News Analysis

CXMT Threw Huawei's Engineers Out of Its Cleanroom

SemiconductorsMemory ChipsHuaweiFact Check
CXMT Threw Huawei's Engineers Out of Its Cleanroom

On July 24, 2026, Reuters published an exclusive: back in June, a team of engineers performing equipment maintenance inside the cleanroom of CXMT's core R&D campus in Hefei was abruptly told to "pack up your tools and leave the floor immediately," and was never allowed back in. Those engineers worked for SiCarrier — the domestic semiconductor equipment maker widely regarded as the one closest to Huawei. Reuters gave a simple cause: CXMT had been raising memory prices on Huawei for months, Huawei asked for relief, and CXMT refused to move.

Three days later, on July 27, ChangXin Technology (688825.SH) listed on the STAR Market, closed its first day up 465.82%, and reached a market capitalization of 3.28 trillion yuan — the most valuable company on China's A-share market.

Put side by side, the two events form a picture that is hard to ignore: the memory maker that spent years living on state capital, held up as the model of "domestic substitution," is now too expensive even for Huawei. But before running with the sentiment, let's take the story itself apart and ask how much of it is actually established.

1. The bottom line: how much of this can you trust

The verdict in one sentence

The core account rests on a single original source — the Reuters exclusive of July 24, 2026, based on two anonymous people familiar with the matter. CXMT, Huawei and SiCarrier all declined to comment, neither confirming nor denying. Every subsequent report, Chinese or English, without exception recycles that one story. This is "an exclusive from a credible outlet," not "a cross-verified fact" — there is a real difference between the two, so read it with that discount applied.

We assessed each element of the circulating version separately:

Claim-by-claim fact check

ClaimVerdict
Location was the Hefei core R&D cleanroom✓ Matches original
Those removed were SiCarrier's on-site service engineers✓ Matches original
No warning, immediate exit, no re-entry afterwards✓ Matches original
Trigger was a breakdown in memory price talks✓ As originally attributed
It happened in June 2026✓ Not 2025
Independent second source corroborates it✗ None; all recycled
Confirmed by the parties involved✗ All three silent
"Huawei staff caught scraping raw data"⚠ Later embellishment
Method: the Reuters piece of 2026-07-24 and its licensed reprints, compared sentence by sentence against the versions circulating in Chinese-language communities.

That last row deserves its own paragraph. From July 26, an "enhanced" version began circulating in Chinese-language communities (V2EX, X, various forums): that CXMT expelled the on-site team because it had caught Huawei employees scraping raw data. That claim does not appear anywhere in the Reuters piece. It originated as speculation in a Taiwanese outlet that stitched the Reuters story together with an earlier Financial Times report (on Huawei and SiCarrier "sharing staff and technology"); two hops later it had hardened into a statement of fact. Along similar lines, some Chinese outlets rendered SiCarrier's Chinese name as that of S2C — an entirely unrelated company that makes EDA prototyping tools.

Both distortions make the same point: with exclusives about state-backed national champions, every hop from the original English copy to the Chinese trending topics grows a new "detail."

2. Replaying the year: how we got here

From the price cycle to the showdown

From September 2025

AI servers and HBM crowd out capacity; DRAM spot and contract prices take off together, and the industry enters a memory supercycle.

December 2025

ChangXin Technology files for a STAR Market listing, then discloses its first full-year profit, for 2025.

Q1 2026

Conventional DRAM contract prices jump 93%–98% in a single quarter, the largest quarterly increase on record. CXMT posts quarterly revenue of 50.8 billion yuan, up more than sevenfold year over year.

April 21, 2026

At a State Council Information Office briefing, MIIT publicly addresses memory-driven phone price increases for the first time, pledging to "guide firms to tighten channel management and crack down on hoarding."

June 2026

The showdown. Without warning, CXMT ejects SiCarrier's resident engineers from its Hefei site and bars them from re-entering the R&D area. SiCarrier executives read this as a direct result of the CXMT–Huawei standoff.

July 24, 2026

Reuters breaks the story; all three parties decline to comment. The "data theft" variant appears in Chinese communities the next day.

July 27, 2026

ChangXin Technology closes its debut up 465.82% at a market cap of 3.28 trillion yuan, the highest on the A-share market, on 141.2 billion yuan of turnover.

Laid out on a timeline, the "June showdown" stops looking like an isolated outburst and starts looking like the inevitable friction sandwiched between an epic price surge and an epic IPO. To understand it, you first need the scale of those two things.

