[Oct 31] Semiconductor News Briefing — Samsung’s record memory quarter, Google TPU sales double its cloud revenue

삼성전자 D램 칩이 들어간 노트북용 메모리 모듈(SO-DIMM)

Key takeaways

  • Samsung Q3 operating profit hits record ₩107T
  • Google Cloud revenue doubles as TPU sales kick in
  • Marvell reportedly in talks to buy Lumentum
  • US DOJ scrutinizes Nvidia's Hugging Face deal
  • TSMC buys AUO LCD fab in Taichung for NT$38.8B
  • China's CXMT pushes HBM3E despite 20-30% yields
  • 2027 server CPU shipments forecast to jump 46%

Watch the video briefing (in Korean)

More briefing videos on our YouTube channel.

1. Samsung Q3 operating profit hits a record ₩107 trillion on rising memory prices

A Samsung laptop memory module (SO-DIMM) with Samsung DRAM chips
A Samsung laptop memory module (SO-DIMM) with Samsung DRAM chips · Photo: Raimond Spekking / Wikimedia Commons, CC BY-SA 4.0

Samsung Electronics reported final third-quarter results on October 29. Consolidated revenue reached ₩195.3 trillion and operating profit ₩107.4 trillion, both all-time highs. The semiconductor (DS) division alone earned ₩106.5 trillion on ₩147.1 trillion in sales, accounting for nearly all of the profit.

On the earnings call, Samsung said its average DRAM selling price rose about 20% from the prior quarter and NAND prices climbed in the high teens. Prices rose across every application, led by servers, and demand keeps outrunning added supply. The company expects memory supply to stay tight in 2028, said next year's HBM output is already sold out, and plans to start mass production of HBM4E in the first half of next year.

Why it matters: Higher prices, more than higher volumes, drove the profit. With the world's largest memory maker expecting shortages through 2028, cost pressure on PCs, phones and other memory-hungry devices may persist.

  • ASP (average selling price): The average price of units sold over a period, a key gauge of memory market conditions.
  • HBM (high-bandwidth memory): Stacked DRAM that moves large amounts of data at once, mounted next to AI accelerators.

2. Google Cloud revenue up 103% as TPU sales contribute; inventory swells to $13.3B

A Google TPU v4 board carrying four liquid-cooled TPU packages
A Google TPU v4 board carrying four liquid-cooled TPU packages · Photo: Norman P. Jouppi, George Kurian, Sheng Li, Peter Ma, Rahul Nagarajan, Lifeng Nai / Wikimedia Commons, CC BY 4.0

Alphabet reported third-quarter results on October 28. Revenue rose 25% to $127.6 billion, and Google Cloud revenue more than doubled to $30.7 billion. The company said sales of complete systems built on its in-house TPU AI chips were a meaningful contributor to growth.

Inventory reached $13.3 billion at the end of September, up from $2.4 billion at the end of 2025. Executives said TPU inventory is built only after sales contracts are signed and is all tied to committed deals, with a meaningful revenue ramp expected in 2027. They cited work with Anthropic as part of the TPU business but did not break out margins for TPU sales. A day earlier, on October 27, Oracle announced it will offer TPUs in its cloud starting in 2027.

Why it matters: Google is becoming a merchant chip supplier and a direct rival to Nvidia. Because it ties up cash buying hardware in advance, how profitable the business is remains to be proven.

  • TPU: Google's custom chip designed for AI workloads.

3. Marvell reportedly in advanced talks to acquire Lumentum

Marvell's headquarters in Santa Clara, California (2011)
Marvell's headquarters in Santa Clara, California (2011) · Photo: King of Hearts / Wikimedia Commons, CC BY-SA 3.0

Bloomberg reported on October 28 that chip designer Marvell is in advanced talks to buy optical components maker Lumentum. An all-stock deal is under discussion and could be announced as early as the first week of November. Lumentum is valued at roughly $83 billion.

Neither company has confirmed the talks, and Bloomberg noted they could still fall apart. Marvell makes optical DSP chips for data centers, while Lumentum supplies lasers and optical switches.

Why it matters: Demand for optical links between AI servers is surging. A deal would let Marvell sell everything from DSPs to lasers and compete more directly with Broadcom.

  • DSP (digital signal processor): A chip that converts and cleans up high-speed signals carried over fiber.

