If you thought the memory market would settle down once initial model training clusters were built, Samsung just killed that dream.
During its calendar Q2 2026 earnings call, the Korean semiconductor giant issued a stark warning: the memory supply crunch isn't easing anytime soon. In fact, it is going to get worse in 2027 and will stretch at least through 2028.
The culprit isn't just massive foundation model training anymore. It is agentic AI. As autonomous software agents move into production, token processing workloads are exploding. That continuous demand for real-time inference is soaking up high-bandwidth memory and server-grade DRAM faster than fabs can burn silicon.
Meanwhile, Samsung is counting cash. The company reported calendar Q2 revenue of ₩171.5 trillion ($119 billion)—a 130 percent jump year-over-year. Operating profit exploded by more than 19 times compared to the previous year, reaching ₩89.5 trillion ($62.4 billion).
If you're buying memory for enterprise datacenters, PCs, or edge infrastructure, prepare for years of elevated prices and tough negotiations.
The Three-Year Fab Trap
Building a modern semiconductor fabrication facility is not a quick process. You cannot simply flip a switch or add a third shift to a production line when demand spikes.
Jaejune Kim, Samsung's Executive Vice President of Memory, made this clear during the earnings call. The lead time from breaking ground on a new fab to pulling finished wafers off the line exceeds three years. Because equipment delivery, cleanroom certification, and yield optimization take so long, capacity planned today won't deliver meaningful volume until late 2027 or 2028.
Kim noted that unmet demand from 2026 will inevitably spill into 2027. Supply constraints next year will likely be more severe than what we are seeing today.
Market analyst firm TrendForce agrees with this timeline for DRAM. While several memory makers plan to start new production lines in 2027, construction schedules and equipment installation mean production ramps will be delayed until the second half of 2027. That pushes true relief out to 2028.
Because lead times are locked in stone, memory manufacturers cannot bridge the gap between supply and demand immediately. Fabs are running at capacity, but the wafer allocation is shifting sharply toward high-margin AI components like HBM4, HBM4E, DDR5, and SOCAMM2. Every wafer dedicated to high-bandwidth memory for AI hardware reduces the silicon available for standard DDR5 modules or consumer devices. This reallocation is a hallmark of what analysts are calling the AI memory supercycle—a structural shift that breaks traditional boom-bust cycles by permanently redirecting fab capacity toward inference-grade silicon.
Record Cash and Multi-Year Captive Contracts
When supply shrinks and demand skyrockets, pricing power shifts entirely to suppliers. Samsung is using this leverage to rewrite how it sells enterprise hardware.
Rather than selling memory on quarterly spot markets or standard annual contracts, Samsung is forcing major buyers to sign multi-year supply agreements.
"Customers who want to secure substantial AI service infrastructure are increasingly approaching us for multiyear supply," Kim told investors. He noted that these multi-year arrangements align with Samsung's goal of hedging its mid- to long-term risk. Consequently, Samsung is prioritizing customers who can guarantee committed future captive demand.
Translate that corporate speech into plain English: if you are willing to lock in high prices across multiple years, Samsung will put you at the head of the line. If you want to shop around on the open market, you get whatever scraps remain.
This strategy protects chipmakers from the boom-and-bust cycles that historically gutted semiconductor margins. By locking hyperscalers and major infrastructure providers into multi-year commitments, Samsung secures guaranteed revenue even if macro demand fluctuates later. The broader memory market is experiencing similar dynamics—Micron's recent performance shows how Wall Street is pricing in this multi-year AI memory demand, with investors betting that the supercycle will outlast initial hype cycles.
For enterprise IT departments, however, this creates a nasty lock-in. Buyers lose bargaining power and must absorb higher component costs into their multi-year capital budget forecasts.
DRAM vs NAND: A Diverging Bottleneck
The memory crunch won't hit every storage layer in the exact same way. While DRAM remains severely constrained, the outlook for NAND flash offers a rare bright spot.
TrendForce predicts that NAND flash supplies could begin to ease in the second half of 2027. This relative relief is driven by accelerated migration to higher-layer NAND architectures and the gradual ramp-up of new fabrication plants. As bit density per wafer grows through layer stacking, NAND output growth will outpace 2026 levels.
DRAM tells a completely different story. High-bandwidth memory variants—such as HBM4 and HBM4E—require complex 3D stacking, advanced interposers, and tight thermal packaging. These manufacturing requirements reduce total wafer yields compared to standard DRAM. As long as agentic AI architectures demand continuous high-throughput inference, HBM production will continue to swallow up DRAM fab allocation.
Samsung confirmed it is aggressively expanding sales of high-performance memory products, including HBM4, HBM4E, DDR5, LPDDR5X/SOCAMM2, and enterprise SSDs. Next quarter, Samsung expects bit growth to increase by mid-single digits for DRAM and high-single digits for NAND flash.
The structural divergence is sharp: flash storage capacity will expand, but system memory and high-speed cache will remain bottlenecked by physical wafer capacity through 2028.
What Infrastructure Teams Must Plan For
If you manage infrastructure budgets, waiting for memory prices to drop is a losing strategy. The supply constraints through 2028 demand immediate architectural and procurement adjustments.
First, factor memory inflation into long-term procurement models. As detailed in our analysis of why the memory shortage will outlast the AI funding boom, hardware unit costs will stay elevated. Budgeting based on historical price-per-gigabyte declines will break your financial projections.
Second, optimize software footprint to reduce memory overhead. When physical DRAM is scarce and expensive, software efficiency becomes an infrastructure asset. Engineering teams must invest in footprint optimization, efficient caching strategies, and tighter memory reclamation routines rather than simply throwing more RAM at slow services.
Third, evaluate hardware commitments carefully. Large enterprises may need to consider long-term vendor agreements to guarantee supply, but doing so locks in today's peak prices. Smaller organizations without multi-year leverage should prepare for longer lead times and higher spot market premiums.
According to Samsung's detailed earnings coverage on The Register, the memory shortage will worsen in 2027 before seeing any potential relief in 2028. The 19-fold profit surge at Samsung isn't an anomaly—it is the direct outcome of a fundamental supply deficit that will shape enterprise IT spending for the rest of the decade.
