Chinese AI models close gap with US rivals as Kimi K3 sparks national security debate

OpenAI is scared of open-weight models. Should the US Be?

The impressive capabilities of Chinese lab Moonshot's Kimi K3, the biggest open-weight large language model, has kicked off a debate that conflates two things: the economic possibilities of American AI giants and the future of LLMs as a technology. OpenAI's head of strategic futures, Dean W. Ball, went so far as to argue that the US government should find a pretext to create regulatory fear, uncertainty, and distrust around the new models, since open-weight models must necessarily deter capital spending by the frontier labs. People freaked out, with tech luminaries like Yann LeCun and Martin Casado arguing that open software can accelerate innovation and coexist with proprietary projects. Ball soon retracted his claims that a regulatory crackdown was the White House's "best strategy" and that open-weight models necessarily slow down advances in the technology. However, Axios reports that the Trump administration is considering banning K3 and other advanced Chinese models at the behest of American frontier labs. Another report from Politico said that the Department of Commerce would not take that step anytime soon. The benefit for major AI companies is clear: Open-weight models, running on independent infrastructure or inside major enterprises, offers cheaper intelligence than Anthropic or OpenAI's class-leading models. If users increasingly spend more outside the closed labs, that means smaller return on their massive investments in model training. That view extends far beyond OpenAI. "Strong, frontier-caliber open source models will place a squeeze on the margins and will bring down the prices of the frontier companies," Braden Hancock, the co-founder of Snorkel AI and a former Meta Director of AI, told TechCrunch. "It will not necessarily mean that the amount of AI usage goes down a little bit. You know, obviously, quite the opposite." That's not a problem for people without shares in Anthropic and OpenAI. AI will still proliferate. So what's the justification for the government to block Americans from purchasing something in our ostensibly free markets? Concerns over Chinese models come in several flavors. One is protecting US data from the Chinese government; the US banned the import of modern Chinese EVs over concerns about their data gathering. But experts tend to think that open-weight models run on US servers are unlikely to leak data back to China, although it's not impossible that such a thing could be done. Another is that the models may have implicit bias toward the PRC -- but it's not clear what that might mean for, say, coding tasks. A third common worry is that Chinese models lack the guardrails that the US government has mandated (through an opaque process), which aim to prevent leading US LLMs from being used to exploit closed computer systems or create weapons. However, those same guardrails may make US companies more vulnerable: David Sacks, the venture capitalist and Trump adviser, has been sharing cases of US companies turning to Chinese LLMs to close security gaps when US frontier models refuse to do the tasks. But the most significant motivation for restricting the models is that fear that China will be able to outpace the US if the frontier labs slow down. Sam Bresnick, a China-focused research fellow at Georgetown's Center for Security and Emerging Technologies, says the growing importance of AI to the US military operations gives the US a reason to support continued investment in AI at the frontier labs. But the whole question, he says, is fraught. "Why should the weight of the U.S. government be aimed at protecting these these companies from competitors that are being locked out from the U.S. market based on their origins?" Bresnick asks. Advocates for open AI say that the frontier companies are creating a false binary between innovation and closed models. "The bigger the bigger impact of having these open source models come from China is less that they're sneaking in back doors, and more that they are owning the innovation," Hancock told TechCrunch. "You end up with, effectively, an expanded workforce on your model. PyTorch became the industry standard because it was open source, and so the whole community could contribute to it rather than just one company, and it grew and grew, and all the rest of the deep learning libraries kind of died in comparison." Hancock and other advocates of fear that Chinese LLMs will become the locus of international research. Already, US graduate programs mainly build on open-weight Chinese models, and Hancock says that half of the papers students study are coming from Chinese institutions, with American frontier labs increasingly reticent about sharing their work widely. "Restricting open models wouldn't make AI safer," Clem Delangue, the CEO of Hugging Face, a platform for open AI collaboration. "It would simply hide the risks, concentrate power in the hands of a few and make it harder for the next generation of builders, researchers, academia, non-profits, governments to participate in making AI safer and more beneficial for all." Bresnick says that the real way to slow China would be to focus more on chip export controls. A better way to preserve US AI leadership would be to stop selling Nvidia H200 processors to China. "That," he says, "could potentially keep us out of this thorny debate about banning open source technologies that huge numbers of US companies want to use." Part of the problem is that uncertainy around AI economics. "The open business model, the proprietary business model -- neither one is figured out. AI companies are are struggling to figure out how to make money on their tools, especially as training costs need to go up and up," Bresnick points out. The same challenges that play out in the US are also playing out in China, where AI companies are also struggling to generate revenue and access compute power, and the government is seen as encouraging open releases for policy reasons despite the challenge in capitalizing on them. Some US companies, including Thinking Machines Lab and Nvidia, are trying to make a business around releasing open models. Hancock points out that Nvidia would do better "if there are dozens or hundreds of companies building AI than rather than two or three and two or three that are well capitalized enough to make their own chips," which is one reason behind its investment in Nemotron, a collection of open models. "The main point is the U.S. would be very well served to have its own very capable, much less expensive open models," Bresnick said. "It just clashes with the approach the frontier labs have taken." With additional reporting from Rebecca Bellan.

The US-China AI arms race has taken an unexpected turn | New Scientist

When Chinese company DeepSeek released its open-source R1 model in January 2025, it made headlines around the world. The large language model (LLM) was reported to rival some of the most powerful AIs from US companies, but it was completely free for anyone to download. A trillion dollars was wiped off the value of US tech companies and US lawmakers immediately proposed banning it on government devices. When another Chinese firm, Z.ai, released GLM-5.2 last month, there were similar claims about performance but, surprisingly, none of the panic. The AI arms race between the US and China appears to have taken an unexpected turn. The US and China have been racing to develop a stream of new, more capable AI models, and to create the chips and data centres necessary to train and run them. The US government has also introduced and strengthened export controls on chips to countries like China in recent years, as well as switching on and off foreign access to the latest models. Proponents believe the technology can revolutionise everything from drug discovery to materials science, potentially giving economies a shot in the arm. Its increasing use in war to select targets and deploy weapons means it has become a matter of national security too. There is even talk from US President Donald Trump of taking a public stake in firms like OpenAI to ensure their goals align. However, the way the US and China are approaching it is very different. Chinese firms tend to release models so that users can download and run them locally, at no cost, which is in stark contrast to how most US firms host them in the cloud, provide access for a fee and closely guard their inner workings. Open-source AIs, including those from China, don't tend to perform quite as well as the most expensive premium models, but many people find them adequate. As one UK software developer told New Scientist: "I used a local model [from Z.ai] and a cloud model [from OpenAI] for the same task - it took 30 per cent longer with local, but was fully free." Soon after Z.ai released GLM-5.2 last month it was ranked the most intelligent open-source AI available on the market by Artificial Analysis, an AI benchmarking company. Z.ai says that it outperforms OpenAI's GPT-5.5 on a common benchmark used to test AI software engineering skills. However, Artificial Analysis found that GLM-5.2 performed slightly worse than GPT-5.5 on its intelligence tests - symptomatic of a problem across the AI industry where there are no standard tests of performance. Nevertheless, it appears GLM-5.2 is finding an audience. It currently ranks as the fifth most commonly used LLM on OpenRouter, which captures only a tiny fraction of AI use but is one of the few public sources of such data. DeepSeek's latest model sits at the number one spot with more than twice as much use. Seven of the top 10 are AIs built by Chinese companies. New normal That GLM-5.2 didn't cause the economic shock waves that DeepSeek's previous model did is perhaps testament to how quickly we can adapt to the idea that China has simply caught up on AI, just as it has done dozens of times over with other technology like smartphones, electric cars and robotics. "It traditionally has been the case that Silicon Valley has been more innovative and China has been a fast follower that scales very effectively," says Serge Belongie at the University of Copenhagen, Denmark. "And there's this new dimension to that, which involves the aggressive open-source aspect - kind of attempting to shame the closed-model frontier labs and really put pressure on the West in that regard." Mikhail Belkin at the University of California, San Diego, says that Chinese open-source models are highly capable - perhaps equivalent to the best US models from six to nine months ago - and may even be more stable. American models can be withdrawn or modified at any time, whereas Chinese models can be run by anyone with a sufficiently powerful server, he says. David Shrier at Imperial College London says the company that ends up dominant in AI is likely to benefit from a £60 trillion market, but there are political and tactical benefits for governments in taking part in the race too. "The Chinese models give you different answers to certain questions about political or other sensitive subjects than the US models," says Shrier. The US could play into China's hands if it continues to react to new US models with market-leading performance by withdrawing foreign access. "It could actually be counterproductive because you're forcing people to develop their own ecosystems and own technology," says Philip Torr at the University of Oxford. "You saw it with Huawei and Android: cut them off from Google's ecosystem and they built their own." Shrier is concerned that China can continue to narrow the performance gap between its own models and those from the US by picking apart how Western models operate just as US firm Anthropic has accused Chinese firm Alibaba of doing. "What this means from a practical perspective is that any US advantage gets eroded rapidly," says Shrier. What may save the large US firms in this race is the inertia and caution of business - red tape, says Belongie. Chinese open-source models might be free, run on a laptop and do a decent job, but for big companies with IT departments, risk analysts and cautious boards, a Chinese model downloaded from the internet feels like an unmanageable risk. That's where the big, established technology firms that already supply industry-standard email, office and support software may be well placed. "Why is it that Microsoft is so successful in the enterprise and so many universities and companies use it? It's not because they have the best technology, but they really speak that language of compliance," says Belongie. Torr says Europe, lacking the AI activity of China and the US, is sleepwalking into a national security problem more serious than the nuclear arms race. AI is, he believes, the single most important technology that the human race has ever developed. "We need to have our own Microsofts and Googles, and big tech firms within Europe, within European legislation, paying European taxes, to level the playing field," says Torr. "Do we want to be an AI colony, totally dependent on systems which we don't necessarily have full control over, or do we want to run our own?"

