Aya X Pulse · Semiconductors

Semiconductors · Weekly X Pulse

Published Sunday, September 20, 2026 (Beijing time)
Today's Heat Board

Tickers ranked by buzz on X (Twitter); bars show the bull/bear split.

BullishBearishMixed/Neutral

Tap a ticker to see its KOL chatter →

1Micron78%
2NVIDIA58%
3TSMC62%
4Samsung45%
5AMD60%
6SK Hynix70%
7SanDisk75%
8Applied Materials68%
9IntelMixed
10Broadcom48%
View the full board (23 tickers) →

Semiconductor Sector Weekly (2026-09-20)

Report date: 2026-09-20 | Data cutoff: US equities close 2026-09-18 (Friday); HK/A-shares close 2026-09-18 Sentiment source: real-time X (Twitter) search (window 2026-09-13 to 2026-09-20; industry-event window from 2026-09-06). Price source: eastmoney real-time quote snapshots. Note: All prices and daily moves come from exchange snapshots, never from figures embedded in tweet text.


1. Key Events

  1. Public calls to slow frontier AI development triggered the week's biggest chip selloff — the Philadelphia Semiconductor Index plunged 5.9% in a single session, while the Nasdaq 100 fell 0.8% and the S&P 500 lost 0.5%, with Nvidia and Broadcom among the major decliners. 𝕏 𝕏²
  2. Jensen Huang addressed the AI-safety debate at a summit in Scotland while giving forward shipment guidance: "UNSAFE PRODUCTS MUST BE HELD BACK" and "NVIDIA TO SELL TWICE AS MANY CHIPS NEXT YEAR AS THIS YR." 𝕏 𝕏²
  3. Intel and SK Hynix were reported to be discussing US-based memory chip manufacturing, and both stocks jumped on the report (per @CNBC). 𝕏
  4. AMD is reportedly preparing a ~10% chip price increase tied directly to rising TSMC wafer costs, with new pricing policies planned for Q4. 𝕏
  5. Omdia: global semiconductor revenue hit a record $425B in Q2 2026, up 31.4% QoQ, with memory accounting for more than half of total revenue; Omdia expects Q3 to surpass $500B. 𝕏
  6. Huawei was reported set to unveil new AI technology this week in defiance of US export controls, advancing its ambition to replace Nvidia in China. 𝕏
  7. Applied Materials announced a $5B investment in India over the next decade, including a 140-acre advanced semiconductor research park with cleanroom, and plans to double its India R&D workforce. 𝕏
  8. Compute rental prices rose in lockstep: Nebius is raising on-demand GPU prices effective Oct. 1 — H100 from $3.85 to $4.50 (+17%), B300 from $7.85 to $9.50 (+21%). 𝕏

2. Institutional & Media Coverage

Factual reporting

AI-slowdown fear and the drawdown. @Reuters reported that Wall Street ended down "weighed down by losses in Nvidia and other chipmakers after top executives in US artificial intelligence companies raised safety concerns and called for a slowdown in the development of AI." @financialjuice headlined its wrap "Chipmakers Slide as AI Leaders Back Slower Development." @zerohedge's premarket movers list included Intel and Micron among decliners. 𝕏 𝕏² 𝕏³

Price check (eastmoney, Friday 2026-09-18 close). The panic was mid-week; by Friday the group had largely recovered — iShares Semiconductor ETF (SOXX) closed at $533.07, +2.69%; Nvidia $222.27, +1.34%; AMD $559.82, +2.70%; Broadcom $357.61, +2.97%; TSMC ADR $434.67, +1.02%. The "AI slowdown" crash did not carry into the week's final session, and readers should not treat the single-day declines in media coverage as the week's closing state.

Memory and equipment were the two strongest lines on Friday. SanDisk (SNDK) closed at $1,791.82, +10.99% — the largest single-day gain among the names we verified; Micron $1,015.80, +3.92%; Western Digital $441.36, +4.13%. In equipment: Lam Research +6.98%, Applied Materials +6.51%, Teradyne +5.19%, KLA +4.74%, ASML +3.08%. The notable outlier is Qualcomm, which closed at $177.72, down 5.82%, sharply diverging from the sector. This search surfaced no tweet explaining that move, so we state the price fact only and offer no attribution.

