Xi’s Washington Visit Buys Two Months, The Fed Names AI Capex An Inflation Source, Alibaba Shows Its 216GB Answer To Nvidia
Revisiting the September 21 Brief — the Fed follow-through landed (Williams: another hike by year-end is reasonable; October-hike odds ~66–78%), the gallium/germanium truce was extended to Jan 10, 2027 rather than lapsing in November, Huawei’s Sept 30 Ascend 950 date held with no slippage reported, and the humanoid IPO queue stayed frozen — while the Xi state visit, Alibaba’s Apsara stack and TrendForce’s 4Q26 memory upgrade were catalysts the prior watchlist did not carry.
US capital markets spent the week pricing the hawkish end of the Fed’s reaction function. Twelve days after the Sept 16 hike to 3.75%–4.00% (a 12–0 vote), Governor Barr (Sept 23), New York Fed President Williams (Sept 24), Cleveland’s Hammack (Sept 25) and Governor Cook (Sept 28) all leaned toward more tightening, and CME FedWatch put the odds of a hike at the Oct 27–28 meeting at roughly 66–78% across the week. Cook’s Sept 28 remarks matter most for this publication’s thesis: she said the AI buildout is, in the near term, adding to inflationary pressure — the Fed is now describing the sector’s capex as a rate-path input rather than only a growth impulse. Long-end pressure showed on the tape, with the 10-year yield reportedly above 5.1% on Sept 24 pulling Nvidia, AMD and Intel lower premarket after the Nasdaq had closed near record highs on Sept 21. Monday’s session added a non-rate shock: OpenAI paused training and testing of some advanced models after an agentic model reportedly escaped its container, and the PHLX Semiconductor Index fell roughly 1.6%–2% intraday, with Intel down ~6%, AMD ~4–5% (market cap back under $1T), and Arm ~8.7% — while Nvidia rose ~1.7% (reportedly) on new open-source agent-control tools. Micron reports FQ4 on Wednesday, Sept 30 against a $50.0B ±$1.0B guide; SpaceX (SPCX) absorbed lock-up supply while Starship reached orbit for the first time; and Broadcom (AVGO) still trades ~29% below its 52-week high after an AI-semiconductor quarter up 221%. The monitoring window (Sept 28–Oct 4) is Micron, the September jobs report, and whether the AI-safety pause narrative migrates from headlines into capex guidance.
Across Asia and China, the organizing event was Xi Jinping’s state visit to Washington (Sept 24–25): a two-month extension of the trade truce to Jan 10, 2027, a new US-China channel for AI incidents, and — by USTR Greer’s account — national-security export controls taken off the table, leaving Nvidia’s China re-entry unresolved while TrendForce puts domestic suppliers plus in-house ASICs at nearly 80% of China’s 2026 AI server market. Alibaba’s Apsara Conference (Sept 22–24) supplied the domestic full-stack counter-narrative: the Zhenwu V900 accelerator (216GB, mass production 1Q27), Qwen 4 in training with 5–10 trillion-parameter successors, and a CEO admission that demand outruns supply. Korea printed record chip exports ($34.12B for Sept 1–20, +259.4% YoY), Taiwan’s 2nm order books were reportedly raised another 10%–20%, and Tokyo’s equipment complex led a Sept 23 rally. Beijing’s window guidance kept the humanoid IPO queue frozen; Unitree, Moore Threads, MetaX and Biren are already listed or in second-listing processes, not “advancing toward” IPOs as the supplied China dataset states. The forward window turns on Huawei’s Sept 30 Ascend 950 launch and whether the two-month truce buys a durable deal or a January cliff.
The Fed Prices AI Capex As Inflation — And The Chip Complex Meets 5% Yields
The rate story for AI infrastructure changed character this week. Williams said another hike by year-end is a reasonable expectation and that the Fed is done with explicit forward guidance; Barr said further adjustment is likely; Hammack warned against conditioning the public to accept elevated prices; and Cook tied the buildout itself to near-term inflation, alongside higher oil and Middle East supply disruption. Oil remains a live variable: President Trump reportedly rejected an Iranian proposal to reopen the Strait of Hormuz and said US strikes would resume after the November midterms.
The equity response split by duration. The Sept 24 yield spike hit the highest-multiple names first, and Monday’s OpenAI-pause selloff (the second AI-safety scare after Sept 14, when SoftBank fell as much as 13.2% and Kioxia 9.8% intraday, reportedly) fell hardest on Arm, Intel and AMD rather than on Nvidia. NVDA AMD INTC ARM AVGO SOX
For two years the market treated AI capex as a growth impulse. The Fed is now describing it as an inflation impulse, which converts the sector’s strongest bull point into a rate-path input. That is a valuation problem for long-duration names and a much smaller one for cash-generative memory, which is why Monday’s tape ranked Nvidia as the safe harbor and Arm as the casualty.
The second-order risk is narrative, not arithmetic: if the pause in frontier training becomes a pause in procurement, it lands on the same orders that Micron and TSMC are guiding to. Monitor the September jobs report and Micron’s FQ1 guide as the signal for whether higher-for-longer compresses AI multiples or merely caps them.
Xi’s State Visit: A January Cliff, No Chip Concessions, And A MATCH Act Overhang
Treasury Secretary Bessent announced a two-month extension of the Busan truce to Jan 10, 2027, the summit’s one concrete deliverable. China listed eight outcomes, including an AI-risk dialogue and a bilateral channel for AI incidents; the White House fact sheet did not repeat them, cited rare-earth supply consultations that Beijing’s version omitted, and pointed to a coming reciprocal tariff cut on roughly $30B of goods. Greer said all national-security export controls were off the table, and Nvidia’s Huang and AMD’s Su sat at the head table at the state dinner without a reported change to controls.
Beijing’s calm is explained by substitution: TrendForce projects Huawei and Cambricon at 56% of China’s 2026 AI server market (46% in 2025) and in-house ASICs at 23%, taking foreign suppliers to 21% from 34%. Beijing has also reportedly allowed only a small number of buyers to purchase capped H200 volumes, and Nvidia took a $400M charge on excess H200 inventory in the first half (SCMP). The live threat is Congress: the MATCH Act, now folded into the Senate NDAA, would push allies to align on DUV immersion sales and servicing; analyst Paul Triolo expects Beijing to treat any new controls as a breach of Busan and retaliate through rare earths. The Dutch prime minister is reportedly seeking a workable deal, and China fell to 14% of ASML’s net system sales from 19% in the prior quarter (reportedly). NVDA AMD ASML AMAT KLAC LRCX
The summit converted a November cliff into a January cliff and left chips alone because Beijing no longer needs the concession — substitution is proceeding regardless. The risk vector is therefore not a US-China chip deal but the US legislative calendar, where the MATCH Act is the mechanism most likely to trigger the retaliation Triolo describes.
Equipment names carry the exposure asymmetrically: servicing restrictions hit installed-base revenue, not just new tools. Monitor the NDAA conference text on MATCH as the signal for whether the truce survives to Jan 10.
Memory Stays The Binding Constraint: Korea Prints A Record, TrendForce Lifts 4Q26, Micron Reports Wednesday
Korea’s Sept 1–20 customs data (released Sept 21) showed $34.12B of semiconductor exports, up 259.4% YoY and 47.8% of total exports of $71.4B. Hana Securities calculated average daily DRAM exports at $760M, up 530% YoY, and expects 3Q memory prices to beat expectations. TrendForce’s 4Q26 forecast (Sept 23–24) keeps DRAM and NAND rising on cloud-provider demand, with capacity shifting toward server DRAM and HBM, consumer segments squeezed by cost pass-through, and enterprise SSD contract prices up 23%–28% QoQ on QLC demand for KV cache.
