This piece was written by contributing Tsinghua student Tianlin Ge and University of Oxford student William Robinson for The Future of U.S.–China Relations, a joint special issue by The Yale Politic and Tsinghua Youth Voice.
On Friday, January 16, 2026, OpenAI announced it would begin testing advertisements in ChatGPT for users in the United States. Many perceived this as a tacit admission that even one of the most widely used chatbots in the world was struggling to be profitable. In a viral clip last fall, an interviewer pressed OpenAI’s CEO, Sam Altman, on how the firm, with an approximate annual revenue of 13 billion dollars, could justify spending commitments that add up to 1.4 trillion dollars.
Across the Pacific, a similar story emerged. On Thursday, January 8, 2026, Zhipu AI, a Tsinghua University-incubated Chinese company listed in Hong Kong, became the world’s first public listing of a firm focused specifically on developing AI foundation models. But Zhipu’s prospectus-era reporting revealed steep losses, laying bare what was once seldom mentioned about the AI boom: the company had experienced a net loss of 2.4 billion yuan in the first half of 2025 amid surging research costs and computing costs.
Put side by side, these two moments underscore the shared problem of profitability for both U.S. and Chinese AI companies. Popular articles in 2025 showed the circular arrangement of how Nvidia invests in OpenAI so that OpenAI will, in turn, buy more of Nvidia’s chips, as though they were a sort of financial merry-go-round. Because of the huge amount of money being circulated, these deals have reminded some market watchers of the dotcom bubble in 1999-2000.
The question of whether AI is profitable is also asked in China, and since being open-weight—meaning users can download and run models locally—is one of the most prominent features of China’s AI ecosystem, users can use Chinese AI very cheaply or even freely, thus making it harder for Chinese firms to profit. Consequently, is AI actually profitable in the U.S. and China? What’s more, how does the problem of AI profitability matter to geopolitics and our everyday lives?
The Geopolitics of Profitability
To answer the question of AI’s profitability in the two countries, one needs to understand the overall landscape of the U.S.-China AI race. Both countries have already invested significant resources, research, and finances into developing AI, positioning the technology as a respective national priority. On Wednesday, July 23, 2025, at the Winning the AI Race summit in Washington, President Donald Trump announced that “from this day forward, it’ll be a policy of the United States to do whatever it takes to lead the world in artificial intelligence.” A few months ago, at the 20th Collective Study Session of the CCP Central Committee Politburo, President Xi Jinping said, “China must face disparities head-on, be strongly oriented toward applications, and firmly grasp the initiative in AI development and governance.”
AI is a long supply chain involving many components. Therefore, a more effective way to measure the profitability of AI in the U.S. and China is to examine its different sectors, from hardware infrastructure to consumer applications.
Internet giants like Google and Meta in the U.S. and ByteDance and Alibaba in China have already embedded AI models into their platforms, Meta’s ad-targeting improvements, Google’s Gemini summarizing search results, and Alibaba’s customer service bot, AliMe. These firms can amortize AI R&D across a massive revenue base, stabilizing profits through cross-subsidization.
Companies like OpenAI and Zhipu AI—standalone labs working on foundation models of AI—are, on the other hand, actors visibly associated with the problem of AI profitability. They do not have their own platforms in which they can embed AI advancements and instead have to rely primarily on other monetization methods: standalone subscription models and API licensing (providing model access to third-party developers for a fee), for instance. But so far, these methods alone have struggled to outweigh their costs.
Therefore, concerning foundation models, the U.S. and China share similar characteristics: standalone labs specializing in AI models have difficulty in generating profits, while tech giants embedding AI can achieve relatively stable profits through deploying on already popular platforms.
However, when we look at the AI upstream and downstream industries, different focuses of the U.S. and China’s AI race emerge.
Leading AI enterprises in the U.S. are dedicated to deepening their efforts in the infrastructure layers of the AI industry chain, which includes data centers and chips, etc. As an example, chips are one of the most profitable parts of the U.S. AI industry. They are also the area where the U.S. can most effectively impede China.
NVIDIA, a leading tech company that designs powerful chips, has one of the highest profit margins in the world. By the second half of 2025, NVIDIA was also one of the highest valued companies; its market value was equivalent to about 16% of the U.S. GDP. Since 2018, the U.S. has issued multiple export control regulations, limiting NVIDIA’s reach across China by expanding the scope of sanctions on key semiconductor components. This has forced China to develop its own chips.
In May 2024, China announced a semiconductor fund with a scale of approximately 47.5 billion U.S. dollars to accelerate domestic self-reliance and compensate for the impact of foreign export controls. While there have been notable successes—recently, some models like Zhipu AI’s GLM‑Image have already been fully trained on domestic chips—China’s semiconductor industry continues to face hurdles. Despite this intense domestic push for development, generally speaking, China’s high-end chip capabilities currently still lag behind NVIDIA’s.