Record Q2 revenue. @wallstengine citing Omdia: Q2 global semi revenue $425B, +31.4% QoQ, memory more than half, driven by AI demand and rising DRAM, NAND and NOR pricing. Long-tail account @YoooJJ8cm added on the same dataset that DRAM, NAND and NOR all set simultaneous record sequential gains, and that "AI demand is reshaping memory fabs' production priorities." 𝕏 𝕏²

Cost pass-through downstream. @wallstengine: AMD has notified partners of an expected ~10% price increase tied to higher TSMC wafer costs, landing in Q4. Paired with the Nebius GPU rental hike effective Oct. 1, these are two links in the same chain. 𝕏 𝕏²

China side. @business reported Huawei will unveil new AI technology this week to challenge Nvidia in China. On prices, SMIC H-shares closed at HK$65.15, +4.41%; SMIC A-shares (688981) at RMB 122.00, +2.85%; Hua Hong at HK$115.50, +4.81%; NAURA at RMB 686.51, +3.16%; AMEC +1.05%; Cambricon at RMB 1,113.14, +0.65%. 𝕏

Views / ratings

Within the designated newswire and media whitelist, no analyst upgrade, downgrade or price-target change tweets on semiconductor names were found for this window. The sell-side views cited in Section 3 (RBC, Citi, BofA) come from individual accounts relaying broker research, not first-hand reporting by whitelisted media — weigh them accordingly.


3. KOL Bull & Bear Views

Bullish

@sean_________ (≈303 likes / 45 reposts) relaying RBC on Micron: "Our Asia supply chain conversations point to 80-100% HBM price increases for 2027, which should remain a tailwind to blended DRAM ASPs." 𝕏

@pequityresearch (≈294 likes / 51 reposts) relaying Citi: memory undersupply conditions extending through 2031, HBM bit demand +62% YoY in 2027, characterized as "Severe Market Undersupply." A second post from the same account (≈159 likes / 26 reposts) cites Citi's HBM demand curve "jumping from just 736 to a staggering 126,990 by 2028," with "Nvidia remains the primary driver." 𝕏 𝕏²

@WealthCode99978 (≈58 likes / 15 reposts) on valuation: "$MU at just 6x, $SNDK 7x, $SKHY even 4x... if AI training and inference keep driving HBM, DRAM and NAND demand, these valuations become very compelling." (These are his 2027 forward P/Es. For contrast, eastmoney snapshots show Micron at 22.96x trailing and SanDisk at 24.29x — a different basis; the gap rests entirely on large upward 2027 earnings assumptions readers must judge for themselves.) 𝕏

@BrianRoemmele (≈945 likes / 195 reposts — the highest-engagement semiconductor post in this search) on AMD/Taalas: "AMD bought the company that etches entire AI models into silicon... Taalas does not load model weights from HBM. It is the model... 17,000 tokens per second... No HBM. No CoWoS." Caveat: the deal description is KOL relay; no first-hand whitelisted-media report on this acquisition surfaced in this search, so treat it as unverified. 𝕏

@youngstockuser (≈17 likes) relaying BofA: a $3.2 trillion global semiconductor market by 2030, memory short of capacity and needing expansion, naming $MU, $LRCX and $AMAT — concluding that the AI fight is "gradually becoming about who holds the most capacity." 𝕏

Bearish / Cautious

@alphaticaio (≈19 likes / 3 reposts): "Calls to slow frontier-model development triggered a global semiconductor selloff... The market immediately priced the end of the AI spending boom. But slowing frontier training is not the same as stopping AI demand." — effectively short-term bearish, longer-term constructive, framing the drop as an overreaction. 𝕏

@IamLadyChelle (≈2 likes, minimal engagement — cited only as a sentiment sample): "AI CHIPS EXPECTED DOWN... AI leaders are talking about slowing frontier model development... short-term sentiment is bearish."

