Collusion? A Shortage? The Real Story Behind the Memory Price Surge

Memory prices have exploded. The contract price for commodity DRAM—the fixed quarterly rate that suppliers and bulk buyers set—jumped 93–98% quarter-over-quarter in the first quarter of 2026 alone, and the NAND contract price rose 55–60%. And essentially three companies set that price. Samsung, SK hynix, and Micron held 89.7% of DRAM revenue in the first quarter of 2026—an oligopoly, a market a handful of firms divide and dominate. What is more, these three carry a conviction for collusion—the illegal coordination in which competitors secretly rig prices and volumes. In the U.S. Department of Justice's international DRAM price-fixing case in 2004, criminal fines alone topped $730 million. On the circumstances alone, the answer looks settled. Collusion. But this time it is not collusion. And it is not a pure, AI-made shortage either.

The Circumstances Point to Collusion

Lay out the numbers alone and the collusion case is solid. Start with market structure. DRAM is an oligopoly in which three companies split 89.7% of revenue, and in NAND the top five hold 90.9%. Decide to raise prices, and there are only a few rivals to watch.

Then the record. In the U.S. DOJ case of 2004–2006, Samsung pleaded to $300 million in criminal fines, Hynix to $185 million, Infineon to $160 million, and Elpida to $84 million, and executives served prison time. Only Micron paid no fine, under leniency—the program that exempts or reduces the penalty for the first cartel member to come forward. In 2018, China's antitrust regulator (SAMR) raided the three companies' China offices and said it had secured "massive evidence." The investigation targeted the roughly 47% jump in DRAM prices in 2017.

Then the vertical climb in price sits on top.

Source: TrendForce Press Center, quarterly fixed contract price (commodity/conventional DRAM = general DRAM excluding HBM for AI). Figures are quarter-over-quarter (QoQ) for each quarter, and the rise had already begun in Q4 2025. Q2 2026 is a forecast. Published 2026-06-01. NAND is forecast at +55–60% in Q1 2026 and +70–75% in Q2 2026.

A few dominant players, a record, and a price that nearly doubled. The circumstances of collusion look to be in place.

But This Time It Is Not Collusion

Circumstances are only circumstances. For collusion to hold, the three companies would have to actually rig the price—and in this run of prices, neither the incentive nor the need is visible.

Take the incentive first. Illegal collusion carries an expensive tag: criminal prosecution and fines. If three companies that lived through the 2004 fines and prison terms and the 2018 Chinese investigation can earn a comparable margin legally, there is no reason to wager prison for it.

Then the need. In a market a few firms dominate, prices converge without any secret rigging. This is the core of the theory of oligopoly George Stigler set out in 1964. The real constraint that binds a cartel is not agreeing on price but enforcing and detecting it. A cartel collapses when its members cannot police each other's secret cheating—and the key condition Stigler identified for detection is whether buyers are few and concentrated. In a market of few buyers where large orders are rare, it quickly shows who secretly cut a price. Memory is exactly that kind of market. There are only three sellers, and the buyers are a handful of large customers too—Dell, HP, Apple, the hyperscalers—so cheating is hard to hide. The structure itself breeds discipline. U.S. antitrust law recognizes this. Pricing in lockstep out of a conscious awareness of interdependence, without any communication—"conscious parallelism" (tacit parallel pricing)—is not in itself unlawful. Parallel pricing alone does not establish a violation of Section 1 of the Sherman Act (U.S. antitrust law); a plaintiff must separately prove an "agreement."

Here one counterexample deserves to be met head-on: then what was the collusion of 1998–2002? The answer is that the market structure differed then from now. At the time the DRAM market held many sellers beyond the three—Infineon, Elpida, and others. With many players, matching each other's prices by sight alone is hard, and so explicit collusion was needed. Then the smaller players fell. Qimonda went bankrupt in 2009, and Elpida, after court receivership in 2012, was acquired by Micron. Today's three-firm oligopoly is a structure that hardened only later, through that shakeout. With just three left, each is now in plain view of the others, and tacit discipline alone yields a sufficient margin—no secret agreement required.

A more uncomfortable counterexample is recent. Regulators called a nearly identical surge in 2017—a phase in which supply throttling drove DRAM prices up about 47%—collusion. So why is 2026 different? In fact, even 2017 was never proven to be illegal collusion. China's SAMR investigation claimed "massive evidence" but went dormant with no public fine or conclusion, and the U.S. class action alleging the same supply-restriction conspiracy (Jones v. Micron) was dismissed at trial for lack of standing (the right to bring suit) and, on appeal, for insufficient evidence of an agreement. A dismissal does not prove the absence of collusion. But the fact that even in a phase regulators called "collusion" the agreement was never established to a courtroom standard shows that the difference between 2017 and 2026 is less "was it collusion or not" than "who raised the suspicion." In both phases the mechanism is oligopoly parallel pricing, and neither reached legal proof of illegal explicit collusion.

