the financial meet, and the ownership structures are more varied than at any other layer. Hyperscalers build and own campuses on their own balance sheets; real-estate investment trusts and infrastructure funds develop and lease them; project-finance structures carve off individual facilities; and sovereign investors have moved from financing to owning. The largest datacenter transaction of 2026, completed on July 21, was the $40 billion acquisition of Aligned Data Centers by a consortium of Abu Dhabi’s MGX, the BlackRock-led AI Infrastructure Partnership, and Global Infrastructure Partners—a deal that NYU’s Winston Ma has described as Gulf sovereign funds becoming direct owners of what Huang calls AI factories. [58][62]
The outbound rules reach the datacenter layer in two ways. The first is the supercomputer prohibition, which bars U.S. investment in Chinese entities developing or producing supercomputers above specified performance levels, and which the proposed rule reinforces with a high-performance-computing sector. [31] The second is indirect: BIS’s May 2026 guidance means that a datacenter operator headquartered in China, or whose ultimate parent is headquartered in China, requires a license for advanced computing items regardless of whether the facility is in Johor, Dubai, or Kuala Lumpur, which in turn means that a U.S. investor in such an operator is financing a facility whose chip supply is legally precarious. [23][24] The practical diligence question at this layer is therefore a chain: who owns the facility, who operates it, who procures the GPUs, who the cloud customers are, and where ultimate beneficial control resides. A facility in one jurisdiction with financial ownership in a second and operational decision-making in a third is now the norm rather than the exception, and the rules were written with that norm in mind.
The distinction this paper proposes for the layer is between investment in the physical computing facility and investment in the specific sensitive technologies it may host. A warehouse of racks is not a covered activity. A warehouse of racks owned by an entity that develops a covered AI system, or that is a covered foreign person because of its ownership, is. The regime does not ask whether the building is sensitive; it asks whether the owner is.
3.4 Layer 4 — Models: The Financial Architecture of Frontier Intelligence
The model layer is where the most money is concentrated in the fewest hands and where the proposed rule makes its most consequential change. The financing architecture of the frontier labs in 2026 is a hybrid of venture equity, corporate strategic investment, and cloud-compute agreements that function as capital in kind. OpenAI closed a $122 billion round on March 31, 2026 at an $852 billion post-money valuation, with anchor commitments from Amazon, NVIDIA, and SoftBank; Anthropic closed a $30 billion Series G at $380 billion in February and then a $65 billion Series H at $965 billion on May 28, led by Altimeter, Dragoneer, Greenoaks, and Sequoia, on a reported $47 billion revenue run rate; xAI raised $20 billion in January and was subsequently merged into SpaceX. [63][64] Abu Dhabi’s MGX holds stakes in OpenAI, Anthropic, and the SpaceX-xAI combination, and Saudi Arabia’s HUMAIN invested $3 billion in xAI earlier in the year with a joint datacenter in the Kingdom attached. [60] Google DeepMind and Meta finance their model programs internally, which is why Meta’s capital expenditure is best understood as model financing by another name. [37]
Chinese model developers are financed on an entirely different pattern, and the difference is itself a Capital End-Use observation. DeepSeek reportedly raised approximately 50 billion yuan in June 2026 at a valuation above 350 billion yuan from a group including its founder, Tencent, CATL, and the national AI fund; Moonshot AI raised $3.5 billion in July 2026 with the national AI fund among the leads. [78] The state is not a passive limited partner in these rounds; it is a strategic anchor, and that fact bears directly on whether an American co-investor could ever be regarded as merely passive.
The proposed rule’s treatment of this layer is the single most important change from the existing program. Under the current rule, many transactions involving a covered foreign person that develops an AI system are notifiable, with prohibition reserved for military, intelligence, and mass-surveillance end uses and for systems trained above very high compute thresholds. The proposal instead treats the development of any super intelligence system as a prohibited technology, on Treasury’s stated reasoning that no single measure, benchmark, or threshold reliably distinguishes advanced models and that training compute in particular is no longer a dependable proxy for capability. [6][7] Freshfields and Ropes & Gray both characterized the resulting definition as broad, and the comment period will be dominated by arguments over where the line between “SI system development” and ordinary machine-learning engineering falls. [6][7] The abandonment of the compute threshold is analytically correct—the DeepSeek episode of early 2025 demonstrated that algorithmic efficiency can substitute for scale, and Stanford’s 2026 AI Index reports that the performance gap between the best American and Chinese models had narrowed to 2.7 percent by March 2026 despite the twenty-three-fold investment disparity—but it transfers enormous definitional weight onto terms that the rule must now define with care. [38][39][70]
Two further observations complete the picture. First, the model layer is where export controls and investment controls converge most visibly, because the thing being financed is itself increasingly treated as a controlled item: in June 2026, Anthropic suspended access to its Fable 5 and Mythos 5 models within days of launch to comply with U.S. Department of Commerce export controls, restoring access on July 1 after the Department lifted them—an episode that showed model weights and model access entering the same regulatory vocabulary as accelerators. [72] Second, investors must distinguish three things that look similar on a term sheet: deploying a model, which is ordinary AI adoption and is not covered; developing a covered system, which may be prohibited; and financing an entity whose principal activity is the second while describing it as the first, which is where diligence failures will concentrate.
3.5 Layer 5 — Applications and Agents: Financing Autonomous Capabilities
The application layer is the broadest, the most commercially diverse, and—since October 7, 2026—the layer at which the regime has actually been enforced. Noematrix develops embodied intelligence: the software that allows robots to perceive, plan, and act in the physical world. Its June 2026 financing round, which drew a Shanghai Jiao Tong University AI fund, a research-institute subsidiary, and a state-owned data company, was explicitly described as supporting the use of AI-powered robots on production lines, and the company’s co-founder holds a senior academic post in one of China’s most important AI faculties. [16] Nothing about that profile is exotic. It describes hundreds of companies in China, and dozens in the United States, working on agentic software, industrial automation, autonomous logistics, cybersecurity agents, and AI-driven scientific discovery. The dual-use character of the layer is structural: a manipulation policy that stacks boxes can stack munitions, a planning agent that routes delivery trucks can route other vehicles, and an autonomous-inspection system developed for a refinery is a sensing-and-decision stack with obvious defense applications.
The outbound rules treat AI systems designed for military, intelligence, or mass-surveillance end uses as prohibited and a wider set of AI end uses as notifiable, and the proposed rule’s SI-system category would extend prohibition further into the development stack. [8][31] The Amidi penalty establishes that an embodied-AI developer is within the notifiable perimeter today, and that an accelerator’s small stake, taken through a Chinese fund, triggers the full set of obligations. [1][15] What makes the layer analytically distinctive is that the investment often buys capabilities that no chip or datacenter can supply: proprietary datasets of robot trajectories, engineering teams with rare skills, commercialization relationships with manufacturers, and intellectual property in control policies. A strategic acquisition at this layer can therefore transfer more capability per dollar than at any other, which is why the paper’s outline is right to insist that application-layer investments will become increasingly important even when the recipient manufactures no chips and operates no datacenter.
