China puts the AI business in check: models almost as good as Claude or ChatGPT and far cheaper

The article, written by Eric Levitz in Vox (July 30, 2026), poses an uncomfortable question for the artificial intelligence industry: what if superintelligence ends up being too cheap for anyone to control — neither monopolized commercially nor regulated politically? The central thesis is that the narrative dominant until now, of a handful of US labs (Anthropic, OpenAI) on their way to enormous profitability, may be wrong, and that the real risk is not the tech oligarchy the left feared — hence Senator Bernie Sanders's calls to nationalize the big AI labs — but a kind of competitive anarchy in which frontier models get cheaper and spread faster than expected.

The starting point is the astronomical valuations: capital markets have put Anthropic and OpenAI close to $1 trillion each, betting on future hyperprofitability. That expectation rests on two pillars, the author explains. The first is market size: a technology capable of improving performance in practically any white-collar sector has gigantic revenue potential, and Anthropic's own revenue growth would confirm that Claude is already generating real value. The second pillar, more fragile, is the so-called "moat": the brutal cost of training a frontier model — hundreds of millions of dollars in initial training alone, plus months of fine-tuning ("post-training") with armies of hired experts (computer scientists, doctors, mathematicians) who correct and guide the model's answers — would act as a barrier that no underfunded competitor could overcome.

That second pillar is the one China is said to be eroding. Over the past two months, three Chinese models have appeared that come very close to the capabilities of the American frontier systems, at a fraction of the cost. In June, the Beijing-based firm Z.ai unveiled a model that performed almost at the level of the second-tier Claude and ChatGPT systems on independent benchmarks. Weeks later, Moonshot released "Kimi K3", which supposedly beats all its US rivals except the most recent versions of Claude and ChatGPT. And just a few days ago, Alibaba unveiled a preview of Qwen3.8 Max, which supposedly would surpass even OpenAI's most advanced systems, trailing only Claude Fable. (The article includes a disclosure that Vox Media has a partnership agreement with OpenAI, though it states that its coverage is editorially independent.)

What is truly unsettling, according to Levitz, is not just that these models are almost as good, but how they were allegedly achieved: through "distillation". The technique consists of massively querying a rival model — in this case, allegedly Claude, via 16 million conversations launched from 24,000 fake accounts — asking it to explain its reasoning step by step, and then using those answers as training material so that one's own model learns to imitate that behavior. This would replicate much of a frontier model's capabilities without bearing the cost of human experts or post-training cycles. Neither Alibaba nor Moonshot has admitted to using distillation, but the article notes that OpenAI and Anthropic are said to have detected Chinese distillation attempts earlier this year, and that Kimi K3 has identified itself as "Claude" in conversations with ordinary users, a classic sign of this kind of training. Elon Musk himself, the text recalls, admitted in court that xAI applied distillation techniques to Claude and ChatGPT to improve Grok, which suggests the practice is not exclusive to Chinese labs. Legally, moreover, it is difficult for Anthropic and OpenAI to frame this as intellectual property theft, given that they built their own models by absorbing the public texts of journalists, programmers and lawyers.

On top of that competitive threat comes another, more structural one: Kimi K3 and Qwen3.8 Max are being released as open source, meaning their parameters will be downloadable for free (Alibaba and Moonshot have not published them yet, but say they will shortly). This means any company or hobbyist with enough computing power will be able to run a near-frontier model on their own hardware, adapt it to a specific use and even resell access, without paying Alibaba anything. For most companies, which do not need the world's smartest model but one that is competent at specific tasks — legal research, IT support, working code — a model that performs 90% as well as Claude at a sixth of the cost is very attractive, with the added benefit that running the model on their own servers avoids handing sensitive data to third parties. The article cites a Linux Foundation survey according to which 63% of organizations already use open-source AI systems, and data from the venture capital firm Sequoia Capital according to which most US AI startups already use Chinese open-source systems. DeepSeek, the text recalls, had already overtaken ChatGPT as the most downloaded free app on the iPhone more than a year ago, evidence that this trend is not new but is accelerating.

Levitz's economic conclusion is that, if these near-frontier models keep getting cheaper and spreading, selling cutting-edge AI could become a low-margin business, and the sector's big winners would not be the labs but the chipmakers and the cloud computing providers. But the article does not stop at economics: it also stresses the security implications. Having a handful of companies control the frontier models is problematic in many ways, but it also makes them easier to regulate — the text mentions as an example that the Trump Administration temporarily blocked Claude Fable on cybersecurity grounds. If, instead, the "recipes" for building ultrapowerful AI circulate freely and anyone with modest technical knowledge can modify them, systems willing to help hack government bureaucracies or design biological weapons could proliferate.

The article closes with an ambivalent reading: from one angle, this "almost free AI" scenario could be seen as capitalism working exactly as expected — a small group of very rich investors funded an enormously useful technology hoping for outsized profits, only for competition to erode those returns and spread the benefits among many more companies and consumers. From another angle, that same process of uncontrolled diffusion could also make it easier for a "supervirus" or other serious threats to emerge, precisely because of how hard it is to put the genie back in the bottle once the technical knowledge is dispersed. The author himself wryly concludes that "no system is perfect", leaving open the question of whether the competitive fragmentation of AI power is, on balance, good or bad news for humanity.