Chinese Memory: a Sovereign AI Long Thesis on CXMT

The Trade

Corbanu entered the CXMT (CXMT; Shanghai 688825) perpetual long through Hyperliquid at a venue-native mid of $8.3403 at 20:09:40 UTC on 4 September 2026, sized at 3.5% of capital in the tracker portfolio ahead of the company's first public results briefing on 7 September. We expect investor excitement around the memory build-out and the first management commentary on HBM3E to build into that meeting. That same afternoon Huawei launches the Mate XT2, Xiaomi puts CXMT (CXMT)'s LPDDR6 on stage in the 18 Fold that evening, and the reported Trump–Xi summit on 24 September will, in our view, center on national collaboration and competition in AI, with Xi arriving on the back of a domestic memory champion. The 3 September daily model forecast an 81.1% annualized funding receipt for the long over the next seven days, equivalent to 1.56% of notional over the forecast week. Our carry signal is that forward model output, not any single intraday observation. We have the exposure into the briefing and will reassess immediately afterward.

Fundamental exit. We will close the position after the 7 September briefing unless management explicitly confirms that at least 10% of current or planned DRAM wafer input is allocated to HBM and provides a commercialization or qualification timeline. An allocation below 10%, vague roadmap language, silence, or denial will trigger the exit. Price stops will have no role.

Valuation. At the ¥54.80 Shanghai close on 4 September, CXMT (CXMT) trades at 13.6x T+12m earnings and approximately 7.3x T+12m EV/sales, versus 4–6x and 1.6–4.3x for SK hynix (SKHX), Micron (MU) and Samsung (SMSN) (Table 4 retains the complete 2 September snapshot). Consensus revenue growth slows from 485% in FY26 to 63% in FY27 after 1H26 printed +874%, while Micron (MU) and the incumbents move customers onto five-year contracts. That deceleration leaves room for HBM3E estimate revisions. From the ¥54.80 close, a fast catalyst move to Mizuho's ¥70 or Capital Securities' ¥78 target would represent 28–42% upside. UBS sits at ¥80, Morgan Stanley at ¥88, Mirae Asset at ¥90, Nomura at ¥116 and Goldman Sachs at ¥129. These targets were published between 24 August and 1 September 2026.

If you put the wider market beside these numbers, you get a second valuation anchor: tech trades at 21x T+12m earnings on the Nasdaq-100 (XYZ100) and 50x on CXMT (CXMT)'s STAR 50 board, while the incumbents carry T+12m FCF yields of 12–27% on enterprise value. CXMT (CXMT)'s parent-attributable T+12m FCF yield is 6.6%, against 4.1% for the Nasdaq-100 (XYZ100) and 2.8% for Hang Seng TECH (Table 4). Funding and incremental high-value memory sales provide the catalyst for returns.

Probability-weighted return

After the 7 September briefing, the failed cases close. We keep the position through Huawei Connect, Alibaba (BABA) Apsara and the 24 September summit once there is confirmation of at least 10% HBM wafer allocation and a commercialization timeline. The combined 65% confirmation probability reflects the 31 August report, which said that CXMT (CXMT) had begun small-scale HBM3E production, supplied samples to Alibaba (BABA)'s T-Head and Cambricon, and targeted commercial use in 2027. Between the 27 July public forecast and 2 September, FY28 consensus remained essentially unchanged: revenue moved from ¥773.3bn to ¥762.2bn and parent net income from ¥393.1bn to ¥393.3bn.

Company Backdrop

CXMT Corp—ChangXin Memory Technologies (CXMT; Shanghai 688825)—is China's only scaled DRAM manufacturer and the world's fourth-largest DRAM maker. Founded in Hefei in 2016, it ships DDR4, DDR5 and LPDDR4X/5X wafers and modules into servers, phones, PCs and cars. Its own prospectus puts its share of the global DRAM market at roughly 7.7%, behind Samsung (SMSN), SK hynix (SKHX) and Micron (MU).

CXMT (CXMT) listed on 27 July 2026 at an offer price of CNY 8.66 and raised CNY 57.9 billion (about $8.6 billion), making it the largest Asian IPO of the year. It closed its first session at CNY 49.00, a 466% pop that made it the most valuable company listed in China on day one. The move has held: after reaching CNY 61.80 on 17 August, it closed at CNY 54.80 on 4 September, with a market capitalization of roughly CNY 3.72 trillion and 4.503 billion unrestricted shares available to trade, or 6.63% of shares outstanding, after the greenshoe.