The case studies the outline requested can be stated briefly. A robotic-manufacturing investment is notifiable at minimum and prohibited if the manipulators are designed for a military end use. An autonomous-logistics investment is notifiable, and prohibited if the routing stack is designed for intelligence or surveillance. A cybersecurity-agent investment raises the sharpest question, because offensive and defensive capability are nearly indistinguishable at the code level and the cyber-enabled-capabilities language in the statute is explicit. An advanced-scientific-systems investment—an AI-for-materials or AI-for-biology platform—would have been outside scope under the existing rule unless the AI end-use criteria were met, but the proposed addition of biotechnology brings a substantial portion of that activity into the prohibited ring. [2][6]
3.6 The Domestic Economic Consequences: American States as Alternative Investment Destinations
The final application of the framework turns from the restriction of capital abroad to its redirection at home. Semiconductor firms have announced more than $230 billion in onshore investment since the 2022 CHIPS Act, with the largest shares going to Texas, New York, and Arizona; TSMC’s Arizona cluster, now projected at roughly $165 billion and ultimately ten fabs, two packaging facilities, and a research center, received up to $6.6 billion in CHIPS funding; Samsung’s Taylor, Texas complex received up to $6.4 billion; Micron has raised its U.S. commitment to more than $250 billion through 2035; and SK hynix broke ground on a $3.9–4 billion advanced-packaging and HBM facility at Purdue Research Park in West Lafayette, Indiana, which the Korean ambassador described as
“not merely an investment in semiconductor manufacturing facilities” — Kang Kyung-wha, Ambassador of the Republic of Korea to the United States [66]
but the first AI-memory production and research center in the country. [65][66][67] Intel’s Ohio One campus in New Albany and its Chandler, Arizona fabs, Micron’s Boise and Clay, New York megafabs, and the Texas Instruments Sherman megasite round out a map that did not exist five years ago. [65]
Governors, economic-development agencies, universities, and infrastructure developers are responding to the changed international capital environment in ways that connect directly to the national-security strategy. Arizona and Texas have competed on energy availability and permitting speed; Indiana and Ohio on university partnerships and workforce pipelines; Virginia on existing datacenter density and fiber; Michigan on automotive-adjacent robotics and automation; California on the proximity of the venture ecosystem to the frontier labs. The policy question the outline poses—whether export and investment restrictions create opportunities for domestic supply chains or simply cause foreign projects to seek capital from alternative sources—has an uncomfortable answer: both. The restrictions do redirect allied and domestic capital toward American projects, and the evidence of the fab map is unambiguous. They do not stop Chinese projects from finding capital, because the Chinese state has shown every willingness to supply it. What the restrictions accomplish, at the domestic level, is to align state-level industrial incentives, energy policy, and workforce development with a federal strategy whose premise is that the capabilities financed in Arizona and Indiana are capabilities not financed in Shanghai.
Section 3’s central argument can now be stated. The Five-Layer AI Economy reveals why investment security requires technological differentiation. The relevant question is not simply whether money enters the AI industry, but which capabilities that money supports, who controls them, and how those capabilities relate to national security—and the answer differs at every layer of the stack.
Section 4: The International Consequences of Technology Financing — China, Global Investment Hubs, and Alternative Capital Networks
A domestic regulation that restricts the conduct of U.S. persons has international consequences only to the extent that the rest of the world responds, and the rest of the world has been responding for some time. The purpose of this section is to evaluate how the outbound investment regime—both the rules in force since January 2025 and the broader rules now proposed—may reorganize international financing, strategic partnerships, and technological competition through 2030. The analysis proceeds geographically, because capital has a geography even when it pretends not to, and it ends with three forward-looking scenarios for the emergence of alternative capital networks.
The starting point is a sobering one for any theory of American financial leverage. Stanford’s 2026 AI Index documents that China leads the world in AI publication volume, citations, patent grants, and industrial-robot installations, hosts the largest number of public or public-private AI supercomputers, and has closed the frontier-model performance gap to within three percentage points, all while recording private AI investment one twenty-third the size of America’s. [38][39] China’s Ministry of Industry and Information Technology reported in January 2026 that the country had more than 6,000 AI enterprises, a core AI industry expected to exceed 1.2 trillion yuan in 2025, intelligent-computing capacity of 1,590 exaflops, and a 60-billion-yuan national AI fund in operation. [56] If American venture capital were the binding constraint on Chinese AI, those numbers would look different. The outbound regime must therefore be evaluated not as a tourniquet but as one instrument among several, whose effectiveness depends on substitution, on allied behavior, and on precision.
4.1 China and the Reorganization of Technology Financing
China’s response to American financial restriction has been to reorganize its own financing around the state. The architecture is layered. At the top sits the Big Fund, now in its third phase with 344 billion yuan of registered capital and a mandate that extends from manufacturing and equipment to high-bandwidth memory and other AI-relevant semiconductors. [53][75] Beneath it, and partly carved out of it, sits the National AI Industry Investment Fund, established in Shanghai in January 2025 as a joint venture of the state-backed Guozhi Investment and Big Fund III, with a mandate to “invest early, invest small, invest cutting-edge” across the AI supply chain, with particular emphasis on embodied intelligence and autonomous vehicles. [53][54] Beneath that sits the 100-billion-yuan national venture-capital guidance fund launched by the NDRC and Ministry of Finance in 2025, which operates through a three-tier structure in which the national fund seeds regional funds that seed sub-funds typically managed by private venture firms. [53] Beneath that sit more than two thousand local guidance funds, Shanghai’s 22.5-billion-yuan AI mother fund, Beijing’s and Shenzhen’s municipal vehicles, and a web of bank-loan programs and industrial-park incentives that RAND has described as China’s “layered state support” for AI. [55][78]
The effect of this architecture is that American restriction accelerates Chinese financial separation at least as much as it slows Chinese technological development. ChinaTalk’s analysis of the funding data makes the causal direction explicit: the national AI fund was established days after a January 2025 tightening of U.S. export controls, and state spending has been accelerated by each American measure. [53] The companies at the frontier—DeepSeek, Moonshot, and the embodied-AI developers such as Galbot and Noematrix—now raise rounds in which the national fund or a university-linked vehicle is a lead or anchor, and in which foreign participation is incidental where it exists at all. [16][78] Whether this is a success or a failure for American policy depends on the objective. If the objective was to deny capability, the evidence is mixed at best, and Miller’s argument that Chinese progress has been constrained more by tools than by money is persuasive. [48] If the objective was to ensure that American institutions are not complicit in financing a strategic competitor’s military-relevant technology, the policy has largely succeeded, at the price of ceding whatever influence minority investors once had. If the objective was to signal to allies and to the Chinese leadership that the United States regards the financial dimension of competition as seriously as the technological one, the signal has been received—Beijing’s June 2026 Regulation on Overseas Investment is, in substance, a mirror-image outbound regime designed to keep Chinese capital and Chinese technology from leaving. [51][52]
4.2 Hong Kong and Singapore: Financial Intermediaries in a Changing Regulatory Environment
Hong Kong and Singapore are the two financial centers through which the largest share of cross-border Asian technology investment is structured, and their regulatory positions under the American regime could not be more different. Hong Kong, together with Macau, is included within the definition of country of concern as part of China, so that a Hong Kong-domiciled entity engaged in a covered activity is a covered foreign person without further analysis. [8][31] Singapore is not a country of concern, is not proposed to become one, and remains the preferred domicile for Asian funds, regional headquarters, and—increasingly—the re-domiciled Chinese-origin companies that seek to exit the perimeter. [50]
The Manus episode exposed the limits of that strategy from both directions. Manus re-registered in Singapore; Beijing blocked its acquisition and restricted its founders’ travel on the ground that the company remained Chinese; and Morgan Lewis drew the inference that Washington may now treat a Chinese security review as evidence that an offshore restructuring was cosmetic. [50][52] BIS’s May 2026 headquarters guidance operates on the same logic at the hardware level, treating an entity whose ultimate parent is headquartered in China as Chinese for licensing purposes regardless of where its servers sit. [23][24] The lesson for intermediary structures is that beneficial ownership, control, and intended use of capital must now be documented at the level of the ultimate parent and the ultimate end-use, because neither government will accept the intermediate domicile as dispositive. Singapore’s attractiveness as a hub is not diminished by this—if anything, the premium on credible, well-documented structures rises—but the era in which a Singapore holding company could launder a Chinese capitalization table into an uncovered one is ending.
