Anthropic's Fable Shutdown Accelerates Shift to Open-Source AI and Chinese Models

Anthropic's Fable shutdown is a big moment for open-source AI

Chinese open-source AI names, including Zhipu and MiniMax, surged as investors bet the fight could boost demand for open models. Anthropic's suspension of its top AI models late last week drove home a hard truth for the companies that were counting on them: access can be cut off at any time. It's been a big theme this week, and one that Wall Street will be watching closely as Anthropic and OpenAI gear up for potentially massive IPOs in the coming months. Microsoft CEO Satya Nadella is warning of the risks, even as his company is the principal investor in OpenAI, and backed Anthropic last year to the tune of billions of dollars. He wrote Monday in a post on X that companies need to "build agentic systems that improve over time, while still retaining control over their IP." "The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see," Nadella wrote. Investors have been trading on that theme. MiniMax and Zhipu, the Chinese open-source AI lab, both surged on Monday as the Anthropic fight put a spotlight on downloadable models that companies can run themselves. The Anthropic-Fable news landed at an awkward moment late Friday. Roughly two hours earlier, SpaceX had wrapped up its first day of trading following the biggest IPO on record. SpaceX's xAI unit is a niche player in artificial intelligence, but CEO Elon Musk is an outspoken voice on the topic. Anthropic announced it had pulled access to its Fable 5 and Mythos 5 models to comply with an export control directive from the U.S. government that cited "national security authorities." Anthropic abruptly disabled the models for all of its customers in order to ensure compliance, but said all of its other models would not be affected. For developers who want full control over model access, there's another approach. An open-source model can be downloaded, run on a company's own infrastructure, and customized for its data and needs. When the model lives on a company's own servers, no political fight can switch it off. Yash Patel, CEO of Applied Compute, which helps companies train and run custom models, said the Anthropic fight "highlighted the significance of owning your own model." He said the shift has become much more mainstream of late. "What we've been hearing increasingly, probably more so in the last month than the entire year, is the fact that they want a multimodal future," Patel said. "They don't want to be locked into a single vendor." That could be a problem for the U.S., as the the open models winning the most adoption are from China, just as the world's two biggest economies are in a battle over controlling the future of AI. Models from DeepSeek, Tencent, Xiaomi, and MiniMax all rank among OpenRouter's most-used this month, even against closed competitors. Zhipu framed its latest release as a rebuttal of sorts to Washington. Cutting-edge AI, the company argued, shouldn't belong to a handful of players or be withdrawn at will. Cost could also speed adoption. As the price of state-of-the-art AI climbs, companies are already routing routine work to cheaper models and saving the most expensive ones for the hardest tasks. Patel said customers are reacting to what he called a "token-pocalypse" as AI products move toward usage-based pricing. "The era of token maxing is over," he said. Companies are now looking for "better, cheaper, faster models." That is pushing some enterprises to reconsider models they would have dismissed months ago, including open models from China. "Before it was just kind of like I don't even want to talk about it," Patel said. "Now they're like, OK, how good could it be, and if it's good, we'll figure it out." It's a reminder that the AI market is still in its infancy, with ChatGPT's public release occurring less than four years ago. For investors, that reframes who is actually leading the AI race. The winners may not just be the big closed model labs, despite what their current valuations suggest. Choose CNBC as your preferred source on Google and never miss a moment from the most trusted name in business news.

Axios C-Suite: Open-source AI pits cost against security

Why it matters: Unlike closed models such as Anthropic's Claude or OpenAI's ChatGPT, open-source AI is free to download, cheaper to run, and dominated by China. Between the lines: Microsoft exploring a secured version of China-built DeepSeek V4 to power Copilot Cowork -- its agentic assistant and the most compute-hungry part of Microsoft 365 -- shows that even the richest software company on Earth can't hold the line on closed American-only models. You probably can't either. The security catch: Microsoft's "secured" version would keep your data on its cloud platform Azure, not Chinese servers, which means the open-source question is shifting from if to how. * But many vendors and tools don't offer that protection. That's why every CEO needs to be wary before approving any model. (See next item.) "Open source" means the model weights -- the actual trained intelligence -- are public. Anyone can download, modify and deploy them. The best-performing, most-used ones are Chinese. * DeepSeek V4 and other leading Chinese models are considered a few months behind frontier American models performance-wise. But that gap is closing fast with buzzy new models like Z.ai's GLM 5.2. * Smart CTOs and engineers are already handing off routine work to open-source models and saving pricier models for harder, more complex tasks. The part you may not see: Many CEOs reading this are already running Chinese models somewhere in their stack and don't know it, because their vendors and engineers chose them on price. Worth watching: Flo Crivello, CEO of AI agent platform Lindy, switched his company entirely to DeepSeek V4, citing millions