@cnfinancewatch (≈4.7k views / 1 like) pushing back on the China domestic-substitution narrative: domestic equipment substitution is far harder than headlines suggest — spec sheets and lab runs are one thing, but between a prototype and stable volume production lie enormous amounts of data and long iteration; even when a full tool is built, core components such as RF power supplies, electrostatic chucks and high-precision flow meters remain heavily dependent on overseas supply chains; much of today's localization is forced by external constraints rather than achieved readiness — "don't fantasize about catching up in a year or two; this is a long war of attrition." 𝕏

Trade media @SemiEngineering struck a neutral-to-cautious engineering note: "The difficulty in building a package digital twin goes beyond creating accurate electrical, thermal, or mechanical models"; and "Rising power density and chiplet complexity drive thermal considerations earlier in the design process" — pointing to hardening engineering bottlenecks in advanced packaging. 𝕏 𝕏²

A contradiction worth stating plainly: the dominant X narrative this week was "AI slowdown → chip stocks crash," yet by exchange snapshot the group closed Friday broadly higher (SOXX +2.69%). The crash was a mid-week single-session event, not the week's closing state.


4. Buzz & Sentiment Shifts

US semis overall: discussion volume was clearly higher than last week, but the driver changed — last week's staple was HBM/memory pricing; this week the AI-safety and "slow frontier development" macro-narrative took the headlines, shifting discussion from industry data toward policy and ethics, and directly producing a 5.9% single-day SOX decline. The bull/bear split runs roughly 6:4 bullish: the highest-engagement posts (Citi/RBC HBM pricing, AMD-Taalas) sit almost entirely on the bull side, while bearish posts drew only single-digit to low-double-digit likes — the bearish case is emotional and lacks high-influence backing. Friday's rebound was driven by Huang's "twice as many chips next year" line and continued validation of the memory/equipment pricing chain.

Memory (MU/SNDK/WDC/SK Hynix): the most concentrated and most one-sided theme of the week. Two numbers — 80–100% HBM price increases for 2027 (RBC) and undersupply through 2031 (Citi) — were repeatedly amplified, making "cheap multiple + rising prices" the standard bull template. Prices confirmed the same direction: SanDisk +10.99%, Western Digital +4.13%, Micron +3.92% on Friday.

Equipment (AMAT/LRCX/KLAC/ASML/TER): moderate buzz, clearly bullish tilt, catalyzed by Applied Materials' $5B India investment and by @jukan05's reporting of doubled component lead times (see Section 5). All five equipment names rose 3–7% Friday.

China semiconductor chain: Chinese-language discussion on X was extremely thin — of the China-assets KOL whitelist, only @cnfinancewatch produced substantive semiconductor content this week; the rest had none. Low buzz, cautiously constructive tone: the Huawei launch expectation lifted sentiment, while the single deepest take (@cnfinancewatch) was explicitly cooling the domestic-equipment story. Prices rose broadly on Friday, with SMIC H +4.41% and Hua Hong +4.81%.


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

Date/TimeCompanyEvent typeOne-line eventStatusSource
09-07Samsung FoundryCapacity allocation / HBMOver 50% of 4nm capacity (~15,000 wpm) allocated to HBM4 base die mass production for Nvidia/Broadcom/AMDConfirmed (Korean media / industry)@jukan05
09-12MicronHBM expansionPlans to add 60,000 wpm of HBM capacity by year-end, lifting total to ~100,000 wpm; HBM4 12-High to reach 50% of mixConfirmed (Korean media / industry)@jukan05
09-14TSMC / Apple2nm mass productionApple's A20 Pro entered TSMC 2nm mass production in early 2026, destined for the iPhone 18 Pro lineConfirmed (media)@dnystedt
09-15TSMC / MediaTek / Google / AWSAllocation / advanced packaging3nm, 2nm and CoWoS remain in shortage, mostly locked up by Nvidia and Apple; MediaTek seeking more capacity; Google TPU v9 splitting CoWoS-L orders between TSMC and Intel EMIB-T; AWS expanding 3nm ordersMedia report (parts are industry rumor)@dnystedt
09-15TSMC (capacity data)Capacity ramp2nm capacity to hit 100,000 wpm and 3nm 185,000 wpm by late OctoberMedia report@dnystedt
09-16MediaTek / TSMCProduct launch / nodeDimensity 9600 Pro built on TSMC 2nm; Dimensity 9600 on 3nmConfirmed (launched)@dnystedt
09-16Samsung Electro-Mechanics / TSMC / SEMESAdvanced packaging / glass substrateSEMCO submitted glass-substrate TGV module samples to TSMC; initial-process yields above 90% but subsequent-process yields drop to ~30%; SEMES signed a Single Axis Stage supply deal with HIWIN's Korean subsidiary for next-gen HBMConfirmed (sampling + contract); yields are industry data@jukan05
09-16Samsung / TSMCHBM dual sourcingSamsung working with both Samsung Foundry and TSMC on custom HBM base dies, read as a strategy to avoid supply-chain isolationRumor / industry interpretation@jukan05
09-16Samsung Foundry Taylor fab / TeslaNew capacity rampTaylor fab began operating ahead of the planned November start and recently reached 30% utilization; the customer triggering the early ramp is Tesla, key product the next-gen AI5 chip; full utilization expected around year-endConfirmed (industry); full utilization is expectation@jukan05
09-17Semi equipment supply chainLead timesLead times for essential equipment components more than doubled: deposition parts from four months to as long as 10 months; up to 40 months for critical components from Japan; ~50% longer than usual overallConfirmed (industry)@jukan05
09-18SK hynix / SolidigmNew capacitySolidigm reviewing construction of its first NAND mass production line in the US East Coast; site and investment timeline still under discussionRumor / under review@jukan05
09-18Applied MaterialsCapex / R&D investment$5B investment in India over the next decade, including a 140-acre advanced semiconductor research park with cleanroom; India R&D headcount to doubleConfirmed (company announcement)@dnystedt
09-18Nexperia / Tata ElectronicsFoundry orderNexperia tapped Tata Electronics to manufacture and package its power control chips (MOSFETs) in India — Tata's first major publicly announced foundry + packaging clientConfirmed@dnystedt

