The AI Memory Boom Is on a Boom-Bust Track, and Relief Waits Until 2028

SK Hynix and Micron’s revenues have tripled over the past year, and Samsung’s has roughly doubled, as AI datacenters absorb nearly every gigabyte of DRAM, NAND flash, and high-bandwidth memory (HBM) the industry can produce. The Register’s Tobias Mann argued in a July 12 analysis that this boom is following memory’s familiar boom-bust pattern, just at a larger scale than usual, and that the earliest realistic exit from it is still years away. Mann cites a recent IDC report warning that relief will not arrive before 2028 at the earliest, a timeline separate IDC research on the PC market corroborates.

Why the Shortage Won’t Break Before 2028

IDC has been tracking the same tightness from multiple angles. In a June 22 blog post, IDC’s Soo Kyoum Kim argued that memory has stopped behaving like the cyclical commodity it used to be and has become “a strategic infrastructure input” instead: AI training and inference workloads do not ease off between quarters the way consumer demand traditionally has, so each new deployment adds to a demand baseline rather than fading with the season. Kim’s post says public guidance from major memory manufacturers has been explicit that tight conditions are not a short-term anomaly, and that HBM, high-density DRAM, and enterprise-grade NAND are now allocated through longer, tighter supply agreements than standard components ever were.

A separate IDC analysis of the PC market, published June 2, puts a number on the near-term pain: IDC forecasts global PC shipments will fall 11.3% in 2026, with average selling prices climbing 17% over the same period. “We’re not seeing any relief to the memory shortage situation before the end of 2027, which means prices will continue to rise and PC manufacturers will struggle to maintain full product portfolios for the foreseeable future,” said Jean Philippe Bouchard, IDC’s vice president of devices and consumers. IDC’s Jitesh Ubrani added that “even as memory capacity expands over the next two years, pricing is unlikely to return to 2025 levels.”

Vendors Are Betting Hundreds of Billions on Catching Up

The big three memory vendors are now investing hundreds of billions of dollars to bring new fab capacity online, according to Mann’s reporting. Micron’s contribution includes a July 9 announcement of up to $3 billion to strengthen the US semiconductor supply chain, including $500 million in strategic financing for GlobalWafers Co.’s 300mm raw silicon wafer facility in Sherman, Texas, paired with a 10-year supply agreement for wafer capacity. US Commerce Secretary Howard Lutnick and Trade Representative Jamieson Greer both issued statements welcoming the deal, which Micron says still requires definitive agreements, customary approvals, and closing conditions before it is finalized.

South Korea has gone further. President Lee Jae Myung announced on June 29 an investment plan exceeding $576 billion, anchored by Samsung Electronics and SK Hynix, to expand the country’s chip fabrication capacity and cement what Lee called a “triple axis” of semiconductors, physical AI, and data centers.

None of that capacity arrives quickly. Building a new DRAM or NAND wafer fab means securing financing, winning permits, and installing hundreds of millions of dollars of lithography, wafer transport, and testing equipment before a single chip ships, a process that Mann’s reporting says typically takes at least three years and longer still to reach full production yields.

The Bust Risk Underneath the Boom

Mann’s analysis frames the current shortage as an unusually severe version of a cycle memory vendors have ridden before: manufacturers pour capital into new capacity during a boom, historically only to watch that capacity arrive just as demand cools and prices crash. What makes this cycle different, per the Register’s reporting, is that AI infrastructure demand has kept prices climbing through 2025 and 2026 instead of falling the way the historical pattern would predict, and IDC’s Kim frames that shift as structural rather than a temporary spike.

That divergence cuts both ways. If AI infrastructure demand keeps pace with the capacity vendors are now building, memory makers stay flush and buyers keep paying elevated prices for years. If it does not, Mann warns, the same new fabs built to chase the boom could flood the market right as demand retreats, producing what his piece calls a bust cycle “to end all bust cycles.” For AI labs already spending heavily on GPUs and training runs, sustained high memory prices are one more cost weighing on the timeline for turning years of investment into profit.

Bottom Line

The near-term numbers favor memory vendors: tripled and doubled revenues, an IDC-tracked shortage with no relief before 2028, and manufacturers and governments committing hundreds of billions of dollars to new capacity. That capacity will not land for years, and in the meantime buyers from smartphone makers to AI labs have no real alternative to paying elevated prices. Whether the eventual landing is soft or the bust cycle memory has historically produced depends on whether AI demand keeps growing fast enough to absorb all the new fabs now under construction, a question none of this month’s announcements answer.

Sources

• The Register: Memory makers are slaves to the boom-bust rollercoaster, and the AI boom is the wildest ride of all
• IDC: Why the memory market is still tight: what comes next
• IDC: PC Market Enters Volatile Territory as Memory Shortage Persists Through 2027
• Micron: Micron Announces Up to $3 Billion Strategic Investment to Strengthen U.S. Semiconductor Ecosystem
• The Japan Times: South Korea taps Samsung and SK Hynix in $576 billion AI-chip drive

Featured image: macro photograph of an HBM DRAM die from an AMD Radeon RX Vega 64, showing the stacked memory’s TSV contact grid. Photo by Fritzchens Fritz, public domain (CC0) via Wikimedia Commons; cropped, resized, and converted to WebP.