China’s artificial intelligence industry is attracting rapidly growing user adoption, but the revenue generated by its leading AI models remains a fraction of that of US rivals, raising questions about whether soaring valuations are supported by underlying business performance.

US research firm Rhodium Group estimates that all major Chinese AI models combined generate only about 10% of the revenue reported by OpenAI and Anthropic. The figures, published Thursday, use annual recurring revenue, or ARR, an industry metric that annualizes a recent monthly revenue figure to capture the pace of rapidly growing businesses.

The gap is striking given the attention Chinese AI companies have received from investors this year.

DeepSeek had the lowest estimated ARR among the major Chinese AI companies at about $500 million, according to Rhodium. MiniMax was estimated at $800 million, while Moonshot stood at about $1 billion.

Z.ai told investors on Wednesday that its latest ARR had reached $1.8 billion, according to a transcript seen by CNBC. The company now expects its ARR to reach $3 billion by the end of the year, up from an earlier forecast of $2.4 billion.

Even after including larger technology companies with AI operations, however, China’s revenue base remains considerably smaller. Rhodium estimated ByteDance’s ARR at $4 billion and Alibaba’s at $2.4 billion, compared with $40 billion for OpenAI and $65 billion for Anthropic.

The disparity becomes more significant when revenue is compared with valuations.

“Valuations relative to revenue appear exorbitant for Moonshot and DeepSeek at present,” Rhodium said, estimating revenue multiples of about 50 times for Moonshot and 163 times for DeepSeek.

By comparison, the report put OpenAI’s valuation-to-revenue multiple at 34 times and Anthropic’s at 21 times.

The figures point to a significant disconnect between how investors are pricing China’s emerging AI leaders and the amount of revenue those companies are currently generating. That disconnect could become an issue as several major AI companies move toward public markets, where investors will have greater access to financial disclosures and will be able to compare valuations against revenue growth more directly.

Anthropic is reportedly expected to list in the US next month, while OpenAI has pushed its IPO plans to next year. Moonshot has reportedly filed confidentially for a Hong Kong listing, while DeepSeek is also reportedly preparing for an IPO.

Moonshot said it does not comment on market rumors or speculation when asked about the reported confidential filing. DeepSeek and Anthropic did not respond to requests for comment.

Open-Source Models Create A Different Revenue Challenge

The revenue gap is not necessarily a straightforward measure of the technological progress being made by Chinese AI companies.

Rhodium acknowledged that its analysis relies on the latest available figures from this summer, while usage of Chinese AI models has increased sharply from relatively low levels earlier in the year. The acceleration means current revenue figures could change considerably if adoption continues to translate into paid usage.

Z.ai’s revised year-end ARR forecast illustrates that trajectory. Its new $3 billion target would represent a substantial increase from the $1.8 billion figure it reported this week.

Chinese AI companies also face a different monetization environment because several of their leading models are open source. Developers and businesses can download models and operate them independently if they have sufficient computing infrastructure, meaning the company that develops the model does not necessarily capture revenue every time that model is used.

Rhodium said Chinese AI labs are therefore exploring ways to capture a larger share of the revenue generated by third parties providing access to their models.

The contrast with US companies is significant. OpenAI and Anthropic largely operate closed models and directly control access to their leading systems. That gives them greater ability to monetize usage, although it also leaves them carrying substantial costs associated with training and operating increasingly powerful models.

According to AI comparison firm Artificial Analysis, the cost per task for leading OpenAI and Anthropic models is substantially higher than for Chinese models. Lower pricing can help Chinese models gain users quickly, but it can also make converting that usage into equivalent revenue more difficult.

This results in a central challenge for China’s AI sector: rapid adoption does not automatically translate into equally rapid monetization.

“The financing gap means it will be far more difficult for Chinese frontier AI labs to scale sustainably,” Logan Wright, a partner at Rhodium Group and co-author of the report, told CNBC.

Wright said Chinese AI companies would be heavily dependent on favorable equity-market conditions, adding that relying on China’s equity market has historically been difficult. He also said government support has been useful for expanding computing infrastructure but suggested that direct government financing for frontier AI laboratories could face limits.

The contrast between infrastructure and AI-model funding is important. Rhodium estimated that more than 60% of equity investment in Chinese AI chips and servers came from state-affiliated sources. That indicates a substantial role for government-linked capital in building the hardware infrastructure needed to support the country’s AI ambitions.

But funding the physical infrastructure required for AI development is different from guaranteeing the commercial success of individual AI laboratories. Companies still need to turn computing capacity and model capabilities into recurring commercial revenue.

