AI’s infrastructure boom has a bill. Someone else is paying it.
In late October 2025, Xiaomi unveiled the Redmi K90. Base price: 2,599 yuan — 100 yuan more than the K80 it replaced. For the factory workers and young professionals who buy Redmi phones, 100 yuan is a meal out with friends — small enough to notice, large enough to resent. Consumers in China noticed immediately. On Weibo, the complaints accumulated fast enough that Xiaomi’s president, Lu Weibing, responded personally. Memory chip prices had “exceeded forecasts,” he said. The surge “may continue to escalate.” Xiaomi quietly rolled back the price on the most popular configuration by 300 yuan to absorb the backlash.
Three weeks later, on the Q3 earnings call, Lu said the quiet part out loud: “I anticipate that the pressure will be significantly greater next year compared to this year. Overall, consumers are likely to experience a substantial increase in retail prices for products. While some of this pressure may need to be alleviated through price hikes, simply raising prices will not suffice to absorb all the costs.”
Most people who caught the Bloomberg headline — “Xiaomi Sees Phone Prices Rise in 2026 Due to Memory Chip Crunch” — read it as a supply chain story. Tight market, AI demand, unfortunate timing. That framing is accurate as far as it goes. It just stops exactly where the interesting question begins.
The Story Everyone Is Telling
The consensus narrative on DRAM is clean and, in its essentials, correct.
Artificial intelligence runs on memory. Training a large language model requires moving massive amounts of data between processors at speeds that conventional DRAM can’t sustain. The solution is High-Bandwidth Memory — HBM — stacked chips that sit directly on the same package as a GPU, passing data through thousands of parallel connections simultaneously. Each Nvidia H100, the chip that powered the current AI buildout, contains 80 gigabytes of HBM. Each H200 contains 141. The hyperscalers — Microsoft, Google, Meta, Amazon — need tens of thousands of these cards.
The three companies that make essentially all of the world’s DRAM — Samsung, SK Hynix, and Micron — can make HBM or they can make the commodity memory that goes into your phone and laptop. Making both at scale, on the same fabrication lines, involves real trade-offs. HBM takes up more factory space per chip and is trickier to manufacture — roughly three times the wafer area of DDR5 (the standard memory used in consumer devices) for equivalent bit count, with lower yields. In practice, every HBM wafer a fab runs is two or three commodity DRAM wafers it doesn’t. And HBM is dramatically more profitable.
So the Big Three pivoted. And as they shifted capacity toward HBM, the supply of conventional DRAM tightened. Prices for commodity memory — the kind in every smartphone, laptop, and mid-range PC — began rising in mid-2025 and accelerated sharply in Q4. By January 2026, a 32GB DDR5 kit that cost $82 in August had risen to approximately $310. Conventional DRAM contract prices were projected to rise 90–95% quarter-over-quarter in Q1 2026.
That’s the story. AI ate the memory market. Painful but explicable. Nobody’s fault.
The Part the Story Skips
Here is what the consensus narrative doesn’t explain: why the price spike arrived when it did, not when the AI demand did.
AI’s appetite for HBM was visible and growing through all of 2024. It wasn’t a surprise. And yet, as recently as November 2024, TrendForce — the industry’s most-cited forecasting firm — was predicting that DRAM contract prices would turn downward in 2025, with larger drops in the first half of the year. Their reasoning: high inventory levels, sufficient supply, weak consumer device demand. The market, they said, was balanced.
They weren’t wrong about the demand side. The demand picture didn’t change between that November forecast and the price spike six months later. Something else did.
The timeline tells you what.
On July 26, 2023, Samsung reported a $3.4 billion operating loss in its memory unit — part of a combined $7 billion loss in the first half of the year. On that earnings call, EVP Jaejune Kim announced the company would extend production cuts and double its HBM capacity. The pivot was real and the losses were real. Samsung wasn’t manufacturing a crisis; it was responding to one.
But watch what happens next.
On February 21, 2024 — roughly six months after Samsung’s call — SK Hynix VP Kitae Kim published a statement confirming the company had sold out its entire HBM production for all of calendar 2024. The announcement was widely covered in trade press within a week. SK Hynix’s CEO reiterated the position in early May: sold out for 2024, nearly sold out for 2025.
On March 20, 2024 — 27 days after SK Hynix’s announcement — Micron CEO Sanjay Mehrotra stated on the company’s quarterly earnings call: “Our HBM is sold out for calendar 2024, and the overwhelming majority of our 2025 supply has already been allocated.”
Twenty-seven days. On a call that every investor, analyst, and competitor tracks in real time.
