Beijing Turns Export Controls on Itself, CXMT Debuts at an $85B Valuation, Samsung and SK Ink $950B in Chip Pacts With US Tech
Revisiting the July 19 Brief — CXMT Prices Asia’s Largest 2026 IPO and Opens Trading Today at RMB 579B (~$85B) Valuation, MOFCOM Weighs Barring TSMC From Fabbing Chinese Chip Designs, Samsung and SK Group Ink $950B in Nvidia/Broadcom Supply Pacts, SK Hynix’s Q2 Print Slips to July 29, DeepSeek Preps a STAR Market IPO Filing, Zhipu Ships a Fully Domestic 1GW Data Center
The week of July 20–26 was the forward bridge resolving on schedule, plus a policy shock the July 19 brief did not anticipate. CXMT’s payment window closed July 20 with retail demand at a reported 243.93x oversubscription against softer institutional bidding, and the company opens trading today, July 27, on the STAR Market at an implied valuation of roughly RMB 579 billion (~$85 billion) — Asia’s largest IPO of 2026 and the board’s biggest listing in history. Behind it, the Financial Times reported July 21 that China’s Ministry of Commerce is consulting Alibaba, ByteDance, and Zhipu on export controls that would restrict foreign downloads of Chinese model weights and — more consequentially for the supply chain — bar TSMC and Qualcomm from fabricating advanced chips designed by Huawei, Alibaba, and ByteDance. On the Western side of the ledger, Samsung and SK Group answered with scale: at a July 24–25 San Francisco summit hosted by President Lee Jae-myung, the two companies signed a combined $950 billion in commercial chip-supply commitments with Nvidia and Broadcom — a private-sector deal the Korean government helped broker and showcase, not a state-to-state pact — layering directly on top of the July 19 brief’s ₩800 trillion (~$531 billion) sovereign budget story. SK Hynix’s first earnings print as a dual-listed company slipped from the previously flagged ~July 22 window to July 29, with the stock rallying 14% into the delay on no company-specific news. DeepSeek moved from private fundraising talk to active IPO preparation, targeting a STAR Market filing as early as this year for a 2027 debut. And on China’s domestic AI stack, Zhipu completed a fully domestic-chip 1GW data center and acquired infrastructure team XCore Sigma, while Oriental Computing Core’s DF1000 near-memory chip and Alibaba Cloud’s pay-per-use M890 super node marked the WAIC story’s shift from exhibition to commercial deployment. This brief audits the July 19 forward vectors, integrates this week’s China intelligence and Korea/Taiwan/Japan supply chain layers, and resets the monitoring window through the SK Hynix and CXMT trading debuts, Samsung’s July 30 earnings, and the August cascade.
A structured review of the July 19 brief’s forward risk vectors and how realized market and supply chain outcomes tracked against expectations.
Granular analysis of the week’s most consequential developments across the CXMT debut, the Korea-Washington chip alliance, and China’s new export-control posture.
CXMT Opens Trading Today: 243x Retail Oversubscription, Softer Institutional Demand, an $85B Valuation Test
CXMT’s subscription window closed July 20 with retail demand reported at 243.93 times the shares offered, while institutional bidding came in comparatively soft — a divergence Micron and SK Hynix shares reflected in pullbacks tied to fears that CXMT’s expanding capacity could eventually add supply to a DRAM market currently defined by scarcity. The company’s listing prospectus, published July 23, disclosed Q1 2026 revenue of RMB 50.8 billion (~$7.5 billion, +719% YoY) and net profit attributable to shareholders up 1,688% to RMB 24.76 billion (~$3.6 billion), with H1 2026 revenue guided to RMB 110–120 billion (~$16.2–17.6 billion). Strategic investors confirmed in the offering include Alibaba Cloud Feitian, Meituan, and Xiaomi — notably without Apple, even as the Financial Times reported Apple has begun qualification testing of CXMT’s LPDDR5X chips for China-market devices while separately lobbying Washington to keep CXMT off the Entity List. CXMT begins trading July 27 under ticker 688825 with no price limits for the first five sessions. CXMT 688825.SH
The retail/institutional oversubscription gap is the number to sit with. Chinese retail investors historically treat STAR Market debuts as a lottery-allocation mechanism largely decoupled from valuation discipline, so 244x retail demand tells you little. Softer institutional bidding — from the desks actually pricing DRAM cycles — is a more honest signal, and it lines up with the same investor logic pressuring Micron and SK Hynix this week: every dollar of CXMT capacity that materializes is a dollar of future pricing risk to the scarcity thesis underwriting the entire memory complex. Z-Ben Advisors expects a sharp first-day pop before the stock and the broader STAR 50 (down ~20% from its July 1 peak on subscription-driven cash calls) settle into a new equilibrium. The Apple qualification-testing story is the one to track independently of the listing: a confirmed LPDDR5X design win, even without a formal supply agreement, would be the first Western consumer-electronics validation of CXMT output and a far more durable catalyst than day-one price action.
Korea’s $950B Answer: Samsung-Broadcom and SK-Nvidia Lock In HBM Supply Through 2030
At a July 24–25 San Francisco summit attended by Nvidia’s Jensen Huang, Broadcom’s Hock Tan, and Anthropic’s Dario Amodei, President Lee Jae-myung unveiled combined Samsung and SK Group commitments of $950 billion in AI chip and infrastructure cooperation with US technology firms — structured as forward purchase agreements and MOUs rather than equity or cash transfers. Samsung signed a five-year, $200+ billion MOU with Broadcom covering HBM4/HBM4E memory supply, sub-2nm foundry manufacturing at Pyeongtaek, and 2.3D/2.5D advanced packaging for Broadcom’s next-generation AI accelerators. SK Group separately signed a letter of intent with Nvidia exceeding $500 billion, spanning next-generation HBM co-development and AI factory construction — SK Telecom’s Nvidia DSX-based facility, targeting up to 2 gigawatts of capacity with HBM4-equipped Vera Rubin systems, is slated online in 2027. The combined SK pillar (Nvidia, Microsoft, Anthropic memory supply agreements) totals roughly $750 billion. 005930.KS SKHY NVDA AVGO
Read this against the MOFCOM story below as the week’s real bifurcation trade. Within 72 hours, Washington’s two Korean memory partners locked $950 billion of forward HBM and foundry demand into US accelerator roadmaps through 2030, while Beijing moved to formalize barring its own national champions (Huawei, Alibaba, ByteDance) from using TSMC at all. Both governments are converging on the same instinct — vertically lock the supply chain to the geopolitical bloc — from opposite starting points. For portfolios, the number that matters is the MOU/LOI distinction: these are non-binding forward-purchase commitments, not signed contracts, and the 2030 horizon leaves substantial room for renegotiation if the memory cycle turns before then. The more immediately tradeable signal is Broadcom’s confirmed move to Samsung’s sub-2nm Pyeongtaek line for next-gen AI accelerators — a direct capacity allocation decision, not a five-year aspiration, and one that further validates the Meta/Anthropic-to-Samsung foundry pivot this brief has tracked since early July.
MOFCOM Weighs Barring TSMC From Fabbing Chinese Chip Designs — Beijing’s Export-Control Mirror
The Financial Times reported July 21, independently corroborated by Reuters, that China’s Ministry of Commerce has spent weeks consulting Alibaba, ByteDance, and Zhipu (Z.ai) on a package of controls that would restrict foreign downloads of the training data and model weights behind China’s open-weight AI models — while leaving hosted API access to those models untouched. The more consequential half of the consultation, reported separately, covers proposed restrictions that would prevent TSMC and Qualcomm from manufacturing advanced semiconductors based on designs from Chinese companies including Huawei, Alibaba, and ByteDance. Officials are also said to be weighing whether leaking proprietary Chinese AI technology should be classified a national-security offense, and new limits on who may fund domestic AI startups. Nothing has been finalized; the proposals would be folded into the next revision of China’s export-control catalogue pending industry feedback, and several companies have reportedly warned regulators the rules could slow their own development. Policy Layer TSM
If enacted, the TSMC/Qualcomm fabrication restriction would be the more durable story of the two. Model-weight download limits touch DeepSeek, Moonshot, and Alibaba’s open-weight ecosystem but leave commercial API access intact — a narrowing of possession, not usage, as multiple outlets note. Barring Huawei, Alibaba, and ByteDance designs from TSMC’s leading-edge nodes, by contrast, would force those firms toward SMIC’s less advanced process technology for any chip requiring the most advanced geometries, directly shaping the domestic-substitution timeline the WAIC coverage below documents. It would also guarantee SMIC’s order book independent of price competitiveness — a subsidy by policy rather than performance. Watch for the next catalogue revision draft as the conversion signal from consultation to binding rule.