4.3 Taiwan, Japan, and South Korea: Strategic Technology Partners
The allied semiconductor economies occupy the opposite position: they are the jurisdictions whose capability the American regime is designed to protect and whose capital it is designed to attract. The IMF’s finding that Taiwan and Korea together source 62 percent of net AI computing-hardware trade is a measure of how concentrated the physical stack remains. [41] The investment flows run in both directions. TSMC’s Arizona commitment, Samsung’s Taylor complex, SK hynix’s Indiana facility, and the Japanese equipment makers’ expanding American footprint represent allied capital building capability on American soil; American capital, in turn, finances Rapidus in Japan, advanced packaging in Korea, and the Taiwanese design ecosystem. [65][66][67]
The policy question is whether more coordinated investment-security approaches across these jurisdictions would strengthen the allied ecosystem or impose commercial costs that outweigh the benefit. The European Union began considering its own outbound program in 2025, with member states expected to report by mid-2026; Japan and Korea have each tightened inbound screening and export controls on semiconductor equipment but have not adopted outbound investment regimes comparable to the American one. [73] The case for coordination is that uncoordinated rules create arbitrage—a Korean fund may invest where an American fund cannot—and that the COINS Act itself directs the executive to work with allies toward compatible programs. [73] The case against is that the three economies have different industrial interests: Taiwan’s entire economic model depends on TSMC’s global customer base, including its remaining Chinese customers for mature nodes; Korea’s memory makers have large Chinese fabs that are already constrained by American tool licenses; and Japan’s equipment makers derive a substantial share of revenue from Chinese demand. A coordinated regime that ignored those interests would fracture the coalition it was meant to strengthen. The realistic path, which Section 5’s Scenario C develops, is coordination on disclosure, ownership transparency, and a shared list of covered persons, rather than on identical prohibitions.
4.4 The Gulf States and the Global Search for Technology Financing
No development in the financial architecture of AI since 2024 has been more consequential than the arrival of Gulf sovereign capital as a strategic rather than a passive investor. IE University’s 2026 report on sovereign wealth funds counts 109 funds managing just over $15 trillion and identifies the Gulf funds—Saudi Arabia’s Public Investment Fund, Abu Dhabi’s Mubadala and its AI-focused vehicle MGX, the Qatar Investment Authority, and the Kuwait Investment Authority—as the dominant source of sovereign AI financing, with PIF committing $10 billion to the HUMAIN–Google platform and MGX anchoring the Stargate Project. [57] MGX closed a fund of approximately $49 billion in mid-2026, targets $100 billion of assets, holds stakes in OpenAI, Anthropic, and the SpaceX-xAI combination, and co-led the $40 billion Aligned Data Centers acquisition; HUMAIN, launched under PIF in May 2025, reportedly deployed $36.2 billion in 2025, invested $3 billion in xAI, and plans 600,000 NVIDIA GPUs and a sovereign Arabic model; the Qatar Investment Authority confirmed a $20 billion AI-infrastructure partnership with Brookfield; and the Kuwait Investment Authority joined the BlackRock–Microsoft AI Infrastructure Partnership targeting up to $100 billion. [58][60][62] Global SWF estimated that sovereign investors deployed $66 billion into AI and digitalization in 2025, with Middle Eastern funds leading. [59] Winston Ma of NYU has characterized the pattern as sovereign funds acting not as allocators but as builders, owning the full stack from energy to application. [58]
None of this is within the scope of the outbound rules, because the Gulf states are not countries of concern and the frontier labs they finance are American. But the Gulf’s rise bears on Capital End-Use in three ways. First, it demonstrates that capital outside the American and Chinese systems is now large enough to be decisive at the frontier, which means that a restriction on American capital does not by itself determine who finances what; a Chinese developer barred from American money is not thereby barred from Gulf money unless the Gulf chooses to align. Second, it raises the beneficial-ownership and third-country questions that Sections 2 and 4.2 identified: a Gulf fund with Chinese limited partners, or a Gulf datacenter operator with a Chinese technology partner, presents exactly the diligence problem the proposed rule’s knowledge standard is designed to address, and BIS’s headquarters test is designed to police. [23] Third, it shows the direction of travel. The Gulf is financing American AI and receiving datacenters, chips, and technology transfer in return, under a framework of U.S. government approvals; the regional conflict of spring 2026, in which datacenters came under attack, did not alter those commitments, and the UAE stated that there was no change to its investment plans. [59][61] The legitimate third-country investor, in other words, is being integrated into the American system rather than excluded from it, and the regulatory task is to distinguish that investor from one that is a conduit for a covered foreign person.
4.5 The Emergence of Alternative Capital Networks, 2027–2030
Three scenarios describe the plausible futures for the financing of sensitive technology through 2030. They are not mutually exclusive; elements of each are already visible.
| Scenario | Core dynamic | Early indicators (2025–2026) | Implications for U.S. investors |
| Coordinated Restrictions | Washington and key partners align selected investment-security rules, reducing opportunities for sensitive financing through allied markets | EU outbound review process; COINS Act directive to coordinate; BIS headquarters test; Japanese and Korean export alignment | Narrower arbitrage; shared covered-person lists; higher but more predictable compliance cost |
| Capital Substitution | Restrained companies seek domestic funding, non-U.S. investors, government support, or alternative structures | China’s national AI fund, guidance funds, DeepSeek and Moonshot rounds anchored by the state; Gulf sovereign platforms | American capital loses influence without denying capability; reputational and signaling objectives met |
| Fragmented Finance | Differing national rules raise compliance costs, complicate cross-border acquisitions, and produce distinct financial networks for competing ecosystems | China’s June 2026 outbound regulation; Manus block; divergent allied regimes; model export controls | Dual compliance programs; re-domiciliation ineffective; higher deal friction for all cross-border technology M&A |
The evidence of 2025–2026 points most strongly toward a combination of Capital Substitution and Fragmented Finance, with Coordinated Restrictions emerging slowly and partially among the closest allies. China has substituted state capital for foreign capital at the frontier and has built its own outbound regime; the Gulf has built a sovereign financing system that is aligned with, but not subordinate to, Washington; the European Union is moving toward an outbound framework of its own but on a timeline measured in years; and Japan and Korea have aligned on exports while remaining cautious on investment. The scenario that best explains the interaction among American security objectives, Chinese industrial development, global financial markets, and international AI competition is therefore one in which the world’s capital for sensitive technology divides into two largely separate networks—one anchored by American and allied private capital with sovereign co-investment, the other anchored by the Chinese state—with a contested middle of third-country intermediaries whose position is determined by documentation rather than by domicile.
Section 4’s central argument follows. Restricting U.S. investment does not automatically eliminate financing for sensitive foreign technologies. Its effectiveness depends partly on the availability of alternative investors, the behavior of allied economies, and the ability of policymakers to identify the transactions that materially affect national security, and the first two of those variables are moving in directions that make the third more important rather than less.
Section 5: Strategic Investment Governance, 2027–2030 — Building a Capital End-Use Framework for Corporations and Policymakers
The preceding sections have described a regime in motion: a statute enacted, a rule proposed, a first penalty assessed, and a global financial system reorganizing itself around new perimeters. The purpose of this section is practical. It asks what corporations, financial institutions, and governments should do in the eighteen months between now and the final rule’s effective date, and it proposes an original analytical instrument—the Capital End-Use Assessment—that is intended to help investors and policymakers evaluate a transaction’s strategic consequences rather than merely its legal classification. The instrument is offered as a research and corporate-governance tool, not as a restatement of Treasury’s legal test; actual obligations depend on the applicable statutes, regulations, and transaction facts, and nothing in this section should be read as legal advice.
5.1 The New Responsibilities of Global Investors
The population of regulated parties is far larger than the population that has historically thought of itself as regulated. Venture firms with Asian funds, private equity sponsors with Asian platforms, corporate development teams at technology and industrial companies, pension plans and endowments with international alternatives allocations, family offices, and the global banks that advise and finance all of them are U.S. persons for purposes of the rule or control entities that are. Each must now maintain the capability to answer, for any cross-border technology transaction, whether the counterparty is a covered foreign person, whether the instrument is a covered transaction, whether an exception applies, and what the party knew or should have known.
