China’s open source AI revolution was supposed to be the great equalizer. The country’s labs released powerful models that anyone could download, customize, and deploy, winning hearts and GitHub stars across the globe. But according to Tom Shaughnessy, co-founder of Delphi Digital, the economics of keeping those models open simply don’t add up.

Shaughnessy flagged a core tension: Chinese labs are spending enormous sums on frontier model training without generating direct revenue or building proprietary data feedback loops from those open releases. Inference, the actual use of these models, often happens through third-party providers like Venice or OpenRouter. The labs foot the bill for training. Someone else captures the value.

The open source dilemma takes shape

Models like DeepSeek and Alibaba’s Qwen family attracted global attention, racking up downloads and uploads on platforms like Hugging Face. Through late 2025, Chinese open models outpaced Meta’s Llama on some adoption metrics. The strategy served multiple purposes: spurring domestic innovation, asserting digital sovereignty, and neatly sidestepping US export controls on advanced chips.

The cracks started showing in early 2026. Alibaba launched Qwen3.6-Plus and Qwen3.5-Omni as hosted, cloud-based offerings rather than fully open releases. Z.ai rolled out GLM-5-Turbo as a closed-source model. ByteDance’s and Kuaishou’s video generation models have remained proprietary entirely.

Chinese media have started calling this the “open source dilemma,” a phrase that captures the collision between idealistic openness and the cold reality of investor expectations. Companies need to show returns. Investors want revenue, not download counts.

The $100 billion target driving the pivot

Look at Alibaba’s numbers to understand the pressure. The company is reportedly targeting $100 billion in combined AI and cloud revenue over the next five years. Hitting that target requires high-end proprietary technologies, the kind companies pay premium prices to access.

Meta remains the biggest counterweight, continuing to release Llama models under permissive licenses. But Meta’s motivations are different: it doesn’t sell cloud compute, so open sourcing models costs it less competitively. Chinese cloud companies don’t have that luxury. Every open model release is a potential customer lost to a self-hosted deployment.