China's AI Ecosystem Is More Advanced Than Most People Think

The Assumption That Needs Updating

There is a persistent assumption in Western technology circles that US export controls on advanced semiconductors have meaningfully constrained China’s AI capabilities. The logic is straightforward: frontier AI training requires enormous amounts of compute, compute requires the most advanced GPUs, the US has restricted China’s access to those GPUs, therefore China’s AI development should be falling behind.

This assumption is wrong, or at least far more nuanced than most people realize. China’s AI ecosystem has continued to advance at a pace that has surprised even close observers. Chinese AI labs have released models that perform competitively with Western counterparts on major benchmarks. Chinese companies have deployed AI applications at a scale that rivals or exceeds anything in the West. And the Chinese approach to AI development has diverged from the American approach in ways that may prove strategically significant.

Understanding what is actually happening in China’s AI ecosystem is essential for anyone trying to assess the global AI landscape. The picture is more complex — and in some respects more impressive — than the export control narrative suggests.

DeepSeek: The Lab That Changed the Conversation

No single development has done more to challenge Western assumptions about Chinese AI capabilities than DeepSeek. Founded in 2023 as a research lab affiliated with the quantitative hedge fund High-Flyer, DeepSeek has produced a series of models that have performed at or near frontier levels on standard benchmarks while using architectures and training approaches that differ meaningfully from those of Western labs.

DeepSeek’s V3 model, released in late 2024, demonstrated performance on reasoning and coding benchmarks that was competitive with GPT-4o and Claude 3.5 Sonnet. The DeepSeek-R1 reasoning model, which followed shortly after, pushed further into territory that Western observers had expected to remain the exclusive domain of US and UK labs.

What made DeepSeek’s results particularly notable was the efficiency of its approach. DeepSeek has been open about its use of Mixture-of-Experts architectures, which activate only a subset of the model’s parameters for any given input, reducing the compute required for both training and inference. The lab has also published detailed technical reports describing innovations in training methodology, including novel approaches to reinforcement learning from human feedback.

The efficiency angle matters because it directly addresses the export control thesis. If Chinese labs can achieve competitive performance with less compute per training run — through architectural innovation, training efficiency, or a combination of both — then hardware restrictions become a less effective constraint than their architects intended. DeepSeek’s results suggest that this is happening.

DeepSeek also benefits from a talent pool that, while distinct from the Western AI research community, is deep and technically sophisticated. The lab has recruited heavily from top Chinese universities and from the broader Chinese technology industry, which has its own extensive experience with large-scale machine learning systems.

Alibaba’s Qwen: The Quiet Powerhouse

While DeepSeek captured headlines, Alibaba’s Qwen series has arguably been more significant in terms of breadth and deployment.

Qwen 2.5, released across multiple size points ranging from small models suitable for edge deployment to large models competitive with frontier Western systems, represents one of the most comprehensive model families available from any lab globally. The models span text, code, mathematics, and multimodal capabilities, and they have been released with open weights under permissive licenses.

The open-weight strategy is strategically important. By making competitive models freely available, Alibaba has built an ecosystem around Qwen that extends far beyond its own cloud platform. Developers and companies worldwide — including many outside China — have adopted Qwen models for applications ranging from customer service to code generation. This creates a form of soft power: as the global developer community builds on Chinese-origin models, the technical standards and architectural choices embedded in those models propagate widely.

Qwen’s performance on multilingual benchmarks is particularly strong. While Western models have historically been optimized primarily for English, Qwen models demonstrate strong performance across Chinese, English, and a range of other languages. For applications serving non-English-speaking populations — which constitute the majority of the world — this multilingual capability is a significant advantage.

Alibaba Cloud has also invested heavily in the inference infrastructure needed to serve Qwen models at scale. The company’s domestic cloud platform is one of the largest in Asia, and it has been aggressively expanding its international presence. The combination of capable models and scalable infrastructure positions Alibaba as a credible alternative to Western AI providers for companies and governments seeking to avoid dependence on US technology.

Baidu, ByteDance, and the Broader Ecosystem

DeepSeek and Qwen are the most visible components of a much larger ecosystem.

Baidu’s ERNIE series has been deployed widely within China, powering the company’s search engine, cloud services, and enterprise AI offerings. ERNIE 4.0 demonstrated competitive performance on Chinese-language benchmarks and has been integrated into Baidu’s suite of business applications. While Baidu has received less international attention than DeepSeek or Alibaba, its domestic deployment scale is enormous — hundreds of millions of users interact with ERNIE-powered services daily.

ByteDance, the parent company of TikTok, has built substantial AI capabilities that are less visible because they are primarily deployed within the company’s own products rather than offered as external services. ByteDance’s recommendation algorithms, which drive TikTok and its Chinese counterpart Douyin, represent some of the most sophisticated deployed AI systems in the world. The company has also invested in generative AI, though it has been less public about its frontier model development.

Beyond the major companies, China has a deep bench of AI startups and research institutions. Zhipu AI, backed by Tsinghua University, has produced the GLM series of models. Moonshot AI, 01.AI (founded by Kai-Fu Lee), and Minimax have all released capable models. The Chinese Academy of Sciences and major universities continue to produce research that is competitive with the best Western institutions.

The aggregate effect is an ecosystem that is largely self-sufficient. Chinese companies can train models on Chinese-manufactured hardware (albeit less efficient than the latest NVIDIA GPUs), serve them on domestic cloud infrastructure, and deploy them to a market of over a billion users. The ecosystem does not depend on access to Western technology — it is increasingly an independent technological trajectory.