Investors Are Already Testing The Valuations

The market’s response to China’s listed AI companies has demonstrated how quickly investor enthusiasm can change.

Z.ai shares rose more than 5% in Thursday morning trading, recovering from an earlier decline following news of its second major fundraising round in two months. The Hong Kong-listed stock has fallen back toward levels seen in the spring after briefly more than tripling during the summer.

MiniMax has faced a similar pattern. Its shares have struggled in recent months to maintain gains above their IPO-day levels after a sharp rise earlier in the year.

The volatility indicates that investors are already grappling with the distinction between AI adoption, technological capability and financial performance.

China’s AI sector has demonstrated that its models can attract users and compete with leading US systems, while companies such as DeepSeek have shown that capable models can emerge with different cost structures. The unresolved concern rests on how that technological progress can produce enough recurring revenue to justify the valuations being assigned to some of the industry’s most closely watched startups.

Artificial intelligence is making it easier for employees to produce emails, documents and code at unprecedented speed. But Shopify CEO Tobias Lütke says that productivity gains can come with a less obvious cost: workers are increasingly passing AI-generated material to colleagues without taking responsibility for whether it is useful, accurate, or worth reading.

Lütke described the phenomenon as “slop grenades,” referring to low-value AI-generated work that employees produce quickly and then effectively throw at someone else to process.

“We call those ‘slop grenades’ that people toss at each other,” Lütke said during an interview on “The Knowledge Project” podcast released Tuesday. “And that’s definitely a bad thing.”

The problem is not necessarily that AI-generated content is inaccurate. Rather, Lütke argues that generative AI has dramatically reduced the cost of producing information without creating a corresponding incentive for employees to determine whether that information deserves to be produced in the first place.

That can turn AI from a productivity tool into a mechanism for transferring work from one employee to another.

An employee, for example, might use an AI model to generate an unnecessarily long email and send it to a colleague. The recipient then uses another large language model to summarize the message simply to determine what the original sender was trying to communicate.

“Why did we invent decompression and recompression?” Lütke said. “This is terrible.”

The example captures a growing tension around AI adoption in the workplace. Companies are measuring how quickly employees can generate code, text, analysis, and other outputs, but the quantity of material produced is not necessarily equivalent to productivity.

If AI allows one worker to generate five times as much material but forces other employees to spend more time filtering, checking, and interpreting it, some of the apparent productivity gain can simply be displaced elsewhere in the organization.

Lütke’s criticism comes from an executive who has been unusually aggressive about incorporating AI into the workplace.

Shopify has rolled out AI tools for agents, while the company also uses an internal agent called River. Lütke said River handles a large share of Shopify’s production code pull requests, potentially as much as half.

That makes his distinction between useful AI and low-value AI particularly important. He is not arguing that companies should produce less work simply because it was generated with AI. Instead, he sees the greatest value in systems that improve the quality of human judgment rather than simply increasing the volume of material moving through an organization.

AI can make it easier to draft a proposal, analyze information, or write software. But someone still has to decide whether the result solves the intended problem, whether its claims are correct, and whether it should be sent to another person.

Lütke said AI is most valuable when it makes people’s thinking “clearer and more concise.” Its value falls when it simply adds more material to a colleague’s workload. That is becoming an issue as companies move from experimenting with chatbots toward deploying AI agents capable of producing work with limited human intervention.

The first wave of workplace AI was largely about helping individual employees complete tasks faster. The next phase is likely to involve AI systems producing and routing work across entire organizations. That could magnify both the benefits and the costs.

A useful AI agent can remove repetitive work, identify relevant information, and help an employee make a better decision. An indiscriminate one can create a stream of drafts, notifications, code changes and reports that someone else must review.

The resulting problem is less about whether AI can generate content and more about who remains accountable for the output.

“Humans take responsibility,” Lütke said. “Machines can help us take more responsibility because they can inform us better.”

That principle challenges one of the simplest assumptions behind corporate AI adoption: that more output automatically means more productivity.

For companies, the harder task may be designing workflows in which AI-generated work has a clear owner and a clear purpose. Without that discipline, the technology can reduce the cost of producing information while increasing the cost of consuming it.

In that sense, the “slop grenade” problem is not really a limitation of AI’s ability to generate content. It is a management problem created by giving employees an extremely cheap way to generate more work than their colleagues need. The companies that extract the most value from AI may therefore be those that measure not only how much their employees can produce with the technology, but also how much unnecessary work it prevents from reaching everyone else.