3. Background 1: just how big is this memory price surge

If your last phone upgrade felt expensive, that wasn't your imagination. This is the sharpest memory price surge since continuous data began in 2016.

Conventional DRAM contract prices: quarter-over-quarter change

Q1 2026 (actual)+93% ~ 98%
Q2 2026+58% ~ 63%
Q3 2026 (forecast)+13% ~ 18%
Data: TrendForce. Note that these are quarter-over-quarter figures — compounded across three quarters, prices are now more than triple what they were. Q3 slows not because supply improved, but because downstream buyers can no longer absorb it.

The official numbers are just as stark. Data from the NDRC's Price Monitoring Center shows that in January 2026 the average contract price of a DDR4 8Gb chip hit $11.50, roughly 83% above September 2025; NAND flash rose close to 150% over the same period. The NDRC's own phrasing: "explosive demand growth, cliff-edge capacity shortage."

The cause is not complicated, and fits in one line: AI ate the memory capacity. Samsung, SK hynix and Micron shifted large blocks of their most advanced capacity to HBM (the high-bandwidth memory used on AI accelerators), and HBM consumes roughly three times the wafer area of ordinary DRAM per unit of capacity. Layer on the three majors' 2025 decision to collectively wind down DDR4, and supply of ordinary consumer memory got squeezed from both ends.

Here is how that reaches consumers:

Selected 2026 handsets repriced because of memory costs

OnePlus 15

+1,100 yuan

Whole line repriced; top spec 5,399 → 6,499 yuan

iQOO 15

+1,500 yuan

Cumulative rise on top spec, 5,499 → 6,999 yuan

OPPO Find N6

+1,000 yuan

Foldable entry price raised

Honor Magic V6

+1,000 yuan

Foldable starting price raised

Xiaomi has said publicly that for the "12GB+512GB" configuration alone, memory bill-of-materials cost rose by about 1,500 yuan. Counterpoint estimates that on entry-level phones (wholesale price under $200), memory now accounts for 43% of total BOM cost.

Downstream buyers are finally pushing back. According to Jiemian News and others, several of China's top six handset makers have flatly rejected suppliers' Q3 price increases, and OPPO and vivo formally turned down Samsung's Q3 quotes. That also explains why the Q3 increase suddenly narrows to 13%–18%.

4. Background 2: CXMT from cash-burning marathon to 3.28 trillion yuan

To see why CXMT feels able to push back against Huawei, look at what happened to it over the past two years.

ChangXin Technology net profit: from years of deep losses to windfall

2023 −16.34 billion
2024 −7.15 billion
2025 +1.88 billion (first profit)
H1 2026 +50–57 billion
Data: the IPO prospectus and the H1 2026 earnings pre-announcement. Figures in yuan; bar lengths are proportional to absolute value, red for losses and green for profits. In eighteen months CXMT went from losing tens of billions a year to earning more than 50 billion in half a year.

Revenue tells the same story: 61.8 billion yuan for all of 2025, versus 50.8 billion yuan in Q1 2026 alone, up 719% year over year; H1 revenue is guided at 110–120 billion yuan. Global market share has likewise moved from the fringe to the edge of the top tier.

Global DRAM revenue share, Q1 2026

Samsung 40.5% SK hynix 29.6% Micron 19.9% CXMT 7.7% — fourth globally, first in China Nanya, Winbond and other remaining suppliers
Data: CFM 闪存市场 (CFM Flash Market), Q1 2026. Counterpoint puts the figure at 8% and notes that CXMT's share doubled quarter over quarter, with revenue up more than 700% year over year. Methodologies differ between firms; don't mix the numbers.

More important is a fact that is easy to miss: CXMT is not running a price war. In Q1 2026, the gap between CXMT's average DRAM selling price and those of Samsung, SK hynix and Micron was only 5%–10%; several module makers said at Computex 2026 that CXMT's DDR5 dies cost them barely less than the big three's. Reuters went further, quoting industry figures who said CXMT and YMTC quotes were in some cases already above Samsung's and SK hynix's.

Meanwhile CXMT signed a five-year supply agreement with ByteDance worth over $7 billion, plus a multi-billion-dollar deal with Tencent — all of it for AI-related memory. When your capacity is locked up by AI orders from internet giants and prices are still climbing, a legacy customer demanding a discount is no longer someone you have to accommodate.

