TSMC Guides Past 40, CXMT Prices the Largest STAR IPO Ever, WAIC Puts Domestic Super Nodes on the Floor
Revisiting the July 12 Brief — TSM Books First 2nm Revenue With FY26 Guide Raised Above 40%, SKHY Whipsaws From ₩-15.4% Seoul Crash to +27% Barclays Rally, CXMT Opens RMB 66.6B Subscription, Atlas 950 and DF1000 Sweep the SAIL Awards, NVIDIA Purges Half Its Asian Whitelist
The week of July 13–19 delivered execution on every major forward vector from the July 12 brief — and opened two new geopolitical fronts. TSMC’s July 16 earnings exceeded even the raised expectations embedded in Citi’s catalyst watch: record revenue of approximately $40.2 billion (+36% YoY), net profit up 77.4% to NT$706.56 billion, gross margin of 67.7%, first N2 (2nm) revenue recognized, and full-year 2026 revenue growth guidance lifted to slightly above 40% with capex raised to $60–64 billion. SK Hynix’s first full week under the permanent SKHY ticker was a stress test passed violently: Korean shares suffered a reported 15.4% single-session collapse — their worst in nearly two decades — before the ADR ripped 27% to $186 on July 14 as Barclays became the first major Western house to initiate at Overweight, options began trading, and two 2x leveraged single-stock ETFs launched. In Shanghai, CXMT priced the largest IPO in STAR Market history at RMB 8.66 per share (July 14), opening subscription July 16 for a raise of up to RMB 66.6 billion — surpassing SMIC’s 2020 record — while WAIC 2026 (July 17–20) converted the domestic-substitution thesis from renderings to physical hardware: Huawei’s Atlas 950 super node and Oriental Computing Core’s DF1000 near-memory 3D AI chip both took the SAIL Award, Sugon unveiled a fully domestic 100,000-card cluster, and Moore Threads and MetaX declared domestic silicon training-capable. The new friction points: NVIDIA cut more than half of its previously approved Asian buyers from its compliance whitelist on July 14, and Beijing’s helium export ban continued to ferment. Seoul answered with a record ₩800 trillion FY2027 budget built on AI-chip tax receipts. This brief closes the loop on the July 12 forward vectors, integrates the full China AI and Asian supply chain intelligence layers, and resets the monitoring window through SK Hynix’s Q2 print (~July 22), Samsung’s July 30 full release, the August 6 SPCX earnings, and September 2 AVGO results.
A structured review of the July 12 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 foundry economics, memory capital markets, and the dual-track IPO cycle.
TSMC Q2 2026: The 40% Guide — First 2nm Revenue and a $60–64B Capex Signal to the Entire Tool Chain
TSMC reported record Q2 results on July 16: revenue of approximately $40.2 billion (+36% year-over-year), net profit up 77.4% to NT$706.56 billion — a single-quarter record — and gross margin of 67.7%, above guidance. High-Performance Computing rose 20% sequentially to 66% of wafer revenue, with smartphones falling to 22%; N3 contributed 30% of wafer revenue, N5 33%, and the first N2 revenue was recognized as the 2nm ramp begins. Management raised full-year 2026 revenue growth guidance to slightly above 40% (from 30%+) and lifted the 2026 capex budget to $60–64 billion, with expanded Arizona investment plans reported at approximately $100 billion. TSM
The July 12 brief framed July 16 as a guidance-raise event; the raise landed a full five points above the bull case. Two structural signals matter more than the headline. First, the 2nm era is now a revenue line, not a roadmap slide — N2 ramp economics will define the 2027 margin path, and Rapidus’s aggressive ¥3–3.5M per wafer 2nm pricing (see Japan section) is the first credible pricing attack on that node. Second, the capex raise to $60–64B is a direct, dated order signal for the equipment and materials chain: ASML, Tokyo Electron, Advantest, Applied Materials, KLA, and Shin-Etsu all re-rate on TSMC’s budget before their own prints. Watch the equipment makers’ order books as the first read on capex flow-through, and monitor whether HPC concentration (66% of wafer revenue) starts drawing customer-concentration questions the way HBM concentration did for the memory names.