Outside mainland China, the perpetual on Hyperliquid's TradeXYZ market is the cleanest access. Its ticker is CXMT. The Corbanu CXMT page has the live price, peers and fundamentals.

Simplified Financial Model

The reported figures come from CXMT (CXMT)'s Shanghai Stock Exchange filings: the FY2025 annual report and the unaudited 2026 interim report. Sell-side projections use consensus as of 2 September 2026. Appendix A contains the five complete financial and peer-comparison tables.

The business moved into profit in FY25. In 1H26, gross margin rose from 13% to 85% over twelve months, driven by price and mix. Consensus carries that run-rate forward: FY26E implies 211bn of second-half revenue, up 40% sequentially, at a 51% net margin. FY27–28 then add volume at roughly flat margins.

Consensus's +133% EPS growth from trailing to FY26E breaks down into +84% from revenue, rising from 196.7bn to 361.4bn, and +27% from net margin, rising from 41.6% to 52.7%. The 52.7% forecast carries the 1H26 margin of 51.6% across the full year. Beyond FY26, projected EPS growth comes almost entirely from revenue, with net margin at 51–53%, gross margin at 84–86% and operating margin at 74–76%. HBM3E can lift revenue estimates without requiring further margin expansion.

Peers. The comparison set consists of the three incumbents—Samsung (SMSN), SK hynix (SKHX) and Micron (MU)—which hold the other 90% of DRAM revenue. CXMT (CXMT) ranks fourth, and it is the only one of the four with a growing share.

Valuation. All figures below are in US dollars at spot, and enterprise value is constructed on the same basis for all four names. CXMT (CXMT) requires two qualifications: it has no trading history to compare against, and minority partners in its fab subsidiaries account for roughly 30% of consolidated earnings and cash flow. The attributable rows therefore provide the relevant valuation basis for 688825 holders.

Catalyst Calendar

From 7 September through January, CXMT (CXMT) repeatedly comes back into focus around management's first public Q&A, Xiaomi's flagship launch, Huawei's HBM roadmap, Alibaba (BABA)'s accelerator conference, the reported Xi visit, peer earnings, index decisions and the first float expansion.

HBM3E Catalyst

What it is. High-bandwidth memory takes DRAM and builds it vertically: eight or twelve DRAM dies are stacked on a logic die and connected through through-silicon vias. The assembly is mounted on a silicon interposer beside the processor, with the board-level alternative placing memory across the motherboard. HBM3E is the current shipping generation, providing 36 GB and roughly 1.2 TB/s per stack—ten to twenty times the bandwidth of a DDR5 module. Eight stacks sit around every NVIDIA (NVDA) B200 and feed the accelerator.

Why it matters. The shortage exists because HBM is absorbing wafer supply that would otherwise go to DDR. Sumit Sadana of Micron (MU) said in August 2026 that producing 100 bits of HBM3E requires the industry to give up about 300 bits of DDR wafer supply, with the ratio moving closer to 4:1 by HBM4E. Customers report that "the #1 constraint they have is DRAM", while Micron (MU) cannot fill more than half of data-center demand. HBM wafer input is 18% of DRAM wafers at end-2025, 22% at end-2026 and 30% at end-2027 (TrendForce). SK hynix (SKHX) said on its July call that the balance is unlikely "to improve meaningfully in the near term" because of process complexity and fab lead times. Device makers are carrying the bill. Tim Cook said on 30 July that Apple (AAPL) is "in a 100-year flood on the memory pricing", has "reluctantly raised prices", and is guiding gross margin from 49.3% to 46.5% over two quarters, with more than all of the change explained by memory. He added, "primarily the DRAM market has three suppliers. Obviously if there were more suppliers, that would be good." Every large buyer of memory now wants a fourth supplier, and CXMT (CXMT) is the only candidate.

The scale of the mix shift. CXMT (CXMT)'s current product mix is entirely conventional DRAM: LPDDR for phones accounts for about two-thirds of 2025 revenue, with DDR4 and DDR5 going to PCs and servers. TrendForce sizes that DRAM market at $619 billion in 2026 and $903 billion in 2027, and consensus capitalizes CXMT (CXMT)'s 7.7% share of that market. HBM, which is absent from CXMT (CXMT)'s current sales, represents a $35 billion market in 2025, about $60 billion this year and $100 billion by 2028 on Micron (MU)'s own TAM; Micron (MU) notes that this TAM is larger than the entire DRAM market of 2024. HBM consumes 30% of industry wafer starts by 2027, and its 2027 contract prices reset multiples higher. The HBM market alone exceeds CXMT (CXMT)'s entire FY26 consensus revenue. At SemiAnalysis's 100k wafers-a-month HBM capacity by 2028—roughly 12% of global HBM wafer share—even a yield-handicapped 5–10% of the 2028 HBM dollar TAM would produce $5–10 billion of revenue that current forward numbers omit.

The estimates continue to move higher. Over the past year, blended next-twelve-month consensus revenue for the four incumbents rose from $366 billion to $1,315 billion, while net income rose from $63 billion to $709 billion. The CXMT (CXMT) listing adds a further $76 billion of revenue and $40 billion of earnings. CXMT (CXMT) is entering a market where estimates are rising and pricing power remains intact: it sells DDR5 and LPDDR5X at prices at or above those of the Big Three, and its LPDDR5X talks with Apple (AAPL) reportedly ended in August after CXMT (CXMT) held its price.

The Chinese inference bid. Domestic model usage creates structural Chinese demand for memory. On OpenRouter's leaderboard this week, 13 of the 20 most-used models are Chinese and they process 72% of the top-20 token volume (62 of 87 trillion tokens). GLM 5.3 Flash is tied for first at 11.4 trillion tokens, up more than tenfold week on week, while Z.ai (ZHIPU) says it runs entirely on Chinese AI chips. Every token served reads the model weights and a growing KV cache out of HBM and DRAM, so rising token volumes on domestic accelerators create a direct bid for domestic memory. Z.ai (ZHIPU) itself (2513 HK, listed 8 January 2026, HK$520 billion market cap) carries consensus revenue that triples in 2027 and more than doubles again in 2028; it trades at 28x forward sales, so the market is capitalizing Chinese inference at multiples that assume the memory to serve it exists.

The September set-up. Management answers public questions on 7 September covering HBM3E yields, customers and timing. Ten days later at Connect, Huawei lays out the Ascend roadmap, and the 950 series ships with in-house packaged HBM from an undisclosed DRAM-die supplier. Five days later Alibaba (BABA) holds Apsara; Alibaba (BABA) is a CXMT (CXMT) shareholder, and its T-Head unit is sampling the part. Xi visits Washington on 24 September with AI on the agenda and a domestic memory champion behind him. We expect HBM3E to dominate the briefing; a capacity disclosure followed by a customer name at Connect or Apsara can move a company with a 6.63% unrestricted tradable float into the summit.

The Chinese independent hardware story. The 7 September consumer launch gives a direct read on where CXMT (CXMT) stands: in selected conventional DRAM products it has reached parity. Xiaomi's 18 Fold pairs its own 3 nm XRING O3 with 16 GB of CXMT (CXMT) LPDDR6. That standard is scheduled to reach Samsung (SMSN) and SK hynix (SKHX) alongside Qualcomm (QCOM)'s next Snapdragon, and CCTV Finance has already held up the combination as a template for Chinese semiconductor collaboration. Two days later Apple (AAPL) launches its own foldable with 12 GB of incumbent LPDDR5X stacked on-package while publicly complaining about DRAM cost. Huawei launches the Mate XT2 on the same afternoon as Xiaomi, and it tops out at 16 GB from an undisclosed supplier. Across three foldables launched in three days, the product evidence gives the fourth-supplier argument three price tags. CXMT (CXMT)'s HBM handicap is in stacking, TSVs and packaging yield; the underlying DRAM cell is further ahead.