4. US Justice Department examines Nvidia's $12.9B Hugging Face acquisition

Hugging Face logo
Hugging Face logo · Image: Victor (Hugging Face Staff) / Wikimedia Commons, public domain

Reuters reported on October 28 that the US Justice Department is examining Nvidia's $12.93 billion purchase of Hugging Face, the largest platform for sharing and distributing AI models. Officials are said to be weighing whether the deal would give Nvidia control over a key distribution point for AI models that it could use against rivals.

According to Reuters, the inquiry is part of a broader review of Nvidia's acquisitions, investments and sales practices. Nvidia has said Hugging Face will keep supporting a range of chip and cloud providers, and it aims to close the deal in the first half of 2027. An investigation does not imply wrongdoing.

Why it matters: As AI competition spreads from chips to software and developer ecosystems, regulators are watching chipmakers' software deals more closely.

5. TSMC buys AUO's Taichung LCD fab for NT$38.8 billion

An AUO building in Hsinchu Science Park, Taiwan (not the Taichung fab)
An AUO building in Hsinchu Science Park, Taiwan (not the Taichung fab) · Photo: Weihao.chiu / Wikimedia Commons, CC BY-SA 3.0

TSMC said in an October 29 filing that it will buy factory buildings, cleanrooms and related facilities in the Central Taiwan Science Park in Taichung from display maker AUO for NT$38.8 billion before tax. The buildings cover about 429,000 square meters. AUO's board approved the sale the same day and expects a disposal gain of about NT$18.5 billion.

TSMC said only that the site will be used for operations and production. Some expect it to host advanced packaging such as CoWoS, but the company has not confirmed that.

Why it matters: Buying an existing display fab gives TSMC buildings and cleanrooms far faster than building from scratch. It bought an LCD plant from Innolux in 2024 for similar reasons.

  • CoWoS: TSMC's advanced packaging that places AI chips and HBM side by side on one substrate so they work as a single chip.

6. China's CXMT moves ahead with HBM3E despite 20-30% yields

A DDR5 laptop memory module using CXMT DRAM
A DDR5 laptop memory module using CXMT DRAM · Photo: Padgriffin / Wikimedia Commons, CC BY 4.0

Reuters reported on October 28 that CXMT, China's largest DRAM maker, plans small-scale HBM3E production within two months. The company has reportedly shipped samples to several Chinese AI chip designers.

Initial yields are said to be 20% to 30%, far below the roughly 70% attributed to SK hynix. One source said the part performed closer to SK hynix's previous-generation HBM3 in testing. CXMT has not publicly confirmed the yield figures.

Why it matters: US export controls make it hard for China to buy foreign HBM, so Beijing is prioritizing a domestic supply of AI memory even at high cost.

  • Yield: The share of manufactured chips that work properly.

7. Intel CTO says the process gap with TSMC is down to a few points, but scale still lags

Intel's Fab 12, 22 and 32 complex in Chandler, Arizona
Intel's Fab 12, 22 and 32 complex in Chandler, Arizona · Photo: neepster / Wikimedia Commons, CC BY-SA 2.0

Intel's chief technology officer said the company's process technology now trails TSMC by only a few percentage points, in a special episode of Dwarkesh Patel's podcast recorded during a two-day tour of Intel's fabs and labs. DIGITIMES reported the remarks on October 28.

He acknowledged, however, that matching TSMC's manufacturing scale will take much longer. Yield, capacity and on-time delivery, not just design metrics, decide the foundry race.

Why it matters: Intel wants to build chips for outside customers. To win them away from TSMC, it must prove it can deliver at volume, not just match the technology.

  • PPA: Performance, power and area, the standard yardsticks for comparing chip process technologies.

8. Intel to acquire Israeli quantum software firm Classiq

A full wafer of Intel quantum chips
A full wafer of Intel quantum chips · Photo: Steve Jurvetson from Menlo Park, USA / Wikimedia Commons, CC BY 2.0

Intel announced on October 26 that it will acquire Classiq, an Israeli quantum computing software company. Terms were not disclosed, but Bloomberg put the price at about $2.3 billion. The deal is expected to close in the first half of 2027.

Classiq's platform automatically turns high-level algorithms into quantum circuits. Intel plans to pair it with its silicon spin-qubit hardