On Monday, February 17, the eve of Chinese New Year, Alibaba Cloud released Qwen-3.5 — an open-weight AI model with 397 billion parameters. The timing was deliberate. In the preceding week, nearly every major Chinese AI developer had rolled out new flagship models. ByteDance upgraded its video generation. Zhipu AI launched GLM-5 and immediately hit the top of open-source benchmarks. The demand surge was so intense that Zhipu raised prices 30 percent; their stock jumped 34 percent in a single session.
🎙️ Listen to this article
The South China Morning Post called it the sharpening of the global AI race. CNBC framed it as a shift from chatbots to AI agents. Both descriptions are accurate, but they understate what happened. In the span of seven days, Chinese AI labs collectively demonstrated that open-source models can match or approach the performance of the most expensive proprietary systems from OpenAI, Anthropic, and Google DeepMind — at a fraction of the cost, with weights anyone can download and modify.
The export controls were supposed to prevent this.
The numbers that Washington should be reading every morning
Qwen-3.5's open-source version has 397 billion parameters. Its predecessor, Qwen-3-Max-Thinking, had over a trillion. The new model outperforms the old one on Alibaba's self-reported benchmarks. This is the efficiency pattern that defines China's AI strategy: do more with less, then give it away.
The closed-source version, Qwen-3.5-Plus, claims performance "on par with state-of-the-art leading models" and offers a one-million-token context window — one of the largest in the industry. Alibaba says the model supports 201 languages and dialects, up from 82 in the previous generation. Marc Einstein, research director at Counterpoint Research, told CNBC that this feature "reflects Alibaba's global ambitions."
Zhipu's GLM-5, at 744 billion parameters, sits at the top of the Onyx open-source leaderboard. Below it: Moonshot's Kimi K2.5 at a trillion parameters, MiniMax M2.5 at 230 billion, DeepSeek V3.2 at 685 billion. All Chinese labs. Meta's Llama, which dominated open-source AI throughout 2024, no longer holds the top position on any major benchmark category.
Moonshot AI's pricing tells the cost story. Their flagship model runs at roughly one-seventh the per-token cost of Claude Opus. For a startup choosing between a Chinese open-source model it can fine-tune for free and an American proprietary API that charges by the token, the math is straightforward. According to industry surveys, 80 percent of startups building on open-source infrastructure are using Chinese models. MIT confirmed that Chinese models have surpassed U.S. models in total downloads.
The open-source strategy is a distribution play, not a charity
The American AI business model is proprietary by default. OpenAI, Anthropic, and Google build closed models and sell API access. Revenue comes from per-token pricing, enterprise contracts, and platform fees. The model weights stay locked. The competitive advantage is the model itself.
China's leading labs have adopted the opposite approach: release the weights, let anyone deploy and modify the model, and monetize the ecosystem rather than the artifact. Alibaba makes money through cloud hosting, enterprise services, and the developer ecosystem that forms around Qwen. The model is the distribution vehicle, not the product.
This is the Red Hat playbook applied to AI at continental scale. Red Hat gave away Linux and built a billion-dollar business on support, integration, and enterprise services. Alibaba is giving away Qwen and building cloud revenue on the same logic. The difference is that AI models are more complex than operating systems, the ecosystem effects compound faster, and the geopolitical dimension adds stakes that Red Hat never had to consider.
Every developer who fine-tunes Qwen for a specific use case adds value to Alibaba's ecosystem without Alibaba spending a dollar. Every startup that builds on GLM-5 becomes a stakeholder in Zhipu's success. Open-source models compound through deployment in a way proprietary models cannot replicate because proprietary models can only scale through their vendor's infrastructure.
Export controls optimized for the wrong variable
In October 2022, the U.S. Commerce Department imposed export controls on advanced AI chips to China. The theory was elegant: without access to NVIDIA A100s and H100s, Chinese labs would fall behind on frontier model training. Training requires enormous compute. Restricting compute restricts capability. The logic held for approximately 18 months.