SKHY Week One: From a 15% Seoul Crash to a 27% Barclays Rally — Price Discovery at Full Amplitude
SK Hynix’s first full week as a dual-listed company was a compressed lesson in what US listing actually does. After the July 10 debut (+13% to $168.01), Korean shares reportedly collapsed 15.4% in their worst single session in nearly two decades, briefly halting the Seoul market and dragging the ADR toward $152. Then on July 14 the ADR surged 27% to $186 as Barclays initiated coverage at Overweight — the first major Western financial institution to do so, citing “substantial” upside to the AI memory supply cycle — and South Korea’s record AI budget landed the same day. Options began trading July 14 (~150,000 contracts by midday), and GraniteShares and ProShares launched 2x leveraged single-stock ETFs, pulling Micron up 5% alongside SanDisk and Western Digital. SKHY ~$186 MU
The volatility is the feature, not the bug. The July 12 brief flagged ADR-to-won parity as the proxy for new money entry velocity; what the week revealed is that the US listing has transferred price discovery to New York — the Barclays initiation moved the stock more than the Seoul crash did, and the options/leveraged-ETF complex now adds a gamma layer that Korean-only listing never had. The analyst initiation cascade predicted at listing has begun; each incremental Western initiation is a mechanical flow event while the SKHY-vs-MU valuation compression trade remains live. The near-term catalyst is the Q2 earnings print expected around July 22 — the first as a dual-listed company. HBM4 ramp timing is the swing factor: SK Hynix’s meaningful HBM4 volume has reportedly slipped to Q3 2026 (full-year outlook trimmed to ~4B Gb) while Samsung’s was raised to ~4B Gb with HBM4 sales reportedly already above $1B. A confident HBM4 recovery narrative on the call re-anchors the bull case; a soft one hands Samsung the HBM4 narrative into its July 30 release.
CXMT: The RMB 66.6B Memory IPO — China’s Answer to the HBM Capacity Vacuum Prices at a Record
ChangXin Memory Technologies set its issue price at RMB 8.66 per share on July 14 for an initial issuance of 6.688 billion shares, opening online and offline subscription on July 16. The initial raise of RMB 57.9 billion reaches RMB 66.6 billion if the over-allotment option is fully exercised — surpassing SMIC’s RMB 53.2 billion (2020) as the largest IPO in STAR Market history — with post-listing market value estimated to exceed RMB 2 trillion. Strategic investors include the National Social Security Fund, AMEC, Alibaba, Tencent, and Xiaomi. As the world’s fourth-largest DRAM manufacturer, CXMT is listing directly into the strongest conventional DRAM pricing environment in years: global incumbents have shifted wafer capacity to HBM, leaving the commodity DRAM vacuum CXMT’s expansion is built to fill. CXMT (STAR, pricing set)
Read CXMT and SKHY as one trade in two currencies: within five trading days, both sides of the decoupled memory market executed record capital raises to fund capacity into the same shortage. The strategic investor roster is the tell — AMEC on the cap table alongside Alibaba, Tencent, and Xiaomi means the equipment supplier, the compute buyers, and the state pension system are all positioned upstream and downstream of the raise. The proceeds flow directly into domestic semiconductor equipment, materials, and OSAT demand (AMEC, Naura, TFME, JCET read-throughs), and every month of 10–20% DRAM price increases fattens the revenue model underneath the RMB 2 trillion valuation. The July 20 payment deadline and oversubscription multiple are the immediate signals; first-day performance sets the valuation anchor for the entire domestic semiconductor sector — and for Enflame’s pricing window behind it. The unhedged risk: CXMT’s capacity is precisely the conventional DRAM supply the global shortage thesis assumes stays scarce. If CXMT executes its ramp, it is simultaneously the largest beneficiary and the largest eventual threat to the pricing supercycle it is listing into.
Korea, Japan, Taiwan, and ASEAN supply chain developments mapped to the USINO.AI coverage universe with US ticker read-throughs on every story.