By October 2025, SK Hynix announced all of its DRAM, NAND, and HBM capacity was sold through 2026. On December 3, 2025 — after a brief period in Q1 2025 where conventional DRAM prices had actually declined, suggesting the market was re-balancing — Micron issued a formal press release announcing the complete shutdown of its Crucial consumer brand. Shipments would continue only through February 2026.
| Date | Actor | Action |
|---|---|---|
| July 26, 2023 | Samsung | Extends cuts; announces HBM doubling on earnings call |
| Feb 21, 2024 | SK Hynix | Publicly announces HBM sold out for all of 2024 |
| Mar 20, 2024 | Micron | HBM sold out for 2024; most of 2025 allocated |
| Oct 2025 | SK Hynix | All DRAM, NAND, HBM capacity sold through 2026 |
| Dec 3, 2025 | Micron | Exits consumer DRAM entirely; Crucial brand shut down |
What changed wasn’t demand. What changed was supply discipline.
Each of these decisions was individually rational. Samsung was losing money. SK Hynix and Micron were watching a competitor announce record margins on HBM and choosing to follow. Micron’s Crucial exit came after a period where conventional DRAM margins had temporarily compressed — a clear signal that the commodity end of the market wasn’t worth the capital.
None of this required a phone call. None of it required a back-channel. It required only three companies, watching the same public earnings calls, responding to the same margin signals, in the same direction, within weeks of each other.
The Legal Name for This
There is a term for what happens when companies in a concentrated market make parallel decisions without explicit coordination. Courts call it conscious parallelism — parallel decisions without explicit coordination — and unlike a cartel, it carries no legal exposure. In the United States, it is entirely legal.
The DRAM industry has been here before. In the 2016–2018 cycle, DRAM prices nearly tripled. Samsung cut production in Q3 2016; Micron and SK Hynix followed a quarter later. A class-action lawsuit was filed in 2018 alleging the three companies — then holding 96% of the global market — had conspired to restrict output. The district court dismissed in 2019. The Ninth Circuit affirmed the dismissal in 2022, reasoning that parallel supply cuts in a concentrated market were “more consistent with conscious parallelism than with conspiracy.” Even with eight alleged “plus factors” — simultaneous capex reductions, supply cuts against short-term self-interest, and the fact that executives from the prior convicted cartel were still employed at all three firms — the court found insufficient evidence of a preceding agreement.
The current cycle mirrors that playbook almost exactly. With one upgrade: this time, the capacity reallocation is being announced publicly on earnings calls rather than inferred from market data. Every firm can observe, in real time, what the others are doing. The mechanism doesn’t need to be covert to function.
Samsung’s operating profit margin on standard DRAM hit approximately 40% in Q3 2025. Its HBM margin was running at 60%. When Micron said publicly that HBM has a “3-to-1 trade ratio with DDR5” — meaning every HBM wafer displaces three commodity DRAM wafers — it was quantifying a known, deliberate choice, not describing a physical constraint. This is not a supply chain accident. It is a capital allocation decision, made by identifiable people, on a recorded earnings call. Samsung demonstrated this clearly in December 2025 when it voluntarily reallocated 30–40% of its 1a DRAM process capacity back toward DDR5 server memory, after that product achieved 75% margins. The fabs can run either product. The decision is about which margin to chase.
Three companies. 95% of global DRAM production. No open investigation by the FTC, the EU, or China’s SAMR. And a legal framework that, by design, cannot distinguish between a cartel and three rational firms reading the same earnings calls.
Who Actually Pays
The downstream cascade is moving faster than the headlines are catching it.
TrendForce’s December 2025 forecast contains a sentence that deserves more attention than it received: low-end smartphones will likely return to 4GB RAM base configurations in 2026. That is not a rounding error. For a decade, the defining story in consumer technology was democratization — flagship features migrating down to affordable devices within two to three product cycles. Cheap Android phones in Southeast Asia and Latin America kept getting better, and memory was one of the primary vectors for that improvement.
That trend is now running in reverse. Mid-range phones with 12GB of RAM are projected to decline by over 40% in availability. 16GB variants “nearly disappear.” DRAM represents approximately 20% of a mid-range smartphone’s total bill of materials — which means the price spike hits affordable devices harder, proportionally, than it hits premium ones. Apple can absorb it. A $200 Android phone cannot.
The Xiaomi moment is the human version of that math. A company that built its business on offering near-flagship specs at mid-market prices tried to pass through a memory cost increase, got consumer pushback, partially absorbed the difference, and then went on an earnings call and said: we can’t keep doing that.