China delivers a one-two punch to America's AI dominance

China's leading AI companies are ramping up the pressure on Silicon Valley, as Moonshot and Alibaba unveiled models they claim can go toe-to-toe with the best from OpenAI and Anthropic at a fraction of the cost. The rapid-fire releases suggest America's lead at the AI frontier is increasingly tight, just as the technology is becoming central to national security, economic power, and geopolitical influence. The opening salvo came from Beijing-based Moonshot AI, one of China's leading AI model developers, which unveiled Kimi K3 on Friday. Moonshot claims its own testing ranks it consistently above nearly every US system, trailing only OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5, though it came out ahead on certain benchmarks. Over the weekend, Chinese tech behemoth Alibaba followed with a preview of Qwen3.8, a new model it says is 'one of the most powerful model[s] available today" and "second only to Fable 5," Anthropic's flagship model. Both companies are emphasizing a key difference from the leading US labs: rather than locking their most advanced models behind closed doors, they are making them publicly available. While some US companies, most notably Meta, have taken a similar approach, releasing models that developers can freely download, modify, and build upon has become a growing point of differentiation for China's AI industry. Moonshot describes Kimi K3 as the world's largest open-source AI system, with 2.8 trillion parameters. Parameter counts are measures of a model's complexity during training and offer a rough indication of its scale and performance, though bigger does not always mean better. Alibaba says Qwen3.8 is a 2.4 trillion parameter model and "continuously evolving." Neither OpenAI nor Anthropic disclose exact parameter counts for their leading systems. It remains difficult to assess how capable either Chinese model is until they are fully released and independently tested. Moonshot says it will release full model weights -- the internal numerical learned during an AI model's training period -- a week from now on July 27th. Alibaba says Qwen3.8 is "going open-weight soon." Even so, the releases have already sharpened competition between the US and China in what has been repeatedly characterized as the defining technological race of our time. They have shaken up the industry in a way not seen since DeepSeek unveiled a low-cost model last year that rivaled leading US systems. The models also raise questions about whether the vast sums of money US companies are pouring into chips, data centers, and model training can secure a durable advantage, particularly if Chinese rivals can approach -- or surpass -- that frontier with fewer resources. The prospect of two highly capable Chinese models being released for others to download and adapt also contrasts starkly with the more guarded approach of US labs, whose most advanced systems remain proprietary. That openness emerges even as Washington moves rapidly to restrict global access to the underlying technology. The government has used export controls to restrict China's access to the most advanced chips, as well as to force Anthropic to pull its most capable system from the market over concerns it could help foreign competitors catch up. Whether the new Chinese models live up to their creator's claims remains to be seen. But, like DeepSeek before them, they are likely to sharpen the technological rivalry between the US and China, influence economic and national security policy, and show that America's lead is far narrower than it once appeared.

Trump administration reportedly reviving push to ban Chinese AI models following Kimi K3 launch, citing cybersecurity concerns -- downloadable open weights could make an outright U.S. ban nearly impossible to enforce amid growing adoption

Critics say the ban will stifle innovation and encourage monopolies The U.S. government may be back on track to ban leading Chinese AI models, following the release of Kimi K3 -- a powerful AI model developed by the Chinese startup Moonshot AI -- last week. According to an Axios report released July 20, Trump's administration is reigniting its push for a ban, citing cybersecurity concerns. Chinese models such as DeepSeek and Kimi K3 are open-weight, meaning their trained model weights are published for public download. This gives enterprises the option to keep their data in-house by self-hosting the models on private infrastructure, while slashing inference costs -- characteristics that have led to rising adoption by U.S. companies. Citing several sources close to the administration, the Axios report says that the government had earlier made a series of attempts to curb growth and expansion of Chinese models in the U.S. over cybersecurity concerns, a move critics say will stifle competition and innovation and encourage monopolies. The U.S. Department of Commerce last year considered adding multiple Chinese AI labs, including DeepSeek, to its "Entity List," a trade blacklist maintained by the department's Bureau of Industry and Security (BIS) that limits foreign companies, research institutions, governments, and individuals from purchasing sensitive American hardware, software, or technology. U.S. officials also considered a joint National Security Agency/Office of the National Cyber Director advisory to discourage the use of Chinese AI models, and drafted an executive order holding U.S. companies liable for security breaches involving hosted Chinese models. While these measures were initially paused due to internal pushback regarding market impacts, they have been revived following the release of new Chinese open-weight models. According to the report, critics of the potential ban -- such as former White House adviser Sriram Krishnan and David Sacks, an outside White House AI adviser -- say the move would negatively impact innovation, while handing a monopoly of the market to leading U.S. AI labs, OpenAI and Anthropic, which the report implies may have a hand in the push for a ban. "We are at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition," wrote Sacks in an X post on Sunday. Chinese AI models are being increasingly used by numerous American companies due to their relatively low cost and perceived matching capabilities with domestic alternatives. The open-weight nature of Chinese models such as DeepSeek V4 and Kimi K3 -- which allows companies to download the models locally and host them on private servers -- is driving adoption by giving enterprises data privacy and slashing API costs compared to closed Western alternatives. Conversely, self-hosting shifts the cost of GPUs, electricity, maintenance, networking, and model operations to the company, making it generally most economical for organizations with substantial and sustained AI usage. Chinese open-weight models price their APIs well below comparable U.S. systems, with DeepSeek-V4-Pro charging $0.87 per million output tokens, compared with $50 for Anthropic's frontier Claude Fable 5 model. This aggressive undercutting has caused a massive surge in developer adoption. CEO Brian Armstrong noted that Coinbase runs models like GLM-5.2 and Kimi in production, cutting their overall AI spending nearly in half even as their actual token consumption spiked. Despite the rising adoption, the U.S. government cites cybersecurity concerns as a reason for a ban. Now, the question of whether a ban on Chinese AI models can be practically enforced arises, as blocking open-weight technology presents a technical and regulatory hurdle. For an individual or a small company just wanting to use DeepSeek via the website or the app despite a U.S. block, a VPN works fine. However, limited app availability and payment restrictions remain effective restrictions. For enterprises that self-host rather than use the hosted app or API, enforcement gets harder for several reasons. Unlike closed-source APIs that require data to leave a company's network by sending it to a third-party provider's servers, open-weight models exist as downloadable files mirrored across public repositories like Hugging Face and independent torrents, making them hard to fully recall once released. Once an American enterprise downloads the weights, it can run the model entirely offline inside a private, air-gapped data center, which limits U.S. regulators' ability to monitor which model is running locally. Modifications further complicate enforcement. Companies routinely fine-tune, quantize, or distill these models, blending the Chinese base with domestic corporate data until provenance blurs and it becomes difficult to define where the foreign model ends and a new domestic one begins. Even under strict download bans, firms could host the models through subsidiaries, although that vector runs into know-your-customer rules at cloud providers and the extraterritorial reach of U.S. export controls. However, the U.S. government may not need an outright ban. According to the Axios report, the strategy appears to be getting U.S. firms themselves to drop the models. Axios-cited government sources say that procurement rules, Entity List threats, and public pressure campaigns aimed at the companies using Chinese models may do the trick. The sources also say the government will "push to highlight potential backdoors and lack of security with Chinese models, and the governance issue that brings." Any ban or restrictions would be yet another event in ongoing broad trade tensions between the U.S. and China that have since extended into the AI industry. Washington had earlier placed export restrictions on critical computing hardware and equipment to China. It later eased restrictions, but Beijing now appears to be focused on developing domestic technologies, while urging Chinese companies to utilize them. The Trump administration has also made known its intention for the U.S. to dominate the AI race. Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