Moreover, the U.S. has an advantage in the technical layer. U.S. AI models typically are dominant in the rankings of models with the best comprehensive capabilities in the world: U.S. AI models often account for 7-8 of the top 10 AI models, while China has only 2-3.
As mentioned, a characteristic of Chinese AI is the focus on being open-weight. The aim of this is to attract users despite potential long-term losses. In a December 2025 study conducted by Stanford University, 23 out of the top 25 open-weight models came from China with respect to their benchmark performance. Furthermore, Chinese models like Qwen can save 96%-98% of the cost compared to using the closed-source GPT 4o. Data thus confirms users might favor open-weight models, but it doesn’t answer why China endures the costs of championing open-weight models.
Developing, training, and running AI models requires a continuous and massive inflow of capital, which tests the financial resilience of both private and public sectors. This investment can be in the form of private investment, the way OpenAI receives capital, cross-subsidies within internet giants’ platforms, or support from the government. The problem is, however, whether private investors have enough patience to wait until OpenAI fulfills its 1.4 trillion dollars spending commitment. In an interview, Sarah Friar, the chief financial officer of OpenAI, seems to have hinted at hopes for the U.S. government to help fund OpenAI.
When enterprises are consumers of AI models on a large scale, they decide by factoring in which model is cheaper. Therefore, to lower prices and seize market share, companies further compress their profit space. Only those who can withstand the losses of AI for longer can offer the lowest prices; if they survive such losses, they will shape the global technological dependency relationship and influence the rules of global AI deployment. From this perspective, lack of profitability at some levels of the AI industry is not a market failure, but a strategic investment in infrastructure—capturing the resources of global AI adopters during the loss stage might secure geopolitical and technological influence for decades.
Considering the global nature of the AI industry, the binary between the U.S.-China AI race presents an incomplete picture. AI data annotation—the process of labelling raw data such as text, images, and videos so that they can be recognized by machine learning algorithms—is being carried out on a large scale in the Global South. Countries in Africa, South Asia, and Southeast Asia have become the main centers for this labor force, often referred to as “microwork” or “ghost work”. Since high-quality, manually annotated data is a necessary ingredient for training AI, the social, cultural, and linguistic nuances embedded in these regions’ work directly, and often invisibly, shape the performance and output of AI models developed by the U.S. and Chinese companies. Europe, as a huge user and developer of AI models, which represents approximately 24.2% of the global AI market by revenue in 2025, is also guiding the development of AI through consumption and legislation, such as the obligations for providers of general-purpose AI (GPAI) models under the EU AI Act, which came into effect on August 2, 2025.
Everyday Lives in the AI Era
This raises an important question: what does the problem of AI profitability in the U.S. and China have to do with the general public if they are not on Wall Street or Silicon Valley?
First, let’s look at workers closely related to the AI industry. For plumbers and electricians in the U.S., the AI industry at the infrastructure level may be a gold rush—an opportunity to make quick money. The U.S. lacks electricians, plumbers, and heating and cooling technicians needed to build physical data centers to support AI. It is estimated that between 2023 and 2030, the U.S. will need an additional 130,000 trained electricians, 240,000 construction workers, and 150,000 construction supervisors.
Beyond the software industry, the construction of AI facilities has provided abundant opportunities for the U.S. workforce. An example is Demond Chamblyss, who transformed from running a small gypsum board company in Ohio to overseeing the construction of a large-scale data center. This career change increased his annual income to over $100,000: “I pinch myself going to work every day.” It is estimated that the salaries of employees who have transferred to data center construction have increased by 25% to 30% compared to their previous jobs. However, industry insiders say, “When the project gets done, they’re not crawling with people. If construction eventually tapers off, there might not be enough alternative jobs to go around. But for now, business is great, and how things will end remains anyone’s guess.”
Data annotation is another job that has introduced many people to an AI infrastructure layer they might not have otherwise known. Duo Wei runs a small AI data annotation studio in Guangxi, China, the kind of business that, on paper, fits into Beijing’s push to treat data as an AI input worth building at scale. China’s National Data Bureau selected seven cities to pilot national data-annotation bases in May 2024, and formally recognized “data annotator/AI trainer” work under the national occupational classification. These efforts aim to strengthen the infrastructure that large models depend on.
Wei told The Politic, “Even with government support like reducing rent, we are still operating at a loss.” His team labeled data for autonomous driving, and the tasks reaching his small studio were often what larger firms had already passed over—each layer squeezed the next, until the job reached studios like his, the margins already shaved to almost nothing. One assignment he recounted involved meticulously labeling lane changes across thousands of 3D data frames. It was slow work and it came with strict accuracy demands that did not match the low unit price.
