Infrastructure’s hidden chokepoints
Analysts warn AI scale depends on obscure supply lines — for example, gallium is emerging as a strategic mineral risk for chip makers, and its supply is concentrated in China. At the same time, Nvidia has been taking minority stakes in specialist vendors rather than big acquisitions, and hyperscalers are locking long‑term compute via large contracts like Meta’s roughly $21 billion deal with CoreWeave. (benzinga.com) (invezz.com) (cloudnews.tech)
Artificial intelligence expansion is running into bottlenecks far below the chatbot layer: specialty minerals, niche software, and long-term cloud contracts. (usgs.gov) (coreweave.com) (sec.gov) Gallium is a soft metal used in chip materials such as gallium arsenide and gallium nitride, which show up in microelectronics, integrated circuits, and optoelectronic devices. The United States Geological Survey says low-purity unrefined gallium is not produced in the United States and U.S. demand is met by imports. (usgs.gov 1) (usgs.gov 2) (usgs.gov 3) China tightened export licensing controls on gallium and germanium in 2023, and the United States Geological Survey said in November 2024 that a total Chinese export ban on both could cut U.S. gross domestic product by about $3.4 billion. The same study said gallium carried the highest supply risk among the minerals it analyzed. (usgs.gov 1) (usgs.gov 2) The Department of Energy says China accounts for more than 90 percent of global gallium production, and the Center for Strategic and International Studies has described Beijing’s position as a virtual monopoly. That concentration turns a byproduct of aluminum processing into a strategic pressure point for chip and defense supply chains. (energy.gov) (csis.org) The scramble is not only about raw materials. Companies also need software that decides which artificial intelligence job runs on which graphics processor, the way an air-traffic controller assigns planes to gates and runways. (nvidia.com) (ec.europa.eu) Nvidia has been building around those chokepoints with targeted deals instead of only giant takeovers. It announced a deal to buy Run:ai on April 24, 2024, and its 2026 annual report says it holds private-company investments that are non-marketable and illiquid at the time of purchase. (nvidia.com) (sec.gov) That approach has shown up in cloud infrastructure, too. CoreWeave said on February 24, 2026 that Nvidia invested $2 billion in its Class A stock, and CoreWeave separately disclosed a new $6.3 billion order form with Nvidia in September 2025 for reserved cloud capacity and access to unsold capacity. (sec.gov 1) (sec.gov 2) The buyers are locking in supply years ahead. CoreWeave said on April 9, 2026 that Meta expanded its agreement to roughly $21 billion for cloud capacity through December 2032, with deployments across multiple sites and some initial use of Nvidia’s Vera Rubin platform. (coreweave.com) (sec.gov) Those contracts show how the artificial intelligence race is shifting from buying chips one quarter at a time to securing power, racks, networking, software, and minerals across most of a decade. A missing metal or an unavailable cluster can now delay model launches as surely as a chip shortage. (coreweave.com) (usgs.gov) (nvidia.com) The result is a quieter map of artificial intelligence power. The companies that control obscure inputs, scheduling software, and reserved compute are becoming as important to scale as the chip designers whose logos sit on the servers. (sec.gov) (coreweave.com) (usgs.gov)
