<div><p><em>Do tech restrictions give America an edge, or are they simply minting Asia's next generation of Tech Tigers?</em></p><p>There is an old joke in the audit profession that the only thing more dangerous than an argument between two CEOs is an argument between two CEOs who are also business partners. We are watching that joke play out in real time.</p><p>In January, at Davos, Anthropic's CEO Dario Amodei told Bloomberg's John Micklethwait that selling Nvidia's H200 chips to China was &#34;a bit like selling nuclear weapons to North Korea. Nvidia and Anthropic had announced a $10 billion partnership just two months prior. By mid-May, Jensen Huang fired back at Stanford, publicly calling the analogy &#34;stupid&#34; and &#34;madness.&#34;</p><p>     Two of the most consequential business leaders of our generation are now publicly using the words &#34;crazy&#34; and &#34;stupid&#34; about each other's positions on the most important industrial-policy question of the decade. So, who has the better case?</p><p>Let's assume both men are arguing from real conviction, even though both have substantial financial interests in being right. This is a situation that, analogies to the atomic race aside, has no precedent. The answer will shape what the global technology industry, and even the balance of global power, looks like in 2030.</p><section id="darios-case-lock-lead-while"><h2>Dario's Case: Lock In the Lead While the Window is Open</h2><p>On May 14, Anthropic published a policy paper titled &#34;2028: Two scenarios for global AI leadership.&#34; It is the most comprehensive statement Dario has yet published of what he hopes Washington will do to lock in America's lead at the AI frontier.</p><p>As Anthropic views it, the most important input for frontier AI is compute. American and allied companies — Nvidia, AMD, TSMC, ASML, Samsung — have built a commanding lead in advanced chips. U.S. export controls, sustained across three administrations, have kept that lead intact.</p><p>It’s indisputable that some Chinese labs have world-class software talent. DeepSeek and Qwen are genuinely impressive but are constrained by access to the highest-end chips. Anthropic's paper argues that if the U.S. closes the smuggling loopholes that have allowed restricted chips to reach Chinese labs through third countries, and tightens semiconductor manufacturing equipment controls, American AI companies could lock in a &#34;12-24-month lead&#34; in frontier AI capabilities by 2028.</p><p>Dario's broader framing, repeated in venue after venue, is what he calls &#34;a country of geniuses in a data center.&#34; Imagine, he says, 100 million minds smarter than any Nobel laureate, organized inside a single national infrastructure. The question of which nation builds that infrastructure first, he argues, may shape the global order for decades.</p><p>Anthropic also clearly benefits commercially from a world in which Chinese open-weight models like DeepSeek and Qwen remain compute-starved. Chinese models are dramatically cheaper to run on a per-token basis and can operate in a &#34;forked&#34; environment where a user can take a copy of the codebase and develop it independently from that point forward. &#34;Forking&#34; means that all the safety firewalls can be stripped out and proprietary information does not flow back to the models' inventors.</p><p>Anthropic has built its enterprise franchise, running at a reported $44 billion annual revenue run rate , partly on the premise that its closed-weight, U.S.-trained models are the safer choice for regulated industries. Those financial incentives don't necessarily make Dario's arguments wrong. It means they should be viewed the way you might consider a Boeing executive's views on aircraft tariffs.</p></section><section id="jensens-case-compute-already-there"><h2>Jensen's Case: The Compute is Already There</h2><p>Jensen Huang has been making the opposite argument with growing intensity for over a year.</p><p>At Computex Taipei in May 2025, he called export controls &#34;a failure&#34; that &#34;gave [Chinese companies] the spirit, the energy, and the government support to accelerate their development.&#34; At the FT Future of AI Summit in November, he told reporters that &#34;China is going to win the AI race&#34; — a statement Nvidia softened hours later via an official corporate post, which clarified that China was merely &#34;nanoseconds behind.&#34;</p><p>Nvidia, for all its success, is clearly feeling pain from China export restrictions. </p><p>In April, Jensen told the Special Competitive Studies Project that Nvidia's market share in China had fallen from &#34;90-some odd percent&#34; to zero. China contributed roughly 20% of Nvidia's revenue at peak. Jensen argues that the export-control regime did not stop China from building AI capacity. It simply transferred that chip revenue to Huawei, Cambricon, and other domestic Chinese chipmakers, who now have a guaranteed home market they would not otherwise have enjoyed.</p><p>On the Dwarkesh podcast, Jensen even claimed that Mythos, Anthropic's cyber-offensive model whose capabilities it considered too dangerous to release publicly before U.S. government and financial infrastructure was hardened, did not require sophisticated GPUs to develop. Mythos, he said, &#34;was trained on fairly mundane capacity, and a fairly mundane amount of it. The amount of capacity and the type of compute it was trained on is abundantly available in China.&#34; He asserted that China already has the chips required to train a frontier model and the gap between what Nvidia sells and what China can produce domestically is shrinking.</p><p>Dwarkesh pressed the obvious contradiction. If Chinese chips are already equivalent to Nvidia’s, what is Nvidia really being denied? And if Nvidia chips are materially better, doesn't selling them to China accelerate exactly the capability gain Jensen says is happening anyway? Jensen's answer was not entirely satisfying, and the contradiction is the soft spot in his case.