Korea: A ₩800 Trillion Budget, an HBM4 Race Reset, and Huawei at the Gates
Seoul announced on July 13 that it will draw up record FY2027 spending of more than ₩800 trillion (~$531 billion), financed by surging tax receipts from the AI chip boom, with three “mega-projects” — semiconductors, AI data centers, and physical AI — taking top fiscal priority alongside a new Future Response Fund. The HBM4 competitive picture reset in parallel: SK Hynix’s meaningful HBM4 ramp has reportedly slipped to Q3 2026 (full-year trimmed to ~4B Gb) while Samsung’s expectation was raised to ~4B Gb with HBM4 sales reportedly already above $1 billion; SK Hynix holds ~62% of HBM share with Micron reportedly overtaking Samsung for #2, and all three are fighting for NVIDIA’s 16-Hi HBM4 contracts with diverging base-die strategies. And the week’s most consequential corroboration: Huawei’s Q4 2026 Korea launch of the Ascend 950 series and Atlas 950 SuperPod firmed up — cluster-level deployments through two local distributors including SK Shieldus, positioned at ~2.87× H20 inference performance at roughly one-quarter the cost — matching the WAIC report that Korean cloud vendors locked in the first 2,000 Ascend 950 orders. SKHY 005930.KS MU
Taiwan: The Foundry Oligopoly Extends — Capex Raise Reprices the Global Tool Chain
TSMC’s record quarter and raised guidance (full analysis above) reverberate through the Taiwan supply chain: the $60–64 billion capex budget is discipline-locked into sub-3nm capacity and advanced packaging, with HPC now 66% of wafer revenue. First 2nm revenue recognition marks the node-transition moment competitors have targeted — Samsung Foundry is reportedly courting Meta (an order reported at over ₩10 trillion) and Anthropic for custom silicon, and Rapidus has opened a price war (below). The advanced packaging chain (CoWoS ecosystem) and Taiwan OSAT names remain the direct beneficiaries of the capex raise. TSM ASML
Japan: Rapidus Prices 2nm at ¥3–3.5M Per Wafer — The First Credible Pricing Attack on the Leading Edge
Rapidus announced 2nm pricing of ¥3–3.5 million per wafer (~$18,500–21,500), versus a reported ~$30,000 for TSMC N2 and Intel 18A — an up-to-40% undercut at the most advanced node in production. The company is in discussions with more than 60 potential clients, collaborating with IBM on advanced packaging, and has received ¥631.5 billion (~$4 billion) in government support for AI chip production, with mass production targeted for 2027 in Chitose, Hokkaido. Japan’s materials and equipment complex (Tokyo Electron, Shin-Etsu, Advantest, Sumco) supplies both the Rapidus ramp and TSMC’s raised capex — a rare double exposure. Blackstone’s president separately told Nikkei the firm will invest $30 billion in Japanese AI data centers. 8035.T TEL 4063.T Shin-Etsu 6857.T Advantest
ASEAN & the Whitelist: NVIDIA Cuts More Than Half Its Asian Buyers as the Transshipment Channel Closes
On July 14, NVIDIA implemented stringent compliance vetting for Asian customers — contract validation and end-user interviews concentrated in Singapore, Malaysia, and Japan — with reports that more than half of previously approved buyers failed. The regime follows late-May US rules requiring licenses for advanced-chip transfers to entities headquartered in China or Macau, and criminal prosecutions over misrepresented end-users are proceeding in Singapore. The ASEAN “China+1” premium thesis now carries a compliance discount: the same jurisdictions marketed as de-risked AI capacity are the ones under the microscope. Regional demand remains structurally intact — SKT announced a $91.5 billion Korean data center program, AWS is investing $7 billion in India, and APAC data center investment hit a record ~$11.6 billion. NVDA QQQ
Korea is the week’s pivot geography — and the contested middle of the bifurcating compute stack. In seven days it produced a record sovereign budget built on chip taxes, the most violent price discovery in SKHY’s short public life, an HBM4 ramp reset that hands Samsung a narrative window into July 30, and confirmation that Huawei’s first NPU beachhead in a US-ally market opens in Q4 with Korean cloud pre-orders already reported at WAIC. The whitelist purge completes the picture: as NVIDIA closes the transshipment channel through ASEAN, Chinese silicon shows up at the front door of US allies with a one-quarter-cost pitch. For US portfolios, the read-through is a barbell — the TSMC capex raise re-rates the entire Western tool chain (ASML, AMAT, KLA, TEL) with dated, funded orders, while the Huawei-in-Korea vector is the first live test of whether NVDA’s ex-China Asian TAM is defensible on price. Watch the SKHY July 22 print and Samsung’s July 30 divisional detail as the sequencing trade of the next fortnight.
