TSMC recently posted strong earnings, helped in large part by Nvidia orders, but both companies are grappling with chips inside their own operations: TSMC says they are too expensive, while Nvidia says there are not enough to go around.
To accelerate internal AI adoption and automation in its drive to build smart factories, TSMC chairman C.C. Wei disclosed several years ago that the company had bought large numbers of Nvidia GPU chips. He joked that the chips were extremely sought after and costly, saying TSMC sold the chips it manufactured at NT$600 each, but spent US$200,000 buying AI-related products. He added that even if the price was very high, the company still had to buy them.
That showed how aggressively TSMC has pursued AI-driven smart manufacturing in recent years. Even GPUs produced at TSMC's own fabs must be bought back at great expense, while Nvidia itself has not been able to fully escape the impact of AI chip shortages.
According to a recent report by Business Insider, Nvidia automotive chief Rishi Dhall said the company's auto division still has to compete internally for the GPUs that have helped make Nvidia the world's most valuable company.
As the generative AI wave accelerates, technology companies worldwide are racing to develop large language models and build data centers, sending demand for high-performance AI computing chips sharply higher. Nvidia GPUs have become the core compute engine for the AI industry, and OpenAI, Microsoft, and Amazon have all increased chip purchases, putting long-term pressure on supply.
Dhall said on a recent episode of The Verge podcast Decoder that even within Nvidia, GPU supply for computing across business units is actually limited. His remarks offered a rare look at how Nvidia allocates resources internally.
As demand for Nvidia chips keeps climbing, semiconductor manufacturing capacity has also become another internal battleground. He said different Nvidia teams often have to compete with one another for the compute resources needed to train and test their own AI models.
According to Dhall, Nvidia has built a tiered internal system for allocating compute resources, with staff coordinating every week how to distribute capacity across training, testing, and other scenarios.
In practice, resource allocation balances short-term operating needs with long-term strategic planning, rather than focusing only on immediate returns. Dhall said that when coordination over GPU allocation becomes difficult, Huang himself has to step in.
He added that Nvidia balances current business needs with long-term strategic opportunities, including what Huang calls "the zero trillion dollar business," markets that do not yet exist but could one day grow into businesses worth trillions of dollars.
One of those bets is autonomous driving. Dhall said self-driving is a key long-term strategic area for Nvidia, which believes automation across mobility platforms is an industry trend and is investing heavily in chips, software, AI models, simulation tools and safety systems for autonomous vehicles.
He noted that Nvidia's push into autonomous driving is not simply about becoming a chip supplier to the automotive industry. On June 1, 2026, Nvidia released Cosmos 3, a fully open omni-modal world model that can perform physical reasoning, world generation, and action generation at the same time.
Nvidia defines Cosmos 3 as physical AI infrastructure and treats it as a concept parallel to CUDA. CUDA makes AI developers depend on Nvidia GPUs, while Cosmos is designed to make robot and autonomous driving developers depend on Nvidia's world models.
Dhall stressed that although the auto business is still far smaller than Nvidia's booming data center division, Huang continues to prioritize it. He said both he and Huang are firm believers in the future of autonomous driving, and Nvidia is continuing to invest in that technology and future direction, not only through external compute allocation but also through foundry capacity support.
Article translated by Rodney Chan and edited by Jack Wu
