Structural coordination and supply restraint in the global memory oligopoly, 2022–2026
Between the third quarter of 2024 and the first quarter of 2026, conventional DRAM contract prices rose by an amount that plaintiffs in a federal class action put at roughly 700 percent. Three firms (Samsung, SK hynix and Micron) controlled about 90 percent of global DRAM revenue throughout. This paper argues that the price level is better explained as a coordinated outcome than as a market-clearing one, and it argues this without claiming to have found evidence of an agreement.
The argument rests on six plus factors: supply restraint sustained against an extraordinary price signal; a simultaneous and complete pivot to high-bandwidth memory that vacated the conventional market; parallel withdrawal from DDR3 and DDR4; operating margins converging near 76 percent across firms with different cost structures; parallel non-price conduct in allocation and customer vetting; and (the paper's original contribution) a quantified demonstration that the three makers' upstream supply chains overlap so heavily that each firm's capacity intentions are legible to its rivals years before output moves. Using the MEMORYCHAIN dataset, we find that 20 of the 40 upstream vendors supplying the big three supply all three, and that between 96 and 100 percent of each maker's mapped inbound relationships terminate at a supplier shared with at least one rival.
The paper states the counterarguments in their strongest form (the AI demand shock is real, capex is rising roughly 340 percent, physical constraints are genuine, and entrants are gaining share), and concedes that the capex evidence is in real tension with the thesis. It closes on the sharper claim: that under the prevailing legal standard for conscious parallelism, a market this concentrated can produce cartel outcomes whether or not anyone ever agreed to anything, and that this is the finding which should concern regulators.
On 25 June 2026, seventeen plaintiffs, among them two small PC businesses, filed a class action in the Northern District of California alleging that Samsung, SK hynix and Micron had violated §1 of the Sherman Act by restraining the supply of conventional dynamic random-access memory. The complaint puts the resulting price increase at roughly 700 percent over four years.1 Jefferies, forecasting forward from the same conditions, projects a further 40 to 50 percent in the third quarter of 2026 and 30 to 40 percent in the fourth, with no meaningful relief before 2028.1
Numbers of that size are not, by themselves, evidence of anything. Commodity markets move violently when demand outruns capacity, and the memory industry has always been the most cyclical business in semiconductors, a market where a two-year price collapse is as ordinary as a two-year spike. The interesting question is never whether the price moved. It is whether the supply response looks like what a competitive market would have produced.
This paper argues that it does not. The thesis under defense is this: the three firms that control roughly 90 percent of global DRAM revenue have, since 2022, restrained conventional memory supply in a manner too parallel, too durable, and too contrary to individual short-run incentive to be explained by independent competition alone, and the resulting price level is best understood as a coordinated outcome rather than a market-clearing one.
Two things about that sentence deserve stating plainly at the outset, because the paper's credibility depends on them and because §6 should not read as a retreat from a claim never made.
First, on register. This is an argument to the best explanation, not a proof. No direct evidence of an agreement between the three makers is presented here, because none is publicly available. Everything sourced to the June 2026 complaint is an allegation and is labelled as such throughout. What the paper claims to establish is that the coordinated-outcome explanation accounts for more of the observed facts, with fewer unexplained residuals, than the competitive-market explanation does, and that the burden of producing an alternative account has shifted to the sellers.
Second, on what would falsify it. The thesis is not unfalsifiable hand-waving about oligopoly. It makes a specific claim: that supply restraint persisted past the point at which unilateral defection would have been individually profitable. Evidence that any one of the three materially accelerated conventional wafer starts during 2025 or 2026, meaning conventional DRAM bit output rather than capex announcements or HBM capacity, would damage the argument badly. Evidence that the three makers' capacity decisions were made in genuine ignorance of each other's plans would damage §5.6 specifically. The paper says where to look for both.
The structure follows the shape of an antitrust plus-factor analysis, because that framework happens to be the right analytic tool whether or not litigation is the point. §2 establishes that DRAM has every structural precondition economists associate with sustainable tacit collusion. §3 establishes the base rate: this industry ran a criminally prosecuted price-fixing conspiracy for four years, and the recidivism record is a legitimate prior rather than an ad hominem. §4 sets out the price record and identifies its structurally odd feature: near-uniformity across product categories with unrelated end markets. §5 assembles the six plus factors that constitute the affirmative case. §6 states the counterarguments in their strongest form and concedes where they land. §7 confronts the legal standard the argument is most vulnerable to, and proposes that the sharper conclusion is not that these firms broke the law but that the law as applied cannot reach what they did.
Coordination is not equally available in all markets. The economics of tacit collusion, from Stigler's foundational treatment onward, identifies a recognisable set of preconditions: few sellers, homogeneous product, transparent prices, high barriers to entry, repeated interaction over a long horizon, and some mechanism by which deviation can be detected and punished before it pays.10 DRAM satisfies every one of them, and satisfies several to an unusual degree.
Fig. 1 · Global DRAM revenue share, Q2 2026
Three firms, 90 points of share. HHI 2,876, well above the 1,800 threshold at which the 2023 Merger Guidelines treat a market as highly concentrated.
