Chip and memory shortages are tightening AI costs
Demand for advanced semiconductors and memory is squeezing capacity and pushing prices higher, which constrains how widely firms can afford frontier AI. TSMC is expected to report another quarter of record profit with 2nm and 3nm capacity sold out into 2027, and industry reports show memory supply is being redirected toward AI workloads, increasing lead times and costs for other manufacturers (reuters.com) (globenewswire.com).
The chips that run frontier artificial intelligence are getting harder to buy, and the memory that feeds them is getting more expensive. Taiwan Semiconductor Manufacturing Co. is expected to report a fourth straight quarter of record profit this week, with analysts forecasting a roughly 50% jump in January-to-March net income on demand for artificial intelligence infrastructure. Reuters reported that the company’s 2-nanometer and 3-nanometer capacity is effectively sold out into 2027. Memory is tightening at the same time. TrendForce said on February 9 that limited supply and rising prices could push the 2026 memory market to $551.6 billion, more than double the projected $218.7 billion for the global foundry market. That squeeze is showing up in the parts artificial intelligence servers need most. TrendForce said the shift from training models to serving answers at scale is lifting demand for high-capacity, high-bandwidth dynamic random-access memory and larger solid-state drives in data centers. High-bandwidth memory is a stack of memory chips placed close to a graphics processor so data can move faster, like adding more lanes to a highway. Building it uses more manufacturing capacity than standard dynamic random-access memory, leaving fewer wafers for older products. Micron said on its March 18 earnings call that fiscal second-quarter revenue reached a record $23.9 billion, with dynamic random-access memory revenue of $18.8 billion and gross margin of 75%. The company also said data-center NAND demand exceeds supply “for the foreseeable future.” The same call showed how fast pricing is moving. Micron said dynamic random-access memory prices rose 65% to 67% sequentially and NAND prices rose 75% to 79% sequentially in the quarter ended February 27, 2026. This is not just a problem for the biggest cloud companies buying graphics processors by the rack. TrendForce said high-bandwidth memory production typically consumes about three times as many wafers as standard dynamic random-access memory, which has tightened supply for conventional personal computer and industrial memory too. Taiwan Semiconductor Manufacturing Co. is still making more money from the boom than memory suppliers are from any single product line, but even its growth is bounded by physics and factory build times. TrendForce said advanced nodes make up only about 20% to 30% of total foundry capacity, while mature nodes still account for roughly 70% to 80%. The result is a narrower market for the most capable artificial intelligence systems. Companies that can lock up chip and memory supply years ahead can keep building, while smaller buyers face longer waits, higher bills, or older hardware.
