DeepSeek and the Open-Source Strategy Reshaping China’s AI Industry
Chinese AI labs are betting that open models win the long game
In January 2026, DeepSeek released DeepSeek-V3, an open-source large language model that matched or exceeded GPT-4o on most benchmarks while costing a fraction to train. The release sent shock through Silicon Valley — not because of the model’s capabilities, but because of what it represented: China’s AI labs had found a way to compete by giving away their technology.
The open-source strategy is now the defining feature of China’s AI industry in 2026. Alibaba’s Qwen, Baidu’s ERNIE, Zhipu’s GLM, and Moonshot’s Kimi have all released significant open-weight models in the first half of the year. The logic is straightforward: if you can’t sell the model, sell the infrastructure and services around it.
The economics of free models
DeepSeek’s training cost for V3 was reportedly under $6 million — a fraction of the $100+ million estimated for GPT-4. The company achieved this through engineering innovations in mixture-of-experts architecture, training data curation, and hardware utilization. By releasing the weights openly, DeepSeek drives adoption, which in turn drives demand for its inference API and enterprise services.
“The model is the marketing,” said Liang Wenfeng, DeepSeek’s founder, at a Beijing tech conference in March 2026. “We make money when companies deploy our models at scale and need our infrastructure to run them.”
Impact on global AI competition
The open-source wave from China is compressing the advantage that closed-source Western labs once held. When Meta released LLaMA 4 in April 2026, the open-source community immediately benchmarked it against DeepSeek-V3 and Qwen-3 — and found the Chinese models competitive or superior in Chinese-language tasks and reasoning benchmarks.
This creates a dilemma for US policy. Export controls restrict China’s access to cutting-edge chips, but open-source models allow China to maximize the performance of whatever hardware it has. A team of researchers at Tsinghua University demonstrated in May 2026 that DeepSeek-V3 running on domestic Ascend 910C chips achieved 85% of the throughput of the same model on NVIDIA H100s.
The enterprise deployment wave
Beyond the model competition, Chinese enterprises are deploying AI faster than their Western counterparts in specific sectors. A survey by the China Academy of Information and Communications Technology (CAICT) published in May 2026 found that 67% of large Chinese enterprises had deployed generative AI in production — up from 34% a year earlier.
The fastest adoption is in financial services (82% deployment rate), e-commerce (76%), and manufacturing (71%). The primary use cases are customer service automation, document processing, and code generation.
What to watch in H2 2026
The inference cost war is the next battleground. As open-source models commoditize, the competition shifts to who can serve them cheapest. Chinese cloud providers (Alibaba Cloud, Huawei Cloud, Tencent Cloud) are already offering inference at 30-50% below US cloud prices, driven by lower energy costs and domestic chip subsidies.
The question for global enterprises: does the cost advantage of Chinese AI infrastructure outweigh the regulatory and geopolitical risks of depending on it?
Sources
- RadarAI, “China AI Industry Developments 2026: What’s Actually Changing”
- CAICT, “Enterprise AI Deployment Survey,” May 2026
- Tsinghua University, “DeepSeek-V3 on Ascend Hardware Benchmark,” May 2026
- DevFlokers, “AI News June 2026: Models, Research & Tech Developments”