Concentration. Samsung holds 39 percent of DRAM revenue, SK hynix 26 percent and Micron 25 percent. The Herfindahl–Hirschman Index on those shares is 2,876. The 2023 Merger Guidelines issued by the Department of Justice and the Federal Trade Commission treat a market above 1,800 as highly concentrated;11 DRAM sits at roughly sixty percent above that line, and the three-firm concentration ratio of 90 places it among the most concentrated globally significant markets in existence.
Homogeneity. A DDR5 die of a given density and speed grade is functionally interchangeable across vendors. Module makers qualify parts from multiple suppliers precisely so they can substitute. This is normally described as a competitive feature, and in a market with many sellers it is. In a market with three, it removes the principal obstacle to coordination: there is no product differentiation to argue about, so a shared understanding requires agreement on one variable only.
Price transparency. Contract pricing is negotiated quarterly and reported publicly by TrendForce and its peers within days. Spot prices are published continuously. A deviation from a shared pricing posture is therefore visible to rivals in weeks, which is the condition Stigler identified as decisive, since a cartel's real problem is not agreeing but detecting cheating.
Entry lag. Entry into leading-edge DRAM is physical, capital-intensive and slow in a way that has no close analogue outside semiconductors. SK hynix's approved Yongin Y2 fab does not reach production until mid-2029; its M17 NAND plant is dated late 2028.5 Even a firm that decided today to enter would face a multi-year lag before a single wafer started, and would face it while holding no EUV allocation, no wafer long-term agreement and no qualified process. The relevant horizon over which coordination must be sustained is therefore not quarters but years, and the punishment horizon for a defector is correspondingly long.
Coordination does not require communication when the input market publishes your rivals' plans for you.
The information channel. The precondition least discussed in the literature, and the one this paper treats as its central structural finding, is mutual observability of capacity intentions rather than prices. Memory fabrication depends on a small set of upstream chokepoints: ASML at essentially 100 percent of EUV lithography; Shin-Etsu and SUMCO together above 50 percent of the 300mm wafer market; a Japanese cluster including JSR, Tokyo Ohka Kogyo and Shin-Etsu supplying more than 70 percent of EUV resists; Tokyo Electron at roughly 90 percent of coater/developer tools.7 All three makers buy from the same vendors, under multi-year long-term agreements, with tool orders placed eighteen months or more before the corresponding output appears.
The implication is that each maker's capacity plans are inferable from its purchasing behaviour long before they show up as bits. This is developed quantitatively in §5.6, where the MEMORYCHAIN dataset is used to measure exactly how much of each maker's upstream footprint is shared with its rivals. The answer, nearly all of it, is what makes the tacit-coordination account structurally plausible rather than merely available.
None of this establishes that coordination occurred. It establishes that coordination here would require no smoke-filled room, only mutual understanding among three parties who interact repeatedly, observe each other continuously, and know that no fourth party can punish them for four years at minimum. That is the correct baseline against which the conduct in §5 should be read.
There is a standard objection to the reasoning in §2, and it is a fair one: structural preconditions are not conduct, and every concentrated industry can be described in language that makes collusion sound easy. The reply is that in this particular industry, collusion is not a theoretical possibility inferred from structure. It is a documented historical fact, criminally prosecuted, with guilty pleas and prison sentences.
Fig. 2 · Criminal fines, U.S. DOJ DRAM cartel prosecution (1998–2002 conduct)
More than a dozen executives served U.S. prison time. Micron participated but avoided prosecution by reporting first under the corporate leniency program.
Between 1998 and 2002, DRAM manufacturers coordinated pricing to Dell, Hewlett-Packard, Compaq, IBM, Gateway and Apple. The Department of Justice obtained guilty pleas from Samsung ($300 million, 2005), Hynix ($185 million), Infineon ($160 million) and Elpida ($84 million), along with prison terms for more than a dozen individual executives.3 Micron participated in the conspiracy and was not prosecuted, having reported it first and secured amnesty under the corporate leniency program.
Two of the three defendants in the 2026 action are therefore adjudicated participants in a DRAM price-fixing conspiracy, and the third is an admitted participant who escaped liability by informing. The European Commission reached a cartel settlement covering related conduct in 2010 figure and date unverified, see App. B. A Chinese antitrust investigation into all three was opened in 2018, with officials quoted describing "massive evidence"; the paper does not assert an outcome, because the investigation's disposition could not be established from a reliable source.6
The evidentiary status of this history needs care. It is not proof of present conduct, and a paper that leaned on it as though it were would deserve dismissal. What it establishes is a base rate. In Bayesian terms, the prior probability that three firms in a given concentrated market are coordinating is low; the prior for three firms who previously coordinated in this exact market, with this exact product, against these exact customers, is materially higher. Courts are appropriately reluctant to reason this way, and for good institutional reasons. Analysts are not so constrained, and should not pretend to be.
