China’s open-source large language models are now responsible for nearly 30 percent of global AI usage, propelled by rapid releases from Qwen, DeepSeek, and Moonshot AI. Chinese-language prompts have climbed to the No. 2 spot worldwide by token volume. The message for global investors is straightforward: China has crossed from fast follower to co-leader in AI. The beneficiaries span cloud, chips, servers, consumer devices, and EVs, with revenue leverage building into 2026.
A new empirical analysis of 100 trillion tokens routed through a major model gateway shows China’s open-source LLMs grew from barely 1.2 percent global share in late 2024 to nearly 30 percent in a matter of months this year. Usage remains diversified, with no model above 25 percent share, but the cadence out of China stands out. Qwen and DeepSeek are shipping at a pace that lets developers recalibrate workloads—and costs—faster than most Western peers. The practical edge is efficiency. Open-source reduces switching costs and lets enterprises tune models to local data, including Chinese-language prompts that now account for roughly 5 percent of all requests, well above the language’s 1.1 percent footprint on the public internet. That is a durable demand signal, not a flash in the pan, and it arrives despite U.S. curbs on high-end GPUs.
This jump is not accidental. Beijing’s multi-year innovation policy has aligned compute clusters, data standards, and safety frameworks with commercial adoption. Alibaba Cloud’s engineering muscle, DeepSeek’s cadence, and Moonshot AI’s Kimi are the visible proof points. Beneath them sits a financing backbone that can fund the buildout: ICBC remains the world’s largest bank by assets at roughly 6.7 trillion dollars and the most profitable among Chinese companies, providing the capacity to support next-gen infrastructure at scale. Brand gravity matters too: Tencent, Alibaba, and others rank among China’s most valuable brands, reinforcing distribution across consumer and enterprise channels. Internationally, China’s official financing has been pivoting into developed markets, with the United States emerging as the largest single recipient of Chinese official sector credit over two decades, sharpening avenues for digital infrastructure and services to ride along. That linkage accelerates cross-border AI deployment.
Open-source adoption converts into revenue in three lanes. First, cloud providers monetize fine-tuning, orchestration, vector search, and inference at scale, even if base models are free. Lower model costs expand total addressable usage and can lift gross margins on incremental workloads. Second, server and networking vendors benefit as enterprises build their own stacks to meet cost and sovereignty goals. Third, application-layer champions in e-commerce, fintech, and mobility can fold LLMs into recommendation, customer service, and autonomy features without proprietary licensing drag. The second-half 2025 acceleration in usage points to a 2026 earnings cycle where compute shipments, cloud billings, and AI-attached services all inflect, particularly across Southeast Asia, the Middle East, and Latin America where local language and data localization drive open-source preference.
1. Alibaba Group (BABA) – Qwen leads China’s open-source surge with fast release cycles that keep cost-per-token low for developers. Alibaba’s cloud division grew revenue 34 percent in 2024 to 39.8 billion yuan, positioning it to capture fine-tuning and inference spend as Qwen usage scales. Global impact: Qwen’s accessibility is pulling SMEs across emerging markets into enterprise-grade AI without proprietary licensing fees. 2. Tencent Holdings (TCEHY) – With a market cap of about 593.81 billion dollars and the No. 1 Chinese brand by value, Tencent can distribute open-source AI across WeChat, gaming, and fintech surfaces. Milestone: its consumer reach and payments rails make it a natural monetizer of AI agents, while cloud offerings integrate open-source models for enterprise. Global impact: cross-border gaming and fintech provide data and distribution beyond mainland China. 3. Baidu Inc. (BIDU) – Baidu’s PaddlePaddle is one of China’s most established open-source deep learning frameworks, anchoring a developer ecosystem that feeds both ERNIE stacks and vertical solutions. Milestone: years of open-source tooling give Baidu leverage to win enterprise hybrid deployments. Global impact: regional developer adoption in Asia supports sovereign AI builds. 4. iFlytek Co. (002230.SZ) – China’s speech-AI leader is embedding LLMs into voice assistants, education, and office tools. Milestone: long-standing dominance in Mandarin and dialect speech tech makes iFlytek a natural bridge between text LLMs and multimodal interfaces. Global impact: multilingual voice AI lowers adoption barriers in emerging markets. 5. Inspur Information (000977.SZ) – A top-three global server vendor, Inspur is a direct beneficiary of AI workloads moving on-prem for cost and sovereignty reasons. Milestone: strong share in AI-optimized servers positions Inspur to supply clusters for Qwen and DeepSeek deployments. Global impact: supplies cloud and enterprise datacenters across Asia, EMEA, and Latin America. 6. Lenovo Group (LNVGY) – Lenovo is pushing AI PCs and edge servers to bring open-source models directly to endpoints. Milestone: as the world’s largest PC maker, Lenovo can pre-install model runtimes and accelerators that reduce cloud costs. Global impact: global channel reach turns China’s open-source innovations into worldwide device experiences. 7. Cambricon Technologies (688256.SH) – Cambricon’s NPUs target both datacenter and edge inference, giving Chinese enterprises a domestic compute path. Milestone: shipping accelerators into servers and embedded boxes that run open-source LLMs efficiently. Global impact: diversifies the global AI supply chain away from single-vendor GPU dependence. 8. Kingsoft Cloud (KC) – Offers cost-optimized cloud services integrating open-source model stacks for Chinese and international customers. Milestone: hybrid solutions that bundle storage, vector databases, and model orchestration. Global impact: a natural partner for Southeast Asian clients building sovereign AI. 9. BYD Company (BYDDY) – The EV champion is infusing vehicles with on-device AI for infotainment, copilots, and energy optimization. Milestones: FinDreams Battery holds 17 percent global EV battery share and BYD ranked 143rd on the Fortune Global 500 in 2024. Global impact: exporting AI-enabled EVs at scale gives BYD a powerful data and software flywheel.
The economics line up. Open-source LLMs reduce total cost of ownership, allow language customization, and minimize regulatory friction by keeping sensitive data local. That is why Chinese-language prompts have become the No. 2 global prompt language despite a smaller web footprint: developers are solving real workloads in finance, logistics, and government services across Asia and beyond. The report’s token distribution shows stronger adoption in hubs like Singapore and Germany, with China ranking fourth by token share, underscoring that this is a cross-border phenomenon. Pair that with China’s export scale in devices and autos, and the result is an end-to-end delivery system for AI, from models to chips to endpoints.
Two tensions are in play: pricing pressure as open-source undercuts proprietary margins, and supply constraints in high-end accelerators. Yet the usage surge is broad enough to lift all layers of the stack. For investors, a barbell approach is sensible: own infrastructure names leveraged to compute shipments and cloud billings, and pair them with application-layer platforms that can convert AI into transaction growth. Watch for operating leverage as inference costs fall and usage expands. On the policy side, China’s financial depth and active industrial strategy temper supply risk by subsidizing local ecosystems and accelerating domestic chip adoption.
Expect more dense release cycles from Qwen and DeepSeek, with Moonshot Kimi expanding context windows and tools. The model landscape is now fragmented by design, which benefits routers and orchestration layers while preventing any single model from taxing developers. On earnings, look for cloud backlog growth tied to AI workloads, server shipment guidance, and disclosure on model-tuning revenue. Key metrics include token volumes by language, cost-per-million tokens in inference, and edge AI attach rates in PCs and EVs. Also watch how Singapore, Germany, and Middle East buyers standardize on open-source stacks—markets where Chinese models are already seeing strong pull. The open-source wave is here, and China’s engineering scale is turning it into a durable, global business cycle.
