Chinese AI companies spent years giving away their most advanced models for free, accepting massive losses as a strategic investment. Now that strategy is finally paying off, thanks to an unexpected shift in how people use AI. Instead of one-off chatbot conversations, enterprises are deploying AI agents that run continuously overnight, burning through tokens at rates that dwarf traditional usage patterns. This transformation has turned open-source AI from a loss leader into a genuine profit engine .

Why Did Chinese AI Companies Give Away Their Best Models?

For the past two years, the Chinese AI ecosystem operated on a radically different business model than Silicon Valley. While OpenAI, Anthropic, and other US companies built proprietary systems and guarded them closely, Chinese labs like Alibaba's Qwen, DeepSeek, Moonshot, and others released their models as open-source, allowing anyone to download, use, and build on them for free . This wasn't charity. It was strategic .

The approach served multiple purposes. First, it helped smaller Chinese labs gain global attention and compete against entrenched US players. Second, it weakened the competitive moat that US companies were trying to build through massive infrastructure spending. Third, it aligned with China's tech culture, which has historically resisted paying for software. GitHub, one of the few major Western sites still accessible in China, became a talent pipeline that helped the country produce roughly half of the world's AI researchers .

The strategy worked spectacularly. Alibaba's Qwen family of models captured more than half of worldwide open-source model downloads, becoming the most widely used foundation for researchers building smaller, specialized versions. These models powered everything from Singapore's government AI initiatives to academic research labs across the developing world. A venture capital partner at Andreessen Horowitz estimated that roughly 80 percent of companies pitching them that use open-source AI were employing Chinese models .

But there was a problem: open-source models didn't generate revenue. Alibaba Cloud's computing margins sat in the single digits, while OpenAI and Anthropic booked 40 to 50 percent margins on closed models. Every token call looked identical. A user typed a prompt, the model produced a response, and the interaction ended. Switching costs were near zero, and users had no reason to stay loyal to any particular provider .

What Changed to Make Open-Source AI Profitable?

The answer arrived in late 2025 and early 2026: AI agents. Unlike chatbots that respond to a single prompt, agents plan, decompose tasks, call models multiple times, read results, reason about them, and retry when things break. A single overnight session running an agent can consume more tokens than an entire month of chatbot use by the same person .

The numbers tell the story. Zhipu AI raised API prices 83 percent in the first quarter of 2026, and call volumes jumped 400 percent anyway. Tencent Cloud hiked pricing on its Hunyuan series by more than 400 percent. Moonshot's Kimi K2.5 earned more in roughly 20 days than the company made in all of 2025 . MiniMax's average daily token consumption on its M2 series climbed six-fold between December 2025 and February 2026 .

"Users now leave OpenClaw running while they sleep, an intensive process he described as potentially boosting demand exponentially," noted Zhipu CEO Zhang Peng at a recent industry event.

Zhang Peng, CEO at Zhipu AI

This wasn't a marginal shift. It was a fundamental change in the product itself. Agents create something chatbots never did: lock-in. Once an agent is configured with a specific model's tool-calling format, memory structure, and prompt conventions, switching providers mid-workflow breaks things. Enterprises running agent workloads on top of Qwen or other Chinese models stay put, creating what economists call inelastic demand. Users can't easily switch, they need more tokens per task, and demand stopped responding to price increases .

How Are Chinese AI Companies Monetizing This Shift?

The mask is slipping on the open-source commitment. Several major Chinese AI companies are quietly closing off their most advanced models and restricting commercial use .

• Alibaba: Reorganized five separate AI units into what it now calls the Alibaba Token Hub, reporting directly to CEO Eddie Wu. In his announcement letter, Wu used the word "tokens" three times and made no mention of openness or the developer community.
• MiniMax: Released its latest M2.7 model and amended its terms of use to prohibit commercial use without authorization, blindsiding the open-source community.
• Zhipu, Z.ai, and Baidu: Released some of their newest models as closed-source at launch and hiked prices across models and cloud services.

The shift reflects brutal domestic competition. Start-ups like Zhipu and MiniMax went public, increasing pressure to satisfy shareholders. Alibaba, after reshuffling its AI research team, needed to find ways to offset training and infrastructure costs. Closing off advanced models and charging premium prices for API access became the obvious solution .

Distribution matters enormously. Tencent's QClaw embeds AI agents inside WeChat's 1.3 billion users. Alibaba's Qwen agent reaches 300 million monthly active users across Taobao, Tmall, and Alipay. ByteDance's Doubao has 315 million chatbot users and launched a phone with ZTE in December. When distribution combines with agent integration, it creates a meter spinning whenever customers sleep, generating continuous token consumption .

Is Open-Source AI Dead in China?

Not yet, but the era of pure open-source generosity is ending. China's AI industry has never been monolithic. ByteDance, for example, has always kept its models closed. And while Alibaba's retreat looks like a turning point, the shift is unlikely to happen all at once .

The real test is DeepSeek. The company's highly anticipated next release will reveal whether the national champion intends to preserve the open-source tradition it helped supercharge or accelerate its decline. DeepSeek's breakthrough R1 model spurred a frenetic year in AI, and its next move will signal whether Chinese labs are moving toward a hybrid model, mixing open and proprietary releases, or walking away from openness altogether .

What's clear is that the business model has fundamentally shifted. Chinese open-source AI was a loss leader waiting for something to monetize. Agents turned out to be it. As China's National Data Administration reported, Chinese models held the top six spots on OpenRouter's global weekly rankings for the week ending April 5, with Qwen3.6-Plus setting a single-day record of more than 1.4 trillion tokens. Daily token calls in China rose from 100 billion at the start of 2024 to 140 trillion by March 2026 .

The giveaway finally found a cash register. It took an AI agent to install it.