Here the meaning of the record flips. Conventional wisdom reads it as "a collusion record, so collusion again," but the causation runs backward. When structure alone already settles discipline, the 2004 conviction and the 2018 investigation raise the criminal and fine risk of illegal collusion and, if anything, lower the incentive to reoffend. There is no reason to choose the illegal route again when a sufficient margin is available legally. So the record is not evidence of "collusion again" but a circumstance consistent with today's no-collusion equilibrium. This is not, however, the causal claim that "the 2004 punishment trained the three companies and created the discipline." The sample is a single case, and discipline is settled by structure before any punishment. The record is not a cause but a circumstance that fits the picture.

The Real Story: A Demand Shock Passing Through a Structural Oligopoly

So why did prices explode? The real story is neither collusion nor a pure shortage. It is a shortage amplified and prolonged as AI and HBM demand—HBM being high-bandwidth memory, the fast stacked DRAM bonded to AI accelerators—passes through the three companies' structural oligopoly. Let us separate the strands.

Demand pulled the size of the surge. HBM bit demand grew 130% in 2025 and 70% in 2026, and the 2026 capital expenditure (capex—the upfront outlay on equipment bought all at once) of the four big hyperscalers—Amazon, Google, Microsoft, Meta (the largest cloud operators)—jumps about 77% year-over-year to roughly $725 billion. A base effect sits on top of this. A year earlier, in the first quarter of 2025, the commodity DRAM contract price was in fact a trough, down 8–13%. Measured against that low floor, the year-on-year rise reads larger. Yet the rise was not concentrated in a single quarter. The commodity contract price had already jumped 45–50% quarter-over-quarter in the fourth quarter of 2025, and the first quarter of 2026's 93–98% followed. It did not run straight from trough to peak; it climbed across two quarters from the second half of 2025.

The problem is supply. Demand spikes like this, yet the three companies' supply follows slowly. Three reasons overlap. First, capacity (capa—productive capacity) cannot expand much in the short run. Global DRAM capacity grows only about 10–15% a year, owing to the lead times on fabs (plants) and equipment. As we will see, this ceiling is not purely a physical constraint; it is a value that past investment decisions partly set. Second, HBM eats into wafers (the thin discs on which chips are printed). HBM consumes three to four times the wafers of commodity DRAM to make the same capacity. So the top three's share of wafer input going to HBM rises from 18% of all DRAM wafers (end of 2025) to 22% (end of 2026) and 30% (end of 2027), while the HBM bits those wafers yield come to just 8% → 9% → 13% of the total. It is a structure that eats a lot and yields little, so the more wafers are turned to HBM, the smaller the commodity DRAM share. Third, volume is locked up by contract. HBM supply is effectively sold out through 2026 and pre-allocated—committed in advance by contract—to key customers such as Nvidia. It means the leftover wafers do not reach the open market.

This far, it is a DRAM story. NAND has no HBM cannibalization; instead it tightened as 2025's production cuts unwound late and demand for high-capacity SSDs for AI servers (eSSD) piled on. The big picture—supply following demand late—is the same, but the component where the bottleneck forms is different. The supply mechanism this piece drills into is the DRAM side.

Split this surge into three forces and the picture sharpens. What pulled the size was demand; what bound the supply was structural rigidity; and what slowed the response was the learned discipline we will take up separately in the next section.

Summary: analysis in the text (numerical basis in the sources below). Discipline is not the cause that created the shortage but a modifier that slows the speed of response.

Not an "Engineered Shortage" but a "Slow Response"

Here we step one further. The picture that "the three companies deliberately engineered a shortage to raise prices" is tempting, but it snags on the data. The falsification test that separates it is simple. If the three were throttling supply to prop up prices, they would have to cut wafer input now, when margins are at an all-time high. But the data runs the other way.

Contrast the tenses and it becomes clear. We have already seen what the three do when they genuinely take losses. In the deep-loss phase of 2022–23 (Samsung's semiconductor division about −15 trillion won for the year; SK hynix an operating loss of −7.7303 trillion won and a net loss of −9.1375 trillion won), the three actually cut supply. Samsung reduced output for the first time in 27 years, SK hynix cut that year's capital spending by more than half, and Micron cut wafer input by about 20%. We know they tighten when they lose money.

Yet now, with margins at an all-time high, they are not tightening. SK hynix's commodity DRAM operating margin is projected to exceed 70%—the highest since 1995 (the prior super-boom). Even at such margins, the three are expanding capacity, far from cutting it. They cut in the downturn and expand in the upturn. So the active supply restraint of locking wafers despite the margin, to prop up prices at this very moment, does not show in the data.