The more instructive precedent is the failed one. In 2018, Hagens Berman filed a class action alleging collusion during the 2016–17 upcycle, when DRAM prices roughly doubled and all three makers throttled supply growth in parallel. The complaint was dismissed in 2020; the Ninth Circuit affirmed in 2022, holding the alleged conduct "more likely explained by lawful, unchoreographed free-market behavior."3 The case never reached discovery.
That outcome is the single most important obstacle this paper faces, and §6 and §7 return to it at length. The short statement of the problem is that parallel conduct in an oligopoly is not merely insufficient to prove conspiracy; it is affirmatively expected under competitive conditions, because rational firms in a transparent three-seller market will independently reach similar conclusions about output. The 2018 plaintiffs had parallel supply restraint and parallel price increases and lost, because that is exactly the pattern lawful interdependence generates.
The question the present paper must answer, then, is not whether 2026 looks like 2018. It is what is new: what facts exist now that did not exist then, and whether those facts do analytic work the earlier ones could not. The candidates are the magnitude (roughly 700 percent versus roughly 100 percent), the duration (a restraint sustained across four years rather than six quarters), the margin level (76 percent operating margin, sustained), the HBM allocation mechanism (which had no 2016 analogue), and the supply-chain transparency measured in §5.6. §6, Objection 5, takes this on directly.
The first quarter of 2026 set records in every memory product category simultaneously. TrendForce, having initially forecast a 55 to 60 percent quarterly increase for conventional DRAM, revised the figure upward to 90 to 95 percent, an unusual admission that the market had outrun the industry's own most informed observers.2 PC DRAM exceeded 100 percent quarter-on-quarter. Server DRAM and LPDDR4X/5X each posted the steepest increases in their history. NAND flash, forecast at 33 to 38 percent, came in at 55 to 60. Enterprise SSD set its own record at 53 to 58 percent.
Fig. 3 · Quarter-over-quarter contract price increases, 1Q 2026
Every category set a record. Note the uniformity across products with unrelated end markets.
The magnitude is arresting, but magnitude is the least analytically useful property of these numbers. A demand shock of sufficient size produces arbitrary magnitudes. What a demand shock does not readily produce is the shape visible in Fig. 3.
The uniformity is the anomaly. The six categories plotted serve substantially different end markets with substantially different demand drivers. Server DRAM demand is driven by hyperscaler capex cycles. PC DRAM demand is driven by consumer and enterprise refresh, which in this period was weak. LPDDR demand is driven by smartphone build plans. NAND demand is driven by storage attach rates and enterprise SSD procurement. Enterprise SSD is a distinct procurement channel again. These series have historically diverged, and that divergence is the ordinary texture of the memory market, and it is why the categories are tracked separately in the first place.
In 1Q26 they converged. Every one of them moved to a record, in the same quarter, within a band running roughly from 55 to 100 percent. Demand shocks are category-specific; they propagate unevenly, because the end markets are unrelated. Supply decisions are not category-specific, because a wafer start allocated to HBM is a wafer start unavailable to every downstream category at once. Uniform movement across unrelated demand curves is the signature of a shared supply constraint, not of shared demand.
This is the paper's most economical argument, and it does not depend on any allegation. It requires only the published price data and the observation that the affected end markets have nothing in common except their suppliers.
What it does not establish is why the supply constraint exists. A constraint can be chosen, or it can be imposed by physics. §5 argues for the former; §6, Objection 3, gives the latter its strongest form.
Fig. 4 · DRAM contract price index, Q1 2022 – Q3 2026 Pending data
The spine of the whole argument: the ~700% figure needs a quarterly series behind it, not a single headline number.
The figure above is empty, and the paper is weaker for it. The 700 percent number is currently carried by a single headline figure sourced to press coverage of the complaint rather than by a quarterly series the reader can inspect. Until that series is assembled, §5.1's claim about restraint sustained against a price signal rests on an aggregate the paper has not independently verified. This is stated here rather than buried, and is listed first in Appendix B.
Parallel pricing among oligopolists is lawful and expected. The analytic device for distinguishing lawful interdependence from coordination is the plus factor: conduct that is difficult to explain except by reference to a shared understanding, because it runs against what an individual firm maximising its own profit would do. This section assembles six. Each concludes by stating why unilateral profit-maximisation does not account for it, which is the only thing that makes a plus factor a plus factor rather than a description.
The complaint alleges that all three makers implemented production reductions beginning in 2022 and continued to restrain conventional DRAM supply as prices rose through 2024 and 2025.8 The 2022 reductions are not in dispute and require no allegation: the memory market was in a severe downcycle, prices were below cash cost for some products, and cutting output was the obvious individually rational response. Nothing about 2022 needs explaining.
