AI‑cycle keeps components tight
- Chipmakers tied to AI demand report record profits as demand outstrips manufacturing capacity, keeping component markets tight. - SK Hynix posted a five‑fold rise in quarterly profit and warned AI demand exceeds capacity. - Persistent AI‑led tightness implies standard‑cost, long‑lead commitments, and reserve assumptions should stay conservative for electronics manufacturers (reuters.com) (developer.nvidia.com).
Artificial intelligence servers are still soaking up memory chips faster than suppliers can make them, and SK Hynix says demand is still running ahead of capacity. (news.skhynix.com) (reuters.com) SK Hynix said on April 22 that first-quarter revenue rose to 52.5763 trillion won, operating profit reached 37.6103 trillion won, and quarterly revenue topped 50 trillion won for the first time. The company said sales were driven by high-bandwidth memory, server dynamic random access memory, and enterprise solid-state drives. (news.skhynix.com) (prnewswire.com) Reuters reported that SK Hynix’s quarterly profit rose more than five-fold from a year earlier and matched analyst estimates, while the company warned that demand tied to artificial intelligence was exceeding manufacturing capacity. Reuters also said the company is a key supplier of high-bandwidth memory used with Nvidia’s artificial intelligence chips. (reuters.com) High-bandwidth memory is the fast memory stacked next to a graphics processor so an artificial intelligence model can move data quickly, like widening the lanes beside a highway interchange. SK Hynix said it expanded sales of that memory and other premium products even though the first quarter is usually a seasonal slowdown. (news.skhynix.com) (cnbc.com) The squeeze is not limited to one supplier. Micron said on March 18 that fiscal second-quarter 2026 revenue rose to $23.86 billion, and the company said its calendar 2026 high-bandwidth-memory supply was sold out while it was already booking calendar 2027 capacity. (investors.micron.com) Nvidia and its partners have been using the term “AI factory” for data centers built to turn electricity, chips, and software into a steady stream of tokens, images, and model outputs. Nvidia’s developer blog has used that language repeatedly in 2026 for systems that run large-scale training and inference workloads. (developer.nvidia.com 1) (developer.nvidia.com 2) That build-out pulls on more than the headline graphics processors. It also locks up memory, advanced packaging, power gear, and storage on long lead times, because the finished server only ships when all of those parts arrive together. (developer.nvidia.com) (news.skhynix.com) For electronics manufacturers that buy components rather than make them, that usually means standard costs can stay above old assumptions, delivery windows can stay long, and inventory reserves cannot be set as if memory prices are about to snap back. Micron’s sold-out 2026 high-bandwidth-memory book and SK Hynix’s capacity warning point in the same direction. (investors.micron.com) (reuters.com) The near-term test is whether more wafer capacity and packaging lines come online fast enough to cool the market before 2027 orders fill up too. For now, the companies closest to artificial-intelligence memory are still reporting that the bottleneck has not cleared. (investors.micron.com) (reuters.com)