Chinese AI models narrow cyber gap with US rivals

UK agency warns cheaper open models could leave companies less time to patch critical vulnerabilities AI models developed by Chinese groups are rapidly closing in on rivals from OpenAI and Anthropic, according to new findings that suggest hackers could soon exploit cost-effective AI systems to overwhelm cyber defences. The UK's AI Security Institute (AISI), the world's leading body for testing AI capabilities, said on Friday that the gap between Chinese open-weight models -- whose underlying parameters can be downloaded and modified -- and the latest US frontier models is narrowing faster than previously understood. The agency said Chinese developers were six to 10 months behind US rivals in 2025. That gap has now fallen to as little as four months. The findings were released a day after Chinese start-up Moonshot released Kimi K3, China's largest AI model to date. The company released benchmark results suggesting the new model outperformed Anthropic's Claude Opus 4.8 and OpenAI's GPT 5.5 on most coding and general AI agent benchmarks. Moonshot was still falling short of Fable, a powerful model that Anthropic briefly suspended after the US raised concerns over its hacking capabilities. But its launch has contributed to the fall in US tech stocks on Friday, as investors grew anxious about exuberance over AI and whether the American lead in the technology can be defended. Companies from Silicon Valley to Europe have increasingly turned to China's advanced open-weight models as they contend with rising fees from leading US AI developers. At the same time the latest AI models have begun to exhibit hacking capabilities that exceed those of the most skilled human hackers. "Once open weight models are released, these options are lost permanently," said AISI. The agency warned the narrowing gap in capabilities between open and closed AI models could imperil cyber systems across the globe, shortening the length of time organisations have to secure their systems. Anthropic earlier this year limited access to its Mythos model after finding that it could chain together vulnerabilities faster than humans could fix them. It has said the limited release gave a trusted group of companies and the US government time to find potential hacks using Mythos and patch them. Sam Bresnick, a research fellow at Georgetown University's Center for Security and Emerging Technology, said that even once existing vulnerabilities were patched, the widespread availability of free and highly capable AI systems could present a threat to critical infrastructure. "We might soon be living in a world where there are models open to everybody that have the capabilities to autonomously identify vulnerabilities in code," he said. AISI's study evaluated models with two sets of tests. The first set compared AI models' abilities to complete specific hacking tasks, spanning vulnerability research and exploitation to cryptography. The second judged models on their ability "to conduct end-to-end cyber attacks autonomously". On the first narrow set of tasks, GLM-5.2, released this year by Beijing-based Z.ai, performed "comparably to the most cyber-capable models released four months before it", such as Anthropic's Opus 4.6 and OpenAI's GPT-5.2-Codex. This was despite the fact that US-made models had already made huge advances in their cyber capabilities, even before the release of Mythos. On the second set of tests, GLM-5.2 also showed strong performance in conducting broader cyber attacks compared to US models, while using fewer "tokens" -- the units of data processed by models. The agency said the Chinese open-weight models it tested were far cheaper to run than US frontier models. While open-weight models brought "real benefits", such as data privacy and low cost, AISI warned "they can also carry risk". "The same openness underpinning these benefits precludes many of the safety measures that closed model developers can use to detect and disrupt misuse, iterate on safeguards as vulnerabilities emerge, control user access and withdraw models." Data visualisation by Clara Murray

Why Silicon Valley Can't Stop Looking Over Its Shoulder at China

Chinese companies are offering artificial intelligence that is nearly as good as the leading U.S. technologies, but it's still America's race to lose. The United States, by almost any measure, dominates the world of artificial intelligence. The most powerful A.I. systems are made in America. The United States has far more A.I. data centers than any other country. And a vast majority of the computer chips used to build A.I. were developed by American companies. Perception, however, is a different matter. In recent weeks, China has appeared to erode that sizable lead. The competition came into sharper focus last week when a Chinese start-up released a system nearly as powerful as the leading American technologies but costing a lot less. It was the second time in about a month that a Chinese company managed that feat. In a speech on Friday, China's leader, Xi Jinping, hailed Beijing as the champion of a new global A.I. order. "A.I. development should not be a solo performance by a single country but a symphony of international cooperation," Mr. Xi said. As it has with other cutting-edge technologies, from consumer electronics to electric cars, China is demonstrating that it can produce something almost as good but at a far more affordable price. That strikes a nerve in Silicon Valley, where industry veterans know that when it comes to technology, good enough often beats best if it is a lot cheaper. That has been true since PCs replaced mainframes, and it could happen again with A.I. But worries about China's increasing A.I. prowess appear to be less about what technologists in that country are doing right and more about what the United States could be doing wrong. American companies could be building their product the wrong way. They could be spending too much and charging too much for their products. The Trump administration could also be needlessly meddling in the A.I. market. "This is still the U.S.'s race to lose, but it is getting damn close," said Rehaan Ahmad, a co-founder of the Silicon Valley start-up alphaXiv, who has been using the latest Chinese technologies for the past several weeks. Last month, the A.I. start-up Anthropic shut down its two most powerful systems after the government unexpectedly demanded that the company bar access to foreign nationals, including its own employees. Anthropic's close competitor, OpenAI, also said the administration had interfered with a recent release of its technology. Like some A.I. researchers, administration officials were worried that the technology could drive cyberattacks or maybe even help build bioweapons. The administration lifted its restrictions, but Anthropic and OpenAI still maintain strict control over who can use the technologies and who cannot. Executives at both companies have called on the government to regulate A.I. in some way. On Friday, Mr. Xi specifically mentioned the so-called open source method that the Chinese companies use to develop their technology. It was a remarkable moment, in which a world leader appeared to be taking a side in a decades-long tech industry debate. On one side are people who think new technology should be tightly controlled, and on the other are people who think it should be freely shared and distributed -- the open source approach that Mr. Xi championed. Without open source, it would be extraordinarily difficult for Chinese companies to catch up. "The most authoritarian government is producing the most egalitarian models, and what should be the most democratic government is breeding companies that are the most authoritarian," said Rayan Krishnan, chief executive of Vals AI, a company that evaluates the performance of the latest A.I. technologies. Companies like Anthropic, OpenAI and Google hope to recoup their aggressive spending by selling increasingly useful but often expensive technologies to businesses and consumers. China is offering similar technology free of charge. That means companies anywhere in the world -- including in the United States -- can operate this technology at much lower costs. On Thursday, a Chinese start-up, Moonshot AI, released a new A.I. technology that is nearly as powerful as the leading American model, Anthropic's Claude Fable 5. In line with other Chinese A.I. companies, Moonshot said it would soon open source the technology, called Kimi 3. The start-up also sells a version as an online service, just as American companies do. A day later, the tech-focused Nasdaq fell 1.4 percent and the S&P 500 fell 1 percent, with Alphabet, Google's parent company, and Meta, the maker of Facebook and Instagram, dropping 2 to 4 percent. It was, in the minds of some tech experts, a silly overreaction. "Nothing fundamentally has changed," said Perry Metzger, a software developer who has worked on A.I. for more than two decades. "The problem is that the bulk of investors have no idea what it is that any of these systems actually do, or how any of the companies make their money, or what any of the news means." The Kimi model is at the cutting edge of open source Chinese technologies that are nearly as powerful as top American systems. Last month, another Chinese start-up, Z.ai, released a model that start-ups and independent developers across Silicon Valley rapidly adopted. After its release, six of the 10 most popular A.I. systems on OpenRouter -- a closely watched leaderboard of A.I. models -- were Chinese technologies. In some key areas, the latest systems from companies like Anthropic and OpenAI still outperform Kimi and other Chinese models, according to benchmark tests run by Vals AI and other independent companies. But because the top Chinese models are generally open source, they cost considerably less to use. In early 2024, a third Chinese start-up, DeepSeek, first set off alarm bells among investors and tech executives with the release of a surprisingly effective open source system. Stocks took a similar tumble. Soon, companies like OpenAI accused DeepSeek and other Chinese companies of improperly harvesting data from their A.I. systems to accelerate the development of Chinese technologies. Last month, Anthropic sent a letter to two U.S. senators accusing the Chinese tech giant Alibaba of "brazenly" and "illicitly" trying to copy its technology through 24,000 fraudulent accounts. Using data from one system to train another -- a process called distillation -- is common in A.I. development, including in the United States. But the Anthropic and OpenAI terms of service forbid anyone to surreptitiously harvest data for distillation. Anthropic called on lawmakers and regulators to explore ways of curbing the practice. Some experts believe, however, that distillation will become less important as companies build systems designed to operate as "A.I. agents" -- digital assistants that can use other software to perform tasks. Training agents requires far more than just distillation, these experts say. Other experts have long argued that Chinese systems will always trail the top U.S. models because U.S. export controls limit the flow of the specialized computer chips needed to train A.I. technologies. But Moonshot, Z.ai and DeepSeek spend millions for access to chips in data centers outside China. Various Chinese companies have also started to build their own specialized chips, and A.I. start-ups like Z.ai say they are beginning to use these chips at least as a way of augmenting the chips from the United States that they have managed to acquire. Even as China sloughs off U.S. export controls and other restrictions, many U.S. executives, lawmakers and policymakers continue to call for regulations that tightly control the use of American technologies. But some experts believe that these regulations could ultimately push more people toward Chinese models because they can use these technologies however they want. "It is really hard for U.S. companies to operate when they know that the Chinese models are getting more and more powerful," Mr. Krishnan said. "Something like Kimi 3 is purely open source. There is no way to control how anyone uses it."

