AI memory shortfall risk

Artificial intelligence’s next bottleneck may be memory, not chips: analysts say memory makers are on pace to cover only about 60% of projected demand by 2027. (waredata.com) The crunch centers on high-bandwidth memory, or HBM, the ultra-fast memory stacked next to an AI processor so it can feed data quickly enough to keep the chip busy. Google Cloud called HBM “one of the most common roadblocks” in AI development because a model stops when that memory fills up. (cloud.google.com) WareData, citing Nikkei Asia and Counterpoint Research, said producers would need to lift annual output about 12% in 2026 and 2027 to close the gap. Counterpoint’s estimate for actual growth is roughly 7.5%, leaving a supply shortfall near 40%. (waredata.com) The companies at the center of that gap are Samsung Electronics, SK hynix and Micron, which dominate advanced memory supply. SK hynix said in January it would invest 19 trillion won in a new Cheongju packaging plant, with operations targeted for the end of 2027. (trendforce.com) That timing is the problem. TrendForce said on March 31 that a “clear shortage” is expected in 2026 and that meaningful capacity expansion is unlikely until late 2027 or 2028. (trendforce.com) The shortage is already changing who gets memory first. TrendForce said North American cloud service providers are signing long-term agreements to lock up supply, while suppliers keep reallocating capacity toward HBM and server products. (trendforce.com) That leaves smaller buyers competing for what is left. TrendForce said PC original equipment manufacturers and module makers are getting lower allocation fulfillment rates and are being pushed into higher-priced purchases. (trendforce.com) The price effect is broad, not limited to AI servers. TrendForce projected conventional dynamic random-access memory prices would rise 58% to 63% quarter over quarter in the second quarter of 2026, while NAND flash prices would rise 70% to 75%. (trendforce.com) SK hynix remains the strongest player in the tightest part of the market. WareData said the company holds about 32% of global DRAM and more than half of the HBM segment, giving it outsized influence as buyers race to secure supply. (waredata.com) If new plants arrive on schedule in late 2027 and 2028, the squeeze can ease. Until then, the memory attached to AI chips looks set to stay scarce, expensive and concentrated in the hands of the biggest cloud buyers. (trendforce.com)