This is the moment where a macro structural story becomes a product decision in Shenzhen, and then a purchase decision in Jakarta.
The working hypothesis: The DRAM price crisis of 2025–2026 is the predictable output of a legal oligopoly making rational individual decisions. The people subsidizing AI infrastructure are not the people using it. And the legal and regulatory frameworks that exist to address this kind of market concentration cannot, by design, reach it.
China’s Calculation
There is one country that has both the leverage and the historical motivation to push back on this dynamic. China imported $88.92 billion worth of chips in 2017. When Korean and American memory makers raised prices 40%+ during the 2016–2018 cycle, China’s NDRC called Samsung in for questioning. By mid-2018, China’s SAMR had raided the offices of all three firms in Beijing, Shanghai, and Shenzhen. The investigation didn’t produce formal penalties, but it directly preceded Samsung’s announcement of capacity expansion in February 2018 — which market participants took as a signal the pricing cycle was ending.
In the current cycle, Chinese smartphone OEMs are experiencing the same squeeze. Xiaomi, Oppo, Vivo, and Honor are all absorbing cost pressure that is structurally identical to 2018. The domestic complaint base is there. The leverage is there.
But China’s antitrust motivation has shifted. CXMT — ChangXin Memory Technologies — is now the world’s fourth-largest DRAM producer, with roughly 11% of global capacity by wafer volume. Its client list includes Alibaba, Tencent, ByteDance, Lenovo, Xiaomi, Oppo, and Vivo. China is no longer solely a buyer in this market; it is trying to build a seller. Pressuring Samsung and SK Hynix now means disrupting the oligopoly that CXMT is attempting to enter — on terms that favor the incumbent.
The country with the most leverage to regulate this market has quietly changed sides.
CXMT is not, for the record, a near-term solution to the supply crunch. As of mid-2025, its DDR5 yield rate was hovering just above 50% — commercial viability requires 80–90%. U.S. export controls have now placed a ceiling on its capacity expansion. A veteran DRAM engineer who spent 26 years at SK Hynix told the Korea Herald in February 2026 that the technology gap is “more than five years” — and may be growing. The scenario in which CXMT disrupts the Big Three’s pricing discipline before 2028 requires the company to achieve commercial yields, clear the export control ceiling, and do all of it before the incumbent suppliers lock in another 12–18 months of contracted AI supply. All three, in sequence. It is possible. It is not likely.
The mechanism that broke the last pricing cycle — Chinese regulatory pressure on the Big Three — is now captured by the same dynamics driving this one.
The Working Hypothesis
The DRAM price crisis of 2025–2026 is the predictable output of a legal oligopoly making rational individual decisions. It is not a conspiracy. It does not require one. Three companies control 95% of a critical input to nearly every consumer electronic device on earth. They all observed the same margin signals. They all made the same strategic pivot. They all announced that pivot on public earnings calls where their competitors were listening.
The “AI shortage” narrative is accurate — AI demand for HBM is real, the trade-offs are real, and the individual business decisions were rational. But the narrative stops before asking the next question: who absorbs the cost when three rational actors in a 95%-concentrated market all chase the same margin at the same time?
The answer, in 2026, is a mid-range Android buyer in Jakarta who wanted 12GB of RAM and is getting 4GB instead.
No regulator is investigating. No court has found wrongdoing. The legal framework, as the Ninth Circuit clarified in 2022, cannot reach this. Conscious parallelism is not a crime.
It doesn’t need to be a crime to matter.
What would change my mind
-
CXMT clears all three hurdles in sequence. Commercial DDR5 yields above 80%, sufficient capacity to meaningfully supply global markets, and access to the manufacturing equipment needed to scale — all before the Big Three lock in another round of long-term AI supply contracts. If this happens before 2028 and commodity prices normalize as a result, the oligopoly’s pricing power was more fragile than this thesis assumes.
-
Commodity DRAM prices normalize within 12 months despite no new Big Three capacity additions. If prices fall sharply by Q1 2027 without a meaningful supply increase, the spike is cyclical rather than structurally managed — and the conscious parallelism framing overcounts the firms’ agency.
-
An internal capacity allocation memo becomes public. If a regulator opens a formal inquiry — FTC, EU, or a newly motivated SAMR — and forces disclosure of how the Big Three actually made their reallocation decisions, and those documents show purely independent analysis with no awareness of competitor positioning, this thesis is wrong on the mechanism, even if right on the outcome.
If you found this useful, the best thing you can do is forward it to one person who would push back on it. I’d rather be wrong in public than right in private.