Huawei Technologies says it cannot produce enough artificial intelligence computing equipment to meet demand in China and is limiting overseas sales, underscoring the growth of its Ascend chip business as Beijing pushes to reduce the country’s dependence on Nvidia and other foreign technology.

Huawei’s rotating chairman Eric Xu said Thursday that the company had no plans to expand aggressively into international markets because its existing production capacity was insufficient even for Chinese customers.

“Since we don’t have enough capacity to even satisfy the demand in China, we don’t have a plan to expand into the international market in a fully-fledged way,” Xu told reporters at Huawei Connect in Shanghai.

Huawei does supply some countries where demand is particularly strong, Xu said, but volumes remain limited. The comments suggest that the immediate market for Huawei’s AI infrastructure remains concentrated in China, where U.S. export restrictions have constrained access to Nvidia’s most advanced processors.

The capacity constraint also points to the scale of demand Huawei is seeing from Chinese technology companies developing and training increasingly sophisticated AI models. Xu said testing of Huawei’s Ascend 950DT processor had produced good results and that the company was in extensive discussions with Chinese AI developers. He expects many of those companies to begin training models on systems using the chip next year.

Huawei has emerged as one of the principal domestic alternatives to Nvidia as China accelerates efforts to establish a self-sufficient AI computing ecosystem. The company has faced U.S. trade restrictions since 2019, while Washington has separately tightened controls on exports of advanced AI chips and semiconductor technology to China.

Xu said reliable data on Nvidia’s share of China’s AI chip market was difficult to obtain, but offered his own assessment of Huawei’s position.

“I think Ascend market share should be bigger than Nvidia’s,” he said, without providing data to support the estimate.

Xu explicitly tied Huawei’s AI chip strategy to China’s broader push for technological self-reliance.

“We cannot accept a destiny where we cannot control our fate being determined by others in terms of willingness to sell chips to China or not,” he said.

“No matter if it’s for the Chinese government, industry in China, or for Huawei, it is certainly the way forward to try to push for full self-sufficiency for chips.”

Huawei has increasingly positioned computing infrastructure as a central part of its technology strategy. Guo Ping, chairman of Huawei’s supervisory board, said in remarks released this week that the company regarded AI as its “biggest opportunity” and wanted its computing and connectivity infrastructure to play a role comparable to Nvidia’s.

The company is accelerating its chip development schedule as it attempts to expand the performance of its domestic AI hardware.

Huawei said Thursday that its Ascend 960DT processor will be ready in the first quarter of 2027, three quarters earlier than previously planned. Its Ascend 960PR is scheduled for the third quarter of 2027, one quarter earlier than the previous timetable.

The company plans to release a new generation of Ascend processors each year, with the Ascend 970 and 980 scheduled for 2028 and 2029 respectively.

The faster development cycle comes as Chinese AI developers continue to demand greater computing capacity. Huawei’s approach is not simply to make individual processors more powerful. It is increasingly focused on connecting large numbers of processors so that they can function together as a much larger computing system.

Huawei said clusters containing about 100,000 chips have become standard for training some of the largest AI models. Communication between machines can consume more than 40% of training time in conventional server systems, according to the company, making the speed at which processors communicate an important constraint on overall computing performance.

Huawei’s response is a new architecture called Peerium, which is designed to allow as many as 1 million processors to operate together as a single system. Its UnifiedBus technology is intended to connect processors, memory, storage and networking equipment across those systems.

The company said its new Ascend 960 supernode can connect as many as 4,096 AI processors. Multiple supernodes can then be linked into clusters containing hundreds of thousands of processors, with the largest planned systems supporting as many as 1 million processors.

The strategy could allow Huawei to compensate, at least in part, for limitations in the performance and availability of individual domestic AI processors by combining large numbers of them and improving the efficiency of communication between machines.

Huawei said it has already deployed more than 1,000 systems using its earlier Ascend 910C processors, while Ascend 950 systems have entered commercial use. More than 5,200 developers are active each month on software for Ascend chips, and more than 40 AI models have been trained directly on Huawei’s computing platform, according to company materials.

Nvidia, however, retains a major advantage in software. Its CUDA platform is widely used by developers to build and run AI applications on Nvidia processors, creating an ecosystem that extends beyond the performance of the chips themselves.

Huawei’s challenge is therefore extending beyond producing processors. It must build enough hardware, improve the ability of large clusters to operate efficiently, and expand the software ecosystem needed by Chinese AI developers.

The effort has become necessary as U.S. restrictions limit Chinese access to advanced Nvidia processors and semiconductor manufacturing equipment.