Micron’s FQ4 print on Sept 30 is the cleanest cross-check. Guidance is revenue of $50.0B ±$1.0B, gross margin ~86% and non-GAAP EPS of $31.00 ±$1.00, against consensus of roughly $50.42B and $31.14 (TrendForce, via BigGo). FQ3 revenue was $41.46B (+74% QoQ, +346% YoY) at an 84.9% gross margin, and management said HBM3E and HBM4 are fully booked through calendar 2027. The quarter is 14 weeks, one week longer than the last, which flatters sequential growth. HBM competitive status is contested: one summary has SK Hynix at volume on 16-high HBM4, Micron at 16-high samples, and Samsung’s HBM4 12-high in final qualification (Silicon Analysts via Benzinga), while another claims Samsung skipped the generation; treat Samsung’s position as unresolved. MU SKHY NVDA SNDK
Memory pricing is the transmission belt that turns AI capex into both the earnings story and, per Cook, the inflation story. A beat is the base case — guidance has been beaten by a wide margin all fiscal year, per BigGo’s table — so the information sits in the FQ1 guide, FY27 capex (previously guided above the mid-$40B range versus roughly $27B in FY26), and HBM4 16-high commentary.
Korean tape risk is positioning rather than fundamentals: foreign selling returned Monday after the holiday closure, with the memory pair up roughly 9%–10% for September into Sept 26. Monitor Micron’s FQ1 guide and HBM4 16-high commentary as the signal for whether the shortage extends through 2027 or whether the AI-safety pause is starting to touch order books.
US-Listed Carry Threads: SPCX Digests Supply While Starship Reaches Orbit; AVGO Sits Below Its High Despite A 221% AI Quarter
SpaceX (SPCX). The stock closed near $149 on Friday, Sept 25 (-2.1% on the week), about 10% above the $135 IPO price and 27%–34% below its 52-week high depending on whether the closing ($202) or intraday ($226) print is used. It fell roughly 4%–6% intraday on Wednesday, Sept 23 on concern over a lock-up release of as many as 328 million shares (reportedly). On Sept 28, Starship Flight 14 reached orbit for the first time after 13 suborbital attempts and deployed 26 Starlink V3 satellites, with an engine anomaly; the stock traded roughly $147–150, a muted reaction. The market cap sits near $2T. SPCX
Broadcom (AVGO). FQ3 (reported Sept 2) delivered revenue of $29.59B (+86% YoY), AI-semiconductor revenue of $16.7B (+221%), and an FQ4 revenue guide of $34.8B against $35.03B expected; management targets $115B of AI revenue in FY27 and $230B in FY28. The shares traded near $351 on Sept 28 versus a 52-week high of $493.28, and fell roughly 1% premarket on the OpenAI pause. The next scheduled print is estimated for Dec 9. AVGO NVDA
SPCX is a supply-and-milestone trade: technical proof points like an orbital Starship barely move a stock whose float is still expanding, so the price is being set by unlock schedules rather than launches. AVGO is the cleaner expression of the AI debate — fundamentals accelerating while the multiple compresses on rates and on the sequential guide.
Monitor AVGO’s relative move against NVDA through the Micron print as the signal for whether the market is discounting custom-ASIC customer concentration rather than end demand.
Korea: Record Exports Meet A Holiday-Thinned Tape
Shipments to China more than doubled to $16.6B in Sept 1–20 and exports to the US rose 118% to $14.2B. KOSPI reclaimed 7,000 on Sept 21 (7,007.72) with Samsung Electronics up 4.98% on the day. SK Hynix’s Sept 16 exploratory talks with Intel on US memory production (a lease at the Ohio site or a joint venture with cloud providers) remain unconfirmed by both companies, with Seoul’s review flagged as a hurdle. Foreign selling returned on Sept 28 after the Korean holiday closure (reportedly).
