China is racing to build its own Nvidia alternatives and AI models, but chip shortages, costly computing, and cheap AI services are exposing a problem the entire industry must solve: who is actually making money from artificial intelligence?
China’s AI boom is running into a problem that has little to do with whether its models can answer questions or write code. The country is spending billions on chips, data centers, and computing systems to compete with America, but AI sales still aren’t keeping up with the buildout. The warning signs are piling up, with some reports even claiming China’s AI bubble has burst.
China’s efforts to develop its own chips to power its AI sector may face headwinds, casting a cloud over an industry that’s struggling to turn a profit. Additionally, China’s STAR 50 stock index, which tracks technology and semiconductor firms, has slumped 40% over the last four months, wiping out around $480 billion in market value.
China’s AI ambitions hit a hardware wall
While China is aiming to develop its own alternative to Nvidia chips to power its AI ambitions, progress may not be as smooth as Beijing would like. US export controls prevent China from accessing some of Nvidia’s chips, and China’s own chips can’t yet match up.
About 2 months ago, we posted about Huawei’s Ascend chips being touted as an Nvidia alternative. While Alibaba and ByteDance are reportedly planning to adopt Huawei’s latest Ascend chips, DeepSeek has said it’s working to optimize its models to run on Huawei chips. However, Reuters reported that Huawei and China’s Cambricon have raised prices on some of their chips in response to memory chip shortages. Ascend 950DT chips have reportedly been quoted at more than 250,000 yuan, representing a 20 to 50% increase from prices quoted two months ago.
The US is currently spending around $800 billion on AI-related data centers this year, according to Rhodium Group, and China’s AI-related data center spending is expected to be just 15 to 20% of that.
Despite spending billions of dollars to develop its AI industry, China may be struggling to turn a profit. Alibaba’s, Baidu’s, and Tencent’s free cash flow went from a combined $23.6 billion in 2025 to a $2.2 billion loss in the first half of 2026. Alibaba made around $420 million in revenue from its AI app business in its latest quarter but saw operating losses of almost $2 billion. Alibaba’s Qwen AI models and ByteDance’s Doubao chatbot, which has reportedly racked up more than 200 million daily active users, are popular. But the more people use AI models, the more compute AI developers need, and computing power doesn’t come cheap.
Cheap AI is harder to monetize
Taken together, China’s biggest AI model developers, including DeepSeek, Alibaba, ByteDance, Moonshot and Zhipu, raked in an estimated $10.7 billion in annual recurring revenue as of August 2026, according to Rhodium Group. This is revenue, not profit, and the figure remains roughly one-tenth of the combined revenue reported by OpenAI and Anthropic. More importantly, China’s AI models may be popular, in part because they’re cheaper than their Western counterparts.
In fact, earlier last year we posted about how DeepSeek trains AI for pennies on the dollar compared to its US counterparts. DeepSeek’s R1 AI model showed the world that China can develop state-of-the-art AI models without spending as much as its US counterparts. DeepSeek told Nature that it trained the final version of R1 using 512 Nvidia H800 chips for around 80 hours, costing about $294,000 in compute costs, though that doesn’t include previous training runs, employee costs, data collection, or other infrastructure costs.
However, that also means AI developers have less opportunity to charge a premium for their services. Even DeepSeek reportedly raised API prices for some customers in August as its compute costs rose. Additionally, DeepSeek’s example may also show the challenges facing Chinese AI firms: if everyone can get their hands on powerful AI models for dirt cheap, as DeepSeek, for one, releases open-weight models that anyone can download, modify, and run, then AI developers may find it harder to charge a premium for their models and make money from their endeavors.
America’s AI boom faces the same hard questions
After US midterm elections, the fear that a change in government could derail data center construction became a reality. The Nasdaq 100 slumped by around $2.5 trillion on November 5. Additionally, the five biggest US data center developers tracked by Rhodium Group saw their combined free cash flow plummet from $191 billion in 2025 to $139 billion.
While Wall Street rallied the day after results were in, recovering around $1.5 trillion on November 9, that recovery did not settle the underlying question.
The US and China are on different trajectories on AI, but both countries will have to figure out how to turn billions of dollars’ worth of chips, computing power, electricity, and AI models into actual revenue. While a market correction can erase inflated valuations in days, the harder test is whether the underlying technology businesses can eventually make the unprecedented numbers work.
In case you missed:
- Huawei Is Coming for Nvidia Now. Its New AI Chips Arrive in 2027
- NVIDIA’s Taiwan Problem: The AI Boom Is Now Dependent on One Island
- China Wants a Future Without Windows. It Starts This September.
- You Can Now Host an NVIDIA Data Centre in Your Home. Some Neighbours Already Want Them Gone.
- Goodbye Blackwell, Hello Rubin: Nvidia’s new AI platform is here!
- OpenAI Just Revealed Its AI Went Rogue. The Rest of the Story Is Even Stranger.
- Apple, Tesla, Tata: The Cyberattack That Hit India’s Manufacturing Boom
- Not Just Oil Part 2: How the Iran War is Disrupting Helium, Chips, and the AI Boom
- Aatmanirbhar Push: India Becomes 7th Nation to Crack Gallium Nitride Chip Technology
- While Bezos Bets $100 Billion on Physical AI, His Top AI Picks Walk Away to Model the Universe