Korea, Japan, and Taiwan supply chain developments mapped to the USINO.AI coverage universe with US ticker read-throughs on every story.
Korea: The San Francisco Summit Cements the Sovereign-Budget Thesis; SKHY Print Slips to July 29
The $950 billion Samsung/SK-Nvidia/Broadcom package (full analysis above) lands directly on top of the July 19 brief’s ₩800 trillion (~$531 billion) FY2027 sovereign budget story, giving Korea’s AI-chip fiscal thesis a private-sector demand floor to match the public one. On the earnings side, SK Hynix’s first quarterly report as a dual-listed company has been confirmed for July 29 — later than the ~July 22 window this brief’s prior audit anticipated — with the company using the intervening days to file KOSPI disclosures on July 21–22 denying reports it plans to acquire Intel’s Ohio fab, calling current speculation premature while confirming it “routinely evaluates” such opportunities. Consensus per Visible Alpha models Q2 revenue near $52 billion (+260% YoY: DRAM +265%, NAND +330%), with conventional DRAM ASP growth the dominant driver over HBM, which is forecast to grow a comparatively modest 32% YoY to $6.1 billion as HBM4 mass production timing slips to Q3. SKHY 005930.KS
Taiwan: MOFCOM’s Proposed Fabrication Ban Is the Week’s Direct Taiwan Read-Through
Beyond the TSMC-specific angle of the MOFCOM story above, Taiwan’s foundry oligopoly continues to benefit from the same capacity scarcity that makes the proposed Chinese design ban meaningful: TSMC’s 2nm capacity is reportedly booked by Apple, Nvidia, and AMD through 2027, the backdrop cited repeatedly in Samsung’s Meta/Anthropic foundry courtship this month. A formalized Chinese restriction on using TSMC would not free up meaningful capacity in the near term — Chinese customers are a small share of TSMC’s leading-edge order book relative to Apple and Nvidia — but it would harden the bifurcation of the global foundry customer base along geopolitical lines. TSM
Japan: No Major New Developments — Rapidus 2nm Pricing War Remains the Live Thread
No incremental Rapidus disclosures this week beyond the ¥3–3.5 million per-wafer (~$18,500–21,500) 2nm pricing reported July 19 (~40% under TSMC N2/Intel 18A), which remains the most direct pricing challenge to the TSMC/Samsung leading-edge duopoly heading into 2027 mass production. 8035.T TEL
Korea’s week was about converting fiscal intent into signed paper. The ₩800 trillion budget flagged July 19 was a domestic political commitment; the $950 billion Samsung/SK package is the private-sector mirror, signed in front of the US technology executives who will actually place the orders. The SK Hynix earnings slip from ~July 22 to July 29 is a minor scorecard miss but not a substantive one — the intervening week was consumed by KOSPI disclosure housekeeping (the Intel Ohio denial) rather than any deterioration in the underlying HBM demand picture, and the stock’s 14% rally into the delay suggests the market reads it the same way. The genuine open question remains unchanged from last week: whether the July 29 print confirms Q3 HBM4 mass-production timing or pushes it further right, handing Samsung’s July 30 divisional detail the stronger narrative by default.
With WAIC closed, the week’s China signal shifted from exhibition hardware to policy escalation, platform commercialization, and the next leg of the capital-formation cycle.