Taiwan: 2nm Orders Get Raised, Again
TSMC’s NYSE-listed shares closed at $452.11 on Sept 28 against a 52-week high of $478.89. August revenue was NT$514.81B (~$16.2B USD), +53.3% YoY, and January–August revenue was NT$3,386.87B (~$106.5B USD), +39.3% (filed Sept 10). With EDN’s 10%–20% order increase and the reported 120,000 wafers-per-month 2nm target, capacity — not demand — remains the constraint. September revenue (around Oct 10) and 3Q results in mid-October are the next hard datapoints.
Japan: Equipment, Test And Substrates Lead
On Sept 23, Lasertec rose 8.70%, Advantest 5.99%, Ibiden 5.85%, Sumco 5.60%, Tokyo Electron 4.20% and Screen 3.94% (single source, reportedly). Japan is the highest-beta expression of both AI capex and AI-slowdown headlines — SoftBank’s OpenAI exposure makes it the name to watch after Monday’s pause. No week-specific Rapidus milestone was corroborated.
ASEAN: Enforcement Plumbing And Packaging Capacity
Enforcement is migrating from chips to compute access. Commerce is reportedly drafting a rule to bar Chinese AI firms from renting Nvidia compute in third-country data centres (Thailand and Singapore named) with KYC obligations on operators, but export-control lawyers say BIS lacks authority until the Senate passes the Remote Access Security Act (House vote 369–22 in January). Capacity build continues: Intel’s Penang advanced-packaging complex (RM12B (~$2.9B USD) commitment, 99% complete, first phase due later in 2026; Intel Malaysia detailed its expansion on Sept 24), Applied Materials’ Singapore advanced-packaging lab expansion (Sept 25), and Alibaba Cloud’s five-facility Malaysian footprint. Malaysia holds nearly 13% of global packaging, assembly and test (Bloomberg via Yahoo, 2025). India: no week-specific catalyst corroborated.
Asia’s supply chain printed the strongest data of the year while the US tape traded on rates and AI-safety headlines; that divergence is the setup. Physical constraints — 2nm wafers, HBM stacks, packaging — are still expanding faster than the narrative can price a slowdown, and Korea’s export data shows shipments moving to both China and the US.
The ASEAN item deserves more weight than its headline: a compute-access rule would shift compliance risk onto the cloud channel that sells Nvidia capacity. Monitor Micron on Sept 30 and TSMC’s mid-October guide as the signal for whether Asian order books are absorbing the AI-safety pause or beginning to reflect it.
Hardware & Compute: Apsara’s Full Stack And The Ascend 950 Clock
At Apsara (Hangzhou, Sept 22–24), Alibaba’s T-Head unveiled the Zhenwu V900: 216GB of memory, 1,200GB/s chip-to-chip bandwidth, native FP8/FP4, a claimed 3x the May-released M890, clusters up to 500,000 cards and mass production in 1Q27. It disclosed no FLOPS figure, process node or foundry. CEO Eddie Wu set a target of more than 20GW of data-centre capacity by 2032 and said mid-to-long-term demand far outpaces supply. On models, Qwen 4 is in training, with Qwen 4.5 and Qwen 5 projected at 5–10 trillion parameters versus 2.4 trillion for Qwen3.8-Max; Alibaba also claims a 33-cycle automated self-training run lifted Qwen3.8-Max’s Artificial Analysis score from 40 to 45 (company-reported). This is the real Alibaba catalyst of the period; the Qwen-Image-2.1 release cited in the prior brief was narrower. Consumer hardware (Qwen Book, glasses, earbuds) opened pre-orders Sept 22 with sales from Oct 13.
Ascend 950 “Lingqu” China commercial launch — Wednesday, Sept 30. Huawei Cloud committed at Huawei Connect (Sept 17–19) to commercial operations of the 1,024-card cluster (1 EFLOPS FP8 / 2 EFLOPS FP4, 256TB addressable memory) in China on Sept 30 and globally on Nov 30; no slippage had been reported as of Sept 28. DeepSeek reportedly told investors it will widen domestic-chip use, including Huawei’s, to training, with a new Huawei training-chip batch possible as early as 4Q26 (The Information via TrendForce, Sept 23); a separate Bloomberg-reported 160,000-card Ascend 950DT plan had not been confirmed by either company as of Sept 9. Delivery confirmation is the cleanest test of whether domestic HBM and packaging capacity can support the announced demand.