The Export Control Paradox

US export controls on advanced chips were designed to maintain a technological advantage in AI by denying China access to the most powerful training hardware. The policy has had real effects: Chinese labs have faced constraints in acquiring the latest NVIDIA GPUs, and the workarounds — including purchases through intermediaries and stockpiling before restrictions took effect — are imperfect substitutes for unrestricted access.

But the controls have also produced unintended consequences that partially undermine their strategic intent.

First, the restrictions have accelerated China’s investment in domestic semiconductor capabilities. Huawei’s Ascend series of AI accelerators, while not yet matching NVIDIA’s most advanced GPUs in raw performance, has improved rapidly and is being deployed at increasing scale within China. SMIC, China’s leading chip fabricator, has made progress on advanced manufacturing processes despite being cut off from the most advanced lithography equipment. The timeline for China to achieve self-sufficiency in AI-capable chips has been compressed by the urgency that export controls created.

Second, the compute constraints have pushed Chinese labs toward efficiency innovations that may prove strategically valuable. When you cannot simply throw more hardware at a problem, you are forced to find smarter approaches to architecture design, training methodology, and data efficiency. DeepSeek’s results suggest that this pressure has produced genuine innovations. If these efficiency gains are transferable — and early evidence suggests they are — then Chinese labs may eventually achieve frontier performance with less hardware, not more.

Third, the controls have strengthened the political case within China for technological self-reliance. The narrative that dependence on Western technology is a strategic vulnerability has been reinforced by the export restrictions, accelerating investment in domestic alternatives across the technology stack. This extends beyond AI chips to include operating systems, cloud infrastructure, and development tools.

Application-Layer Divergence

Perhaps the most underappreciated aspect of China’s AI ecosystem is the scale and sophistication of its application layer.

Chinese technology companies operate in a market environment that differs from the West in ways that accelerate certain types of AI deployment. Mobile payment systems are nearly universal. E-commerce platforms handle transaction volumes that dwarf Western equivalents. Social media ecosystems integrate commerce, communication, and content in ways that create rich datasets for AI training and deployment.

The result is that Chinese companies have deployed AI at application scale in domains where Western companies are still in pilot phases. AI-powered customer service, content moderation, recommendation, fraud detection, and supply chain optimization are deployed at a scale in China that reflects years of iteration and optimization.

The Chinese approach to AI regulation has also diverged from the Western approach in ways that affect deployment speed. While China has implemented AI regulations — including requirements for algorithmic transparency and restrictions on deepfakes — the regulatory framework has generally been designed to enable deployment rather than constrain it. This contrasts with the European approach, which has prioritized precautionary regulation, and the American approach, which remains fragmented across agencies and jurisdictions.

The Multimodal and Reasoning Frontier

Chinese labs have been particularly aggressive in pushing multimodal capabilities — models that process and generate text, images, video, and audio.

Alibaba’s Qwen-VL (Vision-Language) models have demonstrated strong performance on visual understanding benchmarks. ByteDance’s research in video generation and understanding builds on the company’s core expertise in short-form video. Baidu has invested in multimodal models that integrate with its search and mapping services.

On the reasoning frontier, DeepSeek-R1 demonstrated that Chinese labs can compete in the development of models with enhanced reasoning capabilities. The model’s performance on mathematical reasoning and coding benchmarks was competitive with the best Western alternatives, and its open release allowed the broader research community to study its approach.

The significance of these advances is not that Chinese models are ahead of Western models — on most benchmarks, the leading Western models retain an edge, though the gap varies by task. The significance is that the gap is far narrower than the export control narrative would predict, and that Chinese labs are innovating on dimensions — efficiency, multilingual capability, open-weight release — that could prove more strategically important than raw benchmark performance.

What Western Observers Get Wrong

Western analysis of China’s AI ecosystem tends to make several systematic errors.

The first is equating hardware access with capability. Advanced GPUs are important, but they are one input among many. Data, talent, algorithms, and deployment infrastructure all matter, and China has significant strengths in each of these areas.

The second is assuming that benchmark performance is the only relevant metric. Chinese AI companies operate in a market of enormous scale and are deploying AI in production at levels that generate operational expertise. This deployment experience — understanding failure modes, optimizing for real-world conditions, building robust serving infrastructure — is a form of capability that benchmarks do not capture.

The third is treating China’s AI ecosystem as a monolith. The reality is a diverse ecosystem with competing companies, distinct research traditions, and a range of strategic approaches. DeepSeek’s research-focused approach differs fundamentally from Alibaba’s platform strategy, which differs from Baidu’s integration strategy. This diversity creates resilience and drives innovation through internal competition.

What to Watch

Several developments will clarify the trajectory of China’s AI ecosystem over the coming year.

First, the performance of domestically manufactured AI chips, particularly Huawei’s next-generation Ascend processors. If domestic hardware narrows the performance gap with NVIDIA’s latest offerings, the long-term impact of export controls diminishes significantly.

Second, the trajectory of DeepSeek’s research output. The lab’s pace of publication and model releases in 2024 and 2025 was remarkable. Whether it can sustain that pace — and whether its efficiency innovations continue to close the gap with Western frontier models — will be a key indicator of the broader ecosystem’s competitiveness.

Third, the international adoption of Chinese AI models. Qwen’s open-weight strategy is designed to build a global developer ecosystem around Chinese-origin models. The extent to which developers outside China adopt these models — particularly in Southeast Asia, the Middle East, Africa, and Latin America — will shape the geopolitics of AI standards and infrastructure.

The world is not converging on a single AI ecosystem dominated by American companies. It is diverging into at least two major ecosystems, each with its own models, infrastructure, and deployment patterns. Recognizing this reality — and understanding the strengths of both ecosystems — is the starting point for any serious analysis of where AI is heading.