The editorial argues that DRAM's market structure — three suppliers controlling 95% of supply, a homogeneous commodity, and public quarterly capex disclosures — makes coordination almost inevitable. Synchronized capex cuts, executive statements about 'discipline' and 'prioritizing profitability over share,' and a doubling of DDR4/DDR5 contract prices during a period without a genuine commodity DRAM demand shock all fit the pattern of signaled supply restraint rather than independent capital allo
The class action complaint explicitly argues that AI server buildouts pulled HBM capacity, not commodity DDR, so the near-doubling of DDR4/DDR5 prices from late 2023 through mid-2025 cannot be explained by an AI demand shock. Instead, plaintiffs frame Samsung's idled legacy fabs, SK Hynix's 'disciplined' wafer-start commitments, and Micron's profitability-over-share posture as a coordinated supply squeeze.
By surfacing the Seoul Economic Daily story to a 381-point, 177-comment HN front-page slot, donohoe amplifies the framing that the three-supplier price curve looks coordinated rather than demand-driven. The submission's traction signals broad developer-community resonance with the collusion narrative.
The editorial highlights that this is the third major DRAM antitrust action in two decades — following the 2002 DOJ case with $300M+ fines and prison time and the 2018 class action that settled for $445M in 2023. With hyperscalers and OEMs now carrying DRAM settlement reserves as a line item on their books, the interesting question is not whether the suit wins but why this market structurally keeps producing the same case.
A proposed class action filed in the Northern District of California accuses Samsung Electronics, SK Hynix, and Micron Technology — the three companies that together ship roughly 95% of the world's DRAM — of coordinating to restrict supply and inflate prices from late 2023 through 2025. The complaint, reported by Seoul Economic Daily on June 29, leans on the same fact pattern that has recurred in this industry for two decades: synchronized capex cuts, public commentary from executives signaling discipline, and a price curve that bends suspiciously upward across all three suppliers at once.
The numbers are the headline. DDR4 and DDR5 contract prices roughly doubled over the suit's class period, with DDR5 modules climbing from the ~$3.50/GB range in Q4 2023 to north of $7/GB by mid-2025 according to TrendForce spot data the complaint cites. Plaintiffs argue this was not a demand shock — AI server buildouts pulled HBM, not commodity DDR — but a coordinated supply response. Samsung idled legacy fabs. SK Hynix publicly committed to "disciplined" wafer starts on earnings calls. Micron's CFO told analysts in early 2024 the company would "prioritize profitability over share." Each statement, in isolation, is normal capital allocation. Together, the plaintiffs argue, they read as a signal.
This is the third major DRAM antitrust action of the modern era. The 2002 case ended with $300M+ in DOJ fines and executives — including four from Samsung — serving prison time. The 2018 class action settled in 2023 for $445M with no admission of wrongdoing. The pattern is now legible enough that institutional buyers (hyperscalers, OEMs) keep settlement reserves on their books as a line item.
The interesting question isn't whether the suit wins. It's why this market keeps producing the same case.
DRAM is a textbook tight oligopoly: three suppliers, a homogeneous commodity, public quarterly capex disclosures, and customers who buy on multi-quarter contracts. Economists call this a "facilitating structure" — you don't need a smoke-filled room when every player can read the others' moves from public filings and earnings transcripts. The 2002 case actually involved emails and phone calls. The 2018 case relied on parallel conduct. This one looks like it will lean even harder on the parallel-conduct theory, because executives have learned not to leave a paper trail.
The AI capex cycle changed the math in a way the complaint underplays. HBM — the stacked memory glued onto Nvidia and AMD accelerators — is genuinely supply-constrained, carries 3-5x the margin of commodity DRAM, and consumes wafer capacity that would otherwise produce DDR5. Every HBM die is roughly three DDR5 dies of foregone output, which means the suppliers can credibly tell a judge that capacity "shifted" rather than "was withheld." That's a real defense, and it will make the plaintiffs' burden harder than in 2018. But it also explains why this suit was inevitable: when the three suppliers all pivot the same wafers to the same higher-margin product at the same time, downstream buyers see a coordinated outcome regardless of intent.