How CXMT (CXMT) caught up. Over three decades, state capital built the foundation. The National IC Fund II owns 7.7% of CXMT (CXMT), while Hefei's municipal vehicles and Anhui's provincial investment arm own a further 45.7%. State-linked holders therefore account for 53% of the shares, against an unrestricted tradable float of 6.63%. CXMT (CXMT) was founded in 2016 with neither a fab nor a research base, yet reached DDR5 production at volume in under a decade. YMTC followed a similar route in NAND and now ranks third globally. The route has also involved allegations concerning Samsung (SMSN) process technology. On 11 August 2026, a Seoul court heard sworn testimony from a former Samsung (SMSN) engineer, sentenced in April to seven years, that he sold CXMT (CXMT) Samsung (SMSN)'s roughly 600-step process recipe for about $2 million. Samsung (SMSN) says the recipe took five years and more than $1 billion to develop. The US prosecution of Fujian Jinhua lasted five years before an acquittal on every count. CXMT (CXMT) appears on the Pentagon's 1260H list, is challenging that designation in court, and remains absent from the Commerce Entity List.

The absence of EUV is what makes this remarkable. The incumbents are building HBM3E on 1a/1b-class nodes patterned with ASML EUV. ASML's July call links rising DRAM lithography intensity to customers replacing multi-patterning with single-exposure EUV, while buying memory tools at a rate up 75% this year. China cannot buy an EUV tool, and ASML's China revenue is 20% of sales, "mainly related to an increased demand in mainstream logic". CXMT (CXMT) is building its stacks on a 16 nm-class DUV multi-patterned node: its dies are about 40% larger, cost per bit is more than 30% above the Big Three, and estimated HBM3 8-high yields are around 25% (SemiAnalysis). HBM3E risk production on that toolset therefore requires brute force and leaves CXMT (CXMT)'s cost curve as the ceiling.

The structure of HBM is different from the rest of the memory business. Commodity DRAM is fungible, sold by the gigabyte, and a new fab expresses itself through price. HBM is qualified for a specific accelerator and packaged with it; customers order it 20–26 weeks ahead under multi-year take-or-pay agreements. Capacity therefore cannot simply be switched on, and the customer cannot switch supplier. HBM output consequently governs Chinese AI compute. Huawei's Ascend 910C relies on stockpiled Samsung (SMSN) and SK hynix (SKHX) HBM2E, which analysts expect to run out around mid-2026. The 950 series moves to Huawei-packaged HiBL and HiZQ memory, whose DRAM dies have no disclosed supplier. Alibaba (BABA)'s T-Head and Cambricon are sampling CXMT (CXMT) HBM3E for possible 2027 integration, and there is no other domestic source. CXMT (CXMT)'s HBM3E remains a generation behind and yield-constrained, leaving it as China's only domestic path beyond foreign stockpiles. Late-August risk production establishes the part, while named-customer volume qualification provides the catalyst.

Long Term Policy Support

Beijing has placed AI hardware inside a state investment mandate. The policy runs on three terms: self-reliance (自立自强), autonomous and controllable (自主可控), and the new whole-nation system (新型举国体制). Under that framework, memory is given a name, funding, and a fast track that extends from the Politburo down to the industrial statistics. The texts that follow move from Party doctrine through the Five-Year Plan to the State Council, with the level of specificity increasing at each stage.

What it bought. On 30 December 2025, the SSE accepted CXMT (CXMT)'s filing at ¥29.5 billion. Seven months later, CXMT (CXMT) listed after raising ¥57.9 billion, the largest STAR IPO in history. The "national team" took ¥7.6 billion of the strategic tranche. The book included 36 Social Security Fund and basic-pension portfolios, the SOE Structural Adjustment Fund II, PICC and China Life. Alibaba Cloud (BABA), ZTE, Transsion, TCL, NIO, Chery, Xiaomi and Meituan were downstream strategic holders. CXMT (CXMT)'s own filing says the funds are directed in line with state policy support for ICs and pay for advanced memory R&D and capacity.¹⁵ Before the IPO, Hefei's platforms held 45%; today, state-linked holders own 53%, with the whole-nation system holding the equity directly.

