Semiconductor Sector Weekly (2026-09-13)

Report date: 2026-09-13 | Search window: X posts from 2026-09-06 (Section 5 industry signals from 2026-08-30) | Market data as of the 2026-09-11 close for US/A-share/HK markets (source: exchange snapshots retrieved via eastmoney) Sources: grok-x live X search across six streams (broad search + semiconductor/AI KOL list + core newswire/media list + semiconductor trade media + China-assets list + 14-day industry-event sweep). All price moves were retrieved from authoritative market data, never taken from post text.

1. Key Events

TSMC's August revenue hit a monthly record, +53.3% YoY, with the company saying it still cannot meet AI demand — NT$514.8bn (~US$16.35bn), +10.1% MoM; 2026 YTD revenue ~US$107.1bn, +39% YoY. Reported simultaneously by @business, @CNBC, @StockMKTNewz, @wallstengine and @FirstSquawk. 𝕏¹ 𝕏² 𝕏³
The year-end HBM capacity map for the big three was laid out for the first time: Korean media report Micron is adding 60k wafers/month to reach ~100k wpm, versus SK Hynix at ~200k wpm and Samsung at ~250k wpm by year-end. @jukan05 𝕏
Chinese AI chipmakers are raising prices on HBM shortages: Reuters reports Huawei lifted the Ascend 950DT price to RMB 250,000, a 20%–50% increase; since the US tightened advanced-HBM export controls to China in December 2024, Chinese makers have leaned increasingly on grey-market channels. Relayed by @jukan05 𝕏
CXMT's HBM yield reportedly stuck at ~25%: "Three out of four are defective," with immature TSV technology blamed, against SK Hynix's 90%+ on the same generation. This was the highest-engagement semiconductor post of the week (~774k views). @jukan05 𝕏
Samsung Foundry has committed over half of its 4nm capacity to HBM4 base dies — roughly 15,000 wpm on the Pyeongtaek S5/S6 lines; HBM4 entered mass-production shipment in February and supply is accelerating from Q3. @jukan05 𝕏
Nvidia may become an anchor investor in Anthropic's IPO (Reuters exclusive) — talks are underway in what could be the largest IPO in history. @Reuters 𝕏
A-share "Nvidia concept" denial: Guangdong Goworld denied a market rumor that its high-frequency boards passed Nvidia certification, following unusual share-price moves; a companion report the same day referenced scrutiny of Nvidia/Huawei supply-chain names after abnormal price action. @financialjuice 𝕏¹ 𝕏²
Fujitsu is preparing AI chip exports to the US and Asia, using technology developed for supercomputers. @FirstSquawk 𝕏

2. Institutional & Media Coverage

Factual reporting

TSMC August revenue (multi-source, merged) Bloomberg's @business framed it as "TSMC reports a 53.3% increase in monthly sales, with the company struggling to meet overwhelming demand from the AI infrastructure buildout." @CNBC supplied the detail: August revenue of NT$514.8bn, +53.3% YoY and +10.1% MoM, a monthly record. @StockMKTNewz converted this to US$16.35bn and added 2026 YTD revenue of ~US$107.1bn, +39% YoY. @wallstengine carried the single most important industrial line: "TSMC says it still can't meet demand even while working on roughly 20 fabs simultaneously." 𝕏¹ 𝕏² 𝕏³ 𝕏⁴ 𝕏⁵

TSM ADRs closed at $433.24 on 11 September, +1.22%, at 32.27x trailing P/E — direction consistent with the news, but a modest move, suggesting a monthly beat is largely priced in at current expectations.

Nvidia and Anthropic (Reuters exclusive) @Reuters: "Anthropic is in talks to bring in Nvidia as an anchor investor in what could be the largest IPO in history." Still at the talks stage, nothing finalized. Nvidia closed at $218.29 on 11 September, -0.03% (essentially flat), 27.6x trailing P/E, ~$5.27tn market cap. 𝕏

A-share semiconductor/AI concept denials and regulatory attention @financialjuice posted twice: Guangdong Goworld denied the Nvidia-certification rumor after its shares had already moved abnormally; a second post referenced attention on Nvidia and Huawei supply-chain names following unusual price action. This pattern is typical of A-share AI compute rallies — thematic diffusion outpaces factual verification. 𝕏¹ 𝕏²

Fujitsu AI chip exports @FirstSquawk: "Fujitsu prepares AI chip exports to the U.S. and Asia using technology developed for supercomputers." 𝕏