in savings and better performance on core use cases -- while keeping it on American soil, hosted by a U.S. provider rather than Chinese sources. * Even so, the migration took months and far more engineering than expected. The bottom line: Open-source Chinese models are good enough for most work at a fraction of the price. But you've got to be clear-eyed about their use. The Chinese way Dominating open-source AI is China's national strategy. Why it matters: Your company likely already faces thorny China-centric compliance questions -- especially if you've staffed your tech team with AI enablers. * Every open-source model your developers run from a Chinese lab ties part of your infrastructure to Beijing, whether you intended it or not. Between the lines: The fear isn't just that Beijing has a potential backdoor into your company. It's the regulatory threat to your entire workflow as U.S.-China competition escalates. * Bringing Chinese open-source models into the fold means you're using a product built by companies legally required to cooperate with Chinese state intelligence. * You can't be sure their future decisions won't be China-centric. They may propose ways to connect with their models that force a decision: Build for open source or tie to a frontier model. * I'm not arguing against open source -- just illuminating the reality we face. We've been here before: Banning Huawei from the telecom networks of the U.S. and its closest allies cost billions to rip out and replace. * Imagine the cost if companies across the West had to reconfigure their AI stacks overnight. Zoom out: This isn't a fringe risk. Chinese companies -- DeepSeek, Xiaomi, MiniMax, Tencent, Alibaba -- now dominate OpenRouter, the industry's most honest real-time scoreboard of developer usage, with DeepSeek at No. 1. * The models you have to worry about are the same ones your team most wants to use. Worth noting: Not all open models are Chinese. Meta's Llama and Europe's Mistral are credible alternatives with no ties to Beijing. 📋 A CEO checklist: All of us should ask and know ... America's competing interests America made the opposite bet of China: We went all-in on our ahead-of-the-pack frontier proprietary models. The big picture: The U.S. incentivized its frontier AI labs to optimize for revenue -- with eyes on $1T+ IPOs. That pushed them toward subscriptions and closed ecosystems where every use can be metered. Meta was the major open exception on American shores. Meta's 2023 release of Llama 2 was a breakthrough, and Nvidia CEO Jensen Huang later called it "probably the biggest event in AI" from that year. * But Meta's main business was already built around ads, allowing it to push for an open AI ecosystem it could shape. * Then Meta got hit with a huge cost of compute -- and shifted its strategy away from the open ecosystem it once championed to play catch-up with Anthropic and OpenAI. * Its newest model, Muse Spark, is closed-weight. And it's now testing paid AI subscriptions. The bottom line: In the U.S., economics mean the AI feedback loop has to be mostly proprietary: Each lab learns from its own users, then sells the improvements back to them. The complicated counterattack Open-source advocates think American frontier labs are begging for regulation in order to corner the AI market. Their claim: Costly compliance requirements -- think: safety audits or liability frameworks -- would be trivial for companies like Anthropic or OpenAI and fatal to any open-source project. * With enough regulation, the market would be forced to consolidate around the companies the government deems safe. Inevitably, that'd be the most powerful frontier labs. What they're saying: Former White House AI czar David Sacks, who still has President Trump's ear on cutting-edge issues, outlined this "regulatory capture agenda" on last weekend's episode of the "All-In" podcast. * "You've got [Anthropic CEO Dario Amodei] who's out there describing these risks in a very hyperbolic way, and it's very clear where their agenda is headed, which is ... banning open-source models," Sacks said. * "We're gonna be stuck with somewhere between one and three companies -- maybe two -- they'll be an AI monopoly or duopoly and they're gonna decide, along with some new government agency, which'll be a revolving door to their companies, who has access to what capabilities," he continued. Yes, but: Sacks was bemoaning the guardrails Anthropic had established around discussing cybersecurity and biology with its powerful Fable model. But the episode was recorded before the Trump administration issued export controls that caused Anthropic to shut down consumer access to it. * Anthropic has had its share of spats with the Trump White House. But the move is a massive alarm bell for the closed frontier labs themselves. * They profit from having the best models, but now risk the government cutting off consumer access when that goal is achieved. The bottom line: For the moment, the frontier labs are feeling the sting of government regulation long before any open-source model creators. The case for open source I asked Misha Laskin, co-founder and CEO of Reflection AI, an American open frontier lab, to make the pro-open-source case to other CEOs. * Reflection, backed by Nvidia, says it has "identified a scalable commercial model that aligns with our open intelligence strategy." * Demand is clearly there. Its valuation jumped from $545M in March 2025 to $8B in October 2025 to a potential $25B in a March fundraising round. It hasn't yet released a model. Open models are a geopolitical priority. They're Trojan horses for the infrastructure they bring with them. When a nation or enterprise adopts an open model, it also adopts the chips and software stack that come with it natively. 📈 If you're a CEO or on a CEO's team: Ask to join Jim's new weekly Axios C-Suite newsletter.

