人工智能间谍丑闻的便利政治
來源 Express Tribune 巴基斯坦论坛快报16 小時前
Summary
This item is a 英文-language original; headline and summary shown as published.
黄仁勋的下台和 Anthropic 的上市是在同一周。
Full Text
Policy rarely needs a villain, but it works better with one. For years Washington has argued over how much of its artificial intelligence stack to sell, share or lock away, and the argument never quite found a face. Then Jensen Huang, the leather-jacketed salesman of American silicon to anyone with a purchase order, resigned last week over an affair with a woman unnamed US officials accused of working for Chinese intelligence. Within days the argument had a face.
Count what has followed. On November 16, the Commerce Department issued an interim rule requiring licences to send advanced chips and servers to data centres anywhere in the world that may supply computing power to companies in China or other restricted countries; operators already serving such customers have 90 days to get one. A bipartisan group in Congress wants chip-export bills folded into the annual defence bill. Nvidia’s interim chief, Colette Kress, says the company will stop selling its H200 chips in China. And Reuters reports that Treasury and the White House science office have discussed forcing American firms to disclose which Chinese AI models they use, a “name and shame” campaign short of a ban.
Days after Huang, the loudest champion of open-source AI in the US, resigned, the best-known maker of closed models went public. Anthropic priced its shares at $83, above the $70 to $78 range it had marketed, in what is being reported as the largest share sale on record. It closed its first day at $97.17. OpenAI, which filed confidentially in June, is expected to follow next year.
The timing is instructive. An industry that has spent the last few weeks warning that its own product could end badly has found that fear, properly disclosed, is a selling point. Anthropic’s prospectus gives 93 pages to risk factors, among them the warning that its models could “pose catastrophic or existential risks to humanity.” Investors read it and ordered several times the shares on offer. The wing of the industry that argued for sharing powerful models rather than renting them has, meanwhile, lost its best-connected advocate in the West, just as the rental model is valued at roughly $1.5 trillion.
Notice what is doing the work. The espionage claim remains an assessment by unnamed officials, with no charges reported. Meanwhile, the licensing regime that let chips through to China was written in Washington. Representative Brian Mast says the Commerce Department approved sales of more than 879,000 AI chips to Chinese companies, though Beijing’s own reluctance kept actual shipments limited, and the administration’s H200 policy took a quarter of the revenue. According to the New York Times, US officials told President Trump in October 2025 that Alice Huang was presumed to be a spy, but he has said he has no regrets about not acting on it. The scandal indicts that regime at least as much as it indicts the man. Yet it is the man who is being asked to carry it.
The open question
The casualty most worth watching is not Nvidia, whose shares rose about five per cent after it posted $115.8 billion in quarterly revenue, beating expectations in the wake of Huang’s resignation. It is open-weight AI, models whose weights anyone can download, modify and run on their own machines, as opposed to the closed systems that OpenAI and Anthropic rent out.
Huang was its loudest defender. Nvidia has announced $26 billion of investment in open models over five years and agreed to buy Hugging Face for $12.9 billion, the New York Post reported. He signed and promoted a letter arguing that American leadership would be judged “not by one frontier AI model, but by whether the US builds a strong, open ecosystem that diffuses into every sector.” The same report quotes anonymous sources who fear that his fall lets critics frame open source as a favour to foreign adversaries, and that lobbyists for the big labs smell an opportunity.
There is a strategic asymmetry here. Nearly the entire Chinese frontier ships open, with DeepSeek and Alibaba’s Qwen released under permissive licences, while nearly the entire American frontier ships closed. Export controls were built to stop physical objects at borders. An open model is a file, and as one analysis notes, it needs no export licence to spread. Sanctions have little to grip.
The commercial subtext is hardly hidden, and Anthropic’s own filing supplies it. Among the risks it lists are open-weight models that could “exert pricing pressure on our offerings”, and by one count the document makes more than a dozen references to China-related risks, including rival labs, mostly Chinese, that train on Claude’s answers in breach of its terms. Chinese open-weight models are cheaper than American proprietary tools and, Reuters notes, perform well enough. The prospectus also records roughly $900 billion in compute commitments over the coming decades. A business with obligations like that has every reason to want the market for its product to stay a market in which people pay.
