China’s leaders were anxious and frustrated. The world’s most promising new technology was being dominated by OpenAI, Google and other American companies.

Chinese tech companies were so far behind on generative artificial intelligence early last year that many were relying on Meta Platforms’ open-source Llama models, which can be downloaded for free. Worse, U.S. restrictions on exports of top-end AI chips threatened to hobble China even further.

So in the spring of 2024, Beijing ratcheted up pressure on tech executives. One leading Chinese AI company told The Wall Street Journal it fielded calls from 10 different government agencies in a single month urging action on native AI models. The country relaxed regulations, rolled out funding and rushed to install computing power.

Nine months later, Chinese startup DeepSeek turned heads in Silicon Valley with a powerful new AI model. Hope began to take hold.

“China finally has a model it can be proud of,” Premier Li Qiang told officials, according to people familiar with the comment. The optimism catalyzed China’s tech industry, unleashed an even bigger geyser of government support and jolted American competition into overdrive.

The escalating AI race is drawing comparisons with the Cold War, and the great scientific and technological clashes that characterized it. It is likely to be at least as consequential.

The contest is already helping underpin a worldwide surge in tech spending that has juiced the U.S. and Chinese stock markets and unlocked new sources of economic growth , even as it fuels fears of a global AI bubble.

It is poised to transform industry , society and geopolitics. It’s pushing leaders to sideline concerns about the dangers of powerful AI models, including the spread of disinformation and other harmful content, and the development of superintelligent AI systems misaligned with human values. “The AI future is not going to be won by hand-wringing about safety,” Vice President JD Vance argued in a February speech in Paris.

Both countries are driven as much by fear as by hope of progress.

In Washington and Silicon Valley, warnings abound that China’s “authoritarian AI,” left unchecked, will erode American tech supremacy. Beijing is gripped by the conviction that a failure to keep pace in AI will make it easier for the U.S. to cut short China’s resurgence as a global power.

Both countries believe market share for their companies across the world is up for grabs—and with it, the potential to influence large swaths of the global population.

The U.S. still has a clear lead, producing the most powerful AI models . China can’t match it in advanced chips and has no answer for the financial firepower of private American investors, who funded AI startups to the tune of $104 billion in the first half of 2025, and are gearing up for more . But it has a massive population of capable engineers, lower costs and a state-led development model that often moves faster than the U.S., all of which Beijing is working to harness to tip the contest in its direction.

A new “whole of society” campaign looks to accelerate the construction of computing clusters in areas like Inner Mongolia, where vast solar and wind farms provide plentiful cheap energy, and connect hundreds of data centers to create a shared compute pool—some describe it as a “national cloud”—by 2028. China is also funneling hundreds of billions of dollars into its power grid to support AI training and adoption.

AI experts say it’s too early to tell who will ultimately have the upper hand in AI. And the race, they add, won’t necessarily be determined by who spends more.

China has a history of letting the U.S. take the lead in frontier technologies, only to catch up later as know-how spreads. In social media, Americans were clear leaders early on, only to see TikTok, created by Chinese engineers, eventually dominate and redefine the industry . The country’s AI playbook shows how it is mobilizing on all fronts to do the same thing again.

“Our lead is probably in the ‘months but not years’ realm,” said Chris McGuire , who helped design U.S. export controls on AI chips while serving on the National Security Council under the Biden administration.

Chinese AI models currently rank at or near the top in every task from coding to video generation, with the exception of search, according to Chatbot Arena, a popular crowdsourced ranking platform. China’s manufacturing sector, meanwhile, is rocketing past the U.S. in bringing AI into the physical world through robotaxis, autonomous drones and humanoid robots .

Given China’s progress, McGuire said, the U.S. is “very lucky” to have its advantage in chips.

Cold War echoes

Industry insiders, evoking the Cold War’s space race, often cast the release of ChatGPT as a “Sputnik moment” for China in its intensifying competition with the U.S.

Given AI’s broad applications, the more apt analogy might be the decadeslong U.S. competition with the Soviet Union to build computers for defense. Though not as memorable as the sprint to the moon, that lesser-known Cold War race—which the U.S. won—pushed America’s military, universities and companies to produce innovations in computing that rippled through the global economy, redefined warfare and re-engineered daily life worldwide.

Leaders in both Washington and Beijing now see AI as a revolutionary technology that could surpass digital computing—and its offspring, the internet—in its potential for disruption.

If AI surpasses human intelligence and acquires the ability to improve itself, it could confer unshakable scientific, economic and military superiority on the country that controls it. Short of that, AI’s ability to automate tedious tasks and process vast amounts of data quickly promises to supercharge everything from cancer diagnoses to missile defense.

With so much at stake, hacking and cyber espionage are likely to get worse, as AI gives hackers more powerful tools, while increasing incentives for state-backed groups to try to steal AI-related intellectual property. As distrust grows, Washington and Beijing will also find it hard, if not impossible, to cooperate in areas like preventing extremist groups from using AI in destructive ways, such as building bioweapons.

“The costs of the AI Cold War are already high and will go much higher,” said Paul Triolo , a former U.S. government analyst and current technology policy lead at business consulting firm DGA-Albright Stonebridge Group. “A U.S.-China AI arms race becomes a self-fulfilling prophecy, with neither side able to trust that the other would observe any restrictions on advanced AI capability development.”

China’s vision

China’s focus on the technology dates back at least to 2017, when Xi Jinping unveiled a national AI development plan that called for China to become the world’s AI leader by 2030.

At the time, Beijing was especially interested in its potential to improve facial recognition, central to the government’s surveillance capabilities . The arrival of OpenAI’s ChatGPT in late 2022 showed that AI, aside from making it easier to track and analyze populations, also had the potential to influence the dissemination and manipulation of ideas.