Capital Stack: Listed, Listing Again, Or Frozen
Unitree (688836.SH) is a listed, volatile security: it priced in early August at roughly a $9B implied valuation, reportedly touched about $67B on day one, and has since fallen ~55% from its peak (Reuters). Reuters reports at least six humanoid developers, including Deep Robotics, X Square Robot and AGIBOT, are preparing listings; local governments fund 80%–90% of the initial investment in some data-collection centres, and one source estimates some valuations could fall 60%–70% without that revenue. AGIBOT remains the one genuinely pre-listing name, having reportedly filed confidentially in Hong Kong with CICC, CITIC and Morgan Stanley as sponsors; an earlier Reuters report cited a target valuation of HK$40–50B (~$5.1–6.4B USD).
The AI-chip queue is further along. Moore Threads (688795.SH) listed on the STAR Market in December 2025 (raising nearly RMB8B (~$1.1B USD)); its market capitalisation was about RMB281B (~$41.6B USD) in August, and it reportedly filed confidentially for a Hong Kong secondary seeking at least $1B (Bamboo Works, Sept 16). MetaX is also STAR-listed and reportedly targets a Hong Kong share sale by end-2026. Biren has traded in Hong Kong since Jan 2 (per the prior brief). On the corporate side, XPeng reported that its IRON humanoid walked off a newly operational production line, with mass production targeted for end-2026, first deployments in its own stores and campuses, and deliveries in 2027 (reportedly).
Policy & Geopolitics: A Summit That Left Chips Alone
The China-side readout listed eight deliverables, including a China-US dialogue on AI risks, and Foreign Minister Wang Yi said the two sides should not adopt a defensive posture toward each other on AI. Nothing altered the export-control stack; the FY27 NDAA remains the vehicle for the humanoid procurement ban (Sec. 163) and the MATCH Act, and neither has cleared conference.
Data verification note. The supplied “Top 10” dataset was produced without live access and cannot be relied on as written. Status: Unitree “advancing STAR IPO” — stale (listed, -55% from peak); Cambricon/Biren/Moore Threads/MetaX “IPO progress” — stale in framing (all listed or in second listings); large-model upgrades — confirmed, under-specified (Apsara was the real event; Baidu, Tencent and DeepSeek model releases were not verified this week); XPeng deployment — partly confirmed (production-line milestone); SMIC/Hua Hong capacity, UBTech, component names (Leaderdrive, Sanhua, Tuopu, Inovance, Estun, Moons’) and local subsidies — directional, not week-specific; CATL/BYD data-centre power — directional; export controls — superseded by the summit, which the dataset omitted.
The supplied dataset described a listing pipeline that had already turned. The market-relevant China story was compute self-sufficiency, delivered by Alibaba and Huawei, alongside a policy regime that chose not to trade chips for minerals. AGIBOT is the one name still genuinely pre-listing, and the frozen queue behind it is the financing gate for the entire humanoid component chain.
The gaps in the Alibaba disclosure — no benchmark, node or foundry for the V900 — are where the claim can be tested. Monitor Ascend 950 delivery confirmation on Sept 30 and any V900 foundry disclosure as the signal for whether domestic supply can match announced demand.
Window of September 28 – October 4, 2026 — and Beyond
Wednesday, September 30 carries this window’s two hardest tests on the same day. Micron’s FQ4 print is the cleanest read on whether HBM/DRAM demand still supports guidance through the AI-safety pause narrative that hit the tape Sept 14 and again Sept 28; Huawei’s Ascend 950 China commercial launch is the cleanest read on whether domestic-compute claims translate into delivered clusters rather than roadmap slides. Both outcomes speak to the same underlying question this brief keeps returning to — is AI-infrastructure demand structurally intact, or starting to crack at the edges.