Chinese labs responded by optimizing for efficiency rather than scale. DeepSeek's mixture-of-experts architecture routes each input to only a fraction of the model's total parameters, dramatically reducing the compute required per inference. Qwen's training techniques achieved comparable performance with three times fewer parameters than the previous generation. Zhipu invested in novel architectures that maximize capability per FLOP rather than capability per dollar of hardware.
The result: Chinese models that match or approach Western frontier performance using hardware that falls below the export control thresholds. The controls restricted access to the best training hardware. They did not restrict access to better training algorithms, more efficient architectures, or the mathematical insights that allow smaller models to match larger ones.
Google DeepMind CEO Demis Hassabis told CNBC in January that Chinese AI models were just "months" behind Western rivals. His framing acknowledged what the benchmark data already showed: the gap, to the extent it still exists, is narrow and closing.
The counterargument: open-source models have limits, and China has vulnerabilities
The strongest version of the opposing case runs like this. Benchmarks are self-reported and often gamed. Alibaba's claims that Qwen-3.5 matches OpenAI and Anthropic were noted by CNBC as not independently verified. Open-source models, while capable, often trail proprietary frontier models on the hardest reasoning and coding tasks. And China's hardware supply chain remains constrained — the export controls may not have stopped progress, but they have slowed the rate of scaling.
There is also the trust problem. Enterprises in the United States and Europe face regulatory and reputational risk from deploying Chinese AI models in production. Data sovereignty concerns, supply chain security requirements, and the general deterioration of U.S.-China tech relations create friction that pure cost comparisons don't capture. An American bank is not going to run its compliance AI on Qwen regardless of the benchmark scores.
These constraints are real. They are also less relevant than they appear for the use case that matters most: the global startup ecosystem outside the U.S. and China.
The real battleground is the 5 billion people who don't live in either country
When 80 percent of startups building on open-source AI use Chinese models, that statistic isn't driven by startups in San Francisco or Shanghai. It's driven by startups in São Paulo, Lagos, Bangalore, Jakarta, and Cairo — the cities where the next billion AI users will come from.
For a developer in Nigeria building an AI product for the West African market, the choice between a Chinese open-source model they can deploy on their own infrastructure and an American API that charges per token and requires a U.S. bank account is not a close call. The Chinese model costs less, offers more control, and doesn't create a dependency on a foreign API that could be restricted, repriced, or shut down at any time.
Alibaba's expansion to 201 languages signals exactly this understanding. Qwen 3.5's agentic capabilities — compatibility with frameworks like OpenClaw for autonomous task completion — signal the next phase. CNBC reported that Chinese labs are preparing for the possibility that AI agents could "upend traditional Internet business models." Einstein told CNBC the consequences "for those who are not prepared will be severe."
The agentic dimension makes the distribution advantage self-reinforcing. Agents that run on open-source models create tool integrations, workflows, and business logic that become specific to those models. The switching cost for an agent ecosystem built on Qwen is high, and it compounds with every deployment. This is how standards win: not through mandate, but through accumulation.
The implication Washington keeps missing
Export controls target training compute. China's competitive advantage is in deployment, not training. The controls restricted the wrong variable.
If the United States wants to compete for the global AI ecosystem, the strategy has to shift from restricting China's capability to increasing America's distribution. That means funding open-source American models at a scale that competes with what Alibaba, Zhipu, and Moonshot are giving away. It means pricing API access for developing-world startups at levels that compete with self-hosted open-source alternatives. It means recognizing that the AI race is not a benchmark competition between frontier models. It is a deployment competition for the next billion users.
Meta's Llama was the closest thing the U.S. had to a counter-strategy. Llama's download numbers, while still large, are now second to Qwen's. If that gap widens through 2026, the default AI infrastructure for most of the world's developers will be Chinese. That outcome has consequences for standards, for interoperability, for security, and for the economic relationships that flow from technology dependency.
The export controls did not prevent China from building competitive AI models. They may have accelerated the development of the distribution strategy that makes those models harder to displace. The question the next round of policy has to answer is whether Washington understands what game is actually being played.