Anthropic's Fable fiasco leaves the door open for open-source AI, particularly cheaper models from China | Fortune

The U.S. government's decision to stop Anthropic from offering its Mythos and Fable 5 models to non-U.S. nationals may end up providing a big boost to the adoption of open-source models, including those from Chinese AI labs like DeepSeek and Moonshot AI. Users can download open-source models and run them on their own computers or cloud networks, effectively sidestepping the ability of both AI developers and governments to control access. These models can also be more easily fine-tuned by developers to tailor them for specific needs. Chinese labs are already claiming a public relations win from the Anthropic controversy. Shares in Knowledge Atlas, a Chinese AI lab better known as z.ai, surged by over 30% in Hong Kong trading on Monday after it released the latest version of its open-source model, GLM-5.2. (Knowledge Atlas's shares are up more than 800% since they debuted in January) "At a time when some frontier models can suddenly become unavailable, we choose to believe in a different path," Knowledge Atlas posted on social media, according to the South China Morning Post. In a clear reference to the Anthropic news, the company added that "frontier intelligence should not belong to only a few people, nor be subject to withdrawal by a handful of rules at any moment." Demand for Chinese models has already overtaken that for U.S. models on OpenRouter, a popular platform for accessing different AI models. Last week, the top four most-used models came from Chinese companies: DeepSeek, MiniMax, Tencent and Xiaomi. The Chinese open source models have proved popular not just within China but also in many other developing countries around the globe, where they are seen as a good trade off between price and performance. The U.S.'s ban on Fable and Mythos may also end up vindicating China's broader move towards tech self-sufficiency, which picked up in 2022 after the Biden Administration placed controls on the sale of advanced chips and chipmaking equipment. "It's a great move for China," says Neil Shah, vice president of research at Counterpoint Research. "Obviously they're not on the cutting edge because of the export controls, but they have their own silicon and their own software." Why go open-source? On Friday, Anthropic revealed that the U.S. Department of Commerce had ordered it to stop providing access to its frontier models to anyone outside of the U.S. The way U.S. export rules are interpreted also means the company cannot offer the models to any "foreign national" inside the U.S., including its own employees. In response to the government order, the company decided to suspend access to these models to all users. Anthropic had previously argued that its Mythos model was too powerful to be released to the public without safeguards, and had embarked on an early-access program, titled Project Glasswing, for key institutions to use the model to uncover security vulnerabilities. Institutions in about 15 countries, including U.S. allies like Japan and South Korea, eventually got access to Mythos through Project Glasswing. But the U.S.'s move against Anthropic raises the possibility that frontier models from other labs, like OpenAI or Google, might also get hit by export controls. In that event, non-U.S. organizations may be completely locked out from accessing the best U.S.-developed models. Open-source models could be an alternative, particularly for governments hoping to invest in sovereign AI, domestically-developed and -controlled AI models and infrastructure. The U.S.'s export controls on Anthropic only highlights the danger governments have from being locked in to one country's AI models. "It is the first time that a government has ordered a model developer to restrict access to a particular model based on nationality," says Paul Triolo, a partner at DGA-Albright Stonebridge Group. "Companies and governments will start reconsidering how they are approaching application development based on a particular model, and for governments, which companies they will want to partner with for sovereign AI deployments." "Until there is further clarity about what criteria the U.S. government will use in assessing and approving frontier models, companies and governments will definitely be exploring options such as non-U.S. origin models," such as those from Mistral, Cohere, and "capable Chinese open-source models," he adds. The Anthropic order will "push scale for Chinese open-source models," Shah says. "But we'll also see lots of ambitious and self-sufficient economies, like in the Middle East, who will try to build their own indigenous software models." Asian governments in particular have made a public push for "sovereign AI." South Korea, for example, launched a national state-backed competition to develop Korean-language AI models. "We need to advance our own technology as quickly as possible and become as self-reliant as we can," Sung Kim, the founder of Korean AI startup Upstage, said at a press conference on Tuesday, adding that AI was now a "strategic national asset." Japan, for its part, is suggesting that it might turn to Anthropic's arch-rival OpenAI to bolster its cybersecurity defenses. How good are China's open-source AI models? Neither OpenAI nor Anthropic make their models available in China, including the Chinese city of Hong Kong (which sits outside Beijing's internet controls). Both Anthropic and OpenAI have accused Chinese labs like DeepSeek of conducting "distillation" attacks, where their models are used to train smaller, more efficient models. Chinese models still lag models from the U.S. DeepSeek's most recent model V4 performs at approximately the same level as Anthropic's Claude Opus 4.6 and OpenAI's GPT 5.4. Those models were released in February and March 2026, respectively. The Chinese startup estimated it was three to six months behind state-of-the-art frontier models. However, while Chinese open-source models are not as powerful as their U.S.-developed peers, they are still significantly cheaper. DeepSeek's V4 Pro cost $3.48 for 1 million tokens of output; Anthropic's Fable 5 model cost $50 for the same output. (A token is the basic data unit that contemporary AI systems, most of which are large language models, process. It is equivalent to about a word-and-a-half of English text.)

China isn't trying to beat the U.S. at AI -- it's playing a completely different game | Fortune