The Great Freakout Over Open-Source AI Has Begun

Remember when the Chinese open-source AI model DeepSeek was first made public, and everyone freaked out? Welcome to Round Two. When Chinese AI lab Moonshot AI dropped its latest model, Kimi K3, last week, it caused quite an uproar -- including a long-winded meditation on the open-source approach from Dean Ball, the recently hired "Head of Strategic Futures" at OpenAI, in which he called the non-walled-off approach to AI development a risk that makes AI "ungovernable." Ball said that Kimi K3 is competitive with the top American models, which appears true: benchmarking reportedly puts it on par with Anthropic's Claude Fable and OpenAI's GPT-5.6. But he has doubts about the open-source approach. "I am personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks," he said. "To be clear, I myself might be fine with models presenting this level of marginal risk being open weight, but I am surprised that China is fine with it." Saying that he "might be fine" with models like Kimi K3 being open source feels like it's doing some heavy lifting, given everything else Ball said in the post, including claiming that the only reason Moonshot and other Chinese labs are going open source is that "they are behind, and they know that very few people would pay for sub-frontier models from China," while implying the hidden reason they've chosen this approach is because "they know open-weight models are effectively ungovernable, and they simply like the overall cloak of ungovernability open-weight models create over the whole of AI." Ball's real objection seems to be that open-source models put a dent in his company's bottom line. Ball called open source "inherently decelerationist" -- meaning it'll slow the development of AI -- and claimed, "in the end, open-weight models deter further AI capex." In the end, that would lead to "full AI communism," he claimed: "Rather than a market product, AI is a 'public good' which will ultimately be provided by the state as a kind of 'digital public infrastructure.' This future strikes me as a dystopian hellscape, but I've never met an open-weight models advocate who doesn't ultimately concede this is where things end." Ball's position seems to be that if we allow open source models to persist (which he's totally fine with and not at all threatened by, by the way!), it'll eventually erode the desire of closed-source labs like OpenAI to develop their model by making the technology a public utility instead of a business. How awful! Could you imagine a technology not being a purely profit-driven endeavour? Weird position to take from a company that was founded with a charter that insisted if another company was going to beat it to achieving artificial general intelligence, it would cease operation and lend its resources to assist that company. Anyway, Ball theorized that the Trump administration will likely see the development of Chinese models as a threat and seek to ban them -- though he advised they wouldn't need to go that far. Instead, they could just issue a warning that there "may be backdoors in Chinese AI models," regardless of whether it's true or not. "You just create enough regulatory risk that every regulated enterprise backs off," he wrote. Seems he might be right about that. A recent Axios report implied that the Trump administration may renew efforts to create an effective ban on foreign models. According to the outlet, the admin previously explored placing Chinese AI labs on the Commerce Department's "Entity List," which would basically restrict them from operating in the US without a license. Trump also reportedly considered an executive order that would hold US companies accountable for any risks associated with hosting Chinese models, per Axios. But some folks close to the Trump administration are reportedly pushing back against the idea of such restrictions despite their general hawkishness toward China. David Sacks, the former AI and Crypto Czar to Trump, took aim at Ball over the weekend and said on X, "The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open-source competition." Sacks is an interesting ally to the open-source cause here, given his fearmongering over China in the past -- but he's basically molded his position to fit his true cause: killing regulation. Sacks has long accused Anthropic of trying to run a regulatory-capture campaign -- basically locking in strict guardrails that are designed to keep upstart models from knocking the top labs off their thrones. Now he sees OpenAI attempting to do the same, and he's not having it. Sacks likely doesn't really care about open source as an ideological principle, though. He more just wants to tear down the minimal bumpers that have been set up for AI models in the US. Sacks was reportedly the leader in killing federal AI legislation that the industry itself was pushing for, and supported efforts to keep states from implementing their own regulations. More than open-source models, Sacks wants to open the AI floodgates. Not surprising for a guy who reportedly has nearly 500 stakes in companies with ties to AI. He's a powerful ally to the open-source movement given his proximity to the Trump administration, but he should be viewed with just as much skepticism, even if he might be aligned in this instance of pushing back against an industry leader pretty blatantly posturing for fear-based restrictions on a competing model because it's cheaper and "foreign."

Companies turn to Chinese AI models to cut costs

Companies from Silicon Valley to Europe are turning to Chinese AI models as they try to cut the cost of using the technology and reduce their dependence on US frontier labs. DoorDash, Siemens and Airbnb are among the groups that have adopted AI tools built in China, drawn by models that are cheaper, increasingly capable and, in some cases, easier to run on their own infrastructure. Chinese AI models from groups such as DeepSeek and Z.ai have rapidly overtaken US rivals in token consumption this year, according to OpenRouter, a platform that tracks the units of text, code or data processed by large language models. The shift has been driven largely by cost, as companies try to curb ballooning AI bills. But in Europe it has taken on a sharper geopolitical edge after the Trump administration last month imposed export controls on Anthropic's Mythos and Fable models, forcing businesses to confront the risks of depending on US technology. Chinese models are "the elephant in the room", said Eugene Cheah, chief executive and co-founder of Featherless AI. "Enterprises are starting to realise, 'Hey, we don't need the best model, we can use the faster, cheaper models'." DoorDash co-founder Andy Fang said last week that the food delivery group now delegated "lower-level work" to Kimi K2.6, a model by Chinese start-up Moonshot AI, and reserved Anthropic's Fable for only "the hardest work". The new combination "vastly outperform[ed] . . . at a cheaper cost" than a previous set-up that used only US frontier models from Anthropic, he said on X. German engineering group Siemens told the FT it wanted "flexibility" with its AI models. It uses a broad range, including tools from China's DeepSeek and Z.ai alongside models from US frontier labs and Nvidia as well as French AI group Mistral. Some companies have gone further, switching entirely to Chinese models. San Francisco-based start-up Lindy has moved from Anthropic's AI tools to DeepSeek's V4 model. Founder Flo Crivello last month on X hailed the shift as "transformative", saying it had saved the company millions of dollars and improved performance in "many core use cases". The shift has been accelerated by US-based AI groups including Anthropic and OpenAI moving some enterprise services from flat subscriptions to usage-based billing, which has dramatically increased the cost of using their models. At the same time, China's top models have improved, especially on coding tasks. The release in June of Z.ai's GLM-5.2 was praised by many Silicon Valley technologists and signalled that the gap between US and Chinese models was starting to narrow. "Many smart people/AI insiders are saying GLM-5.2 is the first Chinese AI model to match and often beat the American big lab public AI models with no compromises," wrote Marc Andreessen, co-founder of US venture capital group Andreessen Horowitz, in a post on X. "Enterprises have an incentive to shift some of their workload to cheaper models. Why would you pay a premium for Anthropic, OpenAI models when for a lot of the workloads you need, the Chinese models are generally workable?" said Sam Bresnick, a research fellow at Georgetown University's Center for Security and Emerging Technology. Another draw is that many of the leading Chinese AI tools are so-called open-weight models whose parameters are released publicly, meaning they can be hosted on company-managed servers and fine-tuned for specific uses. Airbnb said it used "a limited number of China-origin models" and was able to protect its data and operations by running them "only through approved US-based service providers". Proprietary models like OpenAI's ChatGPT and Anthropic's Claude tools are largely accessed through their creators' systems or third-party enterprise platforms. The best open-weight models are between 10 and 60 times cheaper than their proprietary equivalents, said Vipul Ved Prakash, chief executive and co-founder of Together AI, a cloud provider that helps companies access these tools. "Companies want to deploy them because they have more control and they can adapt the models to their own data," he said. In Europe, companies cite last year's US trade wars and the export controls on Anthropic's models as factors in moving away from US AI tools. While the export ban was overturned, it "changed the perception of the market forever," said Ben Grinnell, chief AI officer at Newton, a UK consultancy firm. "You can put Fable back in the market, but you can't put the genie back in the bottle." Tom Sheridan, US vice-president at venture capital firm RTP Global, said his advice to European start-ups has changed. "For European companies, a self-hosted Chinese model is the most secure choice versus the US one." Zoltan Bettenbuk, chief executive officer of German human resources start-up Timebutler, said that about six months ago his business started to offload some tasks from Anthropic's Claude to Alibaba's Qwen models in order to reduce its dependence on US frontier labs. "I still rely on the most capable flagship models currently because they do a great job, but if all hell breaks loose, I need to have another plan," he said. Platforms offering open-weight models say demand has risen in recent months. Featherless AI's Cheah said it had seen "exploding interest" since the US ban on exports of Fable, particularly from Europe. "People came banging on our door." He added that one customer had "origins near the Greenland area", which the US has threatened to take over. "He was like, 'I don't want to build on top of closed models, because who knows what happens with the geopolitics'." Aidan Gomez, chief executive of Canadian AI group Cohere, said companies were now realising the importance of sovereign AI for their business. "The Mythos ban was certainly the most tangible event, and people having their access revoked. It exposes the risk of relying on any one single entity for any of your workloads," he said. "Two years ago the main worry was China. Right now the bigger worry in Europe is the US," said Per Roman, founder of European venture capital firm Bullhound Capital. "That is staggering."