Robotics is the second memory-intensive mandate. Three weeks after CXMT (CXMT), Unitree listed on the same board as the first humanoid stock and finished 460% higher. The Plan designates embodied intelligence a future industry, while NBS put 2025 industrial and service robot output up 44% and 80%, respectively.¹⁶ In June 2026, Micron (MU)'s Sanjay Mehrotra said a humanoid carries roughly ten times the memory of an L2+ car, that automotive and industrial grades price above mobile DRAM, and that this demand drives a decades-long cycle beginning around 2030. NVIDIA (NVDA)'s Jetson Thor humanoid board already carries 128 GB of LPDDR5X, against 12–16 GB in a flagship phone.¹⁷ CXMT (CXMT)'s ¥9 billion of proceeds earmarked for HBM and next-generation LPDDR5X are being read domestically as aimed at exactly that market, and NIO, Chery and Xiaomi are in its strategic book. The 17,000 humanoids shipped globally in 2025 carried about 2 PB of DRAM, against industry shipments of tens of exabytes. Robotics adds a long-dated option on 2030 volumes beyond the FY27–28 consensus window.

Why the premium can hold. Every other name in the DRAM peer set is cyclical, and the market pays 4–6x forward earnings for these businesses at the top of the cycle. CXMT (CXMT) carries a different status as a state-designated strategic asset, with only a 6.63% unrestricted tradable float, while its board trades at 50x forward earnings. Within a month of CXMT (CXMT)'s listing, a robotics IPO was priced at hundreds of times earnings. Our numbers in Table 4 place CXMT (CXMT) at 13.6x T+12m earnings: above SK hynix (SKHX), Micron (MU) and Samsung (SMSN); below STAR 50, Hang Seng TECH and the Nasdaq-100 (XYZ100); and at a fraction of Unitree. The policy record makes the domestic multiple the relevant anchor, while the state's demonstrated willingness to use extraordinary measures to keep the asset funded through the cycle reinforces that valuation.

Positioning

Why there is carry. The company's post-greenshoe filing shows 4.503 billion unrestricted shares, equal to 6.63% of shares outstanding. For 688825, the Hyperliquid perp is the only available hedge: there is no tokenized spot, ADR, options market, or usable securities lending. Institutional IPO allottees hold a further 1.521 billion shares at ¥8.66, locked until 27 January 2027; once those shares unlock, the current unrestricted tradable pool will rise by roughly 34%. That creates an unusual funding arrangement relative to every other TradeXYZ name: retail longs pay funding elsewhere, while here locked holders short the perp and the shorts pay.

Where it accrues. Since listing, the perp has paid longs 20.9% in funding, with 18.8 points accruing during Shanghai cash hours (01:00–08:00 UTC), equivalent to about +620% annualized for the long. Funding during the US session was flat, and the timing is consistent with hedging against the live cash market.

The historical basis trade through 2 September. Over the first 27 sessions, the long perp / short stock position returned 24.87%. Funding supplied 20.93 points of that result, while 3.94 points came from the perp discount moving from −4.1% to flat. Weekly funding reached 8.4% of notional at its peak and produced 0.7% in the seven days ending 2 September. The hedge on locked holders stays in place until January.

Expected carry. The 3 September daily model forecast a funding receipt to the long of 81.1% annualized over the next seven days, equivalent to 1.56% of notional over the forecast week. The forecast is produced by the validated seven-day estimator described in the appendix. Intraday observations feed the next scheduled model run; they do not independently constitute a reversal signal. Carry is forecast rather than locked, and the scenario table keeps it separate because the hold cases extend beyond the model horizon.

Risk Management

Realized volatility. The stock has run at 65% annualized since its second session across 28 trading days, with a maximum drawdown of 12% from the 17 August high. The perp also trades through the 17 hours each day that Shanghai is closed; it has run at 90% on daily closes and 78% over the last 30 days. The memory peers give you the longer regime history on public exchange data: Micron (MU) ran at 80% volatility, SK hynix (SKHX) at 90%, and SanDisk (SNDK) at 115% over the trailing twelve months. For position size, I would rely on those histories and treat CXMT (CXMT)'s first five weeks as a limited sample.