Three distinctions help to organize the work. The first is between legal compliance under the existing rule and preparation for the proposed one: the current regulations govern until a final rule takes effect, which law firms expect by mid-March 2027, and the proposal’s broader sectors and prohibitions should inform transaction planning now without being treated as currently binding. [5][10] The second is between diligence on the immediate counterparty and diligence on the chain: the Amidi case was a controlled-foreign-entity case, and the proposal’s preamble language on limited-partner commitments to funds with existing covered holdings extends the chain further. [1][6] The third is between the cost of compliance and the cost of its absence: the Amidi penalty was modest in absolute terms, but the proposed program will be staffed with $150 million a year, will have market-monitoring capability, and will have a statutory basis that invites escalation. [31] The changing rules will affect transaction timing, because notification and feedback processes take time; legal costs, because diligence now extends to beneficial ownership and technical activity; investor confidence, because uncertainty about the SI-system definition will chill borderline deals; and access to opportunity, because some foreign managers will decline to provide the documentation American limited partners now require.
5.2 Beneficial Ownership and the Identification of Financial Exposure
The single most important operational capability the regime demands is the ability to see through structure. Ownership chains, subsidiaries, special-purpose vehicles, pooled funds, and controlled foreign entities must be mapped to the level at which a person of a country of concern, or an entity with specified ownership by or revenue from such a person, can be identified. Screening only the immediate recipient is insufficient for the reason the Noematrix case made vivid: the company was on no list, and its status as a covered foreign person arose from its activity and its location, not from a designation. [15]
A workable procedure has four steps. Identify ownership and control at each layer, including the investment adviser’s role as parent of the funds it advises. Identify the technological activities of the ultimate operating entity against the covered-technology definitions, using technical as well as legal review. Identify the investor’s own knowledge—what documents were obtained, what representations were received, what feedback was sought—and record it contemporaneously, because the reasonable-and-diligent-inquiry standard will be applied retrospectively. And calibrate the depth of inquiry to the risk: a $5 million seed position in an embodied-AI company in Shanghai warrants more inquiry than a $500 million passive position in a diversified Asian buyout fund, and the proposal’s exceptions are designed to permit exactly that proportionality. [4][9] Confidentiality concerns, the reliability of documentation obtained from foreign counterparties, and the resource constraints of smaller investors are real, and the comment period is the appropriate venue for proposing safe harbors and model representations that would reduce the burden without reducing the protection.
5.3 Cross-Border Mergers and Acquisitions in Sensitive AI Industries
Acquisitions raise the stakes because they involve control and integration. The risks concentrate where the target holds semiconductor intellectual property, specialized AI models, industrial software with autonomous features, robotics, or advanced scientific systems, and they are compounded when the target has a Chinese subsidiary that will become a controlled foreign entity of the acquirer. The analytical challenge is to evaluate the target’s future technological trajectory without assuming that every general-purpose AI company presents the same national-security risk, because the proposed SI-system category, read broadly, could sweep in companies whose work is far from the frontier.
The outbound rules are transaction-level and apply to the U.S. acquirer; they operate alongside, and must be distinguished from, CFIUS review of inbound acquisitions, merger control under the antitrust laws of multiple jurisdictions, export-control licensing for technical data shared in integration, and—now—the target country’s own security review, as the Manus block demonstrated. [50] Lawful structures remain available: carve-outs that exclude a covered subsidiary from the transaction perimeter, licensing arrangements that transfer rights without equity, commercial partnerships that stop short of joint ventures, and minority positions below the thresholds for notification where the activity is not prohibited. The preservation of commercially valuable international collaboration depends on investors using these tools deliberately rather than abandoning cross-border technology M&A altogether.
5.4 Introducing the Capital End-Use Assessment
The paper’s principal original contribution is a five-dimensional assessment designed to be applied before a transaction is structured, not after it is signed. It is deliberately broader than Treasury’s legal test, because its purpose is to help decision-makers understand what a transaction does in the world, which is the question that the legal test approximates.
| Dimension | Question | What to examine | Why it matters |
| 1. Capital Origin | Who provides the money, and which individuals or institutions exercise investment authority? | Limited partners, advisers, investment committees, co-investors, sovereign or state participation | Determines who is a U.S. person, who may be “knowingly directing,” and whether the investor’s own capital base creates exposure |
| 2. Capital Destination | Who ultimately receives or controls the investment, including through subsidiaries or intermediaries? | Beneficial ownership to the ultimate parent; domicile versus headquarters; controlled foreign entities | Determines whether the recipient is a covered foreign person; defeats cosmetic re-domiciliation |
| 3. Technological End-Use | What technologies, research activities, facilities, or operational capabilities will the investment support? | Five-Layer mapping; product roadmap; compute and data access; research partnerships | Determines the covered-technology category and whether the activity is prohibited, notifiable, or outside scope |
| 4. Strategic Capability | Could the financed activities materially advance sensitive military, intelligence, surveillance, cyber, or other national-security-relevant capabilities? | Dual-use analysis; customers and end-users; state or defense relationships; talent and data concentration | Distinguishes capability transfer from financial return; informs proportionality of diligence and the decision to proceed |
| 5. Governance and Accountability | What rights, knowledge, reporting duties, exceptions, and legal restrictions apply to the investment? | Instrument rights; board and information rights; notification obligations; feedback sought; exceptions claimed | Determines the compliance record that will be judged under the reasonable-and-diligent-inquiry standard |
The assessment is designed to be used iteratively. A transaction that scores low on Strategic Capability and high on Governance—a passive, well-documented, excepted position in a non-sensitive activity—can proceed with light diligence. A transaction that scores high on Strategic Capability and low on Governance—an undocumented minority stake with information rights in an embodied-AI developer financed by a state fund—should not proceed at all, regardless of its size. The space between those poles is where most real transactions fall, and the assessment’s value is in forcing the decision-maker to articulate, on each dimension, what is known and what is assumed.
5.5 Three Regulatory and Economic Scenarios Through 2030
The regime’s trajectory through 2030 will be determined by how Treasury resolves the definitional questions in the comment period, how aggressively it enforces, and how allies and adversaries respond. Three scenarios bound the range.
In Scenario A, Targeted Investment Security, restrictions remain concentrated on clearly identified sensitive technologies and covered foreign persons; Treasury narrows the SI-system definition in the final rule, publishes a usable covered-person database, and uses the feedback process to provide practical guidance; companies adapt through improved diligence and transaction planning; and international collaboration continues in areas that do not trigger restrictions. This scenario preserves both security and openness, and it is the one Treasury’s own rhetoric about a strong, open investment environment implies. [2]
In Scenario B, Expanding Financial Separation, investment restrictions broaden as advanced AI capabilities become harder to distinguish by technical thresholds; the SI-system category is finalized broadly; biotechnology and additional sectors are added over time; compliance obligations become more complex for venture, private equity, and multinational corporations; more research and development relies on geographically separated financing networks; and the strategic costs of fragmentation—duplicated capacity, slower diffusion, higher cost of capital—increase. The evidence of 2026, including Beijing’s mirror-image regulation and the model-level export controls of June 2026, suggests that this scenario is at least as likely as the first. [51][72]
In Scenario C, Coordinated Technology Investment Governance, the United States and selected partners develop compatible approaches to high-risk technology financing; investor disclosure, ownership transparency, and institutional coordination improve; domestic industrial investment and allied technology partnerships grow in strategic importance; and policymakers reduce security risk without blanket restrictions on legitimate international investment. This scenario requires the most diplomatic effort and the most institutional patience, and it is the one most consistent with the paper’s framework, because Capital End-Use is a question that can only be answered reliably when multiple jurisdictions share the same picture of ownership and activity.
5.6 A Decision Framework for Corporate and Government Leaders
The responsibilities the regime creates are distributed, and they are best stated by actor.
Federal policymakers should define proportionate restrictions, with technological criteria clear enough to be applied by a compliance officer rather than only by a litigator; should resolve the preamble-versus-rule-text ambiguities on limited-partner exposure; should publish and maintain the covered-person database; should operate the feedback process on a timeline that respects transaction calendars; and should enforce transparently, explaining in each case what factors were weighed, so that the industry learns from penalties rather than merely fearing them. [6][14]
State governors and economic-development agencies should recognize that the federal strategy depends on domestic destinations for redirected capital, and should compete on the variables that matter to the Five-Layer stack: energy availability and transmission, permitting speed, university partnerships, and workforce pipelines for fabrication, packaging, and datacenter operations. [65][66]
Corporations should integrate investment-security analysis into strategic acquisitions, international partnerships, and capital allocation at the board level, should apply the Capital End-Use Assessment before structuring rather than after signing, and should treat controlled foreign entities as extensions of their own compliance perimeter. [1]
Financial institutions should evaluate beneficial ownership, investment purpose, legal coverage, and compliance obligations as part of credit and advisory work, should require documentation from foreign managers as a condition of American limited-partner capital, and should recognize that the knowledge standard reaches the adviser as parent of the funds it advises. [10]
Universities and research institutions should protect sensitive research while sustaining responsible international scientific cooperation, and should understand that their endowments are regulated investors whose commitments to Asian funds are within the chain the rule contemplates. [6]
Section 5’s central argument is that effective investment governance requires more than restricting money. It requires a credible method for assessing the strategic capabilities that particular financial transactions may enable, and that method must be shared among regulators, investors, and allies if it is to work.
Section 6: What Have We Learned? Seven Pillars of Capital End-Use
The argument of this paper has moved from a $92,478 investment in a Shanghai robotics company to the architecture of global technology finance, and it is worth pausing to extract what the journey has taught. The five pillars that follow correspond to the paper’s five analytical sections; two further pillars—on the convergence of regimes and on the adversary’s mirror—draw out conclusions that emerged in the course of the analysis and that the original outline had not anticipated. Together they constitute the Capital End-Use framework in its most compressed form.
Pillar 1 — Capital Is More Than Money
Investments can provide financing, institutional knowledge, commercial credibility, management expertise, recruiting networks, and access to technological ecosystems, and the national-security consequences of capital must therefore be evaluated beyond the monetary amount of a transaction. The Amidi case is the proof: a sum that would not cover a single engineer’s annual compensation produced the first civil penalty in the program’s history, because what Plug and Play brought to Noematrix’s capitalization table was not the money but the brand, the network, and the signal to subsequent investors. [1][16] The CSET finding that American participation in Chinese AI concentrated at the earliest stages, where intangible benefits are greatest, remains the most important empirical fact in the field. [28]
Pillar 2 — The Final Capability Matters More Than the First Recipient
International investments travel through subsidiaries, funds, jurisdictions, and corporate relationships, and understanding where money initially goes is insufficient without examining the covered entities and activities it ultimately supports. The proposed rule’s treatment of the investment adviser as parent, its preamble language on limited-partner commitments to funds with existing covered holdings, BIS’s headquarters-based licensing test, and Beijing’s own refusal to accept Manus’s Singapore domicile all point in the same direction: the intermediate recipient is no longer dispositive, and the ultimate beneficial owner and ultimate end-use are what both governments will examine. [6][10][23][50]
Pillar 3 — The Five-Layer AI Economy Requires Differentiated Investment Analysis
Energy, Chips, Datacenters, Models, and Applications and Agents perform different economic functions, attract different investors, and sit at different distances from the covered activities in the rules. A gigawatt of generation is not a covered activity; a growth investment in HBM stacking is prohibited; a datacenter is covered because of who owns it rather than what it contains; the development of a frontier model is on the way to categorical prohibition; and an embodied-AI startup is where enforcement has actually begun. [1][6][31] Investment risk must be examined according to actual activities, ownership, and capabilities, with particular attention to the legal scope of covered technologies at each layer, and a regime or a compliance program that fails to differentiate will misallocate both scrutiny and capital.
Pillar 4 — Financial Restrictions Reshape International Competition
Investment restrictions change where sensitive technology companies seek financing, how international partnerships operate, and which countries become alternative sources of capital. China has substituted a layered state-capital system for the foreign venture capital it has lost; the Gulf has built sovereign platforms that finance the American frontier; Singapore has become the contested middle; and the allied semiconductor economies have become the destinations for redirected capital. [53][57][58][65] The strategic effectiveness of restriction must be evaluated against substitution, fragmentation, and economic cost, and the honest assessment of 2026 is that restriction has succeeded in disentangling American institutions from Chinese military-relevant technology while having only a marginal effect on Chinese capability itself. [47][48]
Pillar 5 — Investment Security Must Balance Protection and Innovation
An effective regulatory framework should prevent harmful support for sensitive foreign capabilities while protecting legitimate commerce, research, and investment. Precision in the covered-technology definitions, transparency in enforcement, enforceability through a workable knowledge standard, and cooperation with allies are the conditions of that balance. Secretary Bessent’s statement that the proposed rule would
“prevent our adversaries from using U.S. investment to develop critical technologies” — Scott Bessent, Secretary of the Treasury [74]
is the protective half of the mandate; Treasury’s simultaneous insistence on a strong and open investment environment is the other half, and the final rule will be judged by whether it honors both. [2][74]
Pillar 6 — Export Controls and Investment Controls Are Converging into a Single Regime
The analysis repeatedly found the two instruments reaching toward each other. BIS’s headquarters test mirrors Treasury’s covered-foreign-person definition; Treasury’s prohibition on financing supercomputers mirrors BIS’s controls on the chips inside them; the June 2026 controls on model access placed the output of the model layer in the same regulatory vocabulary as the accelerators that produce it; and the proposed rule’s abandonment of the compute threshold acknowledges that the thing being controlled—capability—can no longer be measured in the hardware units that export controls were built to count. [6][23][72] Charles Wang’s 2025 study of American national-security review of Chinese investment anticipated this convergence, observing that the tech war had progressively collapsed the distinction between reviewing capital and reviewing technology. [71] The practical consequence for investors is that an outbound-investment compliance program and an export-control compliance program are now two views of the same risk, and should be managed as one.
Pillar 7 — The Adversary Builds the Mirror
The most significant development of 2026 that the original outline did not foresee is that China has constructed its own Capital End-Use regime. The June 1 Regulation on Overseas Investment adds national-security review, restricts indirect technology and data transfer, and gives Beijing control over Chinese-origin technology wherever it travels; the Manus block and the travel restrictions on its founders showed the regime in operation before it was formally promulgated. [50][51][52] Stefanie Kam’s analysis that the regulation marks a shift from enabling outbound capital to actively steering it describes a state that has concluded, as Washington did, that money and capability cannot be separated. [51] The consequence is that cross-border technology transactions now face two sovereigns each asserting a claim on the capability being financed, and that the space for the purely commercial transaction—the one in which neither government has an interest—is narrowing from both sides.
Conclusion: Capital End-Use and the Emerging Geography of Financial Power
The American artificial-intelligence economy is entering a period in which technological development, financial investment, and national-security policy have become inseparable. The extraordinary growth of advanced semiconductor manufacturing, hyperscale computing, frontier models, autonomous agents, and robotics has created investment opportunities on a scale that the IMF has compared to the information-technology boom of the turn of the century, with American IT investment as a share of output at its highest since 2001 and the four largest hyperscalers alone committing three-quarters of a trillion dollars in a single year. [35][42] Yet those same developments have raised, with new urgency, the question of when American capital contributes to technological capabilities that conflict with American security. Stanford’s HAI–Hoover working group on AI and geopolitics has funded research on exactly this intersection of U.S.–China competition and financial and talent flows, and the fact that the question is now the subject of a Federal Register notice rather than only an academic agenda marks the transition from analysis to governance. [69][70]
The Treasury Department’s October 22, 2026 proposed rule is the most important development in that transition since the original program took effect. Building on an outbound regime already operating for nearly two years, it reflects a congressional determination—enacted with bipartisan support and anchored in the Defense Production Act—that certain financial relationships require scrutiny alongside restrictions on sensitive technology exports. [29][31] Its implications extend beyond China to a wider set of designated countries and technology sectors, and its treatment of super intelligence systems, biotechnology, limited-partner exposure, and the knowledge standard will shape the behavior of American investors for the remainder of the decade. Nevertheless, the outcome of the rulemaking remains uncertain. The comment period closes December 7; the final rule is due by mid-March 2027; the definitions that matter most are the ones Treasury has acknowledged to be broad; and the economic consequences will depend on the scope of the final regulations, the vigor of their enforcement—which White & Case has predicted will include a structured non-notified process modeled on CFIUS—and the responses of investors, allies, and foreign governments. [5][79]
Through the Five-Layer AI Economy, this paper has established that investment cannot be treated as a uniform economic input. Capital supporting electricity generation, semiconductor fabrication, datacenter development, frontier model research, and autonomous applications produces very different technological and strategic outcomes, and the relationships among those investments—the energy that powers the datacenter that trains the model that animates the agent—are often as important as the individual transactions. The United States therefore faces a balancing challenge of unusual difficulty. Restrictions that are too narrow will fail to address the risks that the Noematrix case made concrete. Restrictions that are unnecessarily broad will discourage the productive investment that the hyperscaler capital-expenditure numbers represent, will complicate the allied cooperation on which the Arizona and Indiana fabs depend, and will weaken the innovation networks that produced the $965 billion and $852 billion private-market valuations of the two leading American labs. [63][80] Policymakers must distinguish sensitive capability development from ordinary commercial activity while recognizing that both technology and financial structure will continue to evolve faster than any rule can be amended, which is why the knowledge standard and the feedback process—adaptive mechanisms rather than fixed lists—may prove to be the proposal’s most durable features.
This is why I have chosen the title Capital End-Use. The expression shifts attention from the movement of money to the technological purposes, institutional relationships, and strategic capabilities that financial investment ultimately supports. Traditional investment analysis asks where capital is deployed, what returns it may generate, and which assets investors will own. Capital End-Use adds a further question that the traditional analysis never needed to pose: what capabilities does the investment help to create, and what are their implications for national security? Chris Miller’s observation that historical export controls significantly altered the playing field, and that without them China might already have secured a leading position in AI, is a claim about the end-use of technology; the outbound regime is Washington’s attempt to make the equivalent claim about the end-use of capital. [49]
By 2030, the competition for artificial-intelligence leadership may be shaped not only by who manufactures the most advanced semiconductors, develops the most capable models, or operates the largest datacenters, but also by who finances technological advancement, which financial relationships remain permissible, and how governments govern the intersection of investment and strategic capability. The next chapter of the U.S.–China technology conflict will therefore unfold as much through investment agreements, corporate acquisitions, limited-partnership commitments, and financing decisions as through semiconductor factories and GPU export licenses—and it will unfold on both sides of the Pacific simultaneously, as each government builds the mirror of the other’s regime. Capital End-Use is the framework for understanding that transformation: a world in which the destination of money is important, but the capabilities it ultimately helps to create may be more consequential still.
Footnotes and Endnotes:
[1] U.S. Department of the Treasury. “Treasury Announces Enforcement Penalty for Violation of Outbound Program.” Press Release, October 7, 2026 (penalty assessed July 2026; statement of Secretary Scott Bessent). https://home.treasury.gov/news/press-releases/sb0652
[2] U.S. Department of the Treasury. “Treasury Issues Proposed Rule to Implement Comprehensive Outbound Investment National Security Act.” Press Release, October 22, 2026. https://home.treasury.gov/news/press-releases/sb0667
[3] U.S. Department of the Treasury. “Outbound Investment Security Program” (program page; regulations effective January 2, 2025; COINS Act NPRM, comments due December 7, 2026). https://home.treasury.gov/policy-issues/international/outbound-investment-program
[4] U.S. Department of the Treasury. “Fact Sheet and FAQs: Treasury Department Issues Notice of Proposed Rulemaking on Implementation of the COINS Act.” October 2026. https://home.treasury.gov/system/files/206/Fact-Sheet-FAQs-COINS-Act-NPRM.pdf
[5] Jenner & Block LLP. “Treasury Proposes Expanded Outbound Investment Restrictions Under the COINS Act.” Client Alert, October 27, 2026. https://www.jenner.com/en/news-insights/client-alerts/treasury-proposes-expanded-outbound-investment-restrictions-under-the-coins-act
[6] Ropes & Gray LLP. “OISP 2.0: Treasury Previews Broader Outbound Investment Regime Under the COINS Act.” October 2026. https://www.ropesgray.com/en/insights/alerts/2026/10/oisp-2-0-treasury-previews-broader-outbound-investment-regime-under-the-coins-act
[7] Freshfields. “Treasury’s COINS Act Proposed Rule: Key Changes from the Current Outbound Investment Program.” A Fresh Take, October 26, 2026. https://www.freshfields.com/en/our-thinking/blogs/a-fresh-take/treasurys-coins-act-proposed-rule-key-changes-from-the-current-outbound-investm-102o6jd
[8] Kelley Drye & Warren LLP. “Treasury Proposes Expanded Rule on U.S. Outbound Investment.” Trade and Manufacturing Monitor, October 26, 2026. https://www.kelleydrye.com/viewpoints/blogs/trade-and-manufacturing-monitor/treasury-proposes-expanded-rule-on-u-s-outbound-investment
[9] Guerrero, Francesca M.S., Samir D. Varma, and David Chieng (Thompson Hine LLP). “Treasury Issues Notice of Proposed Rulemaking to Expand Outbound Investment Security Program.” SmarTrade, October 27, 2026. https://www.thompsonhinesmartrade.com/2026/10/treasury-issues-notice-of-proposed-rulemaking-to-expand-outbound-investment-security-program/
