While focus is on massive supercomputers for training next-gen models, the real supply chain constraint will be 'inference' chips—the GPUs needed to run models for billions of users. As adoption goes mainstream, demand for everyday AI use will far outstrip the supply of available hardware.

The primary constraint on US AI leadership relative to China isn't the ability to build models, but the slow pace of developing necessary compute and energy infrastructure. China faces fewer regulatory barriers, allowing it to scale these critical inputs more rapidly.

A critical, under-discussed constraint on Chinese AI progress is the compute bottleneck caused by inference. Their massive user base consumes available GPU capacity serving requests, leaving little compute for the R&D and training needed to innovate and improve their models.

China's push for open-weight models is not just ideological but a strategic necessity. Lacking compute for large-scale inference and facing a tough market for their closed models, open-sourcing is a way to gain traction, talent, and influence where US controls have limited them.

The performance gap between Chinese and American frontier AI models is not due to a lack of talent or different training techniques. Instead, it is primarily constrained by access to massive-scale compute and the capital required to procure it.

AI expert Noam Brown suggests the strategic high ground in AI is moving from simply possessing model weights to having the massive inference capacity to deploy them. This implies that even if a model is stolen or distilled, the ability to run it at scale becomes the true competitive advantage and geopolitical chokepoint.

China's open-source model ecosystem is structurally unstable. The billion-dollar fixed costs for training frontier models are unsustainable for Chinese tech giants who lack a clear AI revenue narrative and cannot match the compute budgets of Western labs like OpenAI or Anthropic.

Faced with limited access to top-tier hardware, Chinese AI companies have been forced to innovate on model architecture to compete. They've developed superior techniques in memory management and multi-token prediction, making their models highly efficient and formidable competitors despite hardware constraints.

Previously, the biggest constraint in AI was compute for training next-gen models. Now, the critical bottleneck is providing enough compute for *inference*—the real-time processing of queries from a rapidly growing user base.