In early 2025, a quiet evolution emerged out of Hangzhou -- a city of thirteen million famed for its historic West Lake, misty mountains, and poets who depicted it as "Heaven on Earth." Yet it was here that a Chinese company -- under-resourced by Silicon Valley standards -- released a reasoning-focused large language model in January 2025, a few steps behind that era's ChatGPT 4-class models. DeepSeek had been trained, its parent company claimed, for only $6 million -- pocket change compared to OpenAI or Google expenditures. For a global industry convinced that US large language models (LLMs) were unassailable -- built by companies with oceans of compute, elite talent, large capital infusions, and company valuations the size of small-nation GDP -- DeepSeek upended assumptions. The Chinese model emerged at a time that China's AI sector seemed beleaguered by chip bans and a slowing economy and against a backdrop in which private sector investment into the US sector was twelve times China's and twenty-four times the UK's. Analysts found that DeepSeek's parent, High-Flyer, hadn't spent its way into contention; instead, it had been crafty: focusing on targeted domains, lean training, and aggressive energy efficiency. While critics whispered about secret Nvidia chips, covert ChatGPT usage, or hidden costs, the signal was clear: US companies wouldn't hold a monopoly on advanced LLMs. This "DeepSeek moment" ignited a debate: "Are Chinese companies catching up in AI?" That framing misses the point. The AI stack is broad -- spanning energy, infrastructure, chips, foundational research, and application deployment -- and across that landscape the US and China possess different strengths. The real picture is that the world's two dominant AI markets aren't competing in the same race; they're running parallel ones shaped by different economic constraints, cost structures, and market demands -- and that divergence will shape global AI over the next decade. "If you look at the future of deep tech, it's clear that the US and China are like the two sides of [the traditional Chinese martial arts practice] tai chi, with its black-and-white yin and yang symbol -- each with unique strengths, each pushing the other forward," says Soul Capital's Herry Han, "It's not just competition. It's a dynamic balance." In the US, AI development follows a capital-intensive path: massive data centers, soaring talent costs, and billion-dollar frontier science bets. Well-funded US leaders will keep pushing toward artificial general intelligence (AGI) and increasingly sophisticated agentic systems designed to pursue goals through multi-step planning and self-directed action. Chinese companies face a different competitive reality: relatively tighter capital, more limited access to high-end computing power, and a smaller domestic profit pool. Its companies respond by leaning hard into open-source, cost efficiency, fast application layer innovation, and global markets for monetization. China's AI story isn't about catching up or winning the frontier model race but rather by industrializing the use of AI. Chinese AI firms are building an alternative ecosystem that's leaner by necessity, more open by design, and working to wire into the economy. This is prompting its adoption across start-ups, universities, and mid-market firms worldwide. As global industry leaders continue a preference for keeping their most capable models "closed-source" -- which keeps source code proprietary or secret -- Chinese models are diffusing through easy access and utility, quietly shaping global AI practices and creating a new, technical form of soft power. China's AI Efficiency Machine: Low Costs and Constraints An equivalent dollar in China goes farther -- a fact that reframes the AI investment picture. While the US dominates total AI funding by a wide margin, Chinese firms often produce more output per unit of capital. This efficiency begins with input costs. AI engineers earn about 402,000 RMB (about $57,000) per year, far below US salary norms. China also trains one-and-a-half to two times as many AI-relevant PhDs as the US, and many trained researchers are returning home, creating a large, affordable talent pipeline. Data centers benefit from cheaper electricity, discounted land, and aggressive local subsidies -- reinforced by national policies that treat large-scale computing as strategic infrastructure. In some provinces, electricity costs are halved for facilities using chips. These advantages are further sharpened by the constraints that characterize China's AI ecosystem. Foreign direct investment into China has fallen more than two-thirds since 2019, access to Western markets has tightened, high-end domestic GPUs lag behind imported ones, and regional grids are straining under data center loads. Moreover, China's business-to-business (B&B) market is characterized by a low willingness to pay for services; firms routinely build software tools in-house rather than buy them. The result is a domestic software market roughly one-quarter the size of the US's $237 billion industry, making it difficult for many AI start-ups to reach scale or charge premium prices. In addition, large corporation culture -- especially state-owned enterprises -- prizes headcount and incremental change, resists automation, and blunts AI's potential impact. Angel investor Jun Xu captures a core monetization constraint: AI's total addressable market -- the revenues available for a company's product or service -- tracks the cost of white-collar labor. "And that pool is simply much larger in the US and other developed markets than in China because salaries are much higher," Jun says. "China's AI problem isn't chips or models or supply -- it's demand. Demand is cheaper and smaller." These pressures have pushed Chinese firms toward efficiency and ingenuity. Nvidia's Jensen Huang said in 2025 that restricting US chip sales to China would only accelerate China's domestic push. "Local companies are very, very talented and very determined," he said, "and the export control gave them the spirit, the energy, and the government support to accelerate their development." Altogether, the ecosystem has created a set of Chinese open-source LLMs operating along the "efficient frontier," delivering leaner architectures and stronger reasoning at modest compute levels. By late 2025, DeepSeek's parent announced that one million units of output could be had for about 3 RMB -- that's about fifty cents and a twentieth the cost of ChatGPT at the time. This combination of high efficiency, thin margins, and a smaller domestic monetization pool shapes how Chinese AI companies scale. Many are becoming what we call "skinny athletes" -- lean, fast, and relentlessly efficient -- which not only affects how they compete at home but also how and where they grow. For many, that increasingly means serving customers abroad, and the use of generative AI means that language, localization, and customer support are no longer decisive obstacles to global expansion. "All AI start-ups have an international strategy from day one in China, and Southeast Asia is top of the list," says Cindy Chow, CEO of the Alibaba Hong Kong Entrepreneurs Fund, which recently launched a new fund dedicated to AI application start-ups. "In fact, many Southeast Asian conglomerates and financial institutions are keen to invest with us, as they believe that the region won't catch up in AI development on their own, and that accessing advancements from China is key to staying competitive." The Chinese AI unicorn 01.AI shows what global expansion looks like when delivering value is the organizing principle. Founded by Kai-Fu Lee, an ex-Google China chief and president of leading tech VC Sinovation Ventures, the company started as a pure LLM developer in 2023 but quickly shifted toward enterprise AI agents, or what they call "super-employees" designed for functions such as insurance