Even Nvidia, the company whose GPUs power the generative AI boom, cannot meet its own internal demand for computing resources. The head of its automotive division, Xinzhou Wu, disclosed this week that teams across the company compete weekly for limited GPU supply, and that CEO Jensen Huang is sometimes called upon to settle disputes. The admission, made on a technology podcast, offers a rare window into the severity of the global AI chip shortage, which has forced the world’s most valuable chipmaker to ration access even for its own engineers training and testing AI models.

The internal scramble mirrors the external battle for Nvidia’s chips. In recent months, the company has intensified a crackdown on its supply chain across Asia, physically inspecting data centres and disqualifying more than half of its approved customers in key hubs. The move, driven by Washington’s pressure to close smuggling loopholes into China, has cut off many “neocloud” providers—specialist cloud firms renting AI computing power. Beijing’s own export restrictions have compounded the squeeze, blocking even downgraded Nvidia chips and pushing its market share for top-tier processors in China to zero, even as domestic AI development strains under a compute shortage.

The resource war is reshaping competitive dynamics. Google has been pitching its custom Tensor Processing Units (TPUs) to neoclouds, offering to backstop data-centre financing in a bid to break Nvidia’s dominance. Nvidia has reportedly responded with financial incentives to keep customers loyal, a tactic that could attract regulatory attention. SoftBank founder Masayoshi Son, speaking in Tokyo, dismissed bubble fears as “foolish,” projecting $5 trillion in annual global infrastructure spending by 2040 and that AI-related industries could replace 20% of world GDP. Nvidia’s Huang, also in Tokyo, similarly rejected bubble fears, citing “incredibly strong” demand and a decade-long buildout ahead.

The contrasting narratives—of desperate chip rationing and of boundless long-term optimism—are now converging on a handful of concrete tests. Nvidia’s next-generation Vera Rubin platform is scheduled for large-scale deployment from 2027, while SoftBank’s Stargate project and Google’s TPU push will require tens of billions in fresh capital. The immediate milestone, however, is Nvidia’s next quarterly earnings report, which will offer the clearest signal yet on whether the supply-demand imbalance that has defined the AI chip market is beginning to ease or is set to deepen further.