WAIC 2026 (July 17–20, Shanghai) converted the domestic-substitution thesis into delivered hardware: 1,100+ exhibitors, 3,000+ exhibits, 300+ global debuts, and mainstream domestic GPU makers gathered on one floor for the first time, as the State Council released the “Action Plan for Cooperative Development of Artificial Intelligence.”
WAIC Delivers: Atlas 950 and DF1000 Sweep the SAIL Awards — and Domestic Silicon Claims Training Capability
Huawei’s Atlas 950 super node made its first physical appearance July 17 with 1,024 Ascend cards, expandable to 8,192 Ascend 950DT chips on the Lingqu interconnect, and won WAIC’s highest SAIL honor. Oriental Computing Core’s DF1000 — the world’s first software-defined near-memory-computing 3D AI chip (14nm, 520 teraFLOPS, 6.4TB/s memory bandwidth, debuted July 13) — took the SAIL Award as well, alongside a full-stack toolchain and a product line from single cards to liquid-cooled super nodes. Sugon released the Sugon 8000 “Dengfeng,” China’s first fully domestic 100,000-card AI super cluster. At the July 18 “Token Era” forum, Moore Threads CEO Zhang Jianzhong declared domestic models can be safely trained on domestic chips, MetaX claimed integrated training-and-inference coverage at 10,000+ card scale, and JD Cloud announced a domestic inference platform partnership with Moore Threads. Huawei (unlisted) Moore Threads
The July 12 brief flagged customer deployment announcements as the primary verification signal for the Ascend 950 debut — and they arrived: South Korean cloud vendors reportedly locked in the first batch of 2,000 Ascend 950 orders, the first material export order for a Chinese super node platform into a US-ally market. The DF1000’s significance is architectural: 520 teraFLOPS on 14nm with 6.4TB/s bandwidth, requiring neither advanced nodes nor restricted large-capacity memory — a “de-Americanized” compute path that routes demand into mature-node foundry and domestic advanced packaging instead of the restricted chokepoints. Watch post-WAIC order and mass-production disclosures (conference closes July 20) as the conversion signal from exhibition to revenue.
The Capital Stack Goes Vertical: Kunlunxin Targets $50B, DeepSeek Opens at $71B Pre-Money, Zhipu/MiniMax Deploy HK$50B
Baidu’s Kunlunxin entered cornerstone-investor communications for its HK IPO at a target valuation of ~$50 billion (nearly RMB 340 billion) — up more than 16× from its ~RMB 21 billion Series D post-money of July 2025, on Q1 2026 revenue of RMB 1.2 billion (+350% YoY) with external sales now exceeding internal supply; Baidu retains 57.67% control. DeepSeek began contacting investors July 14 (per the Financial Times) at ~$71 billion pre-money, and has approached chip designers, foundries, and memory suppliers while quietly expanding its chip-design team. The nearly HK$50 billion raised by Zhipu (HK$31.4B placement) and MiniMax (HK$16B placement plus convertibles) in the prior week’s window is now deploying — Zhipu’s proceeds are explicitly earmarked for GLM iteration, self-developed AI chip R&D, and cluster expansion, with both companies having spent 93% and 77% of IPO proceeds respectively. Kunlunxin (pre-IPO) DeepSeek (private)
Policy Layer: Helium Ban Ferments, Compute Subsidies Extend to 2028, an RMB 80B Materials Fund Lands
Beijing’s temporary helium export ban (announced July 10 by MOFCOM and Customs) continued to ferment, with MOFCOM restating the policy basis on July 16 — helium is irreplaceable in semiconductor manufacturing and HBM, global semiconductor-grade demand is expected to grow 9% in 2026, and domestic prices hold above RMB 300/cubic meter. The seven-ministry “Three-Year Action Plan for Hard Tech (2026–2028)” began landing: domestic compute subsidies and semiconductor R&D super-deductions extended to end-2028, up to 30% subsidies for Guizhou intelligent computing centers purchasing domestic AI chips, domestic servers, optical modules, and storage in key-support scope, and an RMB 80 billion new-materials fund established under a rare-metals supplement. Policy Layer