brokering, procurement, and logistics optimization. "Monetization challenges force AI to accelerate faster," says Ning Ning, 01.AI's vice president of international business and AI consulting. "The future of enterprise AI is not about selling technology but rather making AI accountable for business outcomes." In practice, delivering value means embedding "24/7 digital specialists" directly into enterprise operations. In one deployment with a Perth, Australia-based mining company, 01.AI's engineers positioned its AI agents as "teammates" alongside employees: a logistics scheduler to optimize rail and port traffic; a procurement agent that reads vendor emails and generates purchase orders; an operations planner to juggle trucks and crews. Each agent retrains itself as conditions change, says Ning, continuously improving its performance. China's Open-Source Bet: Diffusion and Collaboration China's open-source culture has become an accelerant. Hundreds of teams now openly release their model architectures and weights -- the blueprints and parameters that a model acquires through training -- creating a shared infrastructure anyone can examine, build upon, and improve. Chinese AI firms are pragmatic, says Chloe Fang, a founder in the text-to-image/video space. They aim for AI that delivers value and often use open-source "as a global hook to attract global users," she says. "They start building brand equity and word of mouth -- and then release better closed-source models later on." Chinese models now lead in several key generative media categories, accounting for five of the top ten image-to-video models globally and three of the top ten text-to-video and image-editing models. This open-source approach aligns with national priorities emphasizing "open ecosystems," deep integration with all industries and sectors of the economy and society, and "increased global cooperation"Chinese open-source LLMs -- led by Qwen, MiniMax, and DeepSeek -- now account for one-third of global LLM usage, up from virtually nothing in late 2024. Start-ups and companies from Silicon Valley to Africa and Southeast Asia now run Chinese models because they're accessible, transparent, and far cheaper to run than many US alternatives. You can remove anything you want, add anything you need. That flexibility and openness actually gains trust," says Sinovation Ventures founder Kai-Fu Lee, an ex-Google China chief and an influential AI voice. He points out many US institutions, students, and researchers are using Chinese models "not because they're Chinese but because they're open." Jian Wang, founder of Alibaba Cloud, argues that this moment in AI echoes the late 1990s: Netscape made its browser free and its code publicly available -- an open-source "watershed" that helped catalyze the commercial internet. Today's AI parallel is similar, he says: foundational tools are now open and collectively improvable, so the constraint is no longer whether source code itself is accessible. Rather, the industry is shifting toward what Jian calls " 'open resources,' especially model weights, data, and computing resources, which are indispensable to advancing this industry." As more of these resources become available, the more developers can skip over the massive costs of reproducing work already done by others -- a key variable in accelerating the spread of AI, he says. That said, models must also be "differentiable and sufficiently capable," researchers from Andreesen Horowitz find. In this fast-moving industry, users who find a model unreliable or a poor fit will quickly migrate, "finding value in a wider array of options, rather than defaulting to one 'best' choice. We expect the next wave of momentum to emerge at the application layer, with gains to society realized only as organizations and individuals adapt how they work. The potential of AI is universal; the challenge of capturing it remains distinctly human.

Is AI Having Another DeepSeek Moment?

A little less than a year ago, the AI industry acknowledged wider anxieties about a possible bubble. Sometimes, OpenAI's Sam Altman said, "smart people get overexcited about a kernel of truth." Mark Zuckerberg allowed that a bubble is "definitely a possibility, at least empirically," while Google's Demis Hassabis suggested that "some parts of the AI industry are probably in a bubble." Jeff Bezos, while arguing that AI will "change every industry," pointed to signs of an "industrial bubble." A few months later, this bubble-talk bubble was pricked by the arrival of a new generation of AI coding tools, which were notably more useful for complex, real-world programming, and quickly adopted by the tech industry. This had a few different consequences for AI firms. It refreshed the industry's narrative of inexorable progress, even if they hadn't quite anticipated how things would play out: the lines were still going up, and insiders once again described feelings of acceleration. One month, prominent AI figures were conceding that most AI-generated code was "slop," and the next, they were delegating a significant portion of their work to Claude Code. In addition to a narrative reset, AI firms now had something to sell to other firms, which showed immediate interest -- a plausible business model beyond getting as many users as possible. Lastly, frontier AI coding tools were extremely compute-intensive. They could write code, but required enormous quantities of tokens to do so. This resulted in some sticker shock among AI customers, as well as some unsustainable subsidies on the provider side, but also quelled, at least temporarily, concerns about overbuilding data centers. For the first half of 2026, capacity was scarce, major AI firms were selling as much access as they could provide, and plans for hundreds of billions of dollars of infrastructure buildout regained credibility among investors. Problem solved! For a while. The anxious mood of late 2025 was caused, in part, by worries that model progress was at risk of slowing down. The arrival of agentic AI coding relieved that. But another contributor was the arrival of Chinese open-weight models, which were cheap to use, free to run on customers' own hardware, and anywhere from three to nine months behind the American frontier. In mid-2025, a model called DeepSeek R1 promised similar capabilities to OpenAI's GPT-4, accompanied by claims that it had been trained for a fraction of the price, briefly crashing American tech stocks. Now, in the post-AI-coding era, Chinese AI firms - which American companies have since accused of "distilling," or ripping off, their frontier models -- are once again entering the picture. About six months ago, Anthropic and OpenAI passed a new threshold in AI software development. Now, multiple open-weight competitors, according to AI researcher Nathan Lambert, are passing it too. These include Moonshot AI, which makes a model family called Kimi, and Z.ai, which just released an update to its GLM model: Pretty much everyone I respect among the AI commentariat and researcher class has praised the model after using it personally. Such a focal point of discussion among the community has only been so clear with an open model release once before -- DeepSeek R1. This is not a comparison I make lightly, and when I compared Kimi K2's release to a "DeepSeek Moment," GLM-5.2 has well exceeded that... GLM-5.2, he writes, is the first open-weight model "that feels right in coding harnesses as a general agent," which is to say it's the first of these vastly more affordable AI models to be able to do most of what people expect from using Claude Code with a recent Anthropic model, or the equivalent from OpenAI. For the moment, it appears as though these models are at least as good as anything Google has on offer, if not better, which is an outcome that would have sounded strange just a few months ago. (For its part, newer DeepSeek models are starting to show up in corporate AI spending.) The broader consequences of commoditizing circa-early-2026 agentic coding aren't easy to predict. For one, the frontier labs have kept moving. Since then, Anthropic released (and then was forced by the government to recall) Mythos and Fable, a new generation of models from which another set of new capabilities, this time around cybersecurity, emerged unexpectedly, meaning that frontier models still had something unique, valuable, and expensive to offer to customers. (As of this week, however, OpenAI claims to be able to offer similar capabilities at a lower price, just weeks after Anthropic's release; Z.ai claims its own models will be there by the end of the year.) The arrival of cheaper, good-enough coding models could undercut offerings from American firms, which have raised many times more money than their Chinese counterparts. Alternatively, these cheaper models could end up inducing demand, making heavy coding agent usage more attractive to a wider group of customers without cutting too much into the top end of the market, where the frontier labs could still dominate. Then there's the vaguer prospect of another breakthrough that might render six-month-old capabilities less relevant to the bigger picture: AI companies are talking an awful lot about "loops" these days, and hinting at the possibility that their models are getting closer to "recursive self-improvement," which they argue could solidify their leads in profound ways. It would be wrong to say that stranger things have happened -- a self-improving AI that threatens to slip from human control is about as strange as things get -- but things that are at least pretty strange have, and it would be imprudent to dismiss the possibility that, given the billions of dollars of research money and compute thrown at the problem, the next generation of models might be useful for something else the economy values as much as coding. At the very least, the industry is looking for its next big story. But if you zoom out a bit, the basic dynamic here is punishing, and carries undeniable risks for the industry. With each step-change in AI capabilities, affordable, flexible, and open models are pretty close behind, and what was briefly expensive to deploy becomes quite cheap. American AI firms are openly worried about this sort of thing, which they've framed as a threat to national security, a form of intellectual property theft, and a source of uncontrolled AI risks, but which also clearly complicates their still-nascent business models. Bubble talk isn't quite back, but there are early signs -- among tech leaders, but also in markets, which are suddenly cooling on AI and chip stocks -- that vibes may be shifting once again, as they have a half-dozen times in the last three years.

Cheap Chinese AI models are quickly gaining customers across the US market: 'This changes things'

A flood of high-powered, cheap-to-use Chinese AI models are quickly amassing customers across the US - and experts are sounding alarms that America's lead in artificial intelligence could be in danger. One such new open-source model, dubbed GLM-5.2, was released by China's z.AI on June 16 and specializes in coding projects. The company claims that GLM-5.2 is about as advanced as some of the best models offered by Anthropic, OpenAI and Google. "Genuinely impressed, almost shocked, at how good GLM 5.2 by @zai_org is at coding," Guillermo Rauch, the CEO of US-based AI firm Vercel, wrote on X. "This changes things." Mat Velloso, an AI executive who formerly held senior roles at Meta and Google DeepMind, wrote that he had spent "all day" using GLM-5.2. "First open model that passes the bar as a daily driver," Velloso wrote on X. "Things are not going to be the same." The Trump administration has been increasingly wary about China's breakneck pace in AI development - with officials warning as recently as recently as April that China was engaged in "industrial-scale" efforts to rip off AI technology. OpenAI and Anthropic have accused Chinese firms of using a technique called "distillation" to extract data from American models. Chinese AI is gaining a foothold in the US market. Of the 10 models included on AI marketplace OpenRouter's most popular rankings, six were developed by Chinese tech firms, including DeepSeek, Tencent, Xiaomi and MiniMax. Z.AI's leadership has begun to burnish its public profile. When SpaceX CEO Elon Musk predicted last week that a Chinese firm would catch up to Anthropic's frontier models in "probably Q1" of next year, z.AI founder Jie Tang replied that it "won't take that long." Unlike subscription-based US models, many of the leading Chinese models are open-source -- meaning they are readily available to the public and much cheaper to use for major projects. In fact, the cost of AI tokens, a measure of usage, has become so high for top AI models that leading firms like Meta, Uber and Walmart have imposed or plan to impose limits on how much employees can use them. Cursor, an AI coding firm recently acquired by Musk's SpaceX, admitted in March that its "Composer 2" model was built using an open-source model released by China-based Moonshot AI, whose backers including Alibaba. Elsewhere, Microsoft drew criticism earlier this month after Axios reported that it was considering making a version of China's DeepSeek available on the company's new "Copilot Cowork" tool, which allows users to select from an array of AI models to complete long-term projects. The report drew a fiery response from Sen. Rick Scott (R-Fla.), a China hawk, who said "Communist China wants to destroy our way of life." "American companies have no business selling out our national security by partnering with CCP tech companies like this," Scott wrote in a June 16 post on X. The plan is part of Microsoft CEO Satya Nadella's strategy to offer access a variety of more affordable AI models, not just the expensive leading models offered by American AI giants, according to an interview he gave to the Wall Street Journal. In the interview, Nadella rebuked tech CEOs over how they've discussed AI's potential to shake up the US economy. though he didn't mention any by name. Anthropic CEO Dario Amodei, for example, has warned that AI could cause national unemployment to hit 25% over time - with white-collar jobs hit particularly hard. "You can't say, hey, all white-collar jobs are gone and this could even be a weapon and we will use all the power to build data centers," Nadella told the WSJ.