Realistic max drawdown. In this cycle, SK hynix (SKHX) fell 54% in 38 days from its June high, SanDisk (SNDK) fell 56% in 34 days and Samsung (SMSN) fell 41% in 42 days, a month before CXMT (CXMT) listed; the STAR-board post-IPO analogs produced comparable declines on their own board, with SMIC −41% in 55 days after its 2020 debut and Cambricon −63% over ten months. Across the peers' full histories, the worst 5% of three-month windows show 30–60% peak-to-trough drawdowns, and the median window shows 13–27%. That evidence leads us to plan for a 40–55% drawdown in the stock over the catalyst window as the realistic bad case: a loss of 1.4–1.9% of capital on the 3.5% leg, with a median-path loss nearer 0.7%. No funding credit is included in this drawdown budget.

Beta. We entered a short Cerebras (CBRS) leg, written up separately, against the 3.5% long CXMT (CXMT), with the sizing set to neutralize the pair's beta to the Nasdaq-100 (XYZ100). We measure that beta from Hyperliquid daily candles against the XYZ100 perp. CXMT (CXMT) runs at 1.6–1.7 beta to XYZ100 over 38 sessions, with 0.40 correlation; Cerebras (CBRS) runs at 2.75 over its full 126-session history and 3.8–3.9 over the last 60 sessions. Those readings put the neutral short at 1.5–2.0% of capital, and we are carrying 1.75% short CBRS against 3.5% long CXMT (CXMT), net beta about zero. Since the legs are 0.48 correlated, pair daily P&L volatility is 0.15% of capital, no higher than the CXMT (CXMT) leg alone. The adverse case we accept is a CXMT (CXMT)-specific break with a Cerebras (CBRS) squeeze: CXMT (CXMT) at −50% and Cerebras (CBRS) at +50% costs 2.6% of capital.

Resulting factor exposure. Once market beta is stripped out, the exposures are quite deliberate. The stock is up 6x from its offer price, so every momentum screen puts it at the top; when it breaks, those same screens will sell it. The memory-factor exposure comes through CXMT (CXMT)’s 0.48 correlation with Micron (MU) on the perp; the Cerebras (CBRS) short, itself 0.64 correlated with Micron (MU), hedges part of that exposure. The China exposure is policy and currency exposure through an A-share on the STAR board, paid in CNY, rather than the offshore China-tech factor: CXMT (CXMT) shows no loading there, with a correlation to BABA of −0.01 over the same 38 sessions. The short leg adds exposure against the US AI-compute IPO premium. In aggregate, the pair is approximately beta-neutral, long momentum, long memory, long onshore China, and short US AI-hardware froth; the long leg pays funding to us.

Investor Appetite

To see what drives a low-float name into an unlock, look first at the retail turnover that can be directed into a single instrument. The benchmark for that capacity is the iShares Bitcoin Trust: IBIT is the largest retail flow vehicle of the last cycle and trades $2–3 billion a day. Memory is now trading at that scale and above it in places. The Roundhill Memory ETF (DRAM) launched 2 April with CXMT (CXMT) at about 5% of the book via swap; July turnover reached $4.6 billion a day, 3.4 times IBIT, and August still ran at 1.1x IBIT. The 2x Micron ETF (MUU) went from $32 million a day in 2025 to $1.2 billion a day this year, while Micron (MU) itself went from $3.4 billion to $30 billion a day. Over the same period, IBIT fell from $2.9 billion to $2.2 billion. Memory is therefore drawing bitcoin-ETF-scale retail turnover against CXMT (CXMT)'s 6.63% unrestricted tradable float, with the January unlock giving marginal buyers a deadline.

Appendix A: Financial and Peer Comparisons

The figures and consensus run to 2 September 2026. Table 1 derives EPS from parent-attributable net income and 67.9 billion shares. Table 2 carries forward the published consensus EPS series and reports the net income and margin implied by it. The cash-flow tables retain the modeled bridge alongside the sell-side FCF estimate, with the methodology set out in the footnotes.

Appendix: forward-funding model

The model forecasts the long’s annualized funding cashflow over the next 24 hours and the next 7 days. At ingestion, it converts the exchange sign so that every rate in the article follows one convention: a positive rate means the long receives funding. The inputs are the contract’s realized hourly funding history from Hyperliquid, with trailing means over 24, 72, 168, 336 and 720 hours recomputed at each daily decision point.

