Summary

黄敏达确认：DeepSeek 热潮后，中国背景创业者赴美做 AI、投资人转向 AI 的趋势明显加速，华人在硅谷“前所未有地占据了重要的地位”——从英伟达、AMD 高管到 Meta 首席科学家再到基层工程师。但他同时警示，过去不碰 AI 的投资人、创业者纷纷涌向美国，“大家都冲到行业里面去的时候，可能它就是一个行业快要到顶的前兆”。
Reverse CFIUS（正式名称为 Outbound Investment Rule，对外投资规定，2025 年 1 月生效）有两个“第一次”：第一次全面、体系化地限制美元资本的全球流动，第一次把财务数据纳入“中国企业”认定。即便注册在新加坡、股东全非中国籍，只要总收入、净收入、资本支出、运营支出四项财务数据有一项超过 50% 来自中国，也可能被定义为受管辖的中国企业——这可能解释了新闻中 Manus 不仅搬总部、还要遣散中国员工、压低中国运营支出的做法。
黄敏达对 Manus 闪电搬新加坡的判断是“非常不聪明”：刚被中国官媒捧为“AI Agent 界的 DeepSeek”，五月收到美国财政部一封连法律效力都没有的问询函，七月就与中国干净切割，显然忽略了它可能面临的来自中国的压力。而且新加坡本身是个暧昧的目的地——既不是中国，在美国那边又得不到信任，已有国会议员和智库认为它可能成为中国敏感行业企业规避监管的“避风港”。他严重怀疑这套方案“都是投资人的美国律师教的”，只针对明面规则设计规避，完全没想过中国的反应。
Zoom 案例引出的战略性判断是“实质控制”逻辑：黄敏达认为，美国国安监管者在抽象判断时更关心中美真正交恶、必须选边时，美国政府能不能有信心认为这家公司会站在美国这边，而不只是注册地和创始人国籍。但他也强调，在具体法律规则中，注册地、财务数据等细节仍然重要。创始人中国背景的 Zoom 反而成了美国准基础设施；对创业者的拷问是“我到底是一家中国公司，还是一家美国公司”——选哪边没有正确答案，但态度不能暧昧，适应科技式的“三重身份”表态两边都不讨好。
投资端的实际状态是寒蝉效应下的动态平衡：美国本土投资人天然保守、只投特拉华架构、离红线“隔得足够远”，但 DeepSeek 和 Manus 让他们意识到中国工程师红利“不容错过”。结果不是不投华人项目，而是投的时候做更多切割与风险隔离：核心团队是否在美国、目标市场在哪、股权架构、创始人的身份规划、技术会不会触及管制线——事先申报的比较少，但黄敏达团队正在处理一到两单，“相当于为你的投资买了一份保险”。
审查一旦落地，后果可能严厉：随锐集团 2020 年交割的对 Jupiter Systems 投资，五年后被翻出，因后者有美国军方客户，被勒令 120 天内以任何方式全部退出，公布当天即切断中方人员与系统的一切接触——对投资机构“可以这么说”是血本无归。 CFIUS 一年约四百多个申报项目，真正不批准的大概只有两三个，绝大多数会获批，其中有的附条件批准，可就数据库位置、董事安排、独立合规官等进行协商。
拜登到特朗普的管制思路发生根本转变：拜登式的“法典化”全球规则——AI 扩散规则把模型权重列为出口管制对象、把全世界分成三等国家——被废除，特朗普政府则更偏向点对点、具体化的措施；废除当天又制定了全面禁止全世界所有国家使用华为昇腾芯片的新规则。芯片管制在美国内部无定论：黄仁勋主张只封最高端，否则中国“永远都不会再用美国的芯片”；Anthropic 则是全面封锁派的强力推动者。H20 对华销售传出上缴 15% 在华收入的协议，但据黄敏达判断目前应未实际执行。
2025 年的关键词是“密集的变化”，主线是中美拉锯下前所未有的双向互动：安世半导体（母公司为文泰科技）事件中，中国以暂停对欧盟供货、稀土等方式反制，随后美国暂缓 50% 穿透规则，荷兰也收回对安世半导体股权的接管。稀土产能约占全球七八成到八九成，是中国“至少在短期内、未来几年仍然有很大杀伤力”的反制手段；若干年后回望，“2025 年可能会是一个非常重要的节点，也可能会是一个重要的开始点”。