The hype surrounding AI in Chinese society has also had a direct impact on families’ spending on education. Lixue Huang told The Politic that she did not tell her parents what she actually did for work. When they asked, she used the safest phrasing she could find: “I’m training AI.” It sounded close enough to her degree, and far enough from the truth to spare awkward follow-up questions.
Huang majored in Artificial Intelligence at a private, second-tier university in Guilin. The program was one of the school’s flagship majors, and the tuition reflected that. She said her parents worked hard to pay for it because they believed it would buy her a future in a promising sector. But the job Huang found after graduation was data annotation, not the core technical work she’d imagined.
Huang felt reluctant to share the truth because of the huge gap between the prestigious image of AI and the mundane reality of data annotation. In the division of labor in the AI industry in China, data annotation—especially the task of manually annotating images or texts—is often regarded as the “digital assembly line”. Such roles are often characterized by repetitive tasks, low technical barriers, limited career development, and low salary, especially when compared to the high tuition fees paid for the degrees. Beneath her words lies the pressure of educational investment returns; her family’s best intentions, which were guided by society’s attitude to AI, became a moral constraint, compelling her to hide the truth.
Evidence for Chinese society’s positive attitude to AI can be found in an Edelman poll. In November 2025, people in China were much more optimistic about AI: 87% of participants said they trusted AI, while only 32% said so in the U.S.
In Shanghai, parents are scrambling for spots in an 8,800-yuan “AI youth boot camp” for their children. On secondhand marketplaces, a shady economy opened up: ghostwriters for alleged child AI prodigies. Junior engineers or postgraduates offer one-stop services that cover everything needed to win programming competitions, from choosing the topic to polishing the slide deck and rehearsing the presentation. Even for those children who have won such competitions without ghostwriters, it is unclear whether, by the time they enter the job market a decade from now, AI will have made their computing skills redundant.
While many Chinese families stake their hopes for the future on AI-related education, some American consumers have sought their “AI future” through purchasing high-tech products. For instance, on Friday, February 28, 2025, a small crowd gathered in a Discord voice channel titled “The death of AI Pin.” They mourned a $699 lapel device that had promised to pull them away from screens and make AI feel effortless and humane. When Humane announced it was shutting down the service, customers described the feeling: “We’re super bummed.” The community’s grief was not just about losing a device, but about how the “AI future” they bought into was not something they owned. The Humane AI Pin encapsulates how an “AI bubble” can be felt at the consumer level.
The failure of this niche object raises questions about the potential consequences of larger platforms collapsing. On Tuesday, November 18, 2025, at 11:20 UTC, almost all users suddenly found themselves unable to access ChatGPT. Users saw “5xx” server error pages. From 11:20 to 14:30, major services like ChatGPT, Canvas, X, Spotify, and thousands of other businesses struggled or even stopped working. This outage was triggered by a small configuration error at Cloudflare, a cloud infrastructure provider that secures a vast portion of the global internet. Downdetector recorded over 2.1 million incident reports across the world, which heavily impacted the U.S., UK, Japan, and Germany, resulting in direct economic loss of 180-360 million dollars. Many users also suffered the loss of data that had not been saved due to the outage on ChatGPT and Canvas.
The relationship between the general public and AI profitability is bound to be complex. The AI boom can manifest as a gold rush, like in Chambliss’ case, or it can be an unprofitable investment, like in Duo Wei’s case. For students and educators, it can create hype for particular education paths and result in training businesses such as the AI youth boot camp. For consumers, the AI race between the U.S. and China brings not only constantly updated, better, and cheaper products, but also globalized risks caused by something like outages.
Martin Ma, PhD ’07 (founder and CEO of Happy Universe), was asked the question “What can ordinary people do in the AI era?” at the Yale Center Beijing event “AI Enables Education, Law and Business for Systemic Change” on Wednesday, January 14, 2025. He responded that the AI era needs three types of talent: those who understand AI technology, those who understand business, and experts in a certain field. That is to say, for practitioners in any field, AI brings new possibilities.
However, the accounts outlined here reveal the other side of the story. While the current AI economy still seeks sustainable profit models, the U.S.-China AI race might become a question of who can weather the long-term costs and aforementioned risks on their populations and other people around the world.
Mafeng Xiao, the founder of a Chinese executive search firm that has served over 1,400 AI companies in recent years (ranging from tech giants to autonomous driving startups and robotics enterprises), acknowledged the vulnerability of the industry. “There is indeed a bubble in AI, but I believe our bubble is not as inflated as that in the U.S. Just look at the valuations: Zhipu AI and Minimax have a huge gap compared to their U.S. counterparts,” he pointed out. History, he suggested, offers a valuable perspective. “This is a battle that no country can afford to lose. The internet and mobile internet also had bubbles, didn’t they? Bubbles are not the problem—they will eventually normalize. What matters is who is still standing when they do.”