Zhipu Ships a Fully Domestic 1GW Data Center, Acquires XCore Sigma — Vertical Integration Compounds
Zhipu AI completed a 1-gigawatt-scale domestic AI data center built entirely with domestic chips, while formally acquiring XCore Sigma, an infrastructure team spun out of the Chinese Academy of Sciences. The combination secures both the compute-supply and utilization-efficiency sides of Zhipu’s stack in the same week, following the roughly HK$50 billion (~$6.4 billion, Zhipu + MiniMax) IPO proceeds flagged in the July 19 brief as already 93% deployed toward GLM iteration, self-developed AI chip R&D, and cluster expansion. Separately, Coinbase’s confirmed routing of 1,200+ internal AI agents to Zhipu’s GLM and Kimi models (disclosed in June) illustrates the commercial pull developing around Chinese open-weight models even as the MOFCOM consultation above weighs restricting foreign access to their weights. Zhipu (HK-listed)
Domestic Super Nodes Move to Commercialization: Alibaba’s M890 Goes Pay-Per-Use, Sugon Joins the National Grid
Alibaba Cloud made its M890 super node commercially available on a metered basis at RMB 120/hour (~$17.6/hour), while Sugon’s 100,000-card “Dengfeng” cluster connected to the National Supercomputing Internet — both developments marking a shift from the hardware-sales model that dominated pre-WAIC coverage toward pay-per-use access, lowering the barrier to entry for SMEs and expanding the addressable compute market beyond large state and corporate buyers. Oriental Computing Core’s DF1000 near-memory 3D AI chip, which won a SAIL Award at WAIC on its July 13 debut, is positioned as a domestic-substitution pathway specifically for high-end AI training that reduces dependence on advanced manufacturing nodes, though commercial production timelines remain unconfirmed. Alibaba Cloud Sugon
Regulatory Escalation: MOFCOM’s Model-Weight and Chip-Design Controls (Full Analysis Above)
The MOFCOM consultation with Alibaba, ByteDance, and Zhipu — covered in full in the core updates section above — is this week’s single most consequential China AI policy development, and the first time Beijing has seriously entertained restricting its own companies’ export of open-weight models and chip designs rather than only defending against US restrictions. Policy Layer
The week’s throughline is that every layer of China’s AI stack is now simultaneously trying to lock in supply and lock out access. Zhipu owning both a domestic-chip data center and the infrastructure team that runs it is the compute-supply half; MOFCOM restricting foreign downloads of the model weights those data centers produce is the demand-control half; and the metered-access shift at Alibaba and Sugon is the monetization layer connecting the two. DeepSeek formalizing IPO preparations barely a month after its first outside raise is the clearest signal yet that China’s frontier labs are following the CXMT/Kunlunxin/Zhipu playbook of racing to public capital markets while the domestic AI narrative is at its strongest — and it raises the same execution question this brief flagged for Kunlunxin: whether revenue and governance can catch up to valuation before the listing window closes. For US portfolios, the MOFCOM story is the one requiring an update to priors: previous weeks framed China’s posture as purely defensive against US export controls; this week shows Beijing is now willing to restrict its own companies’ overseas reach — a genuinely new lever, and one that cuts against the assumption that Chinese open-weight models will remain a frictionless, freely downloadable alternative to closed Western frontier models indefinitely.
The decisive catalyst events that will structure portfolio positioning and supply chain verification through the CXMT and SK Hynix trading debuts, Samsung’s divisional earnings, and the August cascade.
First-Day Price Discovery: Trading opens today with no price limits for five sessions. First-day performance sets the valuation anchor for the entire domestic semiconductor sector and the pricing window for Enflame’s queued syndication behind it. CXMT
Rate Decision Backdrop: The FOMC meeting lands directly ahead of SK Hynix’s earnings, adding a macro-rate variable to the memory-sector reaction window alongside the company-specific HBM4 catalyst.
First Print as a Dual-Listed Company: Consensus models ~$52B revenue (+260% YoY) on pricing-driven DRAM strength; HBM4 mass-production timing (guided Q3) and full-year shipment guidance are the swing factors for the bull case into Samsung’s July 30 release. SKHY
Divisional Breakdown & HBM4 Narrative: DS margin split by memory type, HBM4 revenue run-rate, and confirmation detail on the Broadcom MOU and Meta/Anthropic foundry discussions. Lands one day after SKHY, setting up a direct HBM4 narrative comparison. 005930.KS
Model-Weight & Chip-Design Controls: Watch for a formal draft of China’s next export-control catalogue revision, and industry pushback from Alibaba, ByteDance, and Zhipu on the model-weight restrictions specifically. MOFCOM
STAR Market Filing Timeline: Financial statement completion targeted for end-December; any formal filing confirmation would be the first concrete step from a private $71B round toward a public listing. DeepSeek
Carry-Forward — First Public Quarter: Starlink net adds (>10.5M expected), xAI enterprise revenue run rate, and the 20% insider lockup release at earnings. Unchanged from prior framing. SPCX
Carry-Forward — Jalapeño ASIC + Samsung MOU Detail: First call incorporating both the OpenAI custom-silicon relationship and any concrete follow-through on the $200B Samsung MOU signed this week. AVGO