We Spent 10 Days Touring Chinese AI Labs. Here's What We Saw
Key Takeaways
- ▸Chinese AI researchers consistently identify insufficient advanced compute access as the primary barrier to building competitive models, with the gap widening as US companies gain Blackwell chips
- ▸The compute constraint represents a structural disadvantage that may prove difficult to overcome, contradicting optimistic Western assessments of Chinese AI capabilities
- ▸Despite resource limitations, Chinese AI labs maintain ambitious AGI research programs, relying on extreme work practices—researchers laboring during holidays and personal time—to compensate
Summary
Journalists from ChinaTalk and Understanding AI spent 10 days in April-May 2026 touring AI and robotics labs across Beijing, Hangzhou, Shanghai, and Shenzhen, interviewing researchers from leading Chinese AI companies including MiniMax, Moonshot, Z.ai (Zhipu), Unitree, and others. A consistent finding emerged across all interviews: Chinese AI researchers universally cited insufficient access to advanced compute resources as their primary constraint in developing competitive, cutting-edge models.
Despite having world-class talent and abundant training data—matching US AI companies—Chinese labs lack the computing infrastructure needed to train large-scale models. This disadvantage is expected to widen as US companies integrate Nvidia's Blackwell chips, far more powerful than any hardware Chinese companies can legally obtain. When interviewed, the researchers dismissed claims from Nvidia CEO Jensen Huang that Chinese companies already possess sufficient compute resources.
Yet despite these structural limitations, Chinese AI labs pursuing AGI remain deeply committed to the goal. The tour revealed an intense work culture where researchers labor through holiday weeks and weekends, with some postponing personal milestones. One researcher mentioned abandoning plans to have a child due to overwhelming demands after the release of DeepSeek's R1 model. The extreme hours appear to be an attempt to compensate for the compute gap through sheer effort and determination.
The reporting illustrates a fundamental asymmetry in the global AI race: while Chinese researchers match their US counterparts in talent and dedication, they operate within structural constraints imposed by export controls and hardware availability that may prove impossible to overcome through effort alone.
- The US-China AI competition reveals an asymmetry: matching talent and data cannot substitute for hardware constraints imposed by geopolitical trade restrictions
Editorial Opinion
This ground-level reporting reveals a crucial reality obscured by headlines about Chinese AI ambitions: structural constraints may matter more than cultural determination. The haunting detail of a researcher sacrificing family planning to meet training timelines captures the human cost of an asymmetrical competition where effort cannot fully substitute for hardware access. Unless Chinese companies achieve breakthrough efficiency gains or secure alternative supply chains, the current trajectory suggests US advantages in compute will compound over time, making AGI pursuits in China increasingly a matter of passion rather than practical feasibility.
