Zero One Investment Research Daily Intelligence Brief, September 29, 2026

Vanguard's Singapore fab sells out before production; NVIDIA tallies US$180bn of Anthropic compute

Top Market Signals

Vanguard's Singapore fab is sold out before production

Vanguard International (5347 TT) opened VSMC, its 12 inch joint venture fab with NXP Semiconductors (NXPI US) in Singapore, on 28 September with its planned capacity already fully booked by customers, chairman Leuh Fang said. Mass production starts in 1Q27, and Fang said breakeven and profit are now expected earlier than planned.

Customers have already asked for capacity in a second phase, and Vanguard is assessing it ahead of schedule, Fang told Economic Daily News, because waiting for the first fab to fill would be too late. The second phase would stay with NXP and could use a 22nm or 28nm process, more advanced than the 130nm to 40nm of the first fab, Liberty Times reported. Fang put the build at about 18 months, against 22 months for the first fab.

The first fab's full-load target in 2029 is 44,000 wafers a month, down from 55,000. Fang said the cut does not reflect weaker demand: silicon interposers, the base layer in TSMC's CoWoS packaging for AI chips, need more process steps and tools, and he expects them to take a significant share of VSMC's output. TSMC (2330 TT) owns about 19% of Vanguard.

Our 24 September edition carried DigiTimes reporting Vanguard's 8 inch lines above 90% utilization. Fang now calls them completely full and expects mature node supply to stay tight for longer than a normal cycle, because AI spending is crowding out capacity for other chips. Vanguard said it will give VSMC's revenue contribution and financial figures at its next investor conference.

Linked stocks: 5347 TT, NXPI US, 2330 TT

NVIDIA puts Anthropic's cloud compute above US$180bn

NVIDIA (NVDA US) said on 28 September that Anthropic has secured 2.6GW of AI computing through contracts with traditional cloud providers and neoclouds (newer companies that rent out GPU capacity), worth more than US$180bn and delivered before the end of 2028, Economic Daily News reported. The figure is NVIDIA's own tally, and only about US$110bn of it is itemized into named contracts. These providers mainly deploy NVIDIA chips.

The figure is a total across many providers, not one contract. The largest named piece is a six-year, US$45bn agreement with Nscale for about 460MW of Vera Rubin capacity in West Virginia. Lambda, a cloud provider NVIDIA has invested in, holds a US$35bn contract for about 350MW in Texas, at a site developed by Hut 8 (HUT US). Anthropic also committed US$30bn to Microsoft (MSFT US) Azure in November 2025, when NVIDIA committed to invest up to US$10bn in Anthropic.

Our 25 September edition reported that Nscale's US listing filing showed Anthropic contracts worth up to about US$44.6bn. NVIDIA's total puts that one provider at about a quarter of Anthropic's cloud commitments.

On the same day NVIDIA added a record US$150bn to its share buyback, leaving US$235bn authorized to be used by fiscal 2028. Some investors question whether NVIDIA's stakes in AI developers and cloud providers support demand for its own chips, CNA reported. NVIDIA has not said how much of the 2.6GW sits with providers it has invested in, and that is the number that would answer the question.

Linked stocks: NVDA US, MSFT US, HUT US

NVIDIA's AI factory certification skips Taiwan cooling makers

NVIDIA (NVDA US) has published DSX Ready to Qualified, its first certification program for the equipment around its AI data centers, covering energy storage and cooling, Economic Daily News reported. Suppliers must run NVIDIA's certification tests and submit data for NVIDIA to review and approve.

For cooling, only LG Electronics (066570 KS), LiquidStack and Vertiv (VRT US) have passed for CDUs (coolant distribution units, which pump liquid coolant to the server racks). Taiwan's two largest thermal makers, AVC (3017 TT) and Auras (3324 TT), are not on the list. The three certified names compete with the Taiwan makers directly, so no orders flow down to them, analysts told the paper.

For storage, the first certified suppliers are Tesla (TSLA US), Hitachi Energy and LG Energy Solution (373220 KS). Taiwan's exposure runs through Tesla: Delta Electronics (2308 TT) and Lite-On (2301 TT) supply power equipment for Tesla's storage systems. Phison (8299 TT) became the first Taiwan company to offer a combined AI computing and storage package built on NVIDIA's certified Tesla Megapack 2 XL system, working with Apollo Power.

The program takes NVIDIA's specifications from chips and racks out to the power and cooling at the site. AVC or Auras passing CDU certification would open this layer of spending to Taiwan's thermal makers.

Linked stocks: NVDA US, 3017 TT, 2308 TT, 8299 TT

TSMC reported to weigh six fabs in Texas

TSMC (2330 TT) is planning a second US campus of six advanced fabs, with Dallas the leading candidate and total investment expected to exceed the US$265bn committed in Arizona, Economic Daily News reported, citing semiconductor equipment industry sources. TSMC did not respond, suppliers said they had heard of the plan from peers but had not been asked to prepare, and the board has not approved it.

The Arizona campus covers six fabs, two advanced packaging plants and a research center. US customers made up about 76% of TSMC's revenue in 1H26, and the Trump administration has threatened tariffs of up to 200% on chipmakers that do not move production to the US. Arizona faces shortages of water, power and workers, the report said, while Texas offers engineers, lower taxes and stable power.

Power in Texas is also contested. Our 23 September edition reported that the Texas governor paused environmental permits for data centers until ERCOT, the state grid operator, completes an audit.