What needs explaining is 2025 and 2026. By mid-2025 the price signal had inverted completely, and by 1Q26 conventional DRAM was rising 90 to 95 percent in a single quarter. In a competitive market, a price signal of that magnitude triggers a capacity race, not because firms are aggressive, but because the first mover captures the rent before it dissipates and the last mover captures none of it. The race did not occur. New conventional capacity is dated 2028 to 2029 across the board.5
Why unilateral maximisation does not account for it. Consider the position of any one maker in mid-2025, holding roughly a quarter of the market and facing prices rising at an annualised rate in the hundreds of percent. Adding conventional wafer starts would have been extraordinarily profitable at the margin: the added output sells into a market where every bit clears instantly, and the incremental cost of running existing tools harder is low relative to the realised price. The standard reason a firm declines such an opportunity is fear that the added supply will collapse the price, but that fear only binds if the firm expects its rivals to match rather than absorb. In an ordinary three-firm market with a four-year entry lag, the individually optimal play is to defect early, take share while the rent is enormous, and accept a lower price later on a larger base.
None of the three did this. That is the first plus factor: a restraint that outlived the point at which unilateral defection would have been individually profitable, sustained by three independent boards across roughly six quarters of escalating price signal.
This is the strongest and most novel of the six, and it requires the most careful statement, because the obvious misreading of it is a claim the paper does not make.
The claim is not that HBM demand is fabricated. High-bandwidth memory demand is real, enormous and well documented; data centres are projected to consume more than 70 percent of all high-end memory produced in 2026.5 Any maker that failed to pivot toward HBM would be destroying shareholder value. The pivot, considered as a business decision, is correct.
The claim concerns simultaneity and completeness. All three makers are in HBM4 volume ramp for the same customer generation on overlapping 2025–26 timelines.7 All three reallocated capacity away from conventional DRAM on comparable schedules. In 1Q26 the makers additionally reallocated NAND lines toward DRAM "despite robust NAND demand," constraining NAND supply as a second-order consequence.2
The pivot removes conventional supply twice over. HBM is a stacked die with substantially higher wafer-area consumption per usable bit than DDR5, a penalty commonly cited at roughly three times, though the paper flags this as a load-bearing number it has not independently sourced unverified, see App. B. A wafer start moved from DDR5 to HBM therefore withdraws not one unit of conventional supply but something closer to three.
Why unilateral maximisation does not account for it. It substantially does, and this is the honest place to concede the point. Each maker individually had strong reasons to pivot, and would have pivoted alone. What unilateral maximisation does not account for is the absence of any differentiated position. In a market where all three rivals are vacating the conventional segment, the individually optimal strategy for at least one of them is to pivot less: to hold conventional capacity and harvest an abandoned market with three-firm demand and one-firm supply. The counterfactual is not speculative; it was run, by Nanya, and §5.3 gives the result.
The plus factor is therefore not the pivot but the unanimity of its degree. Three firms with different HBM competitive positions, different customer relationships and different cost structures arrived at nearly the same allocation split at nearly the same time, and none took the available contrarian position.
The complaint alleges phased withdrawal from DDR3 and DDR4 by all three defendants on overlapping timelines.8 These are the specific products where substitution is hardest and buyers most captive: embedded systems, industrial controllers, networking equipment and legacy server platforms cannot re-qualify to DDR5 without redesigning the board. The customers most exposed to legacy-node withdrawal are precisely those least able to respond to it.
The counterfactual exists and is dramatic. Nanya Technology, a roughly 2 percent pure-play that remained in legacy DRAM, posted +690 percent year-on-year revenue growth in Q2 2026, riding the gap the big three left behind.7 The abandoned market was not marginal. It was extraordinarily profitable, and a firm with two percent of the industry's capacity was able to demonstrate that fact publicly for four consecutive quarters while three firms holding ninety percent of capacity declined to act on it.
Why unilateral maximisation does not account for it. Nanya's results are a live, observable, quarterly-reported proof that legacy capacity retained would have been highly profitable. Each of the big three could read those results. Each declined to slow its own legacy exit in response. The behaviour is consistent with a shared understanding that the legacy segment is being vacated jointly; it is difficult to reconcile with three firms independently maximising, at least one of which should have noticed a 690 percent signal.
SK hynix reported a 72 percent operating margin in Q1 2026 and 76 percent in Q2 2026, its fifth consecutive record quarter.9 For a capital-intensive manufacturer of a commodity product, a 76 percent operating margin is not a scarcity rent in the ordinary sense. It is roughly what a firm earns when output is held materially below the competitive level and the residual demand curve is inelastic.
The duration matters more than the level. A single quarter at 76 percent is compatible with a demand shock outrunning short-run capacity; that is what a spike looks like, and it should decay as supply responds. Five consecutive record quarters is a different object. It indicates that the supply response which should have compressed the margin did not occur, which is §5.1 restated in accounting terms, and is why the two factors are not independent evidence but two views of the same fact.
This factor is materially weaker than it should be, because only SK hynix's margin series is in hand. The analytically decisive exhibit would be operating margin for Samsung's DS division and Micron plotted against SK hynix's, 2022–2026. Convergent margins across three firms with different cost structures, different node mixes and different customer books is itself a plus factor, arguably a stronger one than any single firm's level, since independent cost structures should produce divergent margins under competition. Until that figure is built, §5.4 rests on one firm and should be weighted accordingly.
Price is the most visible lever and the least useful one for coordination, because it is the lever regulators watch. Allocation regimes are where coordination is easiest to sustain and hardest for buyers to observe: who gets supply, in what quantity, on what terms, and who is refused.