Community reaction on HN (381 points, ~600 comments) split along predictable lines. The hardware crowd argued this is just oligopoly economics doing what oligopolies do — three players, no new entrants since the 1990s shakeout killed Qimonda and Elpida, and the capex to build a new fab now exceeds $20B. The legal-minded pointed out that parallel conduct alone has historically been hard to win on — the Supreme Court's *Twombly* decision raised the pleading bar specifically to dismiss these cases at the motion stage. The most cynical take, and probably the most accurate: this will settle for a number that looks large in a press release and small as a percentage of the price increase the plaintiffs paid.
The deeper structural point is that no one is going to build a fourth DRAM supplier. CXMT in China is the closest thing, and US export controls have explicitly tried to stunt it. Antitrust enforcement against a three-player market with no credible entrant is a tax, not a fix — it redistributes some past overcharge but doesn't change the next cycle.
If you're building anything memory-bound, the suit changes nothing about your near-term BOM. DDR5 is still expensive. HBM is still allocated quarters in advance. The supply picture for 2026-2027 looks tighter, not looser, because the same three suppliers are now also competing for advanced packaging capacity at TSMC.
A few concrete implications:
Capacity planning. If your three-year roadmap assumed memory follows a 2008-2019-style boom-bust cycle (overbuild, glut, crash, repeat), update the model. The post-2020 cycle looks more like a managed equilibrium with a structural premium. Bake in 20-30% higher per-GB costs than the trailing five-year average and you'll be closer to reality.
Architecture choices. The price gap between DRAM and storage-class memory (Optane is dead, but CXL-attached DRAM pools and high-end NVMe with persistent memory semantics are live) is now wide enough to revisit tiering decisions you made in 2022. If your service tolerates 200ns vs 80ns access latency for cold pages, CXL memory expansion at $2-3/GB is a real architectural lever, not a research project.
Negotiation posture. Hyperscalers have already restructured DRAM contracts to lock 18-24 month volumes with price collars. Mid-size buyers got squeezed because they were on spot or short contracts. If you're procuring at >$1M/year of memory, the cost of a procurement specialist who runs proper RFPs against all three suppliers (and the Tier-2 module makers — Kingston, ADATA, Crucial-as-channel) has gone from "nice to have" to "obviously pays for itself."
The suit will likely survive a motion to dismiss — the 2018 case did, and the parallel-conduct evidence here is at least as strong. It will probably settle in 2028-2029 for $300M-$800M, with no admission. The class will get coupons, the lawyers will get fees, and the next cycle will start. The real fix isn't litigation; it's either a fourth supplier (which no one is funding) or a memory architecture that breaks the DRAM monopoly's relevance (CXL pooling, processing-in-memory, or HBM eating down-market faster than expected). Until one of those lands, treat DRAM pricing the way you treat cloud egress fees: a structural cost of doing business, not a market clearing toward marginal cost.
It wouldn't be a first: https://en.wikipedia.org/wiki/DRAM_price_fixing_scandal
HBM is also DRAM. I also think it's kind of a weak argument to say that them discontinuing ddr3 (which while in use still today in industrial/embedded was on the way out for consumers 10 years ago) and ddr4 which last had consumer CPUs for it 3 years ago is meaningful. What we need now is
What happens if Samsung and SK Hynix simply stop selling to US at all? Micron is in US but are the rest still in the US jurisdiction?They are selling the hottest commodity of the day. It’s made outside of the US using non-American tooling.
Everybody in our industry loves fat margins. But god forbid if someone else captures the margins and squeeze them out of easy profits.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
This was attempted before in 2022 but fell apart because plaintiff couldn't show an agreement took place https://cdn.ca9.uscourts.gov/datastore/opinions/2022/03/07/2...