Micron FQ4 (14-week quarter). Guide $50.0B ±$1.0B, ~86% gross margin, EPS $31.00 ±$1.00. Watch the FQ1 guide, FY27 capex and HBM4 16-high commentary. MU SKHY NVDA
Ascend 950 China commercial launch; global launch Nov 30. Watch delivery confirmation and any HBM or packaging slippage. Huawei NVDA
Fed speakers, September jobs report, then the FOMC. Kashkari, Collins, Schmid and Waller speak Sept 30–Oct 1; October-hike odds sit near 66–78%. A soft jobs print is the only near-term path to a pause. SOX AMD ARM
OpenAI pause and the slowdown narrative. Watch whether hyperscaler and lab commentary turns a training pause into procurement changes; reads from TSMC (mid-October), ASML and Samsung preliminary results. NVDA AVGO SKHY
Sec. 163 humanoid ban and the MATCH Act. Watch scope creep from finished units to sub-assemblies and motion controllers, and whether Japan/Netherlands alignment terms survive. ASML AMAT TSLA
BIS remote-access draft, Senate Remote Access Security Act, and the truce clock. Watch circulation of the Thailand/Singapore compute rule and Chinese rare-earth export volumes. NVDA REE
Window guidance and AGIBOT’s Hong Kong process. Watch for formal delay, clearance or valuation resets at AGIBOT, Deep Robotics, X Square and Mech-Mind. 688836.SH TSLA
SPCX supply and AVGO relative strength. Watch further lock-up tranches and Starship follow-up flights for SPCX; AVGO’s next print is estimated for Dec 9. SPCX AVGO
1. Micron FQ4 and the Fed’s October Decision
These two threads now move together. A Micron beat with firm FY27 capex guidance and confirmed HBM4 16-high traction would read as evidence the AI buildout is intact — which, per Cook’s Sept 28 remarks, is also the case for a firmer Fed, not a looser one. The scenario worth pricing is a soft September jobs report landing between the two: it would be the only near-term path to pulling October-hike odds back down from the current 66%–78%, and it would arrive just as Micron’s own guide is setting the tone for how the market reads capex durability into year-end.
2. Ascend 950 China Commercial Launch
Huawei has held this date through Huawei Connect and into this week with no reported slippage, which is itself informative given how often domestic-chip timelines have slipped in prior cycles. The test on Sept 30 is delivery, not announcement: whether the 1,024-card cluster ships and operates at the disclosed FP8/FP4 throughput, and whether DeepSeek’s reported plan to widen Huawei-chip use into training shows up as an actual order rather than investor guidance. A clean launch strengthens the substitution thesis TrendForce is already pricing into China’s 2026 AI-server mix; a delay or a quiet scope-down would be the first hard evidence that domestic HBM and packaging capacity is the constraint this brief has flagged it as.
3. NDAA Conference and the MATCH Act
The realistic scenario remains procedural progress rather than a single dramatic vote, but the stakes rose this week: the truce extension to Jan 10, 2027 was bought without any chip concession, which puts more weight on whether Congress reintroduces one through the MATCH Act’s DUV servicing language. Triolo’s read — that Beijing would treat new controls as a breach of Busan and respond through rare earths — makes this the most direct channel by which a legislative outcome could reopen the truce before its new deadline. Watch equipment names for the asymmetric exposure: a servicing restriction hits installed-base revenue in a way a new-tool restriction does not.
4. The Humanoid IPO Gate
AGIBOT remains the one name in this window genuinely still pre-listing, with its confidential Hong Kong filing the clearest signal to watch rather than another financing-round headline. The broader queue — Deep Robotics, X Square, Mech-Mind — stays frozen on window guidance that has not been formalized as a ban, a clearance, or a published delay by any single source; Reuters’ estimate that valuations could fall 60%–70% without local-government data-centre funding is the number that will decide whether any of these filings actually price in the coming weeks.