5. Background 3: who SiCarrier is, and why "thrown out of the cleanroom" is such a heavy blow

SiCarrier was founded in Shenzhen in 2021 and is wholly owned by Shenzhen Major Industry Investment Group, which traces up to the Shenzhen SASAC — legally, it is a Shenzhen municipal state-owned enterprise, not a Huawei subsidiary. Huawei has repeatedly denied any affiliation in public.

The industry does not see it that way, and the reasons are concrete. Reports say SiCarrier's technical core came out of Huawei's Starlight Engineering Lab, set up in 2012, with thousands of engineers spun out of the Huawei system into a standalone entity in 2022; the U.S. Commerce Department lists "Huawei Starlight Engineering Department" outright as one of SiCarrier's aliases on the Entity List, where it was added in December 2024. At SEMICON China in March 2025, SiCarrier unveiled 31 semiconductor tools already in volume production at once — etch, thin-film deposition and metrology covering nearly the whole flow, with five flagship platforms named after Chinese mountains (Wuyi, Changbai, Putuo, Ali, Emei) — everything except a lithography machine. By September 2025 it said its order book had passed 10 billion yuan, covering 80% of China's mainstream fabs.

Why an eviction from the cleanroom is a drastic move in this industry

Before a domestic tool can enter a production line, it has to survive a long stretch of on-site qualification: the vendor's engineers live in the customer's cleanroom for months, tuning parameters, repairing the tool and tracking yield data alongside the line. That is the only path to a mature tool — without real production data from a customer's line, equipment never gets past lab grade. So for an equipment maker, which customers' R&D cleanrooms you can get into is itself the most valuable asset you own. Throwing the engineers out and barring their return unilaterally severs that path to product maturity. It is far more serious than pausing an order.

Put differently: if the Reuters account holds up, CXMT did not play a routine card in a price negotiation — it applied pressure at exactly the point where it hurts most.

6. Why a clash was inevitable: three structural conflicts

Reading this as merely "talks broke down" isn't enough. Three tensions in the Huawei–CXMT relationship all came due in the same year.

① Bargaining power flipped completely

CXMT used to be the side that needed support, with Huawei a scarce anchor customer. Now CXMT's fabs are full and orders are queued, and Huawei has gone from benefactor to one buyer among many.

② The "Huawei orbit" became a liability

CXMT's customers include international names such as Apple, and it now faces overseas supply-chain scrutiny on top of its filing. Being tightly bound to an Entity-Listed, Huawei-linked toolmaker keeps amplifying its own sanctions risk.

③ The IPO changed who it owes duties to

After July 27, CXMT answers to 9.42 million subscription accounts and the public market. Under disclosure rules, a standing discount for one customer is a related-party transaction that needs explaining, not industrial solidarity.

1. CXMT is no longer the company that needed Huawei's help

One number in the prospectus says it all: the average selling price of CXMT's main DRAM products rose 55.08% year over year in 2024 and 33.69% in 2025. Both increases are far below the moves in spot and contract prices over the same period — meaning that through the most frenzied stretch of the surge, CXMT was in fact holding prices down for its established customers. When that discount was withdrawn in 2026 under earnings pressure and shareholder expectations, the customer that felt it hardest was inevitably the one buying the most volume and negotiating the hardest.

2. Being "Huawei-adjacent" is a liability on CXMT's customer list

CXMT is already on the U.S. Defense Department's list of "Chinese military companies," but not yet on the Commerce Department's Entity List. That delicate position means it still has something to lose. International customers including Apple have reportedly sought assurances about whether CXMT might be added to the Entity List. In that situation, keeping Entity-Listed SiCarrier engineers permanently stationed inside your most sensitive R&D site becomes an increasingly hard arrangement to explain to reviewers.

3. The IPO swapped out the people CXMT owes duties to

The timing here is exceptionally tight. The showdown in June; the listing on July 27. The last thing a company on the eve of an IPO needs is a documented history of supplying a related party below market price. Whatever the immediate trigger for the price increases, pulling Huawei's pricing back up to market levels was, from a compliance standpoint, something that had to be done before listing.

7. Knock-on effects: who pays for this surge

This conflict matters because it marks a precise turning point for China's domestic supply chain: in the second half of "domestic substitution," the players no longer face outward together; they start redistributing profit among themselves along market rules.