There is no evidence that any lab arranged Huang’s fall, and Dario Amodei has said Anthropic has never advocated a ban on open-weight models. Interests do not need conspiracies to align. A company that sells access to a model prefers a world in which downloading a rival’s carries a stigma, a disclosure form or a licence. Nvidia hardly argues from altruism either. Its critics note that more AI use means more chip sales. But an industry that has long asked governments for rules should expect to be asked whose rules, and who benefits.
The Huawei playbook
We have seen this film play out before. In May 2019 the Trump administration put Huawei on the Entity List, cutting its access to American technology, while Washington lobbied allies to keep it out of their 5G networks. The effort was eventually branded the Clean Network, and its advocates say dozens of allies and firms signed up, although at the outset even the United Kingdom declined. Huawei, for its part, launched its Ascend AI chip and MindSpore framework that August.
Washington drew one lesson, that a security case plus alliance pressure can force the world to choose. Huang drew another, self-servingly but not wrongly, that locking Chinese firms out of American products pushed them towards local alternatives. He claimed Nvidia’s share of the Chinese market fell from almost 95 per cent to about 50. Huang’s detractors argue that Huawei still cannot make enough chips to replace Nvidia. Both can be true for a while.
Now look at the AI version. Since July 2025 the White House has promoted “full-stack” export packages covering hardware, models, software and applications, with federal financing and expedited licensing; critics at the dialogue platform AI Frontiers call the strategy borrowed from 5G. Add a rule obliging data centres to know their customers, and a possible disclosure regime for Chinese models, and the outline appears – if you want American chips and American models, do not mix them with Chinese ones. No such condition has been announced, but a spy scandal is exactly how such a condition becomes thinkable.
An older arms trade
The deeper precedent is older than 5G. In 1949 the United States and its allies formed CoCom, a club for keeping strategic technology out of the Soviet bloc. Its great scandal came in 1987, when Toshiba Machine and Norway’s Kongsberg were found to have sold milling machines that helped Moscow make quieter submarine propellers. Toshiba’s chairman and president resigned, lawmakers took sledgehammers to a Toshiba radio, and Congress weighed banning the company’s exports for years. Cold War arms sales ran on a simple premise: a buyer of American weapons accepted American conditions on resale and use, and the supplier sold a relationship along with the hardware. A control regime is only as strong as its weakest supplier, so every supplier must be made to sign.
Nobody should push the parallel too far, but the place where it breaks is the part to note. A propeller-milling machine weighs over 200 tonnes and can be counted, tracked and embargoed. A set of model weights is a file. You cannot embargo a download. You can only stigmatise it, license it or make it costly to use. That is why the likelier instrument is not a ban but a condition attached to the sale of the hardware.
Open models do carry real risks. A NIST-backed evaluation found DeepSeek-based agents were 12 times more likely than US frontier models to follow malicious instructions, and Amodei claims guardrails are harder to enforce once weights are out. But for countries without a hyperscaler – one of the handful of giant cloud companies, such as Amazon, Microsoft and Google, that own the data centres AI runs on – open weights are the nearest thing to AI sovereignty. A regime that demands buyers of American chips prove their purity offers them a choice between two landlords.
What to watch
Huang said on November 17 that he would be “honoured” to return as chief executive if asked. Nvidia’s board has said it does not intend to reinstate him. He is due before three House committees for a closed-door interview on December 1, and has agreed to appear before the Senate Banking Committee. The questions will be about a woman, a restaurant and a board. The answers will be pulled towards policy.
The thing to watch is which problem the new rules are written to solve. A spy can steal secrets. It cannot steal what is already open. Whoever wants it closed has just been handed the occasion.
The writer is a freelance journalist and media scholar who writes about politics, security, technology and media narratives
All facts and information are the sole responsibility of the author
Full text provided by the source Express Tribune 巴基斯坦论坛快报 via RSS; copyright belongs to the original publisher.
本站僅提供標題與摘要索引,完整內容與版權均歸原發布方所有,請點擊原文連結閱讀。
前往原文- 所屬頻道
- 國際
- 所屬地區
- 印度
- Language
- English
- Item ID
- world-382615be
Comments 0 comments
No comments yet. Be the first.