WAIC week resolved the question the July 12 brief posed — whether domestic substitution is advancing toward price/performance parity or remaining cost-prohibitive — with a third answer: it is advancing toward export. The 2,000-unit Korean order report, if confirmed in post-WAIC disclosures, is the first time Chinese AI infrastructure has been sold into a US-ally cloud market, and it lands precisely as NVIDIA’s whitelist purge raises the cost of the alternative. The capital formation loop is now fully closed at every layer: memory (CXMT, RMB 66.6B), training silicon (Kunlunxin at $50B, Enflame queued), models going vertical into chips (DeepSeek at $71B, Zhipu’s placement earmarks), and the state underwriting demand through 2028. The “domestic chips capable of training” declaration from Moore Threads and MetaX is the claim to stress-test — training at 10,000-card scale is a systems problem (interconnect, memory bandwidth, failure recovery), not a FLOPS problem, and the DF1000/Atlas ecosystem answer to that is exactly what the SAIL Awards were signaling. For US portfolios the uncomfortable read-through stands: every layer of this stack that works erodes the assumption that export controls cap China’s AI compute, while the helium ban demonstrates Beijing’s willingness to answer chokepoints in kind.
The decisive catalyst events that will structure portfolio positioning and supply chain verification through the CXMT listing, the Korean earnings sequence, and the August earnings cascade.
Oversubscription Multiple & Listing Setup: Offline and online payment closes July 20. The oversubscription multiple is the single cleanest read on domestic institutional appetite for the memory chain; first-day performance sets the valuation anchor for the entire semiconductor sector — and Enflame’s pricing range behind it. CXMT
Order Flow & Mass-Production Disclosures: Watch follow-up orders and production schedules for the Atlas 950 and DF1000, confirmation of the 2,000-unit Korean cloud order, and supply chain awards behind the 300+ global debuts — interconnect, optical modules, liquid cooling, and complete-machine names. Huawei
First Print as a Dual-Listed Company: HBM4 ramp timing (now Q3), full-year shipment guidance, and capex against the $26.5B war chest. A confident HBM4 recovery narrative re-anchors the bull case; a soft one hands Samsung the narrative into July 30. Watch the options market’s first earnings cycle for the new gamma regime. SKHY
Divisional Breakdown & H2 Guidance: DS margin split by memory type, HBM4 revenue run-rate versus the reported $1B+, and any commentary on the reported Meta (₩10T+) and Anthropic foundry discussions. Margin maintenance signals structural scarcity through Q4; any cut flags peak memory pricing. 005930.KS
NVIDIA Asia Compliance Regime: Which Singapore/Malaysia/Japan buyers lose allocation, progress in the Singapore prosecutions, and whether displaced demand reroutes to domestic Chinese silicon or compliant channels. Watch for any formal BIS follow-through codifying the whitelist standard. NVDA
$71B Pre-Money Syndication: Investor list composition and any disclosed chip-design partnerships (design services, IP licensing, foundry allocation). Terms will calibrate the market’s read on model vendors going vertical — and the durability of external compute procurement from the domestic chip chain. 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 the July 12 framing: the float constraint binds until this window opens. SPCX
Carry-Forward — Jalapeño ASIC Contribution: First call incorporating the full OpenAI custom silicon relationship. The custom-ASIC thesis gains urgency as model vendors on both sides of the Pacific (OpenAI, DeepSeek, Zhipu) go vertical into silicon. AVGO