Deep dive

1. The Chinese AI Founder Wave Is Accelerating—but “Everyone Rushing In” May Signal a Top

黄敏达证实了卫诗婕询问的趋势：中国背景创业者赴美做 AI 的现象 2024 年就已出现，但当时还很少；2025 年初 DeepSeek 出现后，这一趋势愈演愈烈，连从未在美国读书或工作过的人也开始在美国创业，过去投资消费或其他行业的投资人也纷纷转向 AI。
黄敏达坦言担忧：“每一次大家都冲到这个行业里面去的时候，可能就是一个行业快要到顶的前兆。”
卫诗婕补充说，她在硅谷看到不少投资人来打听她客户的情况，但律师不能透露。投资人和创业者的热情，尤其是投资人争抢那些所有人都在投的项目，也进一步加剧了拥挤感。

2. Reverse CFIUS: Two “Firsts”

Its formal name is the Outbound Investment Rule. It took effect in January 2025 and was one of the Biden administration’s signature China-tech compliance measures to emerge after its election loss. The core restriction is that US persons and US capital may not invest in Chinese companies that meet specified technology thresholds and operate in designated sectors, including AI, semiconductors and quantum computing.
The first “first” is a comprehensive, systematized constraint on the global movement of dollar capital. The dollar used to be the world’s currency, able to fund companies anywhere; now even China’s most profitable AI companies may be unable to access dollar capital.
The second is the first time financial data has been incorporated into the definition of a “Chinese company.” Previously, regulators focused mainly on the place of incorporation, sometimes looking through to shareholder nationality; now a Singapore-incorporated company with no Chinese shareholders can still be deemed a covered Chinese business if more than 50% of any one of its total revenue, net income, capex or operating expenses comes from China.

3. Did Benchmark’s $75M Manus Investment Cross the Line? The Three-Part Test

Jurisdiction requires all three conditions to be met: the investor must be a US person or dollar capital, the target must qualify as a Chinese company, and its technology must exceed the regulatory floor for a specified industry. It is unclear whether Manus qualifies as a Chinese company; 黄敏达 raised the hypothetical that more than 50% Chinese ownership could be enough. That may explain reports that Manus not only moved to Singapore but also planned to dismiss Chinese employees and minimize its China operating expenses, although outsiders cannot know the specifics.
The technology threshold is defined by training compute: 10^23 floating-point operations triggers a filing with the Treasury Department when the other conditions are met, while 10^25 operations bars US persons from investing. 黄敏达 said he had heard from clients or industry participants that 10^23 operations might correspond to a model with roughly 50B to 100B parameters, although the answer depends on the algorithm.
卫诗婕 noted that 50B to 100B-parameter models are already fairly common. 黄敏达 responded that when the rule was written in the second half of 2024, that size was still considered large, but by 2025 it was no longer a particularly high bar.
The threshold may sweep in projects regulators did not initially intend to restrict, raising transaction costs. Conversely, a company that distills its own model from someone else’s may use far less than 10^23 training operations, potentially allowing more advanced models to evade the rule. 黄敏达 is also unsure whether Manus clears the technical threshold, which depends on its algorithms and other technical details.

4. Chilling Effect: Two Dollar-Capital Markets, Two Risk Appetites

黄敏达 describes the effect as a “chilling effect.” The rule targets a limited set of investments, but US investors on the ground tend to be more conservative in practice. Chinese-background dollar funds historically used VIE structures to invest in Sina and other internet companies; more than 20 years later, the legality of VIEs under Chinese law remains unsettled, which may have made those funds more accustomed to policy uncertainty.
US investors are comparatively cautious. Many will invest only in Delaware corporations and will not look at Cayman or Singapore structures at the outset. They have no shortage of opportunities and can rotate among crypto, Web3 and AI, leaving little incentive to approach a regulatory red line: they want to stay “far enough away.”

