Grace Shao on What the World Should Know About Chinese AI

What It Covers

Independent AI researcher Grace Shao joins Odd Lots in Hong Kong to explain how Chinese AI labs like DeepSeek, MiniMax, Moonshot, and Zhipu AI operate under compute, capital, and talent constraints — and why their open-source, utilitarian approach is generating real revenue while closing the gap with US frontier models.

Key Questions Answered

• •Open-source as business strategy: Chinese labs initially open-sourced models not for ideological reasons but to build trust with Western developers — a branding decision. This created a collaborative ecosystem where labs share breakthroughs, congratulate each other publicly, and build on shared foundations, reducing duplicated R&D spend under tight capital and compute constraints.
• •Revenue model for open-source AI: MiniMax and Zhipu AI are projected to reach $1–1.2 billion ARR, with monthly revenue now matching their entire prior-year total. The profit center is managed inference services — enterprises pay for API access to avoid self-hosting costs including GPUs, deployment, security, and monitoring, mirroring how open-source software companies have always monetized.
• •Post-training optimization as catch-up strategy: Lacking resources for broad R&D exploration, Chinese labs wait for US frontier labs to establish direction, then concentrate entirely on post-training. They also time proprietary dataset purchases strategically — waiting out 3–6 month exclusivity windows to buy the same data at roughly one-tenth the original price paid by OpenAI.
• •DeepSeek V4's hardware signal: DeepSeek delayed its V4 release by 3–4 months specifically to re-engineer inference onto Huawei chips. By releasing this as open-weight, other Chinese labs could study the implementation and begin shifting inference workloads to domestic hardware — a deliberate step toward reducing ecosystem dependence on NVIDIA and CUDA.
• •Hybrid model deployment at application layer: US AI-native companies including Harvey and Cursor are already building on Chinese open-source models like GLM and Kimi for cost efficiency, while using closed models like Claude Opus selectively as a judge or evaluator. This hybrid approach lets application-layer companies cut token costs without sacrificing quality on critical tasks.