But let us honestly note the limits of this test too. If oligopoly discipline really does prop up prices, it works not by locking production this quarter but by cutting investment ahead of time in the downturn to thin the supply baseline itself years later. The output cuts and capital-spending reductions of 2022–23 were exactly that underinvestment, and the capacity thinned that way burst into today's tightness two years on. If so, the 10–15%-a-year capacity ceiling is not a purely physical constraint either. It is a value that past investment decisions partly set. The trouble is that whether this downturn underinvestment is a supporting player that amplified today's tightness or the lead that drove it cannot be separated from upturn data alone. That is the real thing to watch in the three companies' capex guidance.

So why is commodity DRAM still tight? Not because of throttling now, but because HBM devours the added wafers. That is the cannibalization seen earlier. And even though HBM's margin premium inverted below commodity DRAM (under 1x) heading into 2026, HBM expansion does not stop—not because the margin is good, but because the volume is already locked by contract. In sum: the active restraint of the current quarter is not visible, and the trigger of the shortage is demand. Discipline remains a modifier that slows the response by working to lower the supply baseline through downturn underinvestment—but the data gives no ground to declare it the sole cause that created the shortage.

The Word "Collusion" Has Three Layers

If the story so far sounds blurry, it is because the single word "collusion" actually lumps together three legally distinct things. Separating them is what brings the identity of this price into focus.

Summary: Stigler (1964, theory of oligopoly) · U.S. Supreme Court precedent (Theatre Enterprises 1954, Brooke Group 1993, Twombly 2007) · U.S. DOJ 2004–06 case.

This price is ①, and partly ②. It is not ③. There is no direct evidence pointing to recent explicit collusion, and the 2018 investigation, too, went dormant with no public fine or conclusion. The precise phrasing is "looks like collusion, but the mechanism is legal structure and discipline." And this distinction changes the answer to the next question.

So What Ends This Price?

Not, at least, by aiming at the mechanism now on view. What antitrust regulation can touch is illegal collusion (③), but this price's mechanism is legal parallel pricing and structural discipline (①②), which are not in themselves unlawful, so there is nothing to aim at. Of course, if a new investigation actually draws out evidence amounting to ③, the story changes. Nor does the fact that the 2018 investigation went dormant without a conclusion establish innocence. But on the picture now public, regulation is not the lever that ends this price regime. The market fears the wrong target.

What ends this price is whichever of two comes first.

(A) An entrant outside the discipline. China's CXMT has entered DDR5 (the latest-standard commodity DRAM), with its bit-shipment share climbing from 9% in 2025 to 12% in 2027, and YMTC is aiming for 15% of the global NAND market by the end of 2026. But it is slow. And it starts from the low end. On the advanced side, HBM, EUV (extreme-ultraviolet lithography, the key lithography equipment that etches fine circuits) has been blocked from China since 2019, leaving China about three years behind the leaders—so shaking the discipline of the commodity market will take more time.

(B) A reversal in AI demand. If the hyperscaler capex that drove the surge ($725 billion, +77%) bends, even commodity cools all at once. This side is fast and across the board. And since demand led this surge, logically (B) can come sooner, and larger, than (A).

What to watch. The three companies' capex guidance and any lifting of output cuts (the signal that discipline is loosening), CXMT's commodity bit-shipment share, renewals of pre-allocated HBM volume, and any mention of slowing AI demand.

Let me also nail down the condition on which I would myself find my thesis wrong. If CXMT passes 15% commodity bit shipment and the three expand capex again, yet the commodity contract price holds its peak through 2028, then this price was not "the discipline of a structural oligopoly" but simply a vast AI demand cycle. The supply-side term of my thesis would be falsified. Conversely, if the commodity price collapses on an AI capex reversal alone, before China meaningfully enters, then demand was everything. Either way, do not extrapolate to "it is collusion, so regulation breaks it."

Finally, who foots the bill for this price? Right now the downstream does. Memory has soared to 30–40% of a smartphone's cost and up to 35% of a PC's, Dell raised finished-product prices 15–20%, and AWS raised cloud-compute prices for the first time in its history, passing it to consumers. But once discipline loosens and demand bends, that burden returns—as in 2022–23—to the three companies and to investors who rode in at the peak. Discipline does not erase this burden; it only delays it and smooths it out.

Where this surge sits in the cycle I took up separately in the semiconductor supercycle; how the price is reflected in the KOSPI, in Samsung, Hynix, and the KOSPI; and whether the AI-chip price-target forecasts held up, in the AI-chip price-target post-mortem.

The question of hunting the culprit who raised the price was wrong from the start. What must be asked is not "who raised it" but "what ends it." If the reality of this collusion-like price is a demand shock and legal structure, then what we should watch is not the regulators but China's fabs and the hyperscalers' spending plans.