The complaint alleges "similar customer-vetting practices adopted by all three manufacturers."8 Separately, TrendForce reports that cloud service providers are securing supply through long-term agreements rather than open-market purchase, a structure which, whatever its merits for the buyers, has the effect of letting each seller observe the shape of its rivals' committed book well in advance.4
Why unilateral maximisation does not account for it. Convergent pricing among three firms facing a common cost shock is unremarkable. Convergent vetting criteria are harder to explain, because vetting criteria are not driven by any common external variable; they are internal policy choices about which customers to serve during scarcity, and there is no cost-side reason for three firms to select the same ones. This is also the factor with no 2018 analogue: the earlier case had parallel pricing and parallel output restraint to point at, and nothing on the allocation side.
The evidentiary status here is weak, however, and should not be overstated. This factor currently rests entirely on a single unadjudicated allegation with no independent corroboration, and it will remain the paper's softest until buyer-side testimony or discovery materials substantiate it.
The remaining five factors are, in various proportions, restatements of publicly reported conduct. This one is an original quantitative finding, computed from the MEMORYCHAIN dataset, and it supplies the mechanism the other five presuppose.
The tacit-coordination account requires that each maker be able to observe the others' capacity intentions early enough to respond. If capacity plans were genuinely private until output appeared, coordination would be unstable: a defector could add supply and enjoy the rent for the eighteen-plus months before rivals could react, which is long enough to make defection worthwhile. The stability of the observed restraint therefore depends on whether the makers can see each other coming.
They can, and the degree is measurable. Restricting the MEMORYCHAIN graph to inbound supply relationships terminating at Samsung, SK hynix or Micron yields 98 mapped edges from 40 distinct upstream vendors. Of those 40 vendors, 20 supply all three makers, including every one of the ten mapped capital-equipment vendors (ASML, Applied Materials, Lam Research, Tokyo Electron, KLA, ASM International, SCREEN, Advantest, Teradyne, DISCO) and seven of the mapped materials vendors (Shin-Etsu, SUMCO, JSR, Merck Electronics, Entegris, Air Liquide, ADEKA). A further 18 vendors supply exactly two of the three. Only two of the 40 supply a single maker without also supplying a rival.
Fig. 5 · Upstream supplier overlap among the big three
Share of each maker's mapped inbound supply relationships that terminate at a vendor also serving both rivals, and at a vendor serving at least one rival.
The proportions are the finding. Between 96 and 100 percent of each maker's mapped inbound relationships terminate at a supplier shared with at least one rival: 37 of 37 for Samsung, 34 of 35 for SK hynix, 25 of 26 for Micron. Restricting to vendors shared with both rivals still gives 54, 57 and 77 percent respectively. There is essentially no private upstream footprint anywhere in the big three's supply base.
What this means operationally: an EUV tool order, a wafer LTA renegotiation, a resist volume commitment or a coater installation is placed with a vendor that also serves both rivals, typically twelve to twenty-four months ahead of the output it enables. Equipment vendors publish backlog and regional revenue mix; wafer suppliers disclose contracted volumes; installation schedules are visible to anyone tracking cleanroom construction. A maker planning a genuine conventional-capacity expansion cannot execute it without transacting, repeatedly and at scale, inside a vendor network its rivals are watching.
The input market functions as a disclosure regime that no one designed and no one can opt out of.
The contrast with the challengers sharpens the point. CXMT has three mapped inbound relationships, only one of which (Shin-Etsu) is shared with the big three; its remaining suppliers are NAURA and AMEC, domestic Chinese equipment vendors outside the Western tool network. Nanya has six, four of them shared. CXMT is structurally outside the information channel, which is a large part of why it could grow 716 percent without being pre-empted, and is the reason §6's Objection 4 requires a careful answer rather than a dismissive one.
Why unilateral maximisation does not account for it. Strictly, this factor is not conduct at all; it is a structural property of the industry, and no firm chose it. Its role in the argument is different from the others: it does not itself evidence coordination, it removes the principal objection to the possibility of coordination. The standard reply to a tacit-collusion story is that firms cannot sustain restraint without a detection mechanism, because defection pays before it can be punished. Fig. 5 is the detection mechanism. It shows that in DRAM specifically, the informational precondition for stable tacit coordination is satisfied to an unusual degree, not by agreement but by the shape of the input market.
This is also the paper's most defensible contribution, because it depends on no allegation, no litigation and no contested reporting. It is a property of who buys from whom, and it would be true if the 2026 complaint had never been filed.
| Factor | Evidentiary basis | Strength | What would falsify it |
|---|---|---|---|
| 5.1 Supply restraint | Complaint allegation + dated capacity announcements | Strong | Evidence of materially accelerated conventional wafer starts in 2025–26 |
| 5.2 HBM pivot | Public ramp timelines; complaint framing | Strong | Announcement dates that do not cluster; a maker taking a contrarian allocation |
| 5.3 DDR3/4 exit | Complaint allegation; Nanya counterfactual | Moderate | EOL notices spread widely in time, or driven by common supplier constraints |
| 5.4 Margins | SK hynix reported results only | Moderate | Divergent margin paths across the three once the full series is assembled |
| 5.5 Allocation conduct | Single unadjudicated allegation | Weak | Buyer testimony showing materially different vetting regimes |
| 5.6 Information channel | Computed from MEMORYCHAIN (this paper) | Strong | Procurement records showing substantial private upstream capacity |
Table 1 is included partly as a summary and partly as a discipline. Two of the six factors are currently weak or incomplete, and a reader should be able to see which at a glance rather than reconstructing it from the prose. The affirmative case rests principally on 5.1, 5.2 and 5.6.