The hottest AI models in Silicon Valley face a powerful source of competition

As Anthropic and OpenAI rocketed to popularity based off of closely guarded artificial intelligence models, Chinese AI companies have made a different bet: They would make theirs open and free. Over the past year, the strategy has grown more influential, especially in the wake of the U.S. government's now-rescinded restrictions on flagship Anthropic model Fable. Burdened by rising AI costs, American tech companies have begun to shift to Chinese model families such as Alibaba's Qwen, Z.ai's GLM and Moonshot AI's Kimi -- a landmark adoption of Chinese software at enterprise scale in the United States. "This is the first time that it's happening," said Rafiq Dossani, an economist at Rand. "China's been able to catch up in software at a rate that it couldn't do much earlier, even though [it's] well resourced." ChatGPT, Claude and Google's Gemini are all examples of "closed" AI models: The secret sauce of how they produce text is tightly guarded within their companies, and the only way to access them on an enterprise scale is to buy a subscription or use costlier pay-as-you-go pricing. In contrast, many leading Chinese models are open-weight, meaning that their internal formulas can be inspected and accessed for free. While the companies also offer their own web interfaces and desktop apps, most outside developers choose to just download a copy of the model and, with some configuring, hook it up to their systems as they see fit. That means the main cost of using open-weight models is the computing power needed to generate text -- generally far cheaper than a closed model in a big lab, where training and development costs are priced in. The savings from open-weight models are especially stark following the trend of "tokenmaxxing," when some companies in Silicon Valley urged their engineers to pump up their AI usage as much as possible as a sign of productivity. Tokens are the basic units of input and output in language models. Non-Chinese companies, including Meta, Google and France's Mistral, offer open-weight AI, too. But in the global race for AI dominance, the most popular open options come from China-based companies, which began outpacing American counterparts in downloads in August. On OpenRouter, a popular marketplace to access and run open models, Chinese AI accounted for nearly half of U.S. traffic by tokens in the last week of June, up from 16 percent at the beginning of the year. Greg Osuri, CEO of the cloud computing company Akash, spent almost $12,000 on Claude by himself in May. But after the U.S. government instituted export controls on Fable, Anthropic's flagship model, he started using Z.ai's GLM-5.2, which he said was cheaper and better than Anthropic's models. "I switched to GLM-5.2 because I'm not paying top dollar for a model that's inferior," he said. "And I don't see a reason why I should switch back." Osuri has more than halved his personal AI spending since. On his network, 1 million tokens of output from GLM-5.2 -- about 750,000 words -- costs $4.40. The same workload on Claude Opus 4.8 is $25 when using Anthropic's pay-as-you-go pricing. In addition to being used in internal coding tools, Chinese open-weight models have been used to power customer-facing agents. Major cloud computing providers, including Cloudflare, Amazon Bedrock and Microsoft Azure, are offering access to Chinese open-weight models on their networks (Amazon Executive Chairman Jeff Bezos owns The Washington Post). Ottawa-based Shopify has experimented with using Alibaba's Qwen to power an AI assistant for vendors on the e-commerce platform. The company said in a blog post in April that the open-weight model was 68 percent cheaper than a closed option. Shopify did not respond to multiple requests for comment. Chinese open-weight models first burst onto the scene in early 2025, when DeepSeek released an open-weight model with similar capabilities to those of the major U.S. labs, but at a fraction of the training and deployment cost. The most recent leader has been the Beijing-based Z.ai, whose GLM-5.2 model has posted performance comparable to everyday models like ChatGPT 5.5 and Claude Opus 4.8 on coding benchmarks (although more heavyweight versions of American AI post better performance). In turn, major U.S. labs have accused Chinese companies of copying their models using a technique called distillation, where a large, powerful model tutors a smaller one. "All of these Chinese AI labs were very cognizant from pretty early on that they were catching up," said Kyle Chan, a fellow at the Brookings Institution. "Part of it is the appeal of being able to have this transparent, downloadable and customizable format." Because the internal parameters of open-weight models can be downloaded and modified, developers can customize them to a far greater extent than they can do for powerful closed models such as ChatGPT and Claude (The Post has a content partnership with OpenAI.). The software company Cursor, for instance, has used Moonshot AI's Kimi K2.5 as the base for its most recent family of models specialized for coding. Starting from an open-weight model allows skipping the expensive step of teaching the machine basic language and going straight to more targeted training when making a specialized model. But in Washington, the popularity of Chinese models has generated anxiety. In April, Republican lawmakers launched an investigation into Cursor's then-parent company as well as Airbnb for their use of Chinese AI models, claiming the models posed a national security risk (Airbnb has used Qwen in its customer service agent, although it maintains it uses mostly American models). And even among companies that use open-weight models, some still steer clear of Chinese ones. AT&T is among a growing number of large companies that use technology to automatically switch to a cheaper, less cutting-edge AI that is still useful for a particular task. Jeremy Legg, AT&T's chief technology officer, said that the company will not use Chinese AI for national security reasons, but he added that the company's rapid-fire AI switching technology does include other low-cost, open AI such as Meta's Llama. The open-weight era of Chinese models might not last forever, though. Alibaba has released several versions of Qwen that are closed and proprietary, and Reuters reported last week that Beijing was mulling restrictions on overseas access to homegrown models. For many companies, open-weight models, regardless of where they're made, are the most secure option. "Being in charge of the weights means that nothing will ever change on you. No behavior will be different," said Ahmad Osman, whose company builds custom AI infrastructure based on open models for companies. "Nothing will be seen by anybody outside of your institution." Shira Ovide and Kevin Schaul contributed to this report.