Trade media: the week's industry topic list @SemiEngineering's weekly roundup listed: "12-inch high-NA EUV masks; Arm's edge AI push; Amkor's expansion; 3D memory; FeRAM startup makes waves; $60B custom silicon deal; Ayar Labs' scale-up deal; ... fine-pitch RDL tech for adv. packaging." 𝕏

A second @SemiEngineering item flagged a new cost variable: "Cost per token is causing EDA design budgets to balloon; what comes next isn't entirely clear yet." AI-assisted design compute bills are now entering IC design companies' cost structures. 𝕏

@techinsightsinc raised a supply-chain carbon angle: "Scope 2 (electricity) drives 36%-59% of IC embodied emissions. Plus, legacy nodes (>26nm) on high-carbon grids remain major polluters." This creates medium-term customer-compliance pressure on mature-node-heavy fabs located on carbon-intensive grids. 𝕏

Views / Ratings

Across the 17-account core newswire and media whitelist (@DeItaone, @unusual_whales, @zerohedge, @Reuters, @CNBC, @WSJ, @FT and others), no posts on rating changes or price-target revisions for any specific semiconductor company were found this week. @FactSet published only generic S&P 500 Q3 2026 revenue/earnings growth expectations, with no semiconductor breakdown. This subsection is therefore kept short, with no speculative filler added.

3. KOL Bull & Bear Views

Memory / HBM (MU, SK Hynix, Samsung, Kioxia, CXMT)

Bullish

@jukan05 (~636 likes, 306k views) relaying a sell-side flip from bearish to positive: "Turning positive on DRAM sector after prior bearish notes on July 3. DRAM solid: We believe the US hyperscaler inventories low and likely to be lowered further. Blended DRAM px expected ~10% QoQ in 4Q26 (vs high-teens in 3Q26). NAND soft: High-teen QoQ pricing in 3Q26, then softening to low single-digit in 4Q26." This is the week's most important structural call: DRAM and NAND are diverging. 𝕏
@jukan05 (~375 likes, 68k views): SK Hynix's leading-edge mix is moving up fast — "SK hynix's 1c DRAM share rose from around 10% in the first quarter to the 34% range in the fourth quarter," with 1c becoming the mainstream process early next year, bearing directly on HBM4E cost and supply capability. 𝕏
@orlandorae (3 likes, 608 views, long-tail retail): "Memory is the bottleneck... the spend is real, these names keep eating... Semis choppy today on yields — thesis still standing. HBM is the tell." 𝕏

Bearish / Cautious

@jukan05 (~793 likes, 321k views) on the Kioxia CEO: "Kioxia CEO dismisses the possibility of closer cooperation with SK hynix; Kioxia aims to hold NAND prices steady to sustain AI demand; CEO says he instructed employees not to raise memory prices sharply." A clear cooling signal for the NAND price narrative, corroborating the "NAND soft" call above. 𝕏
@jukan05 (~405 likes, 105k views) quoting @insane_analyst: "The reduction in HBM stack height is partly because the need for it has diminished... Instead of stacking higher, a different dimension of scarcity will be emphasized: HBM's high pin speed... The HBM spec downgrade is real." A direct challenge to the HBM stack-height arms-race bull narrative. 𝕏
@jukan05 (~1,111 likes, 206 reposts, 774k views — the week's highest engagement): CXMT's HBM yield stalled at 25%, "three out of four are defective," versus SK Hynix's 90%+. 𝕏

Micron (MU) closed at $975.26 on 11 September, -0.22%, at just 22.04x trailing P/E; SanDisk (SNDK, NAND-centric) closed at $1,633.35, -3.50% — the NAND-side price action aligns with the Kioxia price-discipline and NAND-cooling calls, a useful market footnote to the DRAM/NAND divergence.

Supply-chain bottlenecks and capacity (industry-wide)

Neutral, tilted toward supply constraint

@jukan05 (~494 likes, 501 bookmarks, 134k views): "Even if you secure additional GPUs or memory, total system output cannot increase if other segments such as substrates, MLCCs, and wafer testing are in short supply... meaningful new capacity ramp and production will begin in 2028, not 2027. That means next year, 2027, could be the most supply constrained year in history." Strongly positive for the picks-and-shovels layer (substrates, passives, test), but a ceiling on system/server unit shipments. 𝕏
@dnystedt (~43 likes, 5.3k views): "The scramble for mature-node chips is locked in through at least 2028, with some analysts warning tight supplies could stretch to 2030... The AI-related demand surge caused the main issue, alongside rising consumer device silicon content." This runs opposite to the library material from fabricated-knowledge (2025-01-03), which described mature nodes as being in historic overcapacity; applying the "more recent source wins" rule, this week's information indicates mature-node supply/demand flipped from surplus to tight during 2026, driven by AI demand spillover and rising silicon content per device. The divergence is worth monitoring. 𝕏
@dnystedt (~70 likes, 8.5k views): "Taiwan's top 5 semiconductor fab and fab systems builders are seeing the strongest capacity expansion in history, with a combined order backlog over NT$880 billion (US$27.8B) and visibility through 2027, mainly for key clients TSMC and Micron." 𝕏

