Why China’s AI Suddenly Looks So Strong

The New York Times reported that DeepSeek released its models as open source, meaning others could freely use and modify them. That stood in contrast to OpenAI and Anthropic, whose leading models remained proprietary. The same report said the episode showed that an open-source system could come close to the performance of closed versions .

That does not mean China has leapt ahead everywhere. CSIS cited Chinese researchers noting that China still cannot access the most advanced chip process technology, and The Decoder, reporting on a Stanford analysis, said tests by the U.S. government center CAISI found DeepSeek models were, on average, 12 times more vulnerable to jailbreak attacks than comparable U.S. models .

1. The talent pool was already mature

DeepSeek was not an isolated miracle. CSIS argues that Chinese researchers have been at or near world-class level in many AI research domains for years, while DeepSeek marked the first time a Chinese AI lab was widely viewed globally as a frontier competitor .

Stanford HAI also published a May 2025 policy analysis focused specifically on DeepSeek’s talent base, underlining that the source and development of AI talent has become central to understanding the company’s competitiveness . In plain terms: DeepSeek’s breakthrough looks less like a lucky shot and more like the visible result of a long research and engineering build-up.

2. Chip limits pushed engineering efficiency into the spotlight

U.S. export controls on advanced chips are a major part of the background. CSIS cited Li Guojie of the Chinese Academy of Engineering, who said in February 2025 that, because of U.S. government restrictions, China was unable to obtain the most advanced chip process technology .

That does not mean export controls automatically caused China’s AI progress. The causal story is more complicated. But when access to top-end compute is harder, model teams have a stronger incentive to squeeze more performance from training, inference and deployment. That is why DeepSeek-R1 landed with such force: it was not only about model capability, but also about the claim that similar performance could be achieved more cost-effectively than with OpenAI’s comparable model .

3. Open models multiplied the impact

DeepSeek’s release strategy mattered almost as much as the model itself. The New York Times described DeepSeek’s open-source approach as a sharp contrast with the proprietary route taken by OpenAI and Anthropic .

Open models spread differently. Researchers, developers and companies can test, adapt and deploy them without waiting for a single vendor’s API roadmap. According to the same New York Times report, Chinese firms released dozens of other open-source models in the months after DeepSeek, and by the end of 2025 those models accounted for a significant share of global AI use .

That diffusion effect is one reason China’s AI suddenly seemed everywhere. Open availability can turn one high-profile model into an ecosystem signal.

4. Cost discipline changed the adoption math

Frontier AI is not only a race for the highest benchmark score. It is also a race to run models reliably at a price businesses can justify. DeepSeek-R1 drew attention because it connected capability with cost: DeepSeek said R1 was more cost-effective than OpenAI’s similar model .

For companies, that changes the procurement question. If an open model is good enough for a particular task, the buyer may reassess whether it needs to rely on a single proprietary vendor . That does not make cost claims automatic proof of lower total cost. Latency, security, hosting, private deployment, support and maintenance can all change the real bill. But DeepSeek helped make cost efficiency a front-line issue in AI strategy.

5. A broad application ecosystem helps models become products

INSEAD places DeepSeek within the rise of a broader Chinese AI ecosystem, arguing that China has built a robust AI base capable of challenging U.S. dominance . RAND, meanwhile, analyzes China’s AI industrial policy through a full-stack lens, meaning the relevant story is not only one model company but the wider set of industrial capabilities around it .

That matters because models become more valuable when they can be embedded into products, workflows and sector-specific tools. Once a model reaches a usable threshold, a dense ecosystem gives it more chances to be tested, adjusted and commercialized. The competition is therefore not just chatbot versus chatbot. It is model capability plus deployment infrastructure, business use cases, developer adoption and policy resources .

6. Policy attention and dense competition accelerated iteration

China has treated AI as a strategic industry for years. RAND describes China’s AI industrial policy as an evolving full-stack approach, focused on more than any single model or company .

After DeepSeek-R1, that confidence became more visible. Carnegie’s analysis says the early-2025 release transformed the global AI landscape and gave Chinese leaders new confidence in domestic AI development. It also notes that Chinese leaders invited AI pioneers to high-level meetings, encouraged local governments to speed AI deployment in critical infrastructure and promised to improve AI laws and policies .

Market competition also matters. The New York Times reported that, after DeepSeek, Chinese companies released dozens of additional open-source models . A crowded open-model field pressures companies to lower adoption barriers, improve deployment and respond faster to developers.

The caveats: strong does not mean unbeatable

Advanced chips remain a bottleneck. Chinese AI teams have made major efficiency gains, but CSIS cited Chinese researchers saying China still lacks access to the most advanced chip process technology .

Open models approaching closed models is not the same as across-the-board superiority. The New York Times reported that open-source systems could perform almost as well as closed versions; it did not show that Chinese models win every frontier task. OpenAI and Anthropic’s leading models remained proprietary .

Safety and governance still need scrutiny. The Decoder reported that CAISI tests found DeepSeek models were, on average, 12 times more vulnerable to jailbreak attacks than comparable U.S. models . For sensitive uses, that is not a footnote — it is part of the risk calculation.

What this means for companies and developers

The practical result is simple: buyers and builders have more model options. Open models make it easier to test, modify and deploy alternatives, while DeepSeek’s cost story forces organizations to rethink whether the most expensive or most closed option is always the right one .

In practice, the model’s country label matters less than its performance on your actual workload. Teams should:

• Benchmark models on their own data, languages and tasks, not only public leaderboards.
• Compare accuracy, latency, stability, total cost, licensing terms and deployment options.
• Run jailbreak and safety tests for high-risk or public-facing applications, especially when sensitive data is involved .
• Consider chip and hardware supply constraints if the business depends on high-end compute or a specific infrastructure stack .

Bottom line

DeepSeek is not the only reason China’s AI became competitive. It is the event that made the accumulation visible. China’s AI looks suddenly strong because several conditions reached a tipping point together: talent depth, efficiency pressure from compute constraints, open-model distribution, cost discipline, application ecosystems and policy support .

The clearest judgment is not that China has won the AI race. It is that China has become extremely competitive in open models, cost efficiency and fast deployment — while advanced chips, some closed-frontier capabilities, safety testing and trust in high-risk settings remain unresolved questions .