For a Chinese leadership that had built the world’s most sophisticated system of information control, that prospect was both enticing and terrifying.

Leaders’ first instinct was to err on the side of caution. Months after ChatGPT’s release, China imposed the world’s first comprehensive restrictions on deepfakes . A short while later, it introduced rules censoring the input and output of generative AI models, a tightening of earlier rules that required tech companies to disclose extensive details about their algorithms.

But as American generative AI systems grew more powerful, so did China’s fears of missing out on the next great technological leap.

Chinese tech executives lobbied to relax the country’s rules, which at one point required each company to prepare as many as 70,000 questions to test whether its models produced safe answers before they could be approved for public use. Regulators simplified the process, including allowing companies with good track records to skip reviews of training data, according to people involved in the process.

Officials were convinced that AI developers needed further support after the Biden administration tightened export controls on advanced chips in late 2023. The cyberspace regulator started loosening rules for companies that wished to tap computing power abroad for training AI.

Beijing also revved up the engine of state capitalism. More than a dozen local governments began offering researchers access to computing power at subsidized prices through state-run data centers. Some contained restricted American chips that Chinese authorities bought from resellers who got them through underground channels , people familiar with the activity said.

Authorities built up public data sets for training and set up marketplaces where state agencies and companies could trade data. Local governments held roadshows to help raise funds for startups.

When China’s big breakthrough finally came, in early 2025, it wasn’t from a company that benefited heavily from state largess. DeepSeek was largely bankrolled by founder Liang Wenfeng ’s hedge fund and marched to its own beat . But after its problem-solving model R1 nearly matched OpenAI’s top offering in performance at a fraction of the cost, Liang’s little-known startup became the focal point of Beijing’s plan to catch the U.S. in AI.

DeepSeek effect

A month after the release of DeepSeek’s model, Xi convened a meeting with Liang and other Chinese tech executives. Xi told them to lock in on AI, according to people familiar with the meeting, saying the technology could determine China’s ability to compete globally.

That meeting served as a catalyst for the announcement by Alibaba in February that it planned to invest $53 billion over the following three years to pursue artificial general intelligence, according to the people. It also signaled to the industry that Beijing was done playing defense.

In the U.S., OpenAI had already conjured the specter of Chinese competition, arguing in mid-January the U.S. needed to act to direct global resources such as chips and financing away from Chinese-backed projects and toward “democratic AI shaped by the values the U.S. has always stood for.” DeepSeek’s success amplified that argument.

Worries about China were also evident in “ AI 2027 ,” a widely read report published in April by U.S.-based researchers that imagined the results of a global race to achieve artificial superintelligence.

In the researchers’ baseline scenario, fear of China catching up drives the U.S. to press ahead with AI development in 2027 despite evidence its most powerful model is growing out of human control. American political leaders, cowed by warnings about China’s progress, resist tightening AI safety, allowing superintelligence work to continue. By 2030, U.S. and Chinese superintelligences conspire together to wrest control of Earth from humanity, using biological weapons to wipe out all but a few submarine crews and people hiding in bunkers.

A recent slowdown in AI progress has led many in the industry to push their predictions for the arrival of superintelligence further into the future .

Even so, the Trump administration’s “ AI Action Plan ,” unveiled in July, made it clear that Washington was growing more wary of China. The plan promised to investigate the degree to which Chinese models advance Communist Party priorities, and called on federal agencies to counter Chinese influence on multinational bodies involved in setting global standards for AI.

Xi released his own sprawling AI blueprint, “ AI Plus ,” not long after. It doesn’t mention the U.S. explicitly, but proclaims Beijing’s intention to use AI to “reshape the paradigm of human production and life.” It aims for AI to be used in 70% of China’s economy by 2027, and 90% by 2030.

Chip ‘swarms’

Industry participants believe China may be as much as a decade away from making microchips that match America’s best products, mainly due to U.S. restrictions on its access to advanced chipmaking technology. The shortage of high-end computing power has delayed Chinese companies such as DeepSeek in developing their next-generation models, people close to the firms said.

Beijing has escalated efforts to mobilize its tech champions to develop their own semiconductor supply chain , pressing companies such as ByteDance to suspend Nvidia purchases and work with Chinese chip companies to build AI.

Huawei has collaborated with thousands of local firms, often coordinated by different levels of government, to develop advanced semiconductor technologies , including systems that bundle up to one million chips to boost computing capabilities. Huawei said that approach can match Nvidia’s best systems, despite consuming more power, leading some in the industry to label the strategy “swarms beat the titan.” Local governments have started subsidizing data centers’ electricity bills if they use domestic chips, industry people said.

Huawei also plans to more than double its chip capacity by next year, people familiar with the matter said.

In October, Xi laid out a five-year plan to achieve higher-level technology self-sufficiency, pledging extraordinary measures to pursue breakthroughs in key technologies such as semiconductors.

The U.S. is still making its own advances, as Nvidia and others push the boundaries of chip technology. To draw even, China would need to erase America’s more than half-century head start.

“To say that China can snap their fingers and indigenize this because they’ve indigenized other things in the past ignores that this might be the single hardest thing in the world for them,” said McGuire, the former NSC official.

A key question is whether simply boosting computing power with ever-better chips will be enough to keep yielding more-powerful AI models, says Helen Toner , director of strategy for Georgetown’s Center for Security and Emerging Technology and a former OpenAI board member. If performance plateaus despite all the spending by OpenAI and others—a growing concern in Silicon Valley—China has a chance to compete.

“We don’t actually know which way the technology is going to go,” she said.

Write to Josh Chin at Josh.Chin@wsj.com and Raffaele Huang at raffaele.huang@wsj.com