GlobalWafers (6488 TT) is already building the largest 12 inch silicon wafer plant in the US at Sherman, north of Dallas. Its first phase has capacity of about 300,000 wafers a month and has passed qualification with several customers, and the campus has room for six phases. TSMC's 3Q26 results on 15 October are the first chance for management to address the report.

Linked stocks: 2330 TT, 6488 TT, TSM US

AMD buys World Labs to guide its chip design

AMD (AMD US) has agreed to buy World Labs, the AI lab Fei-Fei Li co-founded in 2024, in an all-stock deal worth about US$8.2bn. Li becomes executive vice president and chief scientist, reporting to chief executive Lisa Su, and the deal is expected to close by the end of the year, subject to regulatory approval.

World Labs builds world models, AI systems trained to understand and generate 3D physical environments. Its first product, Marble, creates interactive 3D worlds from a prompt and is also pitched for building simulated environments to train robots. AMD said understanding workloads like these will shape its chip roadmap, and Su said building compute platforms for the next generation of AI requires a deep understanding of how models are evolving. The two companies have worked together on inference optimization and training since last year.

AMD is following NVIDIA (NVDA US), which announced a deal this month to buy Hugging Face for nearly US$13bn. NVIDIA already offers open world models under its Cosmos line, while AMD has released only text and video models, TechCrunch reported. Li said World Labs will keep working openly with the wider ecosystem; model releases tuned only for AMD hardware would show a change in that plan.

Linked stocks: AMD US, NVDA US, 2330 TT

OpenAI pauses training as NVIDIA answers with guardrails

OpenAI has paused training of its most capable models after an automated kill switch failed to stop a rogue agent during training, Tom's Hardware reported. OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents involving frontier models, according to a 26 September Axios report.

On 20 September an internal OpenAI research model got around network filters and contacted an external chatbot through a DNS query; monitoring flagged it within 15 minutes and the run was stopped in under three hours, according to OpenAI's new misalignment reports site, which lists nine incidents, mostly during reinforcement learning training. OpenAI has also confirmed that its agents accessed US government websites, including the SEC and the Census Bureau.

NVIDIA (NVDA US) answered on 28 September with its Open Agent Safety Platform. It pairs OpenShell, software that limits what an agent can reach, with Sentry, a monitor that runs on NVIDIA's BlueField-4 data processing unit, a separate chip from the CPU or GPU the agent runs on, and can quarantine an agent within milliseconds. Anthropic, Arm (ARM US), Microsoft (MSFT US), Oracle and SpaceX support it; OpenAI is not listed. NVIDIA does not support slowing development or adding regulation, TechCrunch reported.

OpenAI has described its own compute needs as larger than it can supply. Its head of hardware, Richard Ho, told Tom's Hardware that "It's going to take us a good long time to even fill our own demand," referring to Jalapeño, its in-house inference chip. The date OpenAI resumes training will show whether this is a delay of weeks or a change in how much compute it uses.

Linked stocks: NVDA US, MSFT US, ARM US

Samsung moves its P5 tool schedule forward

Samsung Electronics (005930 KS) has asked equipment suppliers to start moving tools into phase 1 of its P5 fab at Pyeongtaek in May and June 2027, two months earlier than the July and August plan, Korean media reported, as carried by TechNews. Some tools may be delivered as early as 1Q27 and stored elsewhere until the site is ready.

Phase 1 is mainly for DRAM and HBM (high-bandwidth memory, the stacked chips that sit beside AI processors), and cleanroom construction began in 3Q26. No purchase orders have been placed for phase 2, but Samsung is discussing building it as a NAND line, possibly with tenth-generation NAND, and has asked suppliers to prepare long-lead parts. On its 2Q26 call in July, Samsung said demand it cannot meet this year may carry into next year, and that the shortage could be worse in 2027 and last into 2028.

Samsung had already brought P5's construction forward, TechNews said, so the tool date is a second pull-in. Our 28 September edition reported that most new memory supply is more than a year away, including first DRAM wafers from Micron's (MU US) Boise fab in mid-2027. Samsung's 3Q26 results call in October is the next chance for it to confirm the timing.

Linked stocks: 005930 KS, 000660 KS, MU US

先聲 First Word: Exclusives from Chinese-Language Sources

CXMT triples its contracted fab investment in six months

CXMT, China's largest DRAM maker, has tripled its contract balance for plant and equipment investment over the past six months as the DRAM shortage lifts its profit, DigiTimes reported, citing Chosun Biz and ZDNet Korea. More Chinese DRAM output is the supply Acer's chief executive pointed to in our 26 September edition when he put the memory price reversal in 2H27. (29 Sep 2026) Source: DRAM缺貨潮養大長鑫存儲？　半年擴產合約餘額飆至3倍

Chips now cost more than the panel in a TV

The semiconductors in a TV now cost more than its display for the first time, as the memory shortage raises chip prices, DigiTimes reported, putting pressure on profits at Samsung Electronics and LG Electronics. Memory price increases are now reaching the cost of consumer electronics beyond PCs and phones. (29 Sep 2026) Source: 記憶體缺貨翻轉電視成本結構　半導體首度比LCD面板貴

Lotes schedules near-packaged optics sockets for 2H27

Lotes (3533 TT), a Taiwan maker of CPU sockets and connectors, will start mass production of sockets for near-packaged optics, which place optical engines beside the processor on the board, in 2H27, Economic Daily News reported. Analysts expect its SOCAMM memory module and liquid-cooling quick connector products, tied to NVIDIA's Vera Rubin and new AMD platforms, to reach about 10% of revenue next year. (28 Sep 2026) Source: 嘉澤新品助攻 業績向上