Three groups take the hit:

  • AI compute buyers. SemiAnalysis estimates that memory accounted for about 30% of hyperscaler capital expenditure in 2026, versus roughly 8% in 2023–2024. Chinese cloud providers have raised AI inference instance prices in several rounds, some by more than 30%. For memory-hungry AI systems such as Huawei's Ascend line, DRAM cost lands straight on gross margin.
  • Handset makers. Memory's share of a phone's bill of materials has climbed from 10%–15% to above 20%, approaching 30% on mid- and low-end models. This is more than a price increase — realme was told by OPPO to pull back from the domestic market, and the industry widely reads memory costs as the last straw.
  • Consumers. The NDRC's Price Monitoring Center expects end-product electronics to rise by 300 to 1,500 yuan. IDC projects that Chinese flagship phone prices will rise more than 30% in 2026.

Worth stressing: regulators have so far done very little in practice. On April 21, 2026, MIIT spokesman Xie Cun addressed memory prices at a State Council Information Office briefing, framing the response as "strengthening supply capacity" and "working with relevant departments to lawfully crack down on hoarding." That was a verbal statement — not a regulatory summons, not a formal case, not price intervention. At the NDRC level there is only a price monitoring analysis. The widely shared claim that "SAMR has made major progress in its antitrust investigation of Samsung, SK hynix and Micron" is a 2018 story that has been heavily recirculated lately; please don't treat it as this year's news.

8. What happens next

Based on what is public today, here are three scenarios and the indicators to watch for each.

Scenario A: quiet de-escalation

Most likely

Neither side ever admits anything happened. Business continues (Reuters explicitly notes the two still trade with each other), while the on-site arrangement is quietly converted to remote support or third-party servicing. As Q3 increases narrow to 13%–18% and supply and demand rebalance in early 2027, the pricing dispute dissolves on its own.

Watch for: whether SiCarrier shows up on the tool bid list for CXMT's next capacity expansion.

Scenario B: Huawei accelerates its move away from CXMT

The Huawei orbit already contains memory plays of its own — SwaySure and PengXinWei, both backed by Shenzhen Major Industry Investment Group and both on the Entity List. If this friction is read as a strategic warning, Huawei may pour more into a second supply system, offsetting its weak bargaining position with capacity built by its own people. The price of that path is duplicated investment, and it does nothing for the near term: standing up a DRAM line capable of volume production takes years.

Watch for: capacity and funding announcements from SwaySure and PengXinWei; any public increase in Huawei's imported memory purchases.

Scenario C: the state steps in to mediate

Both are strategically important, state-backed firms, and a prolonged standoff serves neither side's superiors. But judging by MIIT's April remarks, the ministry currently prefers "more supply" to "administered prices," so direct intervention in two companies' commercial contracts looks unlikely. A more realistic form would be coordination at the industry-association level, or supply-guarantee conditions attached to approvals for CXMT's capacity expansion.

Watch for: the emergence of long-term memory contract pricing mechanisms, a national strategic reserve, or a designated supply-assurance list.

One longer-term thread is also worth watching: none of CXMT's 29.5 billion yuan of IPO-funded projects is dedicated to HBM. For a company worth 3.28 trillion yuan, that is a glaring gap. It means CXMT's current windfall rests on the cyclical dividend in ordinary DRAM rather than on the main battlefield of AI memory — HBM remains firmly held by SK hynix and Samsung, while CXMT's HBM3 is still pushing through qualification, with third parties estimating 8-hi yield at around 25% (against 85%–90% internationally; that figure is modeled, not officially disclosed).

Which means CXMT's toughness toward Huawei has an expiry date. Cycles run both ways: TrendForce thinks DRAM supply and demand cannot begin to balance before Q1 2027 at the earliest, while Morgan Stanley has hinted that Q4 2026 could already be this cycle's price peak. When contract prices go from +93% quarter over quarter to flat or falling, CXMT will be back to needing customers, needing domestic equipment partners, needing everyone to pull together. And the engineers shown the door in June will remember.

9. Sources and verification notes

The central fact in this piece — the eviction — rests on a single original source. We looked hard for independent corroboration and found none. All industry data cited comes from verifiable institutional or official publications, listed below.

Numbers to treat with caution: the 25% HBM3 yield for CXMT is a SemiAnalysis model estimate, not an official disclosure; viral claims such as "DDR4 up 1,800% in a year" and "memory up nearly tenfold" lack any underlying quoted prices and are not used here; estimates of Micron's Q1 revenue vary widely between research firms, so all share figures in this piece use CFM's methodology, as noted.

Share: