Sector deep dive · Memory · the China question · August 4, 2026

The most valuable company in China makes the cheapest chips in the world.

Author Brandon Leon Posted 2026-08-04 Coverage Memory · DRAM/HBM · CXMT · the memory basket we still hold

TL;DR: China's biggest memory-chip maker went public on July 27, rose 466% in a day, and became the mainland's most valuable listed company. The panic that followed wiped more than a trillion dollars off global chip stocks and put a question under the memory fund we still hold. After a week of underwriting, our answer: the fear is real, but it is early — and early is the trade. CXMT makes the ordinary memory in phones and laptops, not the advanced HBM and server DRAM the AI shortage is actually about; export controls and missing tools fence it out of the advanced segment for years; and the single best piece of evidence is its own price list — its chips reportedly sell above Samsung's in China. An entrant about to crash the market doesn't charge a premium. We hold the basket, on watch, with two public tripwires below. Also inside: the honest cost of our own discipline (Alphabet ran 13% without us, exactly the risk we published first), and the week the power gates finally open.

The Monday

On the morning of July 27, almost nobody in American markets could have told you what ChangXin Memory Technologies was. By that evening it was the most valuable listed company in China.

The listing mechanics were simple. CXMT priced its Shanghai IPO at 8.66 yuan and raised about $8.6 billion — Asia's biggest listing of the year. It closed its first session at 49 yuan, up 466%, worth roughly 3.3 trillion yuan. Call it $488 billion — more than ICBC, the state bank that has sat on top of the mainland market for most of a generation. A memory-chip maker from Hefei, seven years into volume production, now outweighs the world's largest deposit-taking bank by market value.¹

What followed was less absurd and more expensive. Over three sessions, global chip stocks shed more than a trillion dollars. Micron — yes, a four-digit stock this summer — fell from a thousand dollars to the low seven hundreds. SanDisk lost a third of its value. Our own memory basket — which we hold, and which is the reason this piece exists — traded 25% below our cost at the bottom. In Washington, a bipartisan group of lawmakers demanded a national-security probe of the IPO itself, on the theory that Beijing had orchestrated the surge to capitalize a strategic asset. The Gang of Eight asked for a classified briefing about a DRAM company.¹

Markets are pricing machines, and what that Monday priced was a single sentence: Chinese memory is not a 2028 story anymore. A $488 billion valuation is not a bet on next quarter's DDR5 shipments. It is the market saying the supply side of the memory shortage — the shortage Samsung says gets worse through 2028, the shortage Tim Cook called a “100-year flood” four days later on his final earnings call — has a new author, and the new author answers to an industrial policy, not a boardroom.²

We sold our Micron position the same week — a pre-placed stop, the repair from an earlier failure of nerve, and that story has been told and graded in the open. But we kept the basket. Which means we owed ourselves, and you, an actual answer to the only question that matters: is the fear right? Not directionally, someday, in the fullness of history — markets don't pay on someday. Is it right about 2026 and 2027, the window our thesis lives in? We spent the week underwriting it. Here is what we found, and where we'd admit we're wrong.

Simply: A Chinese chip company most investors hadn't heard of went public and instantly became China's most valuable company — bigger than its biggest bank. Markets panicked that a flood of cheap Chinese memory chips was coming, and chip stocks worldwide lost over a trillion dollars in three days. We own a memory fund, so this piece answers the only question that matters to us: is that panic justified?

What CXMT actually is

Start by giving the fear its due, because the company is real. CXMT is not a subsidy sinkhole with a ribbon on it. It runs roughly 265,000 wafer starts per month as of last year, on its way toward roughly 350,000 by the end of this year — within shouting distance of Micron's entire global footprint. It skipped a process generation on the way here: rather than grinding through the 1y node, it jumped to a 1z-class process — roughly where today's mainstream DDR5 and LPDDR5X live — and went straight at the memory that goes into current servers and phones. The major motherboard makers have validated its parts. And the financials just inflected in the way state-scaled manufacturing eventually does — once the fabs fill and the yields move, losses flip to surplus almost overnight: from a 1.6-billion-yuan quarterly loss a year ago to roughly 25 billion yuan of net profit in the first quarter of 2026, on revenue up more than 700%.³

The ramp from here is the part that should genuinely widen your eyes: third-party trackers have CXMT reaching perhaps 420,000 wafer starts next year and half a million by 2028 — which would be roughly 17% of global DRAM supply, from a company that was a rounding error five years ago. The new Shanghai fab is planned at two to three times the size of Hefei. This is what it looks like when a country decides an industry is strategic and prices capital accordingly. If you want a finance-native anchor for it: this is the shipbuilding and steel playbook — state-priced capital, capacity added straight through the cycle, incumbent margins doing the dying — pointed at memory. Nobody serious disputes the direction of travel.³

So yes: China's memory champion is real, big, profitable, and growing at a speed the incumbents never had to manage against. If that were the whole story, we would have sold the basket into the panic and written you a very different note.

Simply: The company is no paper tiger. Its factories are enormous and growing, its chips are modern enough for today's phones and servers, and it just flipped from losses to serious profit. By 2028 it could make nearly a fifth of the world's memory chips. Anyone dismissing it entirely isn't paying attention — the question is which chips, which is the next section.

What it isn't — yet

Here is the detail the trillion-dollar unwind skipped past: the shortage and the company live in different neighborhoods.

The memory crisis of 2026 — the one inflating Microsoft's capex line by a reported $25 billion, the one Apple says is flooding its costs, the one that has Samsung locking 60–70% of its capacity into five-year take-or-pay contracts — is overwhelmingly a crisis of advanced memory: HBM stacks feeding AI accelerators, and the leading-edge server DRAM that datacenters are stripping out of every other market — not the commodity DDR5 that powers phones and laptops. That is where the pricing power is, where the contracts are signed, and where our basket earns its keep.²

Chart 2 — One market, two neighborhoods

The shortage lives upstairs. The entrant moved in downstairs.

The 2026 shortage — Samsung's “more severe” 2027, Microsoft's ~$25B component-cost line, Apple's “100-year flood” — is a crisis of the top layer. CXMT's prospectus, tooling access, and estimated yields confine it to the bottom layer into at least 2027. Sources: company disclosures and estimates as footnoted.⁴

And that is precisely where CXMT cannot yet go. Its own IPO prospectus — the document written to make the most seductive possible case to investors — contained no HBM project at all. Its HBM3 exists as samples to Huawei; one respected research house estimates its stacked yields near 25%, which is a science project, not a product line. The company now talks about HBM production by the end of 2026, starting at HBM3 or HBM3E — which, with the incumbents accelerating HBM4, would leave it one to two full generations behind on the day it ships. It has no EUV lithography and no path to buying any; the December 2024 U.S. export controls were written specifically to choke the through-silicon-via and stacking tools that HBM manufacturing requires. CXMT can scale what it has — and what it has is the commodity floor of the market.⁴

Chart 1 — The ramp everyone fears, and the shelf it sits on

CXMT's wafer capacity is exploding. The sliver that could touch the AI shortage — its estimated HBM allocation — is the gold at the bottom of each bar.

Capacity and HBM-allocation figures are third-party estimates (SemiAnalysis; TrendForce), not company guidance — the ramp's slope is the consensus, the exact numbers are not. At ~500K wafer starts, CXMT would represent roughly 17% of global DRAM supply in 2028 — overwhelmingly at commodity nodes.⁴

Then there's the tell — the one number in this whole story that behaves like evidence rather than opinion. If CXMT were about to collapse the memory market, you would expect its chips to be cheap. They are not. Its 64-gigabyte server DDR5 modules have reportedly been pricing above Samsung's inside China. Sit with that for a moment: the feared undercutter is charging a premium, because demand so exceeds supply that even the entrant prices under the shortage's umbrella rather than slashing through it. DigiTimes — which tracks this weekly, and whose data we now watch as a matter of published discipline — flags the China-made price gap narrowing as the thing to watch. Narrowing is not undercutting. Until that premium flips to a discount, CXMT is participating in the shortage, not ending it — and when it flips, the story changes, and we have committed in writing to changing with it.⁵

Chart 3 — Tripwire #1, drawn

Where the entrant prices today, and the flip we watch for every week.

CXMT's 64GB server DDR5 reportedly prices above Samsung's inside China (DigiTimes) — a price-taker under the shortage, not an undercutter through it. The gap narrowing is noted; the gap flipping is tripwire #1. Deliberately schematic: the position on the spectrum is directional, not measured.⁵

The incumbents' own behavior says the same thing louder. In the two weeks since the debut: Samsung reported a record quarter and told analysts the shortage gets more severe in 2027 and persists into 2028. SK hynix responded to CXMT not by cutting price but by expanding long-term agreements with more than ten major customers and pulling its HBM4 schedule forward — running up the stack, away from where CXMT competes, faster. Companies that fear a price war hoard cash and hedge. Companies that fear losing a technology race accelerate. In our framework, that shift — from hedging to accelerating — is exactly what you'd expect in a world where the shortage is real and persistent, not about to be crushed by new commodity supply. Watch what they do.²

The most valuable company in China sells the cheapest kind of memory, at a premium, into a shortage it didn't create and can't yet reach the top of. Both halves of that sentence are true. The trade lives in the gap between them.

Simply: There are really two memory markets: the ordinary chips inside phones and laptops, and the advanced ones (called HBM — high-bandwidth memory) that AI datacenters are desperate for. CXMT can only make the ordinary kind — export rules block the tools it would need for the advanced kind, and its early advanced samples mostly fail. The best evidence it isn't crashing anything: its chips sell for more than Samsung's in China. A company about to flood the market doesn't charge a premium — and its rivals are responding by racing upmarket, not by cutting prices.

Where the fear is right, and what would prove it

None of this makes the bear case wrong — it makes it early, and early and wrong are different things that feel identical in your P&L until the timing resolves. So let's write the bear's best case in our own hand, because that's the version worth testing against.

By 2028, the picture genuinely darkens. Half a million wafer starts is enough commodity supply to cap DRAM pricing across the bottom of the market even if CXMT never ships a competitive HBM stack. The next node down — a 1a-class process — is probably achievable without EUV; Micron itself proved that path exists. And the path where the politics break China's way is easy to sketch: the Commerce Department has had a CXMT Entity List designation approved and paused since June, the longest such pause in a decade, while trade talks continue — and Apple, which has been testing CXMT's LPDDR5X for China-market devices, is lobbying against the listing. A trade deal that shelves the designation, plus an Apple qualification, is the bear's opening chapter. And the yield problem is a solvable problem; 25% becomes 60% the way it always does in memory — with money, engineers, and time. China has the first two in state-backed abundance. The argument is only ever about the third.⁴

Our position is not that the fear is fiction. It's that the fear is mistimed. The Monday said Chinese memory is no longer a 2028 story; our underwrite says 2028 is exactly what it still is — and the market charged a trillion dollars, in three sessions, for a risk whose earliest honest arrival date sits beyond the window that Samsung's contracts, the hyperscalers' capex, and Cook's flooded margins are all screaming about. The basket's thesis — own the uncapped beneficiaries of a shortage that worsens through 2027 — survived this underwrite. It did not survive unconditionally. It survived on watch, with the two tripwires we published in the deployment opener now formalized here:

The early warning: CXMT-versus-Samsung server-DRAM pricing in China's spot market flips from premium to discount — the day the entrant stops sheltering under the umbrella and starts cutting through it. Tracked weekly.
The kill headline: verified volume HBM3 shipments to Huawei at workable yields — the day the fence around the advanced segment fails.

Either one, and this thesis gets re-underwritten in public, with the same knife we've used on our own calls all summer — mark-to-market, dates, and receipts included. That's the deal.

The disclosure that keeps us honest

We hold the memory basket this piece defends, from a $59.73 basis, marked around $54 as we publish — underwater, recovering, and conflicted in exactly the way you should assume every author defending a position is conflicted. That's why the tripwires are specific, dated, and third-party-observable rather than vibes. Judge us on whether we honor them; we've spent the summer building the record that we do — including the one time we didn't, which we published too.

Simply: By 2028 the threat gets real: that much new capacity can drag down ordinary-chip prices even if the advanced fence holds, and the politics could swing China's way — Washington has already paused its harshest restriction while trade talks run, and Apple is testing CXMT chips. Manufacturing problems, meanwhile, get solved with enough money and time. So we keep our memory fund, but with two published warning signs: if CXMT's prices in China drop below Samsung's, or if it starts shipping real volumes of advanced HBM chips, we rethink everything — publicly, with numbers.

Coda: the week the gates open

A housekeeping note on the deployment series, because the calendar we published last Thursday starts mattering tomorrow. The refunding statement lands Wednesday; Constellation reports Thursday before the open and Vistra Friday morning alongside the July jobs report — the power lane's gates, exactly as written. Tonight, AMD reports, and the proof-versus-faith exam we watched the hyperscalers take moves down the stack to the companies that sell them silicon: Palantir sat the same exam last night — revenue up 93%, guidance raised again — and got paid 15% after hours for it. The axis keeps working.⁶

And one honest mark, three trading days old, exactly as we promised to keep them: the first risk we listed in the opener's “what makes this framework wrong” card was that the melt-up would resume without us while our gates kept us in bills. It promptly did. Alphabet — which we rated wait-for-price with an alert near $318 — never came within ten dollars of the alert, caught a Morgan Stanley defense of the very funding story we flagged, and now trades near $376, about thirteen percent above where we passed. We wrote that this was the price of the discipline, and we're paying it in public in week one. The gates don't move. If that costs us Alphabet, it costs us Alphabet — the framework isn't a prediction that patience always wins; it's a decision about which mistakes we're willing to make.⁶

Simply: China's new chip giant is real and growing fast — but it makes the ordinary memory chips, not the advanced ones the AI shortage is actually about, and its own prices (higher than Samsung's!) prove it isn't undercutting anyone yet. We're keeping our memory fund, with two specific public warning signs that would change our mind. Meanwhile: we admit Google ran away 13% while our discipline kept us out — a cost we accepted in writing before it happened — and this week the power-company earnings we've been waiting for finally arrive.