Five objections stand against the thesis. Each is stated first in its strongest form, without hedging, because a paper that steel-mans badly is not making an argument. One of them lands substantial damage, and this section says so.
The objection. Data centres are projected to consume more than 70 percent of all high-end memory produced in 2026.5 This is not a marginal shift in a mature demand curve; it is the arrival of an entirely new demand category, at scale, against a capital stock sized for the previous decade. A genuine and unprecedented demand shock striking fixed short-run capacity produces exactly the observed price path, with no agreement whatsoever. Everything in §4 follows from arithmetic.
Response. The shock is conceded entirely, and the paper's argument does not require disputing it. The case does not rest on the price level, which a demand shock explains adequately, but on the supply response, which it does not.
Two residuals survive the demand-shock account. The first is the capacity timeline. A demand shock explains why prices rose in 2025; it does not explain why the capacity response is dated 2028–29 in a market where the rent from arriving in 2027 instead would be measured in tens of billions. The second is Fig. 3's uniformity. AI demand is concentrated in HBM and server DRAM. It does not, by any direct channel, raise PC DRAM by more than 100 percent in a quarter, or LPDDR4X by 90 percent, or enterprise SSD by 55. Those categories moved because supply was withdrawn from them to serve the first two, which is a supply-allocation decision, and the question is whether three firms made it independently.
The objection. This is the strongest of the five. Combined industry capex is projected to rise roughly 340 percent between 2024 and 2027. SK hynix has approved ₩54.3 trillion ($38.3 billion) in spending through 2031. Micron has committed at least $200 billion to U.S. expansion and $24 billion in Singapore.5 Cartels restrict output; they do not commit a quarter of a trillion dollars to expanding it. The observed behaviour is the behaviour of firms racing each other, and the paper's core factual premise of restraint is simply false.
Response, with a concession. The paper's reply is that capex and conventional-DRAM wafer-start capacity are not the same variable, and that the announced spending is directed substantially at HBM, advanced-node conversion and greenfield facilities whose volume production is dated 2028–29. On that reading, the capex boom and the conventional-supply restraint are consistent: the industry is spending enormously on capacity that arrives after four more years of scarcity rent, and arrives for all three at roughly the same time.
That reading is coherent, but the paper cannot currently demonstrate it, and pretending otherwise would be exactly the failure §1 disclaimed. To sustain the response the paper needs a decomposition of announced capex into HBM/advanced-node versus conventional wafer-start capacity, by firm and by year, which it does not have. This objection is presently the strongest evidence against the thesis, and a reader who weighted it heavily would be reasoning correctly. It is listed second in Appendix B.
One observation partially survives the gap. Even taking the capex figures entirely at face value, the timing is uniform: all three firms' major conventional additions arrive in the same 2028–29 window. Independent firms racing for rent should arrive at staggered times, because the value of arriving first is enormous and the cost of arriving third is total. Synchronised arrival is a weaker signal than restraint, but it is not nothing.
The objection. Memory expansion is bounded by things no board can vote to change: cleanroom construction schedules, grid power, water, EUV tool lead times, wafer supply, resist supply, and the availability of qualified process engineers. Every one of those chokepoints is documented in MEMORYCHAIN itself. The makers are not withholding supply; they are queuing behind ASML, Shin-Etsu and their own construction contractors. The 2028–29 dates are not a choice; they are how long it takes.
Response. This objection is largely correct, and the paper's own dataset is its best evidence, which is precisely why it must be answered rather than avoided. Three points reduce but do not eliminate it.
First, physical constraints explain the floor on response time, not the observed behaviour above that floor. The constraint set does not explain why no maker bid aggressively for reallocated tool slots, accepted worse terms for earlier delivery, or converted existing NAND capacity toward conventional DRAM rather than toward HBM, all of which are levers inside the constraint envelope, and all of which were individually available.
Second, the constraints are shared, which is doubled-edged for the objection. Common constraints do produce common timelines without coordination, but they also constitute the observation channel of §5.6. A maker that did try to bid up tool allocation would be visible to its rivals precisely because the constraint is shared, which converts a physical limit into a coordination-supporting mechanism.
Third, the objection does not reach §5.3 at all. Continuing to run existing DDR4 lines requires no new cleanroom, no EUV allocation and no wafer expansion. Nanya demonstrated the profitability of doing so, at 690 percent year-on-year. No physical constraint explains three firms declining that.
