GPU shortages shift sourcing

Companies building artificial intelligence systems are treating graphics processing units less like standard servers and more like scarce industrial inputs that have to be locked up years ahead. (nvidianews.nvidia.com) A graphics processing unit is the chip that does the heavy math for training and running large models, and the newest systems need thousands of them linked together in a data center. OpenAI said on July 22, 2025 that its Stargate build-out with Oracle would bring more than 5 gigawatts of capacity under development, enough to run more than 2 million chips. (openai.com) Nvidia said on February 26, 2025 that fourth-quarter revenue reached $39.3 billion, including $35.6 billion from data center products, and Jensen Huang said Blackwell demand was “amazing.” Nvidia also said Blackwell brought in $11 billion in its first quarter on the market. (nvidianews.nvidia.com) The shortage is not just about Nvidia’s chip die. These accelerators also need high-bandwidth memory and advanced packaging, the step that stacks memory and chip parts together so the processor can move data fast enough for artificial intelligence workloads. (tsmc.com) That has pushed sourcing decisions up the org chart. On April 9, 2026, CoreWeave said Meta expanded a long-term cloud capacity agreement to about $21 billion through December 2032, including some initial deployments of Nvidia’s Vera Rubin platform. (coreweave.com) A day later, CoreWeave said Anthropic signed a multi-year agreement for capacity to support Claude models, with compute coming online later in 2026. On April 15, Reuters reported that Jane Street had also committed about $6 billion for CoreWeave cloud services and taken a $1 billion equity stake. (investors.coreweave.com) (money.usnews.com) The pattern is that companies are buying access to compute, not waiting to buy hardware when they need it. OpenAI said parts of Stargate I in Abilene, Texas were already running by July 2025, and Oracle had started delivering the first Nvidia GB200 racks there. (openai.com) Memory suppliers are signaling the same constraint. In October 2025, SK Hynix said its 2026 supply of key memory products was sold out, a sign that the bottleneck extends beyond the graphics processor itself to the memory stacks that sit beside it. (techspot.com) The result inside companies is a shift from engineering requests to procurement strategy: reserve capacity, sign cloud contracts, spread workloads across providers, and accept that the fastest chip may not be the one available this quarter. The next fight in artificial intelligence is still about compute, but more of it is being decided in supply agreements than in model code. (coreweave.com)