AI supply crunch

The AI hardware shortage has shifted from a chip story to a factory story: customers can order Nvidia’s newest systems, but many still cannot get them quickly. (nvidianews.nvidia.com) Nvidia said on February 25 that fiscal 2026 revenue rose 65% to $215.9 billion, with data center revenue reaching $62.3 billion in the January quarter alone. Jensen Huang said customers are “racing to invest in AI compute,” even as Blackwell systems keep ramping. (nvidianews.nvidia.com) TrendForce said on April 8 that Blackwell is expected to make up 71% of Nvidia’s high-end GPU shipments in 2026, up from 61% in its earlier view. The firm also cut its 2026 high-end GPU shipment growth estimate to about 26%, citing supply-chain changes and Rubin delays. (trendforce.com) A modern AI server needs more than a graphics processor. It also needs high-bandwidth memory, a stacked memory package that sits close to the chip, plus advanced packaging and network parts that tie whole racks together. (trendforce.com) That memory is tight. Micron said on March 18 that tight industry supply helped drive record fiscal Q2 2026 results, with revenue of $23.86 billion, and it described memory as a “strategic asset” for AI customers. (investors.micron.com) Micron had already told investors in its December 17 fiscal Q1 2026 earnings call that its 2026 high-bandwidth memory volume was sold out and that customer negotiations for 2026 volume and pricing were complete. SK hynix also said late last year that its 2026 supply of key memory products was effectively booked. (fool.com, techspot.com) The bottleneck runs through Taiwan Semiconductor Manufacturing Co., the foundry that makes many leading-edge AI chips and related dies. TSMC reported on April 16 that first-quarter 2026 revenue rose 35.1% year over year to NT$1.134 trillion, while 3-nanometer chips accounted for 22% of wafer revenue and 5-nanometer chips for 36%. (investor.tsmc.com) TSMC told investors on the same day that it is stepping up capital spending to increase 3-nanometer capacity for a multiyear pipeline of artificial-intelligence demand, including high-performance computing and high-bandwidth-memory base dies. That is the plainest sign that the constraint is no longer only chip design, but how fast suppliers can add the exact factory steps AI hardware needs. (investor.tsmc.com) Reports circulating in industry media have put a number on the squeeze, including claims of a Blackwell backlog in the millions and foundry utilization above 95%, but those figures have not been confirmed in the official company filings reviewed here. What is confirmed is that Nvidia, Micron, SK hynix, and TSMC are all describing the same market from different angles: demand is arriving faster than the supply chain can expand. (nvidianews.nvidia.com, investors.micron.com, investor.tsmc.com, trendforce.com) That leaves cloud companies, start-ups, and corporate buyers competing for the same scarce mix of GPUs, memory stacks, packaging slots, and rack parts. In 2026, the question is not whether companies want more AI hardware; it is which supplier can deliver it first. (trendforce.com, nvidianews.nvidia.com)