The objection. CXMT grew 716 percent year-on-year to roughly 7 percent share; Nanya grew 690 percent, both precisely because the incumbents left a gap.7 Successful entry into an allegedly restrained market is the classic signature of a cartel that is not working, and a cartel that is not working is not producing the price level under discussion.
Response. The growth rates are real and the paper does not contest them, but rate and constraint are different things. CXMT at roughly 7 percent and Nanya at roughly 2 percent are together below a tenth of the market; a firm at 7 percent cannot discipline the pricing of a firm at 39 percent regardless of its growth rate. CXMT is additionally constrained by U.S. export controls to pre-EUV process nodes, which limits it in the server and high-density segments where the price increases were steepest; it competes at the low end of the conventional market, not across it.
The more interesting point is the one from §5.6. CXMT has almost no supplier overlap with the big three: one shared vendor of three mapped, with its equipment sourced domestically from NAURA and AMEC. It is structurally outside the observation network, and it is the one participant that grew explosively into the abandoned segment. That pattern is what the thesis predicts. Firms inside the shared-supplier network behaved in parallel; the firm outside it did not. Objection 4, properly examined, is closer to corroboration than to rebuttal.
The objection. A materially identical theory, on materially similar facts, was held to be "more likely explained by lawful, unchoreographed free-market behavior."3 The 2018 plaintiffs had parallel restraint, parallel price increases, the same three defendants and the same cartel history, and the case did not survive the pleadings. Repeating the argument with larger numbers does not answer it.
Response. Three things are factually new, and each does work the 2018 record could not.
The first is the HBM allocation mechanism, which had no 2016 analogue. In 2016 the parallel conduct was a straightforward output decision, the kind of conduct the Ninth Circuit correctly identified as ordinary oligopoly behaviour. In 2026 there is a specific, dateable reallocation of a shared productive input toward a segment that removes roughly three units of conventional supply per unit added, executed by all three on overlapping schedules. That is a more particular fact than "they all produced less."
The second is duration and margin. The 2016–17 episode ran about six quarters with margins in a range consistent with a cyclical peak. The present episode has run substantially longer with a 76 percent operating margin sustained across five consecutive record quarters. Interdependence theory predicts convergence to the joint-maximising outcome; it does not obviously predict that outcome persisting for years with no defection at all.
The third is §5.6. The 2018 plaintiffs asserted that the defendants could observe each other. This paper measures it, and finds essentially no private upstream footprint anywhere in the big three. Whether that measurement would matter to a court is a separate question; §7 suspects it would not, and treats that as the more interesting finding.
This response is built from press summaries of the Ninth Circuit's 2022 disposition, not from the opinion itself. The paper's treatment of Objection 5 cannot be regarded as sound until the opinion is read directly and the court's actual reasoning on plus factors is engaged on its own terms. See Appendix B.
The paper now faces the distinction it is most vulnerable to, and the honest course is to concede it rather than argue around it.
Section 1 of the Sherman Act reaches contracts, combinations and conspiracies in restraint of trade. It does not reach independent parallel conduct, however anticompetitive its effects. Theatre Enterprises established that conscious parallelism alone does not establish agreement; Matsushita requires that a plaintiff's theory make economic sense before an inference of conspiracy may be drawn from ambiguous conduct; Twombly requires that a complaint plead factual matter suggesting agreement rather than merely conduct consistent with it, and did so in a case where the parallel conduct was, as here, entirely undisputed.12 The doctrine is coherent and its rationale is sound: in a three-seller market, firms that independently reason about each other's likely responses will converge on similar output decisions, and punishing that convergence would punish rational behaviour rather than agreement.
Measured against that standard, the six factors of §5 are not obviously sufficient. Factors 5.1 through 5.4 describe conduct that a sophisticated defendant will characterise as interdependent oligopoly behaviour and, on the current record, will characterise successfully, since the 2022 affirmance did exactly that on a similar record. Factor 5.5 rests on a single unadjudicated allegation. Factor 5.6, the paper's strongest contribution, is in a sense the most legally unhelpful of all: it establishes that the makers could observe each other's plans without communicating, which is nearer to a defence than a plus factor. If capacity intentions are inferable from the structure of the input market, then each firm can reach the coordinated outcome by pure inference, and pure inference is lawful.
The strongest version of this paper's finding is not that these firms broke the law. It is that they did not need to.
This suggests the more defensible thesis, and the paper adopts it. Rather than arguing that the plus factors are sufficient to infer agreement, a claim it cannot substantiate and which a court has already rejected on similar facts, the paper argues that the conduct is economically equivalent to cartel behaviour in its effects on price, output and consumer welfare, and that the gap between the economic and legal characterisations is the finding that matters.
Three observations follow.
The doctrine's premise is that parallelism without agreement is unstable. Conscious parallelism is tolerated in part because economics has traditionally held that tacit coordination decays: without an enforcement mechanism, defection eventually pays, and the coordinated outcome erodes. That premise is empirically testable, and the present episode tests it. Five consecutive record-margin quarters with no observable defection, in a market where defection was individually profitable and where §5.6 shows defection would have been detected in advance, is evidence that the premise does not hold under these structural conditions.