Get the next note by email → All writing

Sources & footnotes

The debut: CXMT priced at ¥8.66, raised ~¥57.9B (~$8.6B, Asia's largest 2026 IPO), closed its July 27 first session at ¥49 (+466%), market cap ~¥3.3T (~$488B), displacing ICBC (~¥2.6T) as the mainland's most valuable listed company (CNBC, July 27). The >$1T three-session global chip-sector unwind: CNBC (July 29). Congressional response: bipartisan letters seeking a national-security probe of the IPO and a Gang of Eight briefing request, as reported. Basket drawdown per verified closes ($59.73 basis; $44.85 low on July 29). ↩
The shortage's location: Samsung Q2 call (July 30) — DRAM shortage “more severe” in 2027, persisting through 2028; 60–70% of capacity into five-year supply agreements. Apple FQ3 call (July 30) — Tim Cook on memory-cost inflation as a “100-year flood” pressuring September-quarter margins. Microsoft's ~$25B component-cost attribution within its capex framing, as reported with its FQ4. SK hynix response to the CXMT listing: expanded LTAs with 10+ clients and accelerated HBM4 (Seoul Economic Daily, July 30). ↩
What CXMT is: ~265K wafer starts/month (2025) trending toward ~350K exiting 2026; 1z-class node reached by skipping 1y; DDR5/LPDDR5X validated by major motherboard makers; Q1 2026 revenue ¥50.8B (+719% y/y) with net profit ~¥25B vs a ¥1.6B loss a year earlier; 1H26 revenue guided ¥110–120B; Shanghai fab planned at 2–3x Hefei's scale (IPO prospectus figures and coverage: CNBC, SemiAnalysis, BigGo/Reuters syndication). ↩
What it isn't: no HBM project in the IPO prospectus (Tom's Hardware); HBM3 samples to Huawei with stacked yields estimated near 25% and mass production reported unlikely on schedule (SemiAnalysis estimate; Korean trade press); company HBM target now end-2026 at HBM3/HBM3E — one to two generations behind as incumbents accelerate HBM4 (CNBC, July 31; Seoul Economic Daily); no EUV access; December 2024 U.S. controls specifically cover HBM manufacturing/packaging tooling (TSV etch, stacking); capacity path ~420K (2027e) and ~500K wafer starts (~17% of global DRAM, 2028e) per SemiAnalysis/TrendForce estimates; SemiAnalysis models the DRAM market remaining supply-constrained through 2028 even with CXMT's full ramp. The bear's chapter: G5/1a-class without EUV precedented by Micron; Commerce's approved-but-paused Entity List designation (June, longest pause in a decade); Apple reported testing CXMT LPDDR5X for China devices while lobbying against listing; DoD Chinese-military-company designation with procurement ban effective June 30. ↩
The pricing tell: CXMT 64GB server DDR5 reportedly priced above Samsung's in China (DigiTimes); DigiTimes (August 4) flags the China-made price gap narrowing as the monitored indicator. Our tripwires as first published: the deployment-series opener. ↩
The coda: Palantir Q2 (Aug 3 AMC) — revenue $1.94B (+92.8%), U.S. commercial +149%, FY guide raised to ~$8.15B, +~15% after hours (Fortune/CNBC). AMD reports Aug 4 AMC (consensus ~$1.61/$11.31B; options pricing ~8.5%). Alphabet: closed near $376 (Aug 4) after a Morgan Stanley defense of its capex funding (Motley Fool, Aug 3) — ~13% above our published pass; the wait-for-price rating and ~$318 alert as published in the opener. Gates this week: refunding statement Wed Aug 5; Constellation Thu Aug 6 pre-open; Vistra + July jobs Fri Aug 7. ↩

Methodology note: capacity, allocation, and yield figures for CXMT are third-party estimates, labeled as such — the company discloses selectively and the trackers disagree at the margin; we treat the slope as consensus and the point-estimates as approximate. Tripwires are pre-committed and third-party-observable. Nothing here is a price target. Single-source claims are attributed inline.

Nothing on this page is investment advice. We work in scenarios and trip-wires, not predictions — everything above is for thought and process, not for trading. See disclaimer.

Disclosure: I/we have beneficial long positions in a memory-sector ETF (ticker DRAM), ASML, and a diversified quantum-computing ETF (QTUM) through stock ownership. I/we have no position in Micron, CXMT (which, as a Shanghai STAR Market listing, is in any case not directly accessible to ordinary U.S. retail accounts), Samsung, SK hynix, Alphabet, AMD, or Palantir. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it. I have no business relationship with any company whose stock is mentioned. This commentary is for informational and educational purposes only and does not constitute investment, tax, or legal advice or a solicitation to buy or sell any security. Past performance is not indicative of future results. Readers should conduct their own research and consult a qualified financial professional before making investment decisions.