AI infrastructure strain
Cloud customers are trying to book massive AI capacity as enterprise demand outpaces supply, with reports that some AWS customers are attempting to buy out large portions of the provider's AI capacity. (cio.com) The supply squeeze shows up across the stack — Nvidia is expanding partnerships internationally and component suppliers like Samsung Electro‑Mechanics are being named as substrate providers for inference chips — while investor coverage warns to separate long‑term AI opportunity from near‑term deployment frictions. ( )
Amazon Web Services is turning away requests from customers that want to reserve enormous chunks of its chip capacity for 2026. (aboutamazon.com) In his April 9 shareholder letter, Amazon chief executive Andy Jassy said two large Amazon Web Services customers asked to buy all of the company’s Graviton instance capacity available in 2026, and Amazon said no. Jassy also said Amazon’s chips business, including Graviton, Trainium, and Nitro, is running at more than $20 billion in annual revenue and growing at triple-digit year-over-year rates. (aboutamazon.com) Amazon is pushing its own silicon as a cheaper way to run artificial intelligence workloads in the cloud. Jassy said three years into the current artificial intelligence wave, Amazon Web Services’ artificial intelligence revenue run rate topped $15 billion in the first quarter of 2026, and Trainium2 was fully subscribed with 1.4 million chips deployed. (aboutamazon.com (aboutamazon.com)) The bottleneck is not one part. Artificial intelligence systems need chips, advanced packaging, networking gear, power, and data center space at the same time, and shortages in any one layer can slow deployment. (cnbc.com) Nvidia put a number on that demand in March. At its March 16 developer conference, chief executive Jensen Huang said Nvidia sees $1 trillion in orders for Blackwell and Vera Rubin systems through 2027, up from the company’s earlier $500 billion opportunity outlook. (cnbc.com) That demand is pulling suppliers into the build-out. Samsung Electro-Mechanics was reported on April 10 to be supplying flip-chip ball grid array substrates for Nvidia’s Groq3 language processing unit, an inference chip tied to Nvidia’s next-generation Vera Rubin platform, with mass production expected to start in the second quarter. (businesskorea.co.kr) Nvidia is also chasing more capacity outside the United States. Reporting in late February said Yotta Data Services is investing $2 billion in an Nvidia graphics processing unit hub in India as domestic data center capacity there races to catch rising demand for Blackwell Ultra systems. (ainewsinternational.com) Investors are hearing two messages at once. Nvidia’s long-term order pipeline has expanded, but Motley Fool wrote on April 11 that insider stock sales and deployment friction are feeding doubts about how smoothly that demand converts into near-term revenue. (fool.com) Amazon is making a similar argument from the cloud side. Jassy said artificial intelligence costs should fall over time, but current demand is high enough that Amazon is still deploying heavy capital into chips and infrastructure to keep up. (cnbc.com (aboutamazon.com)) For customers, the immediate problem is simpler than the forecasts: the market has plenty of appetite for artificial intelligence compute, and not enough ready capacity to satisfy every large order at once. (cio.com)