The remedy gap is structural, not evidentiary. If the conduct is lawful, better discovery does not fix it. What the analysis points toward is not a §1 case but the structural instruments: merger policy that never permitted the consolidation to this point, and conduct-independent remedies aimed at concentration itself. Those instruments are largely unavailable retrospectively, which is the practical form the problem takes.
The transparency finding cuts against the standard's assumptions. Doctrine assumes that firms which do not communicate face genuine uncertainty about each other's intentions, and that this uncertainty is what makes coordination fragile and agreement necessary. §5.6 shows that in DRAM the uncertainty has been engineered out, not by the makers but by the concentration of the upstream vendor base. A legal standard built on an assumption of mutual opacity applies awkwardly to a market that is structurally transparent.
What this paper has shown is a market with every structural precondition for sustainable coordination; three participants with an adjudicated history of coordinating in this exact product; a price movement whose uniformity across unrelated end markets points to a shared supply decision rather than a shared demand shock; a supply response that did not occur when it would have been individually profitable; operating margins at 76 percent sustained across five consecutive quarters; an abandoned legacy segment demonstrated to be extraordinarily profitable by the one firm that stayed in it; and an upstream supply base so heavily shared that essentially no participant's capacity intentions are private.
What it has not shown, and does not claim, is an agreement. No evidence of communication between the three makers is presented, because none is publicly available. The 2026 action is unadjudicated, its allegations are contested, and its 2018 predecessor was dismissed on facts that were similar in kind if not in magnitude. A reader who weighs the capex evidence of Objection 2 heavily should discount this paper's thesis substantially, and the paper has not given that reader a reason not to.
The conclusion the analysis actually supports is narrower and, in the end, more troubling than the accusation. A market with three sellers, a homogeneous product, a four-year entry lag, published prices, and an upstream vendor base that renders each participant's plans legible to the others will produce cartel outcomes whether or not anyone ever agrees to anything. The coordination does not need to be arranged; it is emergent, and it is stable because the structure makes it stable. Nothing in the conduct described here requires a meeting.
That is precisely why the legal standard, which asks whether a meeting occurred, is the wrong instrument. It was built for a world in which firms that did not communicate could not reliably coordinate, a world in which the fragility of tacit collusion did the regulatory work that doctrine did not have to. The DRAM market of 2026 is evidence that this assumption has failed for at least one globally significant industry, and there is no obvious reason to think it is the only one. Consumers paid roughly 700 percent more for memory over four years. Whether anyone conspired is, from their position, the least interesting question in the case.
The figures in §5.6 and Fig. 5 are computed directly from the MEMORYCHAIN dataset (app/js/data.js), which maps 126 companies across seven supply-chain stages with 236 directed supply relationships.
Procedure. The edge list was filtered to edges whose target is samsung, skhynix or micron, yielding 37, 35 and 26 inbound edges respectively (98 total). The union of edge sources gives 40 distinct upstream vendors. Each vendor was classified by the number of the three makers it supplies. Two derived ratios are reported per maker: the share of its inbound edges terminating at a vendor supplying all three, and the share terminating at a vendor supplying at least one rival. The same procedure was applied to cxmt and nanya for contrast.
| Maker | Inbound edges | → vendor serving all three | → vendor serving ≥1 rival |
|---|---|---|---|
| Samsung | 37 | 20 (54%) | 37 (100%) |
| SK hynix | 35 | 20 (57%) | 34 (97%) |
| Micron | 26 | 20 (77%) | 25 (96%) |
| CXMT | 3 | 1 (33%) | 1 (33%) |
| Nanya | 6 | 2 (33%) | 4 (67%) |
Vendor classification. Of the 40 distinct vendors: 20 serve all three makers (7 materials, 10 equipment, 2 EDA, 1 substrate/module), 18 serve exactly two, and 2 serve exactly one (BESI, GlobalWafers).
Limitations, stated plainly. Three matter. First, MEMORYCHAIN's own documentation describes its edges as simplified for readability; it is a curated map of publicly known relationships, not an audited procurement record, so the proportions describe the mapped graph rather than actual purchase volumes. Second, edges are unweighted: a vendor supplying 90 percent of a maker's wafers and one supplying a specialty gas count identically, which almost certainly understates the concentration of the observation channel rather than overstating it. Third, absence of an edge is not evidence of absence of a relationship; the challenger figures in particular should be read as reflecting lower mapping density for those firms as well as genuinely narrower supply bases, and the CXMT contrast in §5.6 should be weighted with that in mind.
The following are known deficiencies in the present draft, listed in descending order of how much they matter to the argument. This appendix exists so that a reader can calibrate the paper's claims against what actually stands behind them.
Working draft, August 2026. All allegations of coordination described here are contested and unproven; the 2026 class action has not been adjudicated, and the 2018 predecessor action was dismissed. Statements sourced to the complaint are allegations, not findings. Items marked unverified have not been checked against a primary source and are listed in Appendix B. Fig. 4 contains no data. Not legal or investment advice.