Zero One Investment Research Daily Intelligence Brief, October 7, 2026

Marvell ties a US$80bn target to won designs as SpaceX seeks chip debt

Top Market Signals

Marvell's fourfold revenue goal rests on won designs

Marvell Technology (MRVL US) used its 6 Oct investor day to guide fiscal 2028 (the year to January 2028) revenue to about US$20bn, up 67% YoY, and fiscal 2031 revenue to US$70bn to US$90bn. The US$80bn midpoint is four times the fiscal 2028 figure and compares with an analyst average of US$47.04bn.

Stifel's Tore Svanberg said Marvell built the US$80bn figure from its analysis of design wins it has already secured. That ties the target to customer chip programs Marvell has already won rather than to an assumed share of a growing market. Marvell also raised its fiscal 2029 target for custom AI accelerators and related products to US$12bn, from US$10bn at the end of August.

The customer list has widened. Marvell puts the AI chip market it can address at US$400bn by 2030, driven by in-house chip programs at cloud providers including Amazon (AMZN US), Google and Microsoft. Seaport's Jay Goldberg noted that a few years ago the main worry was Marvell's reliance on one product for one customer, Amazon.

The fiscal 2028 target was US$18bn in August, and FactSet's analyst average was US$18.17bn. Marvell shares rose 5.81% to US$287.01 on 6 Oct, and Broadcom (AVGO US) rose about 4% with it. The US$12bn fiscal 2029 custom accelerator figure is the nearer test of whether won designs turn into volume on schedule.

Linked stocks: MRVL US, AVGO US, AMZN US

SpaceX seeks US$40bn of debt for NVIDIA chips

SpaceX is seeking to raise about US$40bn to pay for an NVIDIA (NVDA US) chip order, made up of about US$10bn of bank loans and US$30bn of investment-grade debt, the Financial Times reported. Apollo Global Management (APO US) would lead the deal and sell the debt to a wide range of investors, and PIMCO is one of the few lenders in talks. The deal is planned to complete in 2027.

Elon Musk said on SpaceX's August earnings call that it had decided to use NVIDIA only, because Vera Rubin is the best architecture. Last month he said Colossus 2, the Memphis computing cluster of his AI company xAI, has 110,000 GB200 and 440,000 GB300 chips, with another 220,000 GB300 about to start, 220,000 more due in November, and possibly 220,000 more by the end of December.

This is the third large debt package for AI chips in a week. Our 2 Oct Blade reported Broadcom (AVGO US) agreeing to lend Anthropic up to US$42bn, and on 5 Oct banks began selling down a US$60bn chip financing for Anthropic. In SpaceX's case the chip buyer borrows directly, and its BBB rating, the second-lowest investment grade, lets insurers and pension funds buy the bonds.

Bond investors have already pushed back once. SpaceX sold US$25bn of bonds less than two weeks after its US$86bn IPO in June, and they fell in the following days on worries about its debt and capital spending. Its 2056 bonds now trade at about 85 cents on the dollar, 2.27 percentage points over Treasuries, close to junk bond levels. The pricing on the new debt will show how much more chip borrowing investors will fund for a single buyer.

Linked stocks: NVDA US, APO US, AVGO US

Nanya lifts DRAM contract prices up to 20%

Nanya Technology (2408 TT) has been telling customers it will raise DRAM contract prices again, by up to 20% on some products, industry sources told Economic Daily News. Nanya said it cannot comment on customer pricing.

The increase runs against the market's main worry. Investors had expected new capacity at Chinese memory makers such as CXMT to loosen supply and start a price fight. Commercial Times reported that new capacity takes time to become usable output, while AI servers keep taking the leading-edge capacity of the large memory makers and leave less for standard DRAM. The top of Nanya's range is above TrendForce's forecast of a 10% to 15% QoQ rise in 4Q26 standard DRAM contract prices.

The gain comes from price. Analysts say Nanya's shipments and unit costs are roughly stable, so higher contract prices go straight to revenue and widen the gap between price and cost. Its new 5A fab starts wafers only in 2H27, reaching 30,000 wafers a month in 2028. Our 6 Oct Blade reported Nanya's 11th straight monthly revenue record, with September at NT$45.09bn; this round would extend that price-led growth into 4Q26.

Demand is broad. AMD (AMD US) chief executive Lisa Su said on 6 Oct that memory supply remains tight, from HBM (high-bandwidth memory) for AI servers to PC memory, and that AMD is working with suppliers and customers to secure enough for AI servers, CPU platforms and PCs. One AMD Helios rack carries 72 MI455X GPUs with up to 432GB of HBM4 each, about 31TB per rack. Nanya's 12 Oct results call is the first chance to confirm the size and timing of the increase.

