Building his own large language model (LLM) is out of the realm of possibility for startup founders like Zhang Haiwei. He’d need hundreds of millions of dollars, and he’d be competing with China’s internet giants, who have a long head start. The likes of Baidu and IFlyTek have been working on LLMs—the foundation of artificial intelligence systems that can mimic human intelligence—for years, long before the current AI boom took off.

Instead, Zhang’s motion-capture startup, Chingmu, is using OpenAI’s models trained with its own data to analyze how people and objects move, to use in animation and sports training.

“My view of this year is involution,” Zhang says, applying a popular term in China which describes a cycle of manic competition that leads to everyone working harder and harder for fewer rewards. Newcomers to the sector have little chance of breaking out, he says, even if they have money to burn. Besides the cost, there’s also the question of computing power and graphics cards. “Products are launched faster, there are more competitors, and customer budgets are down,” Zhang says, adding that while pressure was high in China, this is a global problem: “The world’s in recession; everything’s in recession.”

Like their counterparts in Silicon Valley, Chinese startups rode the surge of excitement about artificial intelligence, which has widely been hyped as transformative for the tech industry and the economy at large. But they’ve just as quickly come down to earth. No revolutionary applications have yet emerged, and limited supplies of GPUs (graphics processing units), which are foundational to AI, are curtailing growth. In a challenging economic environment, AI has become less about revolution and more about involution, and many startups are using it to make small efficiency improvements, hoping to gain enough of an edge to stay competitive.

“Startups and big tech companies are now focused on justifying the initial hype and excitement,” says Kevin Xu, a tech investor and founder of the AI newsletter Interconnected.

ChatGPT, the uncanny chatbot that launched the current wave of AI hype when it was released by the US startup OpenAI last year, is officially blocked in China, but many internet users scaled the Great Firewall to try it out, and a black market for the service soon sprang up. WeChat Moments, a timeline for people to post life and news updates, was filled with answers and references to the technology.

As in the US and Europe, ChatGPT sparked a frenzy of interest in AI, a sector that had started to look a bit moribund. “Some large companies had actually fired their teams working on large language models,” says Xie Mingxuan, founder of an AI startup called vrch.io, adding that the companies regret it now, because those teams then went on to found their own startups.

But developing AI models outside of big companies is a lot more challenging in China than it is in the US. American companies, like OpenAI, were able to access huge amounts of data from Google or social media platforms like Twitter and Reddit. But China skipped the open web, going, essentially, from no internet at all straight to apps, which are far, far harder to scrape for data.

That, along with the cost of computing power, makes it difficult for startups like Xie’s to build the kind of huge, sweeping models that their equivalents in the US are trying to create, so most are focusing on the application level, instead of making their own models.

Founded last year, vrch.io is developing an AI-powered voice-entry image generator. In the past, interior designers might have needed to use renderings made in Photoshop to show clients. Now, when people want to redesign a space, they can do so on the spot using generative AI. “For those of us in design,” Xie says, “we used to spend most of our time converting information that was difficult to accurately express in words, into images, and then using those images to communicate with clients.”

Though vrch.io has an investment from Miracle Plus (formerly Y Combinator China), a startup incubator in China, it’s not currently targeting the Chinese market. That’s because of the lack of regulatory clarity.

“As a small company,” Xie says, “we can’t guarantee that every segment of the business, whether it’s the algorithms, the data sources, or the training of the models themselves, is in line with regulations.”

In July this year, China’s Cyberspace Administration released interim guidelines on generative AI that focused on privacy, personal information protection, transparency of algorithms, and intellectual property rights. They didn’t set compliance standards for the technology that were substantively different from existing regulations on technology, but startups like Xie’s are waiting for more details.

“The regulators clearly don't want to overregulate at the outset to discourage innovation and further widen the gap in AI development between China and the US,” Xu says. The rules demonstrate, Xu thinks, that regulators “are willing to incorporate the needs and input of tech companies, allowing for relatively unencumbered development in private settings and business-specific areas, as long as certain red lines are not crossed in the public sphere.”

Vrch.io is more concerned with getting its product out in overseas markets first. It will wait until large models—most likely, those developed by Chinese Big Tech companies—become available before it rolls out in its home market.

The economic environment is also throwing a shadow over the tech sector. Slower growth, falling consumer spending, problems in the real estate market and concerns over local government debt have contributed to an overwhelming sense of uncertainty. The Chinese government has stopped reporting statistics on youth unemployment in urban areas, one indicator of a general economic slowdown.

“Starting a company in this economic environment, I have to pick very specific, low-hanging fruit problems,” says Pei Hao, founder of AI startup Lingua Technologies.

His company is aiming to compete with translation companies in Beijing, and professional editors in the UK and US who charge fees to Chinese scholars to help make their work legible to international audiences.

Hao says that partnerships between Chinese academics and non-Chinese counterparts are often hobbled by the extra workload given to native English speakers. “There’s so much cognitive load associated with fixing these papers, some of which are five to ten thousand words,” Hao says.

This, he thinks, is a problem which is easily solved by AI, or smoothed. Human editors will still need to check the final papers, but AI will be able to deal with time-consuming adjustments to things like formatting. “The biggest thing that will happen is that the costs associated with knowledge work will be driven downwards,” Hao said.

Companies are looking to keep their costs down by using AI to replace laid-off workers, which might be helping to boost startups that can automate labor-intensive functions, or help companies to introduce chatbots to their services.

On June 30, China-focused health care startup Medlinker held a human-versus-robot livestreamed contest in a bid to show that its AI doctor, “MedGPT,” could hold its weight against real doctors. The company is aiming for a public beta release of MedGPT in October, according to company representative Zhang Hongliang.

While Zhang touts this technology as ushering in the next phase for the development of internet hospitals, there are issues. In the contest, MedGPT asked more questions than a human doctor would after establishing that the patient’s situation was critical (though it reached the same diagnosis). The company is banking on its AI chatbot channeling more patients to its existing infrastructure of hospitals. In raising its latest round of investment, Zhang said that Medlinker was “mainly reliant” on MedGPT as the company’s selling point.

A VC from a leading venture capital fund in China, who asked to be referred to as Zhao, since he was not authorized to speak to the media, told WIRED that he was personally more positive about startups that were adding AI components to their existing products, if they could show that they had their own data and could continuously improve their models.

The macroeconomic environment, he says, means investors are “more careful about how they choose investments, because there’s more potential downside.” But he didn’t think that they had moved on. The lull after the hype, in his view, was more to do with the nature of AI and its long cycle. “VC triggers have been pulled,” Zhao says. “We’re waiting to see the competition.”