5. Chinese Engineers’ Unprecedented Position—and the Investor’s Separation Checklist

黄敏达 believes Chinese voices were less prominent in mainstream Silicon Valley investing in the past. This year, however, Chinese participation has reached unprecedented levels—from Nvidia and AMD executives and mid-level technical staff, to Meta’s chief scientist and other core AI talent, to rank-and-file engineers. That creates opportunities for founders, investors and service providers alike.
Domestic funds are not necessarily refusing Chinese-founder projects; they are applying more separation and risk controls. 黄敏达 replied, “That’s our job—if they don’t invest, we have no projects.” His team has handled a substantial number of AI investments this year, with most clients having Chinese backgrounds, helping them operate compliantly in the US and raise capital, including from US investors.
Investors assess the full picture: whether the founders and core team are in the US—not necessarily with green cards, since an O visa or other visa may suffice; whether the target market is China, the US or elsewhere; whether the cap table uses a Delaware structure familiar to US investors; whether the founder plans to obtain a green card within several years; and whether the company’s technology could eventually enter the scope of CFIUS or Reverse CFIUS controls.

6. Manus’s Singapore Move: “Very Unwise”

黄敏达’s verdict is unequivocal: “Very unwise. The US is not the only country with regulators; China has regulators too.” In March 2025, Chinese state media outlets hailed Manus as the “DeepSeek of the AI Agent world,” treating it as at least a world-class advanced company on the same order as DeepSeek; in May, it received a Treasury inquiry, and in July it moved to Singapore.
In 黄敏达’s view, Manus took an extreme separation step after receiving only a single A4 sheet from the Treasury—and one with no legal force, merely an inquiry letter. That response clearly overlooked the pressure it might face from China.
The destination was also a poor choice, in his view. If Manus had truly decided to comply with US regulation and abandon the Chinese market, the apparent destination was the US; Singapore is neither China nor necessarily trusted by Washington. A number of US lawmakers and think tanks already see Singapore as a potential safe haven for Chinese companies in sensitive industries seeking to evade regulation. Moving to Singapore, changing the founder’s nationality and adjusting financial data might avoid one Reverse CFIUS rule, but would not necessarily address future US rules.
黄敏达 “strongly suspects that Manus’s lawyers in the US taught its investors to make this choice”: the lawyers designed an avoidance strategy around the rules on the books while giving no thought to China’s reaction.

7. Zoom as the Counterexample: Strategic Review Cares About Effective Control, but Legal Details Still Matter

The most useful comparison with Manus is Zoom. Its founder may have been a Chinese citizen for the first 30-plus years of his life, later studying and working in the US and founding Zoom. The company grew into an almost infrastructure-level US application, used by government agencies that may discuss sensitive information.
黄敏达 believes that, in a strategic and abstract assessment, US national-security regulators care more about effective control: if US-China relations deteriorated and the company had to choose a side, could the US government be confident that the company would stand with the US? But he stressed that specific legal rules still contain many details, and factors such as the place of incorporation and financial data remain relevant.
He cited examples including Musk, who is a South African immigrant, and 黄仁勋, who was born in Taiwan, to illustrate that nationality and appearance may not be decisive in this strategic assessment. The more important question is: “Will you stand on my side?”
That leads to the question every founder must ask: “Am I ultimately a Chinese company or an American company?” The idea that capital and companies are borderless may survive for another 5 to 10 years, but if the US and China must choose sides, founders need to think through their own choice in advance.
卫诗婕 distilled the point this way: there is no right answer as to which side to choose, but the position must be clear rather than ambiguous. 黄敏达 cited Adapt Technology, arguing that declaring itself simultaneously a Singaporean, Chinese and American company could leave it in the bad graces of regulators on both sides.

8. The Law Changes, the Objective Does Not: Why Gaming a Single Rule May Fail

卫诗婕 relayed 洪家阳律师’s observation: US laws targeting Chinese technology keep changing, but the determination to contain China’s technology sector has not. The rules may continue to tighten dynamically as tensions intensify. An avoidance structure built for Rule A may work in the short term, only to become useless when Rule A turns into Rule B two or 3 years later—or even the following year.
That is why 卫诗婕 believes companies need to think more broadly about the purpose and future direction of regulation, then prepare accordingly, rather than designing a structure around a single rule.
Asked whether Manus would have missed its window or faced worse consequences if it had not moved, 黄敏达 answered, “Of course,” and said that was precisely the mistake it made at the outset. But he also stressed that both confronting the inquiry and relocating are difficult; once Manus became the “DeepSeek of the AI Agent world” under scrutiny from both governments, any future choice could be wrong.
There is room to respond to an inquiry. 黄敏达 said CFIUS has handled more than 400 filings a year in the past, with only about 2 or 3 actually rejected. Most are approved, either outright or conditionally. Companies can negotiate with the US government over issues such as database location, board composition and an independent compliance officer; the potential cost is lost time, delayed investment and impaired follow-on fundraising.