Linked stocks: 2408 TT, 2344 TT, AMD US, 6239 TT

King Slide falls as Vera Rubin nears volume

King Slide Works (2059 TT), the main maker of server rails (the sliding mounts that hold servers in racks), reported September revenue of NT$4.46bn, up 196% YoY but down 30.9% MoM, ending six straight monthly records. Analysts attribute the drop to lower shipments on some server projects as NVIDIA's (NVDA US) Vera Rubin platform nears volume shipments and customers change over from current models.

Other server suppliers kept growing through the same month. AVC (3017 TT), a thermal maker, reported September revenue of NT$21.02bn, up 44.9% YoY and 7.9% MoM, its first month above NT$20bn. Its finance chief said in August that liquid cooling demand is rising both from NVIDIA GPUs and from custom chip customers. Gigabyte (2376 TT) reported record September revenue of NT$63.10bn, up 125% YoY and 33% MoM, with servers now about 80% of revenue, mostly GB300 NVL72 racks that are moving to Vera Rubin.

Reported revenue also understates some demand. Accton (2345 TT), a network switch maker, reported September revenue of NT$40.40bn, up 66.1% YoY but down 1.1% MoM. Analysts say part of the dip comes from Accton moving memory and circuit boards to consigned purchasing, where the parts are bought for it and their value no longer counts as Accton's revenue, as both prices rise.

Analysts expect King Slide's growth to return as Vera Rubin and custom chip servers ramp in 4Q26 and its US plant enters volume production in 2H26. October revenue will show whether September was a gap between platforms or a slower start to Vera Rubin.

Linked stocks: 2059 TT, 3017 TT, 2376 TT, 2345 TT

Microsoft and Meta trim internal Claude use

Microsoft (MSFT US) has cut its internal budget for Anthropic's Claude by more than a third, from an expected US$1bn or more a year, and told staff to use Microsoft's own AI tools instead, according to foreign media reports carried by Commercial Times and TechNews. At Meta (META US), internal users of Claude Code, Anthropic's coding assistant, fell to about 30,000 from about 60,000 earlier this year.

The cuts cover internal use only. Enterprise customers' use of Anthropic models through Microsoft's cloud platform is still growing. Meta's drop partly reflects a cut of about 10% of its 78,000 staff this spring, but mainly its own tools: MetaCode has more than 30,000 internal users and Muse Code, opened to outside testers in August, more than 6,000. Meta still spent more than US$105m on Claude Code over the past 28 days.

The two companies matter to Anthropic's IPO. Its prospectus shows two unnamed customers providing about 25% of revenue, and the market believes they are Meta and Microsoft. Our 3 Oct Blade reported the same filing lists US$518bn of compute and infrastructure obligations against 2025 revenue of about US$4.6bn. Those commitments are still growing: Lambda's backlog rose to US$50bn in September from US$15bn in June, US$35bn of it from an Anthropic deal signed in late August, as Lambda raises up to US$4bn at a US$14.5bn pre-money valuation.

Anthropic's revenue from enterprise customers buying through cloud platforms now has to grow fast enough to offset its two largest customers using less internally.

Linked stocks: MSFT US, META US, AVGO US

先聲 First Word: Exclusives from Chinese-Language Sources

Taiwan Star Tech raises capex to NT$13.43bn for custom AI chips

Taiwan Star Tech (3265 TT), the packaging and test unit of Sigurd Microelectronics (6257 TT), raised its approved 2026 capex to NT$13.43bn, mainly for wafer test, final test and packaging tools, citing strong custom AI chip demand. Sigurd expects optical communications revenue to double in 2027 and AI and other high-end products to pass 30% of revenue, from 24% this year. (06 Oct 2026) Source: 矽格、台星科 看旺明年

Samsung plans more than US$4bn of memory test plants in Vietnam

Samsung Electronics (005930 KS) is reportedly planning to spend more than US$4bn on two DRAM and NAND test plants in Vietnam, with the first line starting as early as 4Q27. The plants would add test capacity for AI-driven memory demand and move part of Samsung's memory testing and export base out of Korea. (07 Oct 2026) Source: 三星再擴越南半導體布局 擬砸40億美元建2座DRAM、NAND測試廠

Taiwan chip designers plan to keep passing on costs through 2027

Taiwan chip design companies say the next round of foundry and packaging price increases is largely set, and they expect to keep passing costs to customers through 2027, DigiTimes reported. Supply chain sources say TSMC (2330 TT) will adjust wafer prices by process node from January 2027, with mature and specialty nodes negotiated case by case. (07 Oct 2026) Source: 晶片漲價近在眼前 台系IC設計難阻2027年繼續反應成本