American AI is expensive. Some startups are turning to cheap Chinese models

SAN FRANCISCO -- Flo Crivello's San Francisco-based startup, Lindy.ai, creates artificial intelligence "assistants" to manage your email and calendar. At first, the company leaned heavily on Anthropic's top-of-the-line AI models. But in meeting after meeting with his finance guy, Crivello said, one thing became clear: "By far, our No. 1 expense was Anthropic," he said. "Like, more than payroll." More than payroll -- for over two dozen employees. More than rent. More than for anything else. So last month, Crivello announced that Lindy had migrated 100% of its traffic to the Chinese AI model DeepSeek-V4. "It was just 10x cheaper," he said, adding that it had saved the company millions of dollars. "So it was a very, very simple business decision." Artificial intelligence has become one of the -- it not the -- fastest-growing costs for U.S. businesses. But for many companies, it's a double-edged sword: necessary but expensive. To survive, a growing number of firms are switching from American models to cheaper Chinese AI. In the race to create the best AI models, U.S. companies like Anthropic, OpenAI and Google lead the world. Experts say Chinese models are six to 12 months behind in terms of capabilities. But China has carved out a niche in open-source models, which are free to download and adapt. "The open-source scene right now is absolutely dominated by the Chinese. It's not even close," Crivello said. He said every founder he knows who is working in the AI space either is thinking about switching to Chinese models or has done so already. And ballooning AI costs are not just a startup issue either. Uber CEO Dara Khosrowshahi spoke about it last month on the Invest Like the Best podcast. "We blew through our AI budget in a quarter, you know, for the whole year, essentially. And it is forcing us to adjust," he said. (Uber did not respond to NPR's request for information about whether it uses Chinese models.) Bloomberg reported Airbnb CEO Brian Chesky as saying that last year the company relied on Alibaba's Qwen model, which was "good," "fast and cheap." Perplexity and Nvidia have also made use of Qwen. Like a Ferrari or a Honda Many companies are wary of trumpeting their use of Chinese models due to political sensitivities, but the models are widely available on AI-model hubs like Hugging Face, on the code hosting platform GitHub and via model aggregators and inference providers based outside China. That includes the San Francisco-based company Featherless, which offers access to some 30,000 AI models. Founder and CEO Eugene Cheah said Chinese models are popular, even if they aren't "frontier" models, or best in class. "It's like the difference between driving a Ferrari and a Honda. You can have the best luxury car, or you can just have a Honda at scale that works," he said. "Actually, a lot of open-source AI groups are perfectly fine being N-1, N being where the frontier is," he continued. "Because as the gap keeps shrinking, at some point the question is: Does it actually matter?" For many, like Lindy, it doesn't matter. The Honda of AI is perfectly good. OpenRouter, another platform where startups can access a range of AI models, reported that use of China's DeepSeek has gone from around 9% to nearly 20% since January. Use of models from the Chinese companies MiniMax, Xiaomi and Tencent have also risen. Some users download and self-host open-source Chinese AI models, but many use them via paid AI-hosting companies, like Featherless and OpenRouter, so that user data is kept in the United States. Victor Su-Ortiz, who does global product marketing at the Shanghai-based MiniMax, attended a recent AI engineers conference in San Francisco. Companies pay to use AI models by paying for tokens, or units of AI work. Su-Ortiz said it all comes down to the cost per token. "A lot of repetitive tasks can be done with a model that's just as performant but has much lower cost per token" when compared with leading AI models, he said. "And this is essentially what has brought these open-weight models into the United States." He said companies are shifting from "tokenmaxxing" -- using as much AI as possible -- to saving costs by limiting usage, switching to cheaper models or routing different types of AI work to different kinds of models. For research or "deep reasoning," for instance, the cutting-edge models may perform better, said Su-Ortiz. "But if you're routing for a coding task that is repetitive, high volume ... that's where one of our models, especially MiniMax M3, will perform exceptionally well at like only one-tenth the cost." Saving a few dollars isn't worth it for everyone For some companies, Chinese models still aren't good enough. Jon Gordner is CEO and co-founder of Comment.io, which was founded just weeks ago and is developing a product that he said is like Google Docs for coders and AI agents. "We need to make as good software as we can as fast as possible. And for us, saving a few dollars on a cheaper model isn't worth it if we have to spend two or three more weeks fixing its mistakes," he said. Gordner said his company is getting value out of Anthropic and OpenAI models in part because both companies are subsidizing users to hook customers. He said monthly subscriptions offer tokens at a huge discount now -- but that probably won't last forever. "Then for us, it's going to make a lot more sense to start evaluating Chinese models and open-source models," he said. Ara Kharazian is lead economist at Ramp, a company that helps businesses track, control and automate spending. It has insight into AI spending, and Kharazian said he thinks U.S. companies will keep adapting -- in other words, they may keep prices in check or introduce high-quality open-source models in a bid to outcompete Chinese rivals. "The rise of these Chinese models is indicative of the fact that businesses want something that is today not being offered by the American model companies," he said. "The only reason why I'm bearish about the Chinese models is because I assume that the American model companies will respond competitively." Gordner, of Comment.io, is less certain. He thinks the major U.S. AI companies may have to start charging more for AI as pressure to demonstrate profitability rises, possibly as they get closer to going public. Both Anthropic and OpenAI have filed confidential paperwork with the U.S. government to get the ball rolling on eventual initial public offerings. "At some point," Gordner said, "the music's going to stop." Anthropic is a financial supporter of NPR.

OpenAI Exec Laments That China Is Giving Away Models So Good That For-Profit Companies Won't Be Able to Compete

Can't-miss innovations from the bleeding edge of science and tech Last week, a Chinese-made large language model called Kimi K3, developed by Beijing-based firm Moonshot AI, burst onto the scene. The open-weight model impressed with early benchmark results, trading blows with some of the most advanced closed-weight ones being developed by the likes of OpenAI and Anthropic -- and at a fraction of the cost. Much like Chinese competitor DeepSeek's AI model upending Silicon Valley in early 2025, Kimi-K3 sent shockwaves across the industry. A major sell-off roiled the tech-heavy Nasdaq, with the S&P 500 slipping after the unveiling. For top closed-weight AI labs, it was the perfect storm. Executives are already balking at rapidly rising costs and desperately looking for cheaper alternatives. Kimi K3 may end up luring them in, making it even harder for the likes of OpenAI and Anthropic to attract new customers -- right as they need to cover at least some of their exorbitant costs. OpenAI's head of strategic futures, Dean Ball, who joined the Sam Altman-led company two weeks ago after helping shape AI policy for the Trump administration, was shaken by what he saw. In a lengthy and controversial tweet on Friday, Ball accused the Chinese state of acting recklessly by allowing models as powerful as Kimi K3 to be open sourced "given potential risks." Ball also argued that "open-weight models are inherently decelerationist," a claim that sparked a raging debate in the comments. In the context of AI, accelerationism advocates for faster AI progress to establish a new world order. Decelerationism, by contrast, argues that the risks outweigh the benefits, calling for a far more careful approach. To Ball, open-weight models hinder the progress of AI by deterring labs from investing more capital in development. As a result, he predicted that the Trump administration would "create large amounts of regulatory risk around the use of open-weight Chinese models," which would in turn generate enough fear, uncertainty, and doubt -- or "FUD," in the lingo -- that would have hyperscalers back off from using Chinese AI. Leaving aside his glaring conflict of interest as an executive at OpenAI, netizens were baffled by his line of argumentation. "The head of strategic futures OpenAI seems a little rattled," one user wrote. "This is really crazy stuff to believe," wrote Alexander Green, founder of AI company Littlebird, who effectively disagreed with Ball on his entire line of thinking. It's not hard to see the motivation behind Ball's defensive stance. OpenAI continues to struggle behind its competitors as AI companies seek massive investments to cover their skyrocketing spending. Powerful AI models like Kimi K3 that subvert the status quo could prove an existential threat without regulatory intervention. "Ball's original post reads less like advocacy and more like the kind of unsolicited geopolitical candor that plays fine on a think-tank blog but lands badly when you are two weeks into a job at the company most likely to benefit," AI agency Lumien wrote in a new blog post. The OpenAI exec's views on how the Trump administration would most likely intervene also didn't convince Trump's AI czar David Sacks. "I'm not sure whether Dean Ball is confessing to a regulatory capture strategy or simply predicting this will happen," he tweeted in response. "Either way, the weaponization of regulatory uncertainty as a competitive tool should be completely unacceptable." "We are at a critical inflection point in AI policy," he added. "The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open source competition." "They have laid their cards on the table," Sacks wrote. "It is time for the rest of Silicon Valley -- the vast majority that still values open competition -- to do the same." Other government officials were far less kind in their assessment of Ball's comments. In one particularly scathing response, US defense undersecretary Emil Michael called Ball the AI industry's "supreme village idiot" -- a shocking, yet somehow unsurprising, tone for a top Trump official.

