The US is winning the AI race but China might’ve found a shortcut
For years, many researchers and experts believed that China simply couldn’t keep up with the U.S. in the race through the maze of artificial intelligence. China just didn’t have enough high-end chips, top-tier talent, or raw computational firepower that U.S. labs were pouring billions of dollars into.
The idea was that more raw compute power would lead to better AI models. For a while, the idea was correct, that was until a Chinese lab that almost nobody in the West had heard of dropped the equivalent of the Trinity Test in January 2025, when it released DeepSeek R1.
The AI chatbot matched and in some cases surpassed the best American models. But the kicker was that the company that created it did so in two months, with less than $6 million, using chips it wasn’t even supposed to have.
DeepSeek’s release was the first major crack in the West’s AI dominance, and data shows that dominance may soon fade. In 2023, China’s best models were failing about a third of the tasks American models could handle. Now, if you put both countries’ best models in front of a real user and asked them to pick the better answer, they’d be nearly indistinguishable — and China spent 23 times less money to get there.
How do you compare AI capabilities?
It’s important to note that comparing two countries’ AI capabilities is not a straightforward task, and which country looks “ahead” depends on what you measure. It’s not apples-to-apples, but it’s not apples-to-oranges either.
Most researchers look at five main dimensions. The first is model performance benchmarks, which are head-to-head tests that see how well a model performs on tasks like coding, reasoning and math. These are typically the most widely cited and used by analysts but are increasingly contested since companies can fine-tune their models to essentially “game” a benchmark.
The next comparison is compute, or a country’s access to chips. This is a test of raw processing power and who makes the best, most cutting-edge hardware on the market. Another way to look at AI capabilities is how much talent a country is attracting, a trend in which the U.S. is currently ahead, but recent changes have affected that trend.
Investment is also a big factor in capability measurements. Finally, there are AI adoption rates and diffusion. These measure how AI is embedded in the actual economy, not just in leading organizations developing the most advanced foundation models and AI technologies.
How does the US compare?
The U.S.’s approach to AI is like its approach to muscle cars: bigger is better. The U.S. has more than 10 times as many data centers as any other country, at nearly 5,500, and it spends more than any other country on AI.
So far, this approach has worked, and the U.S. continues to release the world’s best frontier AI models like Claude, ChatGPT-5 and Gemini.
But the U.S. still faces challenges in the AI space. The recent surge in energy infrastructure by big tech companies is no coincidence. Companies are undertaking these massive projects to address the country’s energy bottleneck. The International Energy Agency estimates that U.S. data center power demand will reach 426 terawatt-hours, or about 9% of total electricity demand, by 2030.
China has the upper hand in this case, since the country can quickly build new power generation. In 2025, China added more than 540 gigawatts of new power capacity, compared to the U.S.’s 54 gigawatts.
Another major issue facing the American AI industry is its ability to attract foreign AI talent. Since 2017, the number of AI researchers and engineers moving to the U.S. has dropped by 89%, with an 80% drop in 2025 alone, according to Stanford. At the same time, China is focusing on repatriating Chinese-born researchers from the U.S., using high salaries and local competition to entice new talent.
Data shows that enterprise AI adoption is broad at 88%, but it’s uneven. Most of this adoption is concentrated in tech, finance and professional services. Industries like manufacturing are notably behind.
How does China compare?
Chinese AI models routinely match the performance of top U.S. models at a fraction of the cost, according to a March report by the U.S.-China Economic and Security Review Commission. They are also getting popular, with Alibaba’s Qwen chatbot receiving 942 million downloads, more than double the combined downloads of the next eight competitors.
AI rollout in China’s manufacturing sector is nearly double at 67%. Other adjacent industries, like logistics, have also begun using AI in their workflows. JD Logistics now offers a 12-hour delivery window in major Chinese cities using AI, while shipping company Cainiao has used the technology to cut cross-border delivery times by 50%, according to AI Frontiers.
While the U.S. government has taken tepid steps toward AI legislation, China has dived straight in. The Chinese government mandated AI integration across state-owned enterprises and set an ambitious goal of 70% AI penetration across key industrial sectors by 2027.
China also has a unique political feature that the U.S. doesn’t: fewer privacy restrictions on data collection. Because of this advantage, the country can give its companies access to large swaths of real-world behavioral data from its citizens. That access is wildly important as companies use this data to train AI models. Experts estimate that leading U.S. AI companies will begin to run out of high-quality publicly available training data as early as this year.
While China’s system of government is good for some things, it’s worse for others. Chinese startups face a thinner investor base and growing pressure to demonstrate commercial success quickly, CNN reports. The lack of funding forces Chinese companies to go public before their American counterparts.
When it comes to AI chips, things get complicated for China. On one hand, Chinese domestic chips make up more than 40% of China’s AI chip market. Before 2023, that number was almost zero, and NVIDIA had more than 90% of the market, according to Brookings.
On the other hand, these chips are dwarfed in power by those used by American AI companies. But these Chinese-made chips are good enough for most general tasks, and Chinese companies have developed their most advanced AI models to work with them.
The U.S. previously approved the export of later-generation NVIDIA chips to China, but that deal has yet to go through. China has yet to approve the deal as they debate using foreign-made, higher-end chips rather than focusing on domestic production.
So, who’s winning?
For now, America is squarely in first place. While there’s no disputing America’s lead, what’s shifting is how scores don’t fully capture the race.
On the U.S. side, it has the very best models, the latest chips and the most capital. But it’s sitting on an aging power grid, a shrinking pipeline of foreign talent and an adoption curve that’s broad in some areas but nearly empty in others.
China, on the other hand, is using its unique circumstances to build for an AI-centered future. Its models aren’t the best, but they’re getting close, and they’re astronomically cheaper to produce.
If the long-term value of AI is measured by how deeply it is woven into a country’s economy, not how impressive the model actually is, China’s plan may prove more durable.
But at the same time, some in China are pushing back on how close they are to overcoming the U.S.
When asked how likely China is to overtake the U.S. in the next three to five years, Qwen AI technical lead Justin Lin didn’t have a positive answer.
“Below 20 percent,” Lin told CNN. “And I think 20 percent is already very optimistic.”
Round out your reading
- Are aliens real? Pentagon’s newly released UFO files suggest the truth is out there.
- A million Americans get this knee surgery every year. But does it work?
- Is the apocalypse now? One man is tracking flights by the rich to find out.
- ‘No Mow May’: A boon for the bees or a well-intended mistake?
- We’re building a new Straight Arrow. Help us shape our future by taking our survey.
Ella Rae Greene, Editor In Chief
