Samsung vs CXMT: Why Banning Chinese RAM Won't Fix $400 DDR5 Prices

0:00 The cheapest 32-gigabyte kit of DDR5 in America right now runs $402. That same kit, same speed, same capacity, has a recorded floor of 72, and nothing about it improved.

0:13 It's five and a half times the price for an identical part, and it isn't scalpers, it's the market clearing. About a fifth of the world's DRAM wafers now go to AI hardware instead of your

0:23 machine, because memory got eaten. The stacked memory that feeds AI accelerators burns about three times the silicon per gigabyte that normal DDR5 does. So, every wafer that moves costs

0:34 you three. That's the whole shortage in one sentence, and it's arithmetic, not malice. That leaves the one question that actually matters. Who is going to go and

0:43 build more of the boring kind? There's exactly one company doing it at speed, and until recently, almost nobody outside China had heard of it. It's called CXMT.

0:53 Last month, it listed in Shanghai and opened up 471% for part of that day. It was the most valuable listed company in the country. It outranked the banks. Then, on the

1:03 11th of August, a court in Seoul heard sworn testimony about where its process came from. Here's what it actually is, because the scale is what people get wrong. Three

1:12 big memory fabs, two in Hefei, one in Beijing, starting somewhere around 300,000 wafers every month between them. The researchers who modeled this expect it to leave this year within about

1:23 25,000 wafers a month of Micron, the smallest of the three Western giants, and the fabs it's building now would roughly double that again. It already supplies Lenovo, Xiaomi, and

1:34 the big Chinese clouds, and its DDR5 now clocks higher on an ordinary desktop board than most enthusiast kits managed 2 years ago. Last quarter, it took about 7% of global

1:44 DRAM revenue and grew 716% against the year before, and nobody else in the industry is anywhere near that curve. Three Western PC brands have started putting its chips in notebooks.

1:56 Not many, and reportedly, none of those machines are sold in America. That detail matters later. So, the relief valve exists, it's real, it's ramping, and it just raised the largest sum any

2:06 company has ever raised on Shanghai's tech board, which is precisely when the story turns. On the 11th of August, a witness took the stand in that Seoul courtroom.

2:15 Korean reporting identifies him only by his surname Jun. He's 56, and he gave 28 years to Samsung. In 2016, he left and joined a Chinese startup that had just been founded to make memory.

2:28 That startup was CXMT. In April, a court sentenced him to 7 years. What he took has a name. Samsung calls it a process recipe plan, and it's exactly what it sounds like. About 600 steps in order

2:41 for turning a blank silicon wafer into working memory. Not a blueprint of the chip, the instructions for the factory. Temperatures, pressures, how long each layer sits in each machine, and which

2:51 machine, the etch conditions, the anneals, the implant settings. A Korean business paper reports that plan took Samsung 5 years and over a billion dollars to work out, and Jun was

3:01 paid about 2 million for it, plus stock options. So, the recipe cost roughly 500 times more to develop than it cost to buy. That ratio is the entire economics of industrial espionage, and it explains

3:13 why this keeps happening. The reason this week mattered isn't the money, though. It's what he told the court about intent, and it's worth quoting rather than paraphrasing. In his words,

3:23 "From the beginning, CXMT as chief executive and chief technology officer, tried to develop memory by stealing Samsung's process recipe plan." And then the sentence that does the

3:33 damage. "The company," he said, "had no research institutes or facilities when it was founded, and no intention of developing technology through its own research."

3:42 That's an allegation from a convicted man, and you should weigh it accordingly. But it's now sworn testimony on a court record, and court records travel. There's a specific

3:51 American door that a foreign court record can open. It's called section 337 and it comes from a trade act written in 1930. It lets America's Trade Commission

4:01 investigate imports built on stolen trade secrets and if it finds a violation, it tells customs to stop those products at the border. Every shipment, no damages, no negotiation,

4:11 just the door shut. It's fast by legal standards, roughly a year and a half start to finish and it doesn't care whether the company has any American presence at all.

4:20 That's the weapon people are pointing at right now. One thing though, Samsung hasn't filed it. Not at the commission, not anywhere. No case exists yet. That idea came from

4:30 prosecutors and reporters and Samsung itself has said nothing at all, which brings up the question everyone skips and it's the one this whole thing turns on. If you ban the copy, does the

4:41 original get cheaper? Three things say no. The first is the one people refuse to believe. CXMT memory isn't cheaper. A tracked listing in China last month put a 64 GB server module built on Chinese

4:53 chips at about $2800 and the identical module on Samsung or SK Hynix chips at 2745. The Chinese part was about 2% more expensive. In a shortage, everybody prices to what the

5:05 market will bear and a new entrant has no reason on earth to undercut a sold out market. Second thing, those chips mostly aren't coming here. The notebooks using them sell outside America. CXMT's

5:17 output is absorbed by Chinese customers who ordered it first and Congress is working to keep it that way. Last month, Chuck Schumer and Jim Banks gave Apple until the 21st of August to promise it

5:28 would never put Chinese memory in any product anywhere. That deadline has passed without a public answer. Third thing, and this one has a scoreboard. America has already run this exact case

5:38 once. In 2018, American prosecutors charged a Chinese memory maker, Fujian Jinhua, with stealing Micron secrets. It was cut off from American suppliers within

5:48 weeks, and the foundry that helped it pleaded guilty and paid $60 million dollars. And in February 2024, after more than 5 years, a federal judge found Jinwa not

5:59 guilty on every count. That was a criminal case with a criminal standard, and a trade commission is a lower bar. But 5 years and an acquittal is the base rate anyone predicting a quick ban

6:09 should have to look at. There's an older rhyme here that almost nobody brings up. In 1986, America accused Japan of dumping memory below cost and forced an agreement. Japan agreed to stop and to

6:20 raise its prices. It worked, and prices went up. They went up everywhere outside Japan, which meant American computer makers, the exact people the deal was meant to protect, paid more for memory

6:30 for years. The opening it created didn't go to America, though. It went to Korea. Over the following 14 years, Japan's share of memory fell from about three-quarters to roughly a fifth, and

6:41 Samsung took the industry. One more thing complicates the villain chasing. In June, in a federal court in California, 17 buyers sued Samsung, SK Hynix, and Micron together, arguing the

6:53 three of them coordinated the squeeze on purpose. Those three control about 90% of the market. Micron denies it, and the case is live and unproven. So, here's where this lands. On the law, Samsung is

7:05 right and should press it. A 600-step recipe that cost a billion dollars and 5 years is exactly what trade secret law exists to protect, and a court has already convicted a man for

7:15 handing it over. But that's a case about theft, not a case about your shopping cart. So, ban the copy and you remove supply from a market that's already short, and

7:24 nothing in that makes memory cheaper. It does the opposite. The idea worth throwing out is the one both sides are selling, that a trade remedy is a price remedy. A ban won't bring prices down,

7:34 and imports won't either. The pricing sheet already proved it. The concession is real, though. If CXMT ever points that capacity at America at scale, section 337 is the only tool that

7:46 actually works, and Samsung would be right to have it built and ready. So, here's the bet on the record. No American exclusion order will be blocking Chinese memory at the border by

7:55 the end of 2027, which leaves the question worth sitting with. If the recipe really was stolen, and shutting the border wouldn't move your price by a dollar, then what's the

8:05 remedy actually for, and who is it for?