China semiconductor chain (SMIC, memory, AI compute)

Bullish

@cnfinancewatch (~6,986 views): argues that "memory chips (CXMT supply tightness) and semiconductor manufacturing (SMIC) are resonating," pointing to a "tech self-sufficiency + import substitution" theme. 𝕏
@cnfinancewatch relaying CITIC Securities' view: "continued optimism on AI development, maintaining a long-term positive stance on both the North American compute chain and the domestic AI compute chain," noting test instruments, optical modules and GPU/ASIC names performed through H1 2026. 𝕏

Cautious

@cnfinancewatch (~4,229 views) turned cautious on 8 September, flagging that "the semiconductor sector is under near-term pressure," while noting the AI-compute/PCB/optical-module resonance remains intact. 𝕏
@cnfinancewatch (13 likes, 31 replies, 11k views) on Innolight gave the most complete framing: "policy is distinctly two-sided: benefiting from new-productive-forces tailwinds while facing tail external risk from 1260H taking effect and the draft FCC equipment ban... a top-tier leader at a cyclical high, with explosive earnings, structural flaws and visible risk." 𝕏

Price check (11 September close, retrieved via eastmoney): SMIC A-shares (688981) closed at RMB 117.41, -1.49%, 139.48x trailing P/E; SMIC in Hong Kong (00981) closed at HK$63.15, -0.47%; Hua Hong (01347) at HK$111.70, -1.15%; Cambricon (688256) at RMB 1,040.00, -0.37%; AMEC (688012) -0.18%; Montage (688008) -1.73%. Note the timing gap versus the posts: the "+2.46%" for SMIC cited in @cnfinancewatch's 7 September post was that day's move, whereas by the final trading day of the week (11 September) A-share semiconductor leaders closed broadly lower, with only Innolight (300308) bucking the trend at RMB 926.00, +4.03% — optical modules outperforming chip manufacturing is the genuine intra-chain divergence in A-shares this week. Re Top (603601), the small cap mentioned by @cnfinancewatch, closed at RMB 10.79, +8.33%, but is a high-volatility micro cap (~RMB 12.3bn market cap, 303x trailing P/E) and should be treated as thematic.

US semiconductor sector overall

Bullish: @jumpingjunebugg (1 like, 106 views, long-tail): "Semiconductors are ignoring the weak tape. $SOX ~+1.6% while the broader market trades lower... As long as $SOX keeps outperforming, AI remains one of my favorite areas." 𝕏

Cautious: @treasureh8nter (4 likes, 1 repost, 543 views) ranked names by drawdown from all-time highs, with INTC and AVGO deepest and NVDA shallowest. @HumeTrader (0 likes, 36 views) counted more decliners than advancers and described "overall downward momentum." Both are low-engagement long-tail views of limited weight, included here only as a sentiment reading. 𝕏¹ 𝕏²

To be explicit: the price figures embedded in those two posts were not used. Against the authoritative 11 September close, US semis were in fact broadly higher that day — ARM +4.17%, Marvell +4.03%, Qualcomm +2.88%, Intel +2.61%, AMD +2.49%, KLA +1.95%, TSM +1.22%, with only Nvidia (-0.03%), Micron (-0.22%) and SanDisk (-3.50%) lower. The posts' description of sector weakness does not match Friday's actual tape; market data governs.

4. Buzz & Sentiment Shifts

Memory/HBM — the clear center of gravity this week, with sentiment shifting from "everything rises" to "divergence." These posts dominated the week's engagement (CXMT yield post 774k views, Kioxia 321k, DRAM upgrade 306k), far outstripping equipment or foundry topics. Sentiment skews bullish on DRAM/HBM (low hyperscaler inventories, ~10% QoQ price gains still expected in 4Q26, all three majors adding capacity), while an unusually cautious voice emerged on NAND (Kioxia proactively holding prices, 4Q26 QoQ gains decelerating to low single digits). This is a material change from last week's blanket "memory prices are rising" narrative, and the tape has started to reflect it: SanDisk -3.50% on Friday versus Micron (22x P/E) down just 0.22%.