Interpretation of the key events

① Doubled equipment component lead times (09-17) — the most under-appreciated signal of the period. Deposition parts stretching from four to ten months, and certain Japanese critical components out to 40 months, means the physical pace of fab expansion is now gated by upstream components. For equipment vendors (AMAT/LRCX/TEL/ASML) this cuts both ways: backlog visibility lengthens and order quality improves, but revenue recognition slips and gross margin is squeezed by component costs. For fabs it is an outright negative — when Micron aims to add 60,000 wpm of HBM capacity by year-end and Samsung's Taylor fab targets full utilization by year-end, a 40-month lead time on critical parts says the bottleneck in this AI capacity cycle is shifting from "will they invest" to "can the tools actually be installed." That, in turn, gives a concrete supply-side foundation to Citi's call for undersupply through 2031: it is not that producers don't want to expand, it's that they cannot expand quickly.

② Samsung shifting >50% of 4nm capacity to HBM4 base dies (09-07) plus dual-sourcing base dies with TSMC (09-16). Taken together, these reveal a structural change in the HBM4 generation: base dies are migrating from legacy DRAM processes to advanced logic nodes, expanding competition from "memory maker vs memory maker" to "memory maker × foundry" alliances. Samsung committing more than half of its own 4nm capacity is trading foundry capacity for HBM4 share; simultaneously engaging TSMC suggests it is hedging on its own foundry yields. For TSMC this is AI demand spilling over from the GPU die into the memory stack; for Micron and SK hynix it means HBM cost structures will increasingly depend on which foundry they lock in.

③ TSMC's triple shortage in 2nm/3nm/CoWoS, with MediaTek squeezed out of the priority queue (09-15). Even with 2nm at 100,000 wpm and 3nm at 185,000 wpm by late October, capacity is insufficient and largely locked up by Nvidia and Apple. This directly explains AMD's ~10% price increase this week — TSMC's pricing power is being forcibly transmitted down the chain, with AMD's hike and Nebius's GPU rental increase as the second and third batons in the same relay. For second-tier designers at the back of the allocation queue (MediaTek et al.), the risk is not demand but access to capacity. For TSMC itself it re-confirms capacity as a strategic asset — consistent with stratechery's 2026-01-21 piece noting TSMC's own admission that it under-invested relative to AI demand.


Disclaimer: This report compiles publicly available information and real-time X platform search results for informational and educational purposes only. It does not constitute a recommendation to buy or sell any security, nor personalized investment advice for any individual's financial situation. Views expressed in cited posts belong to the respective accounts and do not represent this platform's position; items labeled "rumor/expectation" have not been confirmed by the companies involved. Market data may be delayed or contain errors; refer to official exchange data. Investing involves risk.

Want a daily X pulse for your own watchlist?

$10 free credit on signup · Just $5 per Mtoken · 30% off first top-ups

Start Free

This content is AI-generated from public posts on X (Twitter), for reference only and not investment advice. Investing involves risk.