[10] Lexology. “The Outbound Investment Net Widens: Treasury’s COINS Act Proposal Raises the Stakes for Investors.” October 2026. https://www.lexology.com/library/detail.aspx?g=bc265467-b7cd-4bcc-a392-89e7b9e30359
[11] Morrison & Foerster LLP. “Treasury Announces First Enforcement Penalty Under the Outbound Investment Security Program for Failure to File a Notification.” October 15, 2026. https://www.mofo.com/resources/insights/261015-treasury-announces-first-enforcement-penalty
[12] Latham & Watkins LLP. “Treasury Issues First-Ever Enforcement Penalty Under Outbound Investment Security Program.” October 2026. https://www.lw.com/en/insights/treasury-issues-first-ever-enforcement-penalty-under-outbound-investment-security-program
[13] Steptoe LLP. “Treasury Announces First Civil Penalty Under Outbound Investment Security Program.” International Compliance Blog, October 2026. https://www.steptoe.com/en/news-publications/international-compliance-blog/treasury-announces-first-civil-penalty-under-outbound-investment-security-program.html
[14] NYU School of Law, Program on Corporate Compliance and Enforcement. “Treasury’s First Outbound Investment Penalty: Implications for Private Fund Sponsors and Cross-Border Investment Activity.” October 21, 2026. https://wp.nyu.edu/compliance_enforcement/2026/10/21/treasurys-first-outbound-investment-penalty-implications-for-private-fund-sponsors-and-cross-border-investment-activity/
[15] Kharon. “Treasury Issued Its First Outbound Investment Penalty Over a Stake in Chinese AI Company Noematrix.” October 2026. https://www.kharon.com/resources/article/compliance/treasury-oisp-penalty-noematrix
[16] The Epoch Times. “US Issues 1st Outbound Investment Penalty Over Chinese Robotics AI Deal.” October 2026. https://www.theepochtimes.com/china/us-issues-first-outbound-investment-penalty-over-chinese-robotics-ai-deal-6102312
[17] Wiley Rein LLP. “Executive Order Rebrands AI as ‘Super Intelligence’ and Signals Potential Federal Legislative Changes.” October 2, 2026. https://www.wiley.law/alert-Executive-Order-Rebrands-AI-as-Super-Intelligence-and-Signals-Potential-Federal-Legislative-Changes
[18] Arnold & Porter. “What’s in a Name? Executive Order Directs Agencies to Say ‘Super Intelligence,’ Not ‘Artificial Intelligence.’” Enforcement Edge, October 2026. https://www.arnoldporter.com/en/perspectives/blogs/enforcement-edge/2026/10/whats-in-a-name
[19] The White House. Executive Order 14434 of September 29, 2026, “Inaugurating the Era of Super Intelligence,” 91 Fed. Reg. 63129 (October 2, 2026). https://www.whitehouse.gov/wp-content/uploads/2026/09/eo-14434.pdf
[20] U.S. Department of Commerce, Bureau of Industry and Security. “Department of Commerce Revises License Review Policy for Semiconductors Exported to China.” Press Release, January 13, 2026 (statement of Under Secretary Jeffrey Kessler). https://www.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china
[21] Mayer Brown LLP. “Administration Policies on Advanced AI Chips Codified, with Reverberations Across AI Ecosystem.” January 2026. https://www.mayerbrown.com/en/insights/publications/2026/01/administration-policies-on-advanced-ai-chips-codified
[22] Compute Law Blog. “Export Control Compliance for Running Advanced AI Chips.” September 2026. https://computelaw.blog/deals/export-control-compliance-advanced-ai-chips/
[23] Greenberg Traurig LLP. “‘Enforcement Pause’ Has Limits: BIS Clarifies Ongoing License Requirement for Advanced Computing Items to China-Linked Entities.” June 2026. https://www.gtlaw.com/en/insights/2026/6/enforcement-pause-has-limits-bis-clarifies-ongoing-license-requirement-for-advanced-computing-items-to-china-linked-entities
[24] Hogan Lovells. “BIS’s New Enforcement Guidance Clarifies License Requirement for Exports of Advanced Computing Items.” June 2026. https://www.hoganlovells.com/en/publications/biss-new-enforcement-guidance-clarifies-license-requirement-for-exports-of-advanced-computing-items
[25] Morgan, Lewis & Bockius LLP. “President Issues National Security Memorandum on America First Investment Policy.” February 25, 2025. https://www.morganlewis.com/pubs/2025/02/president-issues-national-security-memorandum-on-america-first-investment-policy
[26] Mayer Brown LLP. “Trump Administration’s ‘America First Investment Policy’ Previews New Restrictions on Inbound and Outbound Investment.” February 26, 2025. https://www.mayerbrown.com/ja/insights/publications/2025/02/trump-administrations-america-first-investment-policy-previews–new-restrictions-on-inbound-and-outbound-investment
[27] Congressional Research Service. “Regulation of U.S. Outbound Investment to China.” In Focus IF12629, updated March 26, 2026. https://www.congress.gov/crs_external_products/IF/PDF/IF12629/IF12629.8.pdf
[28] Center for Security and Emerging Technology, Georgetown University. “U.S. Outbound Investment into Chinese AI Companies.” Policy Brief, February 2023. https://cset.georgetown.edu/publication/u-s-outbound-investment-into-chinese-ai-companies/
[29] Covington & Burling LLP. “FY26 NDAA Outbound Investment Provisions Overview.” December 2025. https://www.cov.com/en/news-and-insights/insights/2025/12/fy26-ndaa-outbound-investment-provisions-overview
[30] Herbert Smith Freehills Kramer. “Codifying the Outbound Investment Security Program.” March 2026. https://www.hsfkramer.com/en_US/insights/2026-03/codifying-the-outbound-investment-security-program
[31] Arnold & Porter. “National Defense Authorization Act Introduces New Outbound Investment Regime.” Advisory, December 2025. https://www.arnoldporter.com/en/perspectives/advisories/2025/12/national-defense-authorization-act-introduces-new-outbound-investment-regime
[32] Holland & Knight LLP. “Outbound Investment Screening Rule Goes into Effect.” January 2025. https://www.hklaw.com/en/insights/publications/2025/01/outbound-investment-screening-rule-goes-into-effect
[33] NVIDIA Corporation. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027.” Press Release, August 26, 2026 (statement of Jensen Huang). https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx
[34] NVIDIA Corporation. Form 10-Q for the quarter ended July 26, 2026. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm
[35] Statista. “Big Tech’s AI Spending to Reach $760 Billion in 2026.” July 2026. https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
[36] FutureX Capital. “Q2 2026 Earnings: Big Tech’s AI Capex Race Hits $730 Billion (Microsoft, Meta, Amazon, Google).” October 2026. https://futurex.capital/en/ai-lab/reports/big-tech-ai-capex-2026q2
[37] EffectStory. “Microsoft, Alphabet, Meta, Amazon: Q2 2026 AI Capex Actuals and Raised Full-Year Guidance.” October 2026. https://effectstory.com/en/articles/hyperscaler-ai-capex-q2-2026-microsoft-alphabet-meta-amazon
[38] Stanford Institute for Human-Centered Artificial Intelligence. “The 2026 AI Index Report.” Stanford University, 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report
[39] The Next Web. “Stanford’s AI Index Finds China Has Nearly Closed the Performance Gap with the US Despite Spending 23 Times Less.” April 2026. https://thenextweb.com/news/stanford-ai-index-2026-china-us-performance-gap
[40] OECD. “Venture Capital Investments in Artificial Intelligence Through 2025.” OECD Policy Briefs, February 17, 2026. https://www.oecd.org/en/publications/venture-capital-investments-in-artificial-intelligence-through-2025_a13752f5-en/full-report.html
[41] FourWeekMBA (summarizing International Monetary Fund, World Economic Outlook, October 2026, and Global Financial Stability Report, October 2026). “IMF Holds 2026 Growth at 3.1% as AI Investment Props It Up.” October 2026. https://fourweekmba.com/ai-imf-holds-2026-growth-at-3-1-as-ai-investment-props-it-up/
[42] Reuters (via AJOT). “IMF Sees Steady Global Growth in 2026 as AI Boom Offsets Trade Headwinds.” January 19, 2026 (IMF World Economic Outlook Update; remarks of Pierre-Olivier Gourinchas). https://www.ajot.com/news/imf-sees-steady-global-growth-in-2026-as-ai-boom-offsets-trade-headwinds
[43] Benzinga. “MIT Economist Sounds Alarm on AI Hype: ‘A Lot of Money Is Going to Get Wasted’” (Daron Acemoglu, Bloomberg interview). https://benzinga.com/z/41166528
[44] MIT Shaping the Future of Work Initiative. “BBC: Could the Tech Bubble Burst?” Interview with Daron Acemoglu, February 5, 2026. https://shapingwork.mit.edu/?p=7316