Foundry/leading edge — moderate buzz but the highest density of hard facts, sentiment bullish. TSMC's August revenue was the only hard datapoint reported simultaneously by five-plus outlets this week; "roughly 20 fabs running simultaneously and still short" combined with Taiwanese fab-systems builders' NT$880bn backlog and visibility through 2027 reinforces the view that the capex up-cycle has not peaked. Notably, that discussion has not translated into share-price elasticity (TSM only +1.22% Friday), implying the high growth is already inside expectations.

China semiconductors — buzz materially below US/Korea, and skewed toward theme over fundamentals. Only one of the ten China-assets whitelist accounts (@cnfinancewatch) produced relevant content; nothing substantive surfaced on Hua Hong, Cambricon, Ascend or YMTC. The genuinely informative China facts this week instead came via Korean media relays (CXMT yields, price hikes under HBM export controls). On sentiment, the A-share bull narrative concentrates on "import substitution + compute," yet SMIC, Cambricon, AMEC and Montage all closed lower on 11 September while optical modules (Innolight +4.03%) stood alone. Combined with Guangdong Goworld's forced denial of the Nvidia certification rumor, the theme has clearly reached the stage where diffusion requires factual falsification.

Main drivers of change versus last week: last week's throughline was "AI demand → memory prices rise across the board." This week it fractured into three finer threads — (1) DRAM and NAND diverging; (2) the bottleneck migrating off the chip itself and onto substrates/MLCCs/test (@jukan05 calling 2027 potentially "the most supply constrained year in history"); and (3) early signals of mature nodes flipping from surplus to tight. All three point to supply constraint rather than demand falsification, which is why sentiment fragmented without turning bearish.

5. First-hand Industry Signals (Last 14 Days)

Date/Time (ET)	Company	Event type	One-line event	Status	Source
09-10 02:08	TSMC	Capacity/revenue	August revenue NT$514.8bn, +53.3% YoY, +10.1% MoM, a monthly record; says it still can't meet demand while running ~20 fabs simultaneously	Confirmed	@CNBC / @wallstengine
09-09 21:05	TSMC	New fab / leading edge	Taichung 1.4nm fab cluster in full acceleration (P1 steel complete, P2 pouring concrete, P3 permitted, P4 applied); mass production possibly pulled into end-2027	Rumor (media)	@dnystedt
09-11 (Korean media)	Micron / SK Hynix / Samsung	HBM capacity	Micron doubling HBM capacity to ~100k wpm by year-end; SK Hynix ~200k wpm and Samsung ~250k wpm	Confirmed (industry sources)	@jukan05
09-08 18:58	Samsung	Fab start-up	Samsung's Taylor, Texas foundry fab has begun operations	Confirmed (Korean media)	@jukan05
09-07 01:45	Samsung Foundry	HBM capacity allocation	Over 50% of 4nm capacity allocated to HBM4 base-die mass production (Pyeongtaek S5/S6, ~15k wpm); HBM4 in volume shipment since February	Confirmed	@jukan05
09-11 13:18 (weekly)	Amkor / Arm / Ayar Labs	Expansion/packaging/custom silicon	Weekly roundup cites Amkor expansion, 12-inch high-NA EUV masks, a $60B custom silicon deal, fine-pitch RDL advanced packaging	Confirmed (media roundup)	@SemiEngineering
09-08 (Reuters)	Huawei / Chinese AI chipmakers	Pricing/supply	On HBM shortages, Huawei raised Ascend 950DT pricing to RMB 250,000 (+20%~50%)	Confirmed (Reuters)	@jukan05
09-03 18:01	Micron	HBM expansion	Plans to add up to 60k wpm of HBM capacity by year-end (incl. HBM4 12-high), having procured tools for Taiwan and Singapore sites	Confirmed (industry sources)	@jukan05
09-02 00:57	TSMC	Advanced packaging (CoWoS)	Asked materials/equipment partners to develop 5μm-class microbump bonding and underfill; mass production targeted 2H 2028, no near-term switch to hybrid bonding	Rumor (materials chain)	@jukan05
08-31 02:45 / 08-30 21:22	SK Hynix / Intel Foundry	Large order / capacity allocation	Plans to outsource part of next-gen (HBM4E) base-die production from TSMC to Intel Foundry, building a multi-supplier strategy	Rumor (Korean media/industry)	@FirstSquawk / @jukan05
08-30 02:28	CXMT	Yield	Beijing Fab 1 completed transition to 4th-gen process (~16nm) in 2Q26; standard DRAM yields estimated above 90%, DDR5 up to 5600MHz	Confirmed (broker report)	@jukan05
09-09 (Korean media)	CXMT	Yield	HBM yields stalled at ~25% on immature TSV, versus SK Hynix's 90%+	Rumor (Korean media relay)	@jukan05
09-13 01:49	Fujitsu	New product / export	Preparing AI chip exports to the US and Asia using supercomputer-derived technology	Confirmed (newswire)	@FirstSquawk