Businesses are experimenting with cheaper Chinese AI models as U.S. rivals get more expensive | Fortune

AI has become a focal point within the Trump administration, often framed as a two-player race between the U.S. and China. And while U.S. companies like OpenAI, Google, and Anthropic may have developed some of the world's most advanced AI models, they are among the priciest. As costs associated with token and AI usage rise, now some consumer-facing companies are turning to China's cheaper, open-source models. Take for example DoorDash, which, according to a post on X on Wednesday by co-founder and CTO Andy Fang, will be launching DoorDash CLI, an experimental tool in limited beta that will allow users to order DoorDash through an AI agent, or even directly from the terminal. Earlier this month, Fang said using a model from Chinese startup Moonshot AI is "better quality" and comes at a "cheaper cost." DoorDash is far from the first to turn to Chinese AI companies, or Moonshot for that matter. Cursor, the AI coding startup, used Moonshot's Kimi to help build its Composer 2 coding agent, while fellow startup Lindy has reportedly dropped Anthropic's tools altogether in favor of DeepSeek's V4 models, according to the FT. These companies is joining the likes of Airbnb and Siemens -- both of which are experimenting with moving their daily operations to Chinese AI companies like Alibaba and DeepSeek -- to save on rising AI costs. For Yasir Atalan, deputy director and data fellow in the Futures Lab at the Center for Strategic and International Studies, the shift comes down to three factors: cost, capability and the availability of open-source models. "What we're seeing right now is that it seems like the recent high-quality, high-performance models by U.S. companies seem expensive compared to Chinese models," Atalan told Fortune. "The idea of open-source models is much more exciting for some people, specifically countries other than the U.S. for the reason that people don't want to share their enterprise data." As excitement builds around open-source AI, companies looking for more control over their data are embracing Chinese open-source models. Running these models locally can give companies more control over how sensitive information is handled and reduce the need to send proprietary data to outside providers. "It's better for you to host a local model instead of just a closer model because that means everything will stay in that computer and will not go to any company," said Atalan. "Open-source models give that sort of relief to those people who want to keep their data." The approach comes with tradeoffs. "You need to have a very high-level computer in your company, like you paid $30,000 for GPUs, RAM, storage, etc," he said. Cheap at the cost of secure Others in the industry are more skeptical. While some startups turn to cheaper Chinese AI models to cut costs, experts warn they may be overlooking key security risks. Snehal Antani, co-founder and CEO of Horizon3.ai, said in a statement to Fortune that startups adopting such models "risk severe data sovereignty violations by exposing proprietary code and user data to foreign surveillance," while also overlooking "critical vulnerabilities in model integrity and reasoning." Still, Atalan cautioned against viewing the trend as a wholesale migration to Chinese AI models. Rather than replacing U.S. models outright, he said companies are experimenting with alternative models for different tasks. "A company could try to use one of those open-source models for one task and use Claude for something else. That's very plausible," he said. Few companies have publicly disclosed using Chinese AI models, but they are widely available through platforms like code-hosting site GitHub and AI-model hub Hugging Face, where developers can upload, download and run open-source models. A March 16, 2026 study from Hugging Face found that Chinese open-source models accounted for 41% of downloads. However, lower cost doesn't eliminate risk. Even as companies experiment with cheaper models, questions remain around security, data control and how these systems perform in higher-stakes use cases. For many companies, the decision may come down to cost and capability, not country of origin. As Atalan suggests, if a model is "cheap and capable enough" and can be run locally, businesses are likely to use it regardless of whether it came from the U.S. or China.

Hugging Face uses open weights Z.ai GLM 5.2 to defend against attacker after commercial frontier model refusal

Hugging Face Inc., an open-source artificial intelligence platform often described as the "GitHub of machine learning," found itself forced to use an open-weights model to respond to an agentic AI attack after the safety guardrails on commercial AI models blocked requests. Last week, Hugging Face said it detected a breach from an attacker using an autonomous AI agent system to access a limited set of internal datasets and several credentials used by internal services. The company responded defensively, cut off the attacker and hardened the system. As part of the analysis, Hugging Face went to frontier AI models to assist with log analysis - a completely sensible option given that Hugging Face is a premier access point for numerous AI models. However, this failed: this kind of analysis requires sending a tremendous amount of real attack data and commands, exploit payloads and attack artifacts. These requests were blocked by commercial frontier AI models because their guardrails cannot distinguish between being asked to build exploits for an attacker and a defender trying to detect them. With a need for rapid analysis and speed at the onset, the company switched to Z.ai Co. Ltd. GLM 5.2, a powerful open-weight model with about 753 billion parameters. Unlike a third-party model, it can be run entirely on local or cloud hardware and within a company's protected firewall perimeter, meaning no data exits its controlled infrastructure. Chinese-built AI models such as GLM 5.2 and Beijing Moonshot AI Technology Co. Ltd.'s Kimi K3 have proven that open-source and open-weight models can reach, or even rival, the current forerunner and flagship frontier models built by American companies. GLM 5.2 reaches the capabilities of Anthropic PBC's Fable 5, a Mythos-class model capable of advanced reasoning, coding and even discovering vulnerable code and exploits. It also does so with far cheaper inference costs than delivered by Anthropic. However, to prevent the misuse of these models, Anthropic and other leading closed-source AI developers have put strong safety guardrails in place that trigger higher false positives to avoid misuse. This also makes them far less capable overall for real-world usage in valid cybersecurity roles. Anthropic has noted that the false positive rate is being adjusted as it works to make its frontier models safer for use by researchers. Although companies such as Anthropic and OpenAI PBC Group have voluntarily placed these restrictions on their most powerful models, Mythos 5 and Fable 5 were both pulled last month at the request of the United States government shortly after they first launched. Similarly, the U.S. government asked OpenAI to delay the release of its own frontier model family GPT 5.6, which is now publicly available. Tension between Chinese open source and American closed source models The recent release of Kimi K3, a powerful near-frontier-class open-weight model from Beijing-based Moonshot AI, has fueled rumors that the Trump administration may ban U.S. companies from using Chinese open models. Parts of the administration have already attempted to construct de facto bans on foreign open-source models before, according to Axios. This comes at a time when many U.S. companies are increasingly using Chinese open-weight models because they are cheaper to deploy and come close to rivaling commercial-only models. The U.S. would not need to ban the use of Chinese models outright; instead, the government could use pressure campaigns to steer corporations away from them, citing a lack of security and governance. According to sources, the U.S. Commerce Department also considered adding multiple Chinese AI labs to its "Entity List" last year, which would effectively cut off access to companies without proper licensing. David Sacks, the White House AI and crypto advisor, wrote on X, formerly Twitter, on Sunday: "We are at a critical inflection point in AI policy. The leading closed labs, already a duopoly in terms of AI model revenue, want the government to eliminate their open source competition." In a two-part X post, noting the recent Hugging Face event, Sacks added: "There's no reason to limit American models on tasks that Chinese models handle without issue. We're only making ourselves less competitive." In his commentary, he mentioned that Kimi K3 successfully fixed 15 critical security bugs that he noted OpenAI and Anthropic's models refused to because of "cyber guardrails." The 15-bug fix is a claim from a developer who used the model, not a known benchmark, but it has become a focal point for developers to note that Chinese models are becoming a mainstay over U.S. closed-source models. The developer added that the whole affair also only cost $250. Cost comparisons would be difficult, if impossible, without knowing the exact work done, but Kimi K3 is around one-third the price of Fable 5. Of course, the price distinction doesn't matter if Fable 5 refuses to do the work. Noting that the world is beginning to turn towards its models, China's President Xi Jinping called for global cooperation. "The development of artificial intelligence should not be a solo performance by any single country but rather a symphony of global cooperation," Xi said at the opening of China's annual World Artificial Intelligence Conference in Shanghai. In his speech, he called for opposing the "overstretching of the concept of national security" as it relates to AI. China has been historically blocked from using some of the most advanced AI in the U.S. and restricted from state-of-the-art AI chips. China intends to expand AI cooperation with the Association of Southeast Asian Nations, the League of Arab States, the African Union, the Community of Latin American and Caribbean States, the Shanghai Cooperation Organization and the BRICS countries. Over the next five years, Xi said China will provide 5,000 AI training opportunities for developing countries. He also said China will provide 30 countries access to a Chinese-developed meteorological tool for early warning systems.

