Export-banned Nvidia chips are behind China’s most powerful AI models—uh oh, Jensen! - TechCrunch

Export controls have failed to sever compute access, instead cementing Nvidia's monopoly in China's AI infrastructure. This exposes the futility of physical isolation against deep software ecosystem dependency: leading Chinese models are tightly coupled to the CUDA architecture, rendering short-term alternatives non-viable. This path dependency internalizes compliance costs, forcing firms to operate in regulatory gray zones where performance necessity overrides supply chain security. Competitors like AMD struggle to fill the void due to immature software stacks, unable to disrupt the status quo. Over the next 12-24 months, we will witness the 'compute decoupling' paradox: stricter policies increase the marginal value of existing chips, deepening Nvidia's moat. This is not merely a commercial battle but an early signal of fractured global division of labor. It implies that top-tier AI models will accelerate toward private, customized compute clusters to mitigate single-source hardware risks, marking a structural shift in how AI sovereignty is defined.