Each day the forecast is republished on every Corbanu market page in the "T+7d funding APR" column and fed into the total-return index, which scales prior perp levels by realized hourly funding. Behind that output sit three estimators: a 1-day persistence, a 50/50 blend of the 1-day and 7-day means, and a ridge regression on the 1-day to 30-day means, fitted per contract with walk-forward validation. The selection was made by sweeping every standalone window from 1 to 30 days and 16 coarse-grid blends—1d, 3d, 7d, 14d, 30d, all pairwise 50/50 blends, and their equal-weight average—across 22 contracts with at least 45 common daily observations, scoring the candidates on identical dates and equal-weighting them across contracts. For the 7-day horizon the 7d/30d blend is used; it delivers 80.6% sign accuracy, with flip precision of 66% and recall of 59% for calls in which the sign changes relative to the preceding 24 hours, and a mean absolute error of about 13 percentage points annualized. For the 24-hour horizon the per-contract ridge is used once 45 daily observations exist; before that, the 1d/1w blend is used. CXMT (CXMT), listed on 27 July 2026, remains inside the short-history window for the 24-hour model.

Appendix: sources for the HBM3E section

⁶ Chosun Biz via TrendForce, 24 Dec 2025: Samsung (SMSN) and SK hynix (SKHX) raised HBM3E supply prices by nearly 20% for 2026. ⁷ TrendForce, 29 Oct 2025 and 18 Dec 2025: HBM3E premium over server DDR5 more than 4x in 2Q25, expected to narrow to 1–2x by end-2026. ⁸ Seoul Economic Daily via TrendForce, 1 Sep 2026: 36 GB HBM3E at about $2,100 on the spot market, four to five times LTA; 16-high HBM4 about $3,500. ⁹ TrendForce, 2 Jun 2026: HBM wafer input 18% / 22% / 30% of DRAM wafers at end-2025 / 2026 / 2027; HBM contract prices expected to rise multiples higher in 2027. Conventional DRAM contract price moves: TrendForce quarterly contract-price reports, 1Q26 and 2Q26. Company quotations are from the Apple (AAPL) fiscal 3Q26 earnings call (30 Jul 2026), the Micron (MU) presentation at the KeyBanc Technology Leadership Forum (10 Aug 2026), the SK hynix (SKHX) 2Q26 earnings call (29 Jul 2026) and the ASML 2Q26 earnings call (15 Jul 2026). CXMT (CXMT) HBM yield and node estimates: SemiAnalysis; risk-production report: The Information, 31 Aug 2026; Seoul court testimony: Korean press reports of the 11 Aug 2026 hearing; shareholder and lockup data: CXMT (CXMT) 2026 interim report and IPO result announcement.

Appendix: sources for the policy section

¹⁰ Politburo collective study session on AI, readout via people.cn, 26 Apr 2025. ¹¹ CCP Central Committee recommendations for the 15th Five-Year Plan, Fourth Plenum, 23 Oct 2025. ¹² Outline of the 15th Five-Year Plan, adopted 13 Mar 2026 (Xinhua); NDRC and CPPCC readings of chapters 8 and 12; sell-side plan notes on storage-chip localization (interpretation). ¹³ 2026 Government Work Report, 5 Mar 2026; NBS, Jan 2026 industrial output release. ¹⁴ Xi Jinping, WAIC keynote, news.cn, 17 Jul 2026. ¹⁵ Caijing on the strategic tranche; CXMT (CXMT) strategic-allocation filing and prospectus (SSE), Jul 2026; SSE filing acceptance 30 Dec 2025. ¹⁶ Unitree STAR listing, 19 Aug 2026; NBS 2025 robot output; Galaxy Securities plan note on embodied intelligence. ¹⁷ Micron (MU) FQ3 2026 earnings call, Jun 2026 (Mehrotra on humanoid memory content); NVIDIA (NVDA) Jetson Thor specifications; Unitree prospectus (5,511 units shipped in 2025).

Important Notice

This memorandum sets out research commentary and a model-portfolio decision; it provides no personalized investment advice and constitutes no offer to buy or sell securities or derivatives. CXMT (CXMT) and Cerebras (CBRS) are highly volatile. Perpetuals introduce funding, leverage, liquidity, oracle, counterparty and venue risks. Forecasts and estimates may be wrong.

Corbanu · Updated 4 September 2026 · Entry recorded from the Hyperliquid venue mid at 20:09:40 UTC; other information and estimates are dated as stated in the article.