9. The Cost of Enforcement: The Su Rui/Jupiter Case and the TikTok Precedent

If a transaction is rejected, the question is not whether the investor can exit; the investor must exit. If the investment has already been spent and the company loses value because of the investigation, that loss is not the US government’s concern.
黄敏达 cited Su Rui Group’s investment in Jupiter Systems. The deal closed in 2020 and was not investigated until 5 years later. Jupiter Systems makes large-screen products resembling visualization controllers and had US military and government customers. The US government required the investor to fully unwind the investment by any means within 120 days; on the day the decision was announced, Su Rui Group and all Chinese personnel were barred from entering the site, accessing the system or viewing data. For the investment institution, this was, in his words, total loss of capital.
The investor may still exit by giving up its board seat, severing ties or finding a buyer for the shares. TikTok is a precedent: it entered the US through the acquisition of Musical.ly, which Trump considered a CFIUS violation at the end of his first term and ordered unwound. TikTok sued and at least temporarily blocked the administrative action; after the Biden administration took office, it quietly withdrew the relevant executive order. If a buyer can be found, selling the stake may be the relatively dignified solution.

10. The Pitfalls of Setting Up in the US: Incorporation Is Easy, Operations Are Not

黄敏达 says his team often acts as “the people who clean up after the company.” They have seen companies 1 or 2 years into their operations that still cannot say how many shares each shareholder owns. The US has no equivalent of a unified corporate registry, so ownership must be determined from the company’s internal cap table and records; tax filings, the charter and investor agreements may each show a different figure, leaving “three different ownership percentages for one company.” Lawyers must reconstruct, from incorporation onward, how many shares were supposed to be issued, how many were actually issued, to whom, and what changed in between before the company can raise its next round.
Incorporation itself is usually straightforward: file a one-page document with the state government stating the company’s name and address, and appoint a local registered agent to receive mail. The process can take 1 or 2 days; tax IDs, bank accounts and related steps can generally be completed in about a month. Foreign-investor registration, foreign-exchange controls and additional permits—common questions from Chinese clients—are usually not prerequisites for incorporation.
The real traps come in operating the business. New York and California require certain insurance coverage; failure to buy it may go unnoticed until a penalty arrives the following year, when the fine can be 10 or 20 times the insurance premium or more. Hiring, payroll platforms and payroll frequency are also regulated: California requires payment twice a month or once every 2 weeks.
Once the product reaches consumers, the company may also face litigation or claims involving privacy, consumer rights and telemarketing rules. These issues are usually solvable, but if the company does not understand them in advance, the cost of fixing them later will be much higher.

11. Visa Routes: There Are Options, but Every Option Has a Cost

黄敏达 said that before Biden left office in January 2025, a policy arrangement commonly called “self-employed H-1B” emerged. H-1B no longer had to be used only to work for someone else and could also support founding one’s own company. His team has encountered at least 20 to 30 companies using the route this year; some primarily submitted business plans, and USCIS approved their use of H-1B to launch businesses in the US.
Other routes include O and L visas. The O visa has a high bar for the individual applicant. The L visa is for executives of multinational companies and does not require a lottery, but it requires a multinational corporate structure and proof that the applicant genuinely serves as an executive. Investors generally care more about whether the founder has a visa or status that permits lawful work in the US than about any particular visa category.
Working in the US on a tourist or B visa violates immigration law. 卫诗婕 said she knows many founders of Chinese-background companies who do this and believes a visa is the first challenge, though not necessarily the hardest. 黄敏达 said the ideal arrangement is to keep substantive work outside the US, but in reality many people begin working as soon as they land. Every route carries a cost.
On reports that Musk did not regularly attend classes while holding a student visa, 黄敏达 said that would be an obvious visa violation. 卫诗婕 also stressed that this is not an example to follow; founders ultimately have to weigh the risks and returns themselves.
A past immigration violation may trigger close scrutiny when applying for a new visa or green card. The Trump administration also showed a stronger enforcement posture toward immigration violations; consequences can include visa invalidation, loss of lawful status, removal and, in serious cases, criminal exposure. But immigration law and corporate law are separate systems: a problem with an individual’s status does not automatically make a company founded during that period illegal, and the company and its proceeds still belong to the founder.

12. Biden to Trump: From a “Code” to Point-to-Point Deals

In the final week of the Biden administration, the AI Diffusion Rule treated AI model weights as an export-control item and divided the world into 3 tiers. Tier 3 contained just over 20 countries; even long-standing US friends such as Saudi Arabia and the UAE were placed in Tier 2, triggering diplomatic backlash and questions about enforceability.
黄敏达 compared model weights to a cooking recipe: they are the output of training and are easy to copy. Once an open-source model is released, it is especially difficult to control. He therefore believes the Biden administration may have overestimated America’s ability to control model weights globally.
The Trump administration repealed the AI Diffusion Rule in May while issuing a new rule that broadly barred every country in the world from using Huawei Ascend chips. 黄敏达’s contrast is that the Biden administration favored complex, systematic, global rules, while the Trump administration is more comfortable negotiating point to point with specific companies and countries.
Executive Order 14117, a Biden-era restriction on exporting US personal data, was viewed by 卫诗婕 as a federal-level first. It took effect that October, but as of the conversation there had been no actual enforcement case, leaving a potential gap between tightening on paper and enforcement in practice.

13. The Tariff Hybrid War and the Two Chip Camps: H20’s 15% “Export Tax”

April’s tariff war was no longer just about tariffs. Trump first imposed what he called fentanyl tariffs on China, raising the relevant rate from 10% to 20%, then added tariffs as high as 145% while also raising duties on almost every country in the world and negotiating with them one by one.
The fentanyl narrative expanded from chemical inputs to ordinary pharmaceutical equipment such as pill presses. 黄敏达 believes that after China stopped exporting the relevant inputs, the US accused China of exporting the equipment used to produce the drugs. At bottom, he says, the US government’s failure to control its domestic drug problem required an outlet: blame China, then express that blame through tariffs.
There is no consensus inside the US on chip export controls. 黄仁勋 argues that a blanket ban would only push China to build its own chip industry, after which China might never use US chips again. He therefore favors restricting only the highest-end chips while continuing to sell less advanced products such as H20 to China. That position also serves Nvidia’s commercial interests.
Media reports described the H20 condition as a 15% payment to the US government based on China chip revenue—“revenue, not profit.” 黄敏达 judged that, as of the conversation, the arrangement had probably not been implemented and H20 had probably not yet entered formal large-scale sales in China. That raises questions about whether enough money could loosen controls, whether H20 could be eased and whether H100 could eventually be eased as well; under the Trump administration, all became discussable, but none had a clear answer.

14. Nexperia and Rare Earths: China Found Something It Could Use to Retaliate

The Nexperia chain began in September 2025, when the US Bureau of Industry and Security said that subsidiaries more than 50% owned by an Entity List company would also be treated as subject to Entity List controls. Nexperia is a Dutch company wholly owned by Wentai Technology, a Chinese company already on the Entity List, so its operations were affected.
The Dutch government took control of Nexperia under a national-security law enacted decades ago. A Dutch court then issued an interim injunction preventing Wentai Technology from exercising its shareholder rights. Viewed in isolation, the episode looked like the US and its allies using legal and political tools to seize assets in a Dutch company controlled by a Chinese owner.
China quickly retaliated by halting shipments to the EU market, causing another squeeze in the European auto-chip supply chain; the Chinese government also negotiated with Washington. The US ultimately paused implementation of the “50% rule” issued in September, and the Dutch government withdrew its takeover of Nexperia’s equity after the legal basis for the intervention weakened. 黄敏达 sees this as a case in which Chinese retaliation forced the US back to the table for a settlement and truce.
Rare earths are another retaliatory lever. 黄敏达 explained that rare earths are not all genuinely scarce, but China has mature refining technology and efficient production, historically controlling roughly 70%-80% to 80%-90% of global capacity. If the US bans chips and China restricts rare earths in return, companies on both sides suffer, leaving negotiation as the only practical option.
From China’s perspective, there had previously been few export-control measures that were actively enforced. The relevant laws were built gradually after Trump’s first trade war; active enforcement began with last year’s and this year’s rare-earth measures. Chinese companies have started filing with the Ministry of Commerce when exporting rare earths, while customs has begun checking the shipments. The system is still being built.

15. TikTok and DJI: Trading Time for Space in To C, and Regulators Learning to Strike First

The forced-sale bill for TikTok passed with bipartisan support and was pushed by Biden’s national-security team during the election period. 黄敏达 believed at the time that a sale would alienate more than 100M US users and hurt the election, but Biden’s national-security team considered the issue more important to the US and ultimately pushed the legislation through.
Trump did not repeal the law directly; he used administrative and enforcement mechanisms to delay it. Media reports said the Justice Department sent letters to platforms and vendors potentially subject to fines, saying they could continue working normally with TikTok and would not face enforcement. Trump repeatedly extended the deadline far beyond what the law allowed. 黄敏达 believes that, at least in this case, legislation and enforcement were completely disconnected, with enforcement even contradicting legislative intent. Republicans in control of Congress showed no obvious reaction, but if Democrats take Congress after the midterms, they could use the law again to create political trouble.
黄敏达 believes other companies can learn from TikTok’s strategy of “trading time for space.” When Trump first acted, TikTok had roughly 60M to 70M US users; by the second action, it had 150M. Once the user base became large enough, TikTok could say, “We won’t sell; worst case, we shut down,” creating pressure that an elected US leader might find difficult to bear.
DJI is similar. The US is discussing a nationwide ban, but no comprehensive ban has yet taken effect; because DJI is already embedded in US agriculture, industry and other civilian systems, replacement costs would be high. As a To C product scales, the cost of banning it rises; To B products can still face pressure from the political system on every front.
Regulators learn from these cases too: they cannot wait until a company becomes large before banning it and may move to constrain growth while the company is still at seed stage. Controlling funding is viewed by the US government as a relatively direct management tool; even a highly compliant company may struggle to withstand a prolonged investigation.

16. Anthropic’s Political Gamble and the Year’s Keyword: “Dense Change”

黄敏达 does not want to simply label Anthropic “anti-China.” He sees the company instead as a forceful advocate and supporter of a more aggressive US blockade of China. As 卫诗婕 added, there is no meaningful “pro-China camp” in Washington; the disagreement is mainly over how China should be contained.
Anthropic issued a notice in September restricting Chinese companies, or companies more than 50% owned by Chinese companies, from using its services. 黄敏达 believes Claude does lead in areas such as AI coding, but not by a wide margin and amid intense competition. By taking this step, Anthropic is sacrificing commercial interests on one side while potentially buying political returns and future political support on the other.
With insufficient support at the federal level, Anthropic’s founders may turn to states such as California, where regulation is stricter, to push state-level legislation requiring registration or review of AI Agents and AI products. That could also anger the federal government, which may believe such authority belongs at the federal level.
The commercial logic may be that if US customers cannot use relatively inexpensive models such as Qwen or other Chinese models, US AI companies, Agent companies and major platforms such as Meta and Google will rely more heavily on domestic models. Anthropic could then benefit from a better competitive environment and stronger commercial returns. 黄敏达 believes that if the blockade camp prevails, Anthropic could reap substantial political and commercial gains in the US.
The keyword for the year is “change.” Add one more word and it becomes “dense change”—“dense change.” 黄敏达 believes Trump has done in 1 year what other presidents might have done in 2 or 3, at least in terms of the scale of disruption and impact on the global situation.
Two-way interaction between China and the US has become the new central thread. The US can issue an export-control policy aimed purely at domestic enforcement, yet China can use negotiations to push Washington to withdraw or suspend it within a month. Looking back several years from now, 2025 may prove to be a major inflection point in global affairs and US-China relations—or an important starting point. For now, the direction of travel remains unknown.

20VC: Will SpaceX IPO at $1.5TRN? Cursor vs Figma, Lightspeed's $9BN E219 | Dynastic Feuds and an Acquisition Battle: How Did €14B of Hermès Shares Disappear?