A wave of cheap and powerful Chinese open-weight AI models is causing a scare among investors in the AI boom. For a decent chunk of the tech universe, however, there is less to worry about than meets the eye.
The rapid advance of open-weight models, which make public a set of numerical values governing their behavior, is certainly a concern for closed-model developers. Those companies don't allow changes to how their models operate, giving them tighter control over how they respond to queries.
The leaders in that space, OpenAI and Anthropic, are both planning initial public offerings in the coming months. Any doubt about the durability of their competitive advantage could diminish their appeal to investors.
Google is another closed-model developer with something to lose. Gemini models have fallen a bit behind the cutting edge performance-wise in recent months; better and cheaper open-weight models threaten to pull customers away.
Worry about the impact sent Google parent Alphabet's stock 10% lower in the days after the debut of the open-weight Kimi K3 model from China's Moonshot AI last month. Microsoft, Amazon.com and other AI-infrastructure heavyweights also fell.
Kimi K3 performed on par with, or better than, some leading closed-weight models, and access to it was priced far more cheaply. Other powerful and efficient models have emerged in China this summer, too, including ones from tech giant Alibaba and startups Z.ai and MiniMax.
That has added to concern about U.S. leadership in AI and Big Tech's insatiable spending on AI infrastructure. If Chinese companies can do leading-edge AI better and more efficiently with open-weight models, the reasoning goes, plans to spend untold sums on data centers and chips might need reworking.
The opposite is just as likely to be true if open-weight models take off, though. Should AI become far cheaper to deploy, the so-called Jevons paradox, named after a 19th-century economist, would likely take hold. That would mean people respond to lower costs simply by using the technology more, eating up just as much if not more computing power.
That is one reason chip companies have largely welcomed open-weight models. Andrew Feldman, the chief executive of the Nasdaq-listed AI chip company Cerebras Systems, said open-weight and closed-weight models pushing each other would benefit consumers, but wouldn't hurt chip demand. "There is no reason for chip stocks to go down when open-source models come out," he said.
Nvidia Chief Executive Jensen Huang has made himself an open-weight poster child in the past month, leading a consortium of companies championing the approach and drawing somewhat-dubious connections between open-weight models and the highly successful open-source software movement.
Beyond the chip industry, cloud-computing leaders like Amazon, Microsoft and even Google would also benefit handsomely if open-weight models expand demand for computing power. Open-weight models might be cheaper and more efficient, but while they might be free for anyone to pick up and use, they aren't cost-free to run. That is a win for anyone in the computing-infrastructure game.
As Morningstar analyst Malik Khan said in a recent note, "if an enterprise were to consolidate its entire AI stack on open-weight models, it would still need cloud infrastructure to run those workloads, store data, manage security and access to resources, et cetera, all tailwinds to cloud infrastructure companies."
A large number of corporations are already using open-weight models. They often tune them to excel at narrow tasks, like summarizing documents or answering customer questions on company-specific topics.
That can be a big money-saver compared with paying usage fees for smarter but more expensive closed-weight models. A McKinsey survey last year found nearly two-thirds of companies that had experience with AI were using open-weight models -- mostly less than cutting-edge ones offered by Meta Platforms, Google and France's Mistral.
Yet there are reasons to doubt advanced open-weight models -- particularly those from China -- will make major inroads in the AI race. While there is plenty of demand, there is no obvious profitable business model for them. Which is one reason open-weight model developers haven't attracted a lot of venture-capital interest.
Open models can make sense for big tech companies for whom they are loss leaders, costing them money but drawing users to other profitable parts of their businesses. It is less clear whether independent open-model developers can make enough revenue to pay for the gargantuan cost of training a cutting-edge model.
Cheaper pricing for Chinese models was "a business-model choice, not necessarily a hardware-cost read-through," BofA Securities analyst Vivek Arya said in a note recently. Despite its efficiency gains, Moonshot's Kimi K3 model still needs around the same amount of high-cost memory as an advanced OpenAI open model, he said.
The economics might not matter as much in China, given that all of its leading AI developers have state support and AI is seen there as a strategic priority. But a large-scale migration to China's open-weight models, at least among large Western companies, seems unlikely even if they do remain cheaper.
DeepSeek, the Chinese AI developer that shook markets last January after releasing its own advanced model, quickly faded as a threat to the AI boom. This time could easily be similar, with the possibility of U.S. restrictions on the use of Chinese open models serving as a deterrent for prospective customers -- even if the curbs never come.
The disruptive possibility of open-weight models appears likely to linger, especially for OpenAI and Anthropic. But for others, their impact looks far less severe.
Comments