GLM 5.2 Could Be China's New AI Wrecking Ball

GLM 5.2 may not yet be the wrecking ball itself, but it is another heavy swing at the valuation scaffolding beneath the US AI trade. Takeaways * GLM 5.2 is not DeepSeek 2.0 yet, but it is another clear warning that the frontier-model moat is narrowing while the global AI cost curve is falling. * Goldman Sachs sees the gap between open and closed models continuing to compress. That is the real market signal, not one benchmark leaderboard. * JPMorgan's key distinction matters: mature intelligence should keep deflating, while genuinely superior workflow capability can still command a premium. * UBS's broader work suggests China is closing the capability gap with a far lower cost base, raising harder questions about returns on the US AI capex boom. China's New AI Wrecking Ball GLM 5.2 may not yet be the wrecking ball itself, but it is another heavy swing at the valuation scaffolding beneath the US AI trade. The issue is not whether one Chinese model can claim a temporary victory over GPT, Claude, or whichever frontier model happens to be on top of the leaderboard this week. Markets have become far too mesmerized by the daily benchmark horse race, as though the prize is a trophy rather than the cash flows sitting behind it. The bigger question is whether intelligence is becoming cheaper, more portable, and less geographically captive, just as the US hyperscaler complex is pouring hundreds of billions into an infrastructure buildout priced on the assumption that frontier capability will remain scarce, differentiated, and richly monetizable. That is why Z.ai's GLM 5.2 matters. The model, released by the Beijing-based group formerly known as Zhipu, has quickly become the latest focal point in the open-weight AI debate. The early claims are formidable: a large mixture-of-experts model built for long-context reasoning, coding, and agentic work, capable of operating across enormous code repositories and handling more complex multi-step tasks. But the raw parameter count, the technical feature list and the usual social-media euphoria are all side dishes. The main course is the economics. GLM 5.2 appears to offer frontier-adjacent capability at a price point that challenges the assumptions embedded across much of the Western AI stack. Goldman Sachs' Delta One head Rich Privorotsky captured the market signal cleanly. In Goldman's view, GLM 5.2 is another Chinese open model that appears highly competitive on software-engineering benchmarks relative to the latest private models. It may not yet sit at the absolute frontier, but the gap between open and closed systems is narrowing, and the open availability of model weights matters enormously. Models that can be distilled, quantized and reproduced are not simply software products sitting behind a toll booth. They are potentially portable intelligence. For markets, that is the uncomfortable part. Once intelligence can travel light, the scarcity premium begins to look less like a fortified moat and more like a hotel room in peak season: immensely valuable right up until enough new supply comes online. The first models to close the gap do not need to be perfect. They only need to be good enough to make the next enterprise procurement meeting a much harder conversation. China's frontier models have moved from roughly 60% of leading US model intelligence in 2023 toward around 90% today. The moat remains, but the water level is falling. UBS has been making precisely that broader point. The bank argues that Chinese frontier models have moved from around 60% of the intelligence level of leading US models in 2023 toward roughly 90% today, according to Artificial Analysis data. That is not a marginal improvement. It is a structural compression of the capability gap. China still trails the latest US frontier models in several areas, particularly at the absolute leading edge of coding and general-purpose intelligence, but the gap is no longer a canyon. It is increasingly looking like a river that markets are convinced can be crossed. Closing the gap UBS also notes that Chinese developers have become globally competitive in multimodal applications, including video generation. That matters because the commercial AI market will not be won by one benchmark, one coding test or one model release. It will be won by the ability to deliver increasingly capable intelligence across a broad range of real-world applications at a price point enterprises can justify. The US may still own the frontier, but China is moving quickly across the broader commercial terrain where "good enough" intelligence is often more than enough. Artificial Analysis' early work has placed GLM 5.2 near the top of the open-model pack, while coding and agentic-work benchmarks have reinforced the sense that the model is operating in genuinely rarefied territory. Those results still deserve the usual caution. A new model can look spectacular in controlled tests and still need time to prove itself in enterprise workflows, real software environments and persistent user adoption. Benchmarks are an audition, not a box-office result. But once a model clears the threshold where developers and enterprises view it as a credible daily alternative rather than an interesting experiment, the market conversation changes quickly. The disruptive force here is price. The market already learned from DeepSeek that the cost of "good enough" intelligence can fall much faster than incumbent valuations initially assume. GLM 5.2 pushes that debate into more valuable territory. Reports comparing its token economics with premium Western frontier offerings suggest it can deliver a meaningful portion of the capability at a dramatically lower cost. The precise discount will vary by workload, usage pattern and provider, but the direction of travel is plain enough. Customers increasingly have alternatives. And once customers have alternatives, price becomes a weapon. JPMorgan's work on GLM 5.2 adds a more nuanced and commercially important layer to the story. The bank argues that while the headline listed pricing may look broadly similar to GLM 5.1, the removal of lower-priced tiers means the effective blended price paid by customers should move higher. JPMorgan estimates GLM 5.2's API price is around 13% above GLM 5.1's blended price, even though the model remains within the same broad 744-billion-parameter family, with roughly 40 billion active parameters. In other words, Zhipu appears to be improving output quality and realised pricing without materially lifting its underlying cost base. That is not simply a technology story. It is a margin story. It is the kind of result every model provider wants: better performance, better customer outcomes, firmer pricing and a cost structure that does not need to move in lockstep. JPMorgan's larger framework is the most useful way to think about the model layer. Mature intelligence deflates. Once multiple vendors can offer comparable capability, inference improves, hardware utilisation rises and the market-clearing price falls. DeepSeek remains the cleanest illustration of that force. It has reset expectations around what customers should pay for routine tasks such as standard summarisation, basic content generation, lower-risk coding assistance and straightforward tool use. In those categories, intelligence is sliding toward commodity economics. The model that cannot differentiate gets dragged into a price war. But JPMorgan