[45] Miller, Chris. “The Battle over AI Export Controls.” Hitachi Center for Technology and International Affairs, The Fletcher School, Tufts University, January 27, 2026. https://sites.tufts.edu/hitachi/the-battle-over-ai-export-controls/
[46] Texas National Security Review. “Hard Then, Harder Now: CoCom’s Lessons and the Challenge of Crafting Effective Export Controls Against China” (quoting Chris Miller). September 2025. https://tnsr.org/2025/09/hard-then-harder-now-cocoms-lessons-and-the-challenge-of-crafting-effective-export-controls-against-china/
[47] Benzinga. “‘Chip War’ Author Chris Miller Says China Has Been ‘Underspending’ on AI for Four Years.” June 2026. https://www.benzinga.com/markets/tech/26/06/53109449/chip-war-author-chris-miller-says-china-has-been-underspending-on-ai-for-four-years-beijing-just-doesnt-believe-it-matters
[48] Bloomberg. “China’s Chip Catch-Up May Take Five Years: Miller.” Insight with Haslinda Amin, October 22, 2026. https://www.bloomberg.com/news/videos/2026-10-22/china-s-chip-catch-up-may-take-five-years-miller-video
[49] Carnegie Mellon Institute for Strategy & Technology. “Chips and Chokepoints: Chris Miller on the Geopolitics of the AI Supply Chain.” March 2026. https://www.cmu.edu/cmist/news-archive/news/2026/march/chips-and-chokepoints-chris-miller-on-the-geopolitics-of-the-ai-supply-chain.html
[50] Morgan, Lewis & Bockius LLP. “The Manus Decision: China’s First AI Security Review Block and Implications for Cross-Border AI Investment.” May 2026. https://www.morganlewis.com/pubs/2026/05/the-manus-decision-chinas-first-ai-security-review-block-and-implications-for-cross-border-ai-investment
[51] Kam, Stefanie. “The Chinese State Goes Abroad: Beijing’s New Outbound Investment Regime.” IDSS Paper IP26077, S. Rajaratnam School of International Studies, July 7, 2026. https://rsis.edu.sg/rsis-publication/idss/ip26077-the-chinese-state-goes-abroad-beijings-new-outbound-investment-regime/
[52] Burnham, Jack, and Susan Soh. “China Introduces New Outbound Investment Laws to Prevent U.S. Decoupling.” Foundation for Defense of Democracies, Policy Brief, June 3, 2026. https://www.fdd.org/analysis/china-introduces-new-outbound-investment-laws-to-prevent-u-s-decoupling/
[53] ChinaTalk. “Why China’s AI Bubble Is Also Industrial Policy.” September 2026. https://www.chinatalk.media/p/why-chinas-ai-bubble-is-also-industrial
[54] South China Morning Post. “Tech War: China Creates US$8.2 Billion AI Investment Fund Amid Tightened US Trade Controls.” January 2025. https://www.scmp.com/tech/big-tech/article/3295513/tech-war-china-creates-us82-billion-ai-investment-fund-amid-tightened-us-trade-controls
[55] RAND Corporation. “Full Stack: China’s Evolving Industrial Policy for AI.” Perspective PEA4012-1, 2026. https://www.rand.org/pubs/perspectives/PEA4012-1.html
[56] State Council Information Office of the People’s Republic of China. “China Home to Over 6,000 AI Firms in 2025.” January 22, 2026. http://english.scio.gov.cn/pressroom/2026-01/22/content_118293752.html
[57] IE University Center for the Governance of Change, with ICEX-Invest in Spain. “Sovereign Wealth Funds 2026: AI and Strategic Autonomy in a Fragmenting Global Order.” 2026. https://static.ie.edu/CGC/SOVEREIGN%20WEALTH%20FUNDS%20REPORT%202026%20IE%20UNIVERSITY%20ICEX%20INVEST%20IN%20SPAIN.pdf
[58] Ma, Winston. “The Sovereign AI Architects: How Gulf Region Sovereign Wealth Funds Are Building Jensen Huang’s AI Layer Cake.” NYU Development Research Institute, 2026. https://nyudri.org/publications/the-sovereign-ai-architects/
[59] Semafor. “Data Centers Under Fire Test Gulf Sovereign AI Ambitions.” April 13, 2026. https://www.semafor.com/article/04/13/2026/data-centers-under-fire-test-gulf-sovereign-ai-ambitions
[60] Ghosh, Indranil. “What the SpaceX IPO Reveals About Gulf Money in AI.” Rest of World, 2026. https://restofworld.org/2026/spacex-ipo-gulf-money-ai/
[61] CNBC. “The Middle East War Is Testing the Gulf’s Ambitions to Become an AI Hub.” May 24, 2026. https://www.cnbc.com/2026/05/24/middle-east-war-testing-gulfs-ambitions-to-become-ai-hub.html
[62] Apolonia Capital. “Why Gulf Sovereign Wealth Funds Are Buying Data Centers in 2026: Inside the $40 Billion Aligned Deal.” July 29, 2026. https://apoloniacapital.com/gulf-sovereign-wealth-funds-ai-infrastructure-2026/
[63] CNBC. “Anthropic Tops OpenAI as Most Valuable AI Startup, Nears $1 Trillion Valuation in Latest Round.” May 28, 2026. https://www.cnbc.com/2026/05/28/anthropic-open-ai-startup-value.html
[64] AI Funding Tracker. “AI Funding Statistics 2026: Quarterly Totals and Top Rounds.” October 2026. https://aifundingtracker.com/ai-funding-statistics/
[65] The Korea Herald (Herald Corp.). “Will SK hynix Build a Front-End Fab Too? Samsung, Micron and TSMC Race to Invest in US.” 2026. https://biz.heraldcorp.com/article/10853872
[66] Manufacturing Digital. “US Semiconductor Supply Chain: SK hynix’s $4bn Investment” (statement of Ambassador Kang Kyung-wha). https://manufacturingdigital.com/news/us-semiconductor-supply-chain-sk-hynixs-4bn-investment
[67] State Science & Technology Institute. “Commerce Signs Preliminary Terms with TSMC Arizona and Samsung Under the CHIPS and Science Act.” https://ssti.org/node/80764
[68] Military.com. “COINS Act: Protecting National Security From Overseas Tech Investments” (remarks of Matthew Rabinowitz, Pillsbury). January 14, 2026. https://365.military.com/daily-news/2026/01/14/coins-act-protecting-national-security-overseas-tech-investments.html
[69] Stanford Institute for Human-Centered Artificial Intelligence and Hoover Institution Technology Policy Accelerator. “AI and Geopolitics” (Working Group led by Amy Zegart). https://hai.stanford.edu/policy/policy-efforts/ai-and-geopolitics
[70] Hoover Institution. “Technology Policy Accelerator Newsletter: Condoleezza Rice on the Tech Arms Race; Elizabeth Economy and Amy Zegart on DeepSeek.” 2025. https://www.hoover.org/node/371951
[71] Wang, Charles C. “The United States’ Strengthened National Security Review of Chinese Investment: Implications from the Tech War.” American University International Law Review, Vol. 40, Issue 4, Article 5 (2025). https://digitalcommons.wcl.american.edu/cgi/viewcontent.cgi?article=2191&context=auilr
[72] Anthropic. “Statement on Access to Claude Fable 5 and Claude Mythos 5” (suspension June 12, 2026, to comply with U.S. Department of Commerce export controls; access restored July 1, 2026). https://www.anthropic.com/news/fable-mythos-access
[73] Conventus Law. “U.S. Treasury’s ‘Reverse CFIUS’ Authority Is Codified and Some Restrictions Are Clarified.” 2026. https://conventuslaw.com/report/u-s-treasurys-reverse-cfius-authority-is-codified-and-some-restrictions-are-clarified/
[74] Lexology. “Treasury Proposes Outbound Investment Regulations Under the COINS Act” (quoting Secretary Scott Bessent). October 2026. https://www.lexology.com/library/detail.aspx?g=ed8cdbcf-880e-425a-bc47-8387274201d5
[75] Wikipedia. “China Integrated Circuit Industry Investment Fund” (Big Fund Phases I–III). https://en.wikipedia.org/wiki/China_Integrated_Circuit_Industry_Investment_Fund
[76] Steptoe LLP. “COINS Act Codifies and Expands Outbound Investment Rules.” International Compliance Blog, December 2025. https://www.steptoe.com/en/news-publications/international-compliance-blog/coins-act-codifies-and-expands-outbound-investment-rules.html
[77] Baker McKenzie. “President Trump Signs COINS Act Codifying and Expanding Outbound Investment Regulations.” Sanctions & Export Controls Update, December 2025. https://sanctionsnews.bakermckenzie.com/president-trump-signs-coins-act-codifying-and-expanding-outbound-investment-regulations/
[78] Second Talent. “Top 50+ Chinese AI Investment Statistics [2026]” (compiling Stanford AI Index, Moody’s, Caixin, and fund-registry data). September 2026. https://www.secondtalent.com/resources/chinese-ai-investment-statistics/
[79] White & Case LLP. “NDAA for FY 2026 and the Impending Changes to the US Outbound Investment Security Program.” Alert, 2026. https://www.whitecase.com/insight-alert/ndaa-fy-2026-and-impending-changes-us-outbound-investment-security-program
[80] Pinggy. “Racing to a Trillion: OpenAI and Anthropic’s Funding History.” May 2026. https://pinggy.io/blog/openai_anthropic_funding_history/