No verified posts on tape-outs, sampling, specific 2nm/N2/A16 actions, or lead times were found in this period. Those signal categories are absent this cycle and no speculative filler has been added.

Commentary on key events

① Samsung committing over half its 4nm capacity to HBM4 base dies (09-07, confirmed) This is the period's most directly competitive development. Moving base dies from "memory makers build them in-house" to "built on leading-edge logic" means that from HBM4 onward, a memory maker's HBM competitiveness is partly bound to logic foundry capability. Samsung is the only one of the three with captive leading-edge foundry, and tilting 4nm internally toward HBM4 trades vertical integration for delivery speed — at the cost of that capacity no longer being available to external logic customers, raising the bar for Samsung Foundry's external customer acquisition. The mirror image is SK Hynix: it has no captive leading-edge logic, has relied entirely on TSMC's 12nm-class for base dies, and is now rumored to be weighing moving part of HBM4E base-die output to Intel Foundry. If that lands, it would be Intel Foundry's first external HBM-related order of scale, and the first crack in TSMC's monopoly at the HBM base-die layer. Intel (INTC) closed at $102.94 on 11 September, +2.61%, but trailing P/E remains negative (-49.25) — the foundry profitability inflection has not arrived, so this is narrative improvement rather than realized earnings. To be clear: this item is rumor; neither party has confirmed it.

② The DRAM/NAND split (09-06 to 09-09, views cross-checked against industrial actions) This week broke the vague notion of "memory upcycle" into parts. On the DRAM side: hyperscaler inventories are low, 4Q26 blended prices are still expected up ~10% QoQ, and SK Hynix's 1c mix jumps from ~10% in Q1 to 34% in Q4 — volume, price and process all pointing up. On the NAND side: after a high-teens QoQ increase in 3Q26, 4Q26 is expected to decelerate to low single digits, and Kioxia's CEO has publicly said he will not raise prices sharply, aiming to "hold prices steady to sustain AI demand." A leading supplier voluntarily capping price typically appears when supply is loosening or demand durability is in doubt — a sharp contrast with DRAM tightness. The investment implication: memory should no longer be traded as a single beta. Friday's tape already began pricing this, with DRAM/HBM-weighted Micron (-0.22%, just 22x P/E) versus NAND-weighted SanDisk (-3.50%). Separately, @insane_analyst's point that "the HBM spec downgrade is real" is a second constraint on the slope of HBM ASP gains, and a useful counter-indicator against the bull case.

③ Bottleneck migration: 2027 as potentially "the most supply constrained year in history" (09-09, a view, but grounded in supply-chain structure) @jukan05's core claim is system-level: however many GPUs and memory dies you secure, if substrates, MLCCs or wafer test are short anywhere, total system output cannot rise — and meaningful new capacity only ramps in 2028. Two hard facts from this period corroborate it: TSMC's "~20 fabs simultaneously and still short," and Taiwanese fab-systems builders' NT$880bn backlog with visibility into 2027. If correct, 2027's defining feature would be "demand visible, delivery impossible" — margin-expansionary for the picks-and-shovels layer (substrates, passives, test equipment, fab construction), while physically capping AI server unit shipments and lifting unit costs for downstream OEMs and cloud providers. This also partly explains why US semicap names were steady rather than weak this week (AMAT +0.55%, LRCX +0.07%, KLAC +1.95%, ASML +0.64%) — a longer build-out cycle extends order visibility for tool suppliers. Caveat: this is a KOL view, not a confirmed fact, and the 2028 capacity timeline is itself uncertain.

This report compiles live searches of public X posts alongside exchange market data, and is provided for informational and educational purposes only. It does not constitute a recommendation to buy or sell any security, nor personalized investment advice. Views belong to their original authors and do not represent the platform's position; items marked "Rumor/Expected" have not been confirmed by the relevant companies and readers should verify independently. Markets carry risk.

Semiconductors · Weekly X Pulse