also argues that newly unlocked frontier capability can still command a premium when it improves task completion, reduces retries, saves human time or enables workflows that were previously too difficult to automate. That is where GLM 5.2 becomes more interesting than a simple discount-model story. Better coding agents, enterprise workflow automation, long-context document processing and multi-step task execution are not merely about selling more tokens. They are about delivering completed work. The price of a token matters far less if the model can replace hours of human effort, navigate a complex codebase, complete a workflow reliably or move an enterprise process from suggestion to execution. This is the two-speed AI economy now emerging. DeepSeek lowered the floor. GLM 5.2 is testing whether China can also compete for the ceiling. The mature end of the market is likely headed toward relentless price compression. The harder and more valuable tasks may still support premium pricing, but only for as long as the capability gap remains real. The minute a lower-cost open model can handle those same tasks reliably, the premium starts to look less like a moat and more like a temporary toll booth. The Huawei angle may prove even more consequential than the benchmark story. Z.ai has said GLM 5.2 was trained using domestic Huawei Ascend accelerators rather than Nvidia hardware. That claim needs to be tested carefully over time, particularly around scale, efficiency, training economics and repeatability. But if it proves durable, it carries implications far beyond one model launch. It would suggest China is not simply navigating around export controls at the margins. It would suggest an alternative domestic AI stack is becoming capable of supporting increasingly sophisticated model development without depending on the US hardware ecosystem. That is where UBS's capital-efficiency analysis becomes increasingly uncomfortable for the US AI capex narrative. UBS notes that Chinese listed model developers such as Zhipu and MiniMax have operated with R&D spending that is a fraction of the outlays committed by leading US frontier labs, while Chinese API pricing remains materially below comparable Western offerings. Yet UBS channel checks suggest major Chinese model providers can still generate API gross margins in the roughly 20% to 40% range, with MiniMax's reported margins broadly comparable to those seen at major Western peers. The numbers will continue to evolve, but the directional message is powerful. China is not trying to copy the US model dollar for dollar. It is trying to compress the cost curve. That is a far more dangerous proposition for incumbent valuations than a simple race to produce the largest model. The US AI supercycle is being financed like a modern industrial revolution: data centres, chips, power, networking and ever-larger model budgets, with debt increasingly helping to keep the construction cranes moving. The strategic case is clear enough. AI demand is real, enterprise adoption is broadening, and the winners could own some of the most valuable commercial infrastructure ever built. But markets have a habit of confusing a powerful demand story with a guaranteed return story. A capex boom can create extraordinary technology while still leaving investors holding the wrong end of the economics if the product being built becomes cheaper faster than expected. GLM 5.2 does not break that thesis by itself. One model launch never does. But it sharpens the question the market has been trying to avoid. What happens if AI demand remains extraordinary, yet the value capture migrates away from the model owners and hardware suppliers toward enterprises, developers, low-cost platforms and open ecosystems? What happens if the world consumes vastly more intelligence, but intelligence itself becomes cheaper faster than incumbents can defend their margins? The market has priced an AI gold rush. But the larger the capex cheque, the more sensitive the investment case becomes to the persistence of monopoly-like returns. If Chinese open-weight models can offer 90% or 95% of frontier capability at a fraction of the cost, then the return on every additional dollar of US spending deserves much harder scrutiny. This is not an argument that Nvidia, hyperscalers or leading US labs suddenly become irrelevant. They still retain immense advantages in scale, enterprise relationships, ecosystems, proprietary data, distribution and frontier research. But the market may be overestimating how long those advantages can support scarcity pricing across the full stack. The US is building an AI aircraft carrier. China may be trying to win with a fleet of cheaper, faster boats. That does not mean the aircraft carrier is obsolete. It means the cost of defending the waters may be much higher than investors expect. The listed-market reaction tells its own story. Zhipu's Hong Kong-listed vehicle surged after the GLM 5.2 launch, supported by bullish research from JPMorgan and Bank of America, with investors treating the model as evidence that China's model layer can become both technologically relevant and commercially viable. For now, that looks more like a substitution trade than a wholesale liquidation of US AI leaders. Chinese AI names are being rewarded, while US AI equities have not yet been forced into a full rethink. That is precisely why the risk remains underappreciated. Markets usually price second-order effects only after adoption begins to appear in usage data, enterprise budgets, margin guidance and capex revisions. By then, the easy part of the trade is often over. So is this DeepSeek 2.0? Not yet. The original DeepSeek shock landed because it was cheap, unexpected and sufficiently credible to force investors to revisit the economics of the entire AI hardware and model stack in one violent move. GLM 5.2 is different. The market has spent the past eighteen months becoming progressively more aware that Chinese open models are improving quickly. The surprise is smaller. The direction is not. The genuine DeepSeek-style shock may arrive through a different door. It could come from verified evidence that Huawei-trained frontier models can scale economically. It could come from enterprise adoption moving decisively toward lower-cost Chinese open-weight alternatives. It could come from DeepSeek, Kimi or Zhipu releasing another model that closes the remaining gap in high-value reasoning, coding and multimodal work. Or it could arrive when investors finally begin to question whether the US AI capex complex is generating durable returns or simply financing a race toward cheaper intelligence. For now, GLM 5.2 is another crack in the wall around the AI toll booths. The important question is not whether it edges out the latest American model on a leaderboard. The question is whether intelligence is becoming abundant, portable and cheap faster than the US AI complex can convert extraordinary spending into durable economic rents. That is the real wrecking-ball risk. AI demand may remain enormous, and the technology may yet redraw the commercial map. But markets are beginning to confront a harsher possibility: the gold rush can continue, the shovels can keep selling, and yet the price of the gold can still fall. The next leg of this trade will not be decided by who has the flashiest benchmark. It will be decided by who can still charge a premium once intelligence stops behaving like a scarce asset.