US Companies Are Realizing That Chinese AI Models Are Way Cheaper, Ditch American Ones

Can't-miss innovations from the bleeding edge of science and tech As corporate AI bills spiral out of control, many companies are beginning to ask themselves a simple question: why pay a pretty penny for the US's leading AI models when Chinese ones are far cheaper? Major companies like DoorDash, Airbnb, and Siemens are adopting Chinese AI tools, the Financial Times reports, attracted not only by their lower costs but their "open-weight" approach that allows them to be molded to each company's particular needs. According to data from OpenRouter, a platform that provides all-in-one access to major AI models and tracks their usage, leading Chinese models from DeepSeek and Z.ai have overtaken US equivalents like Anthropic's Claude and OpenAI's ChatGPT. Cost-cutting, it seems, trumps all geopolitical rivalries. Chinese models are "the elephant in the room," Eugene Cheah, CEO of the AI platform Featherless AI, told the FT. "Enterprises are starting to realize, 'Hey, we don't need the best model, we can use the faster, cheaper models.'" US-based AI models have frequently been seen as the most advanced, but that perception is shifting. The release of GLM-5.2 from the Chinese startup Z.ai last month caused a stir in Western tech circles, as major Silicon Valley figures hailed it as capable or nearly as capable as US systems despite being significantly cheaper to use. Cheap Chinese AI couldn't be coming at a more opportune moment. The corporate world, wooed by AI companies' promises of supercharging their productivity, has spent the past year deploying AI across its workforces, and many are being put off by the costs. One organization reportedly blew through $500 million in a month on Claude usage fees. That's an extreme outlier, but recent research from the Ramp AI Index found that the businesses most dedicated to AI are spending around $7,500 per employee every month on AI. Considering the culture around AI, it's no surprise why: some companies like Meta mandate their employees use AI systems as much as possible, factoring it into their performance reviews. Software engineers, now expected to produce more work than ever, often run multiple AI agents at the same time to complete tasks in the background. If companies are unwilling to crank back the AI knob, then the next best choice is to look for cheaper models. DoorDash cofounder Andy Fang said on X last week that it was saving a lot of money by having "lower-level work" performed by a model from the Chinese startup Moonshot AI. San Francisco startup Lindy has completely ditched Anthropic's AI tools in favor of DeepSeek's latest V4 models. "Enterprises have an incentive to shift some of their workload to cheaper models. Why would you pay a premium for Anthropic, OpenAI models when for a lot of the workloads you need, the Chinese models are generally workable?" Sam Bresnick, a research fellow at Georgetown University's Center for Security and Emerging Technology, told the FT. But cost isn't the only consideration: many Chinese models are "open-weight," meaning their parameters or values are entirely visible to the user. That allows an organization to mold a model to its specific needs -- and from a cybersecurity perspective, have more control and insight into how it might process sensitive company data. For foreign companies disillusioned by US leadership, the choice is even easier to make. There's less faith in the US as stewards of AI, especially after the Trump administration suspended access to Anthropic's Mythos model overseas. "The Mythos ban was certainly the most tangible event, and people having their access revoked," Aidan Gomez, CEO of the Canadian AI group Cohere, told the FT. "It exposes the risk of relying on any one single entity for any of your workloads."

Chinese AI: US panic is missing the point on Chinese AI

Since last year, I've been tracking the trend of American startups using Chinese AI. This year, the shift has become harder to ignore. But it didn't emerge in a vacuum. China's low-cost, open-weight push was always going to appeal to developers, the backbone of AI innovation. Washington's wake-up call arrived late. The rising use of Chinese artificial intelligence models in the US has led to growing anxiety among Silicon Valley executives, Washington lawmakers, and even officials in Beijing. But the warning lights have been flashing for a while. Since last year, I've been tracking the trend of American startups using Chinese AI. This year, the shift has become harder to ignore. But it didn't emerge in a vacuum. China's low-cost, open-weight push was always going to appeal to developers, the backbone of AI innovation. Washington's wake-up call arrived late. At the heart of the latest flashpoint is a debate about distillation, or training one model on the outputs of another, and a growing attempt to discredit Chinese competitors as stolen goods. Yet the technique is widely acknowledged as inevitable in the industry. Elon Musk told a federal court earlier this year that it is "standard practice" and something "generally all the AI companies do." Some observers have also questioned whether US AI companies are learning from Chinese competitors. But accusations from US frontier labs that Chinese companies are illicitly copying them in violation of terms of service have given the issue new political force. A method that was once a breakthrough known as "knowledge distillation" is being rebranded as "knowledge theft." Silicon Valley seems to be hoping that this line of attack is enough to stop people from adopting these tools. There is a rich irony here. The same companies that swept up the works of billions to build tools threatening artists, writers and other creatives are now feeling the same anxiety of witnessing their own economic leverage erode. But putting that aside, this battle over distillation is raising a bigger question. In the long run, can knowledge and information be contained by a handful of US AI companies in a technology race increasingly defined by diffusion? A business model built on this looks increasingly fragile as millions of users, engineers and labs around the world are incentivized to learn from and iterate or be left behind. Preventing distillation isn't easy, and current technical defenses are insufficient. As some researchers point out, it's not like physical hardware controls, which means fully restricting it seems almost impossible. Some attempts have already backfired. Anthropic recently rolled back hidden code it embedded in Claude Code to track users' locations after criticism that it was invading the privacy of all. The controversy prompted Beijing to issue a rare warning last week that the tool poses a serious threat. Claude Code, notably, was already barred for use in China. But a grey market of middlemen and so-called "transfer stations" has emerged, Zilan Qian, a research associate at the Oxford China Policy Lab, points out. Whether all of these accounts -- held by students, professors and everyone behind the Great Firewall seeking to try the best AI tools -- are being used for systematic, surreptitious distillation is likely overblown, Qian added. With technical defenses lacking, it makes sense that companies are turning to the US government for help. For now, lawmakers seem to be attempting to inflict reputational harm on Chinese models. But it's very likely that tighter restrictions are being debated on both sides of the Pacific. When Washington abruptly curtailed access to Anthropic's top models for all foreign nationals, including some of the company's own staff, it suggested that more stringent regulation could be on the horizon for Chinese AI. (It reversed that decision soon after.) And Reuters reported last week that Beijing was also considering restrictions on overseas access to China's most advanced AI models. But the reality is that if either side clamps down on America's rising use of Chinese models, the pain will be felt the most by US developers. A strong ecosystem of researchers, startups and talent is something that Washington should be supporting. A short-sighted attempt to protect current American AI leaders could make it harder for the next champions to get built. Rather than banning Chinese offerings, the US should build better alternatives. That means serious investments in open-weight projects, public compute, working with universities and offering startup-friendly options. The cost of these models is appealing, but so is the control they offer. Developers can fine-tune, run and build on them locally without being fully dependent on shifting whims of closed frontier labs. That flexibility is key for developing future AI products and spreading the technology more broadly throughout the economy. Company leaders and investors should also focus on other moats, including institutional trust, computing resources and better support for paying enterprise customers that make users not want to switch over. Labeling Chinese AI models as rip-offs has become a convenient talking point. As I've written before, the argument that these labs can only compete due to copying is tired. Worse, it risks leaving the US caught off-guard again by real innovations, including techniques that let models use compute more efficiently. This exchange of knowledge is helping American labs now, especially amid the outbreak of a price war. Washington can try to protect today's AI giants by rebranding their business-model problems as a national security issue, or it can focus on cultivating tomorrow's champions by supporting open research. It will be hard to do both.

Chinese AI Models Overtake US Rivals as Token Share Among American Firms Hits Record 58%

Chinese artificial intelligence models now account for the majority of tokens processed by U.S. firms on OpenRouter, overtaking American rivals for the first time as developers increasingly adopt offerings from companies such as DeepSeek. Chinese Models Gain Majority Share According to data shared by The Kobeissi Letter on X on Sunday, Chinese AI models accounted for a record 58% of tokens processed by U.S. firms on the platform. The share has nearly tripled since mid-January and briefly reached 63% during the first week of July. At the start of 2025, Chinese models accounted for less than 10% of usage, while U.S. models represented roughly 80%. OpenRouter allows developers and companies to access and compare AI models from multiple providers, making token usage a closely watched gauge of real-world AI model adoption. DeepSeek And Kimi Lead China's AI Push The Kobeissi Letter said DeepSeek has emerged as the most widely used Chinese AI model among U.S. firms in recent months. "Chinese AI models are rapidly gaining market share," the market commentator added. Last week, Chinese startup Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter model at the World Artificial Intelligence Conference in Shanghai. The company said it is the world's largest open-weight AI model, with early benchmark rankings placing it ahead of several leading U.S. models from Anthropic and OpenAI in some capabilities. AI Race Sparks Fresh Debate On Chip Demand The launch of Kimi K3 weighed on semiconductor stocks Friday as investors questioned whether increasingly capable and lower-cost open-weight AI models could reduce the computing power needed to train and deploy advanced AI systems. David Sacks, who chairs the President's Council of Advisers on Science and Technology, called Kimi K3's benchmark performance "concerning," while Billionaire investor Bill Ackman chimed in, saying he "agreed." Disclaimer: This content was partially produced with the help of AI tools and was reviewed and published by Benzinga editors. Photo courtesy: Alexander56891 from Shutterstock Market News and Data brought to you by Benzinga APIs To add Benzinga News as your preferred source on Google, click here.