</p><p>Jensen also has skin in the game. Nvidia just wrote down billions of dollars in inventory tied to canceled China sales. Huawei is projecting $12 billion in AI chip revenue in 2026, a 60% jump, in a market Nvidia used to own outright. Jensen was a prominent member of the U.S. CEO delegation in Beijing this month, and he is publicly hoping for an H200 sales relaxation. Again, it doesn't make him wrong, but he is far from disinterested.</p></section><section id="other-two-ceos-room"><h2>The Other Two CEOs in the Room</h2><p>It is tempting to frame this as a binary argument between Dario and Jensen, but two of the other most influential American AI CEOs disagree with both in revealing ways.</p><p>Sam Altman of OpenAI has staked out a distinctly different position: less focused on restriction than on acceleration. Testifying before the Senate Commerce Committee in May 2025, he argued that the United States should be “exporting our values” through AI rather than ceding the global market to Chinese alternatives. “I think the U.S. winning the AI race is very important,” he told senators. “I think it matters a lot.” On the specific question of compute controls, Altman has been notably agnostic, suggesting that the smarter play is to make American AI so capable and so trusted that allied nations choose it on the merits, rather than through restrictions that may simply accelerate domestic alternatives. </p><p>Elon Musk has gone furthest of all. On Peter Diamandis's Moonshots podcast in January, he predicted that &#34;based on current trends, China will far exceed the rest of the world in AI compute,&#34; because Chinese electricity output will reach roughly three times America's by year-end. &#34;China will figure out the chips,&#34; he said.</p><p>As Elon sees it, the entire chip debate is a sideshow. The binding constraint on AI is power generation, and on that variable, China is simply leapfrogging the United States. As I noted last summer, Musk has been making this point for years: &#34;China power generation looks like a rocket going to orbit and U.S. power generation is flat.&#34;</p><p>Coming from the man building the largest single AI cluster on earth at X.ai's Colossus complex in Memphis, that suggests U.S. policy makers may want to make the power supply the number one priority in the AI race. The question is whether anything America does on chips matters relative to the massive build out of China's power grid.</p></section><section id="what-chinese-side-actually-looks"><h2>What the Chinese Side Actually Looks Like</h2><p>DeepSeek, Qwen, Kimi, and a handful of other Chinese frontier labs have produced AI models that, on many published benchmarks, are competitive with American models — at a fraction of the training cost. Chinese researchers have made genuine contributions to model architecture, distillation techniques, and reinforcement learning. When I wrote about DeepSeek last summer, I came away genuinely impressed by the technical sophistication of what Chinese teams were producing.</p><p>Huawei's recent Ascend generation of chips is not the equal of Nvidia's Blackwell, but it is closer than most American observers realize, and the gap has narrowed over the past eighteen months.</p><p>The Chinese policy response to U.S. export controls has been exactly what one would expect from a country that has executed five-year industrial plans for the better part of seven decades. Beijing has poured capital into domestic chip champions, accelerated wafer fab construction, and turned what was originally a vulnerability into a strategic priority.</p><p>Whether that response would have happened anyway is the central empirical question. Hard to know without rerunning history. What we can say with confidence is that the Chinese chip industry today is materially more capable than it was when the export-control regime began in 2022.</p><p>If Huawei delivers in 2026 and 2027 anything close to what Huang predicts — chips that close 80 to 90 percent of the gap with Nvidia's leading edge, manufactured at scale on Chinese soil — then Jensen's case will look prescient, and the export-control regime will look like an expensive way to subsidize Huawei's customer base. If the Ascend program stalls in yield, in advanced packaging, or in software ecosystem maturity, Dario's case looks stronger.</p><p>And if Elon is right that the ultimate AI constraint is electrons, and not chips or algorithms, and if Chinese power generation continues to expand at three to four times the American rate, then the compute-availability question could become irrelevant. No matter what Washington does on chips, American grid investment, permitting reform, and nuclear restart timelines may end up mattering more to this contest than any single Commerce Department rule.</p><p>In the meantime, the chatbots themselves may be more honest than their CEOs. When I asked DeepSeek last summer whether China could catch up, it offered the answer that both Dario and Jensen are still avoiding: &#34;If China solves its chip problem, their scale and state power could tip the scales. But if the U.S. keeps attracting geniuses and stays ahead in AGI? Game over.&#34;</p><p>Game over, but not necessarily in the way the chatbot implied. The more likely outcome is messier, more human, and arguably more interesting: two technologically sophisticated civilizations, each with genuine strengths, racing to solve the same hard problems in parallel. That race has produced some of the most consequential scientific progress of our lifetimes, and there is no particular reason it has to stop. The CEOs are arguing about the rules. The rest of us will get to use whatever they build.</p></section></div>

英偉達

1 倍做空英偉達 ETF - Direxion

1.5 倍做空英偉達 ETF - Tradr

<p>黃仁勳和達里奧·阿莫代伊在美國對中國的人工智能芯片出口管制問題上公開發生衝突。阿莫代伊認為這些限制鞏固了美國的領先地位，而黃仁勳則聲稱這些限制通過將收入轉移給華為等國內製造商，加速了中國的自給自足。這場辯論突顯了在保持技術優勢與全球半導體市場的商業現實之間的緊張關係</p>

Jensen And Dario Square Off On AI Chip Exports

- 美國科技限制可能影響全球技術競爭格局。  
- Dario Amodei 與 Jensen Huang 分別就美中 AI 政策爭論。  
- 中國在芯片製造和電力生成上的進展或改變競爭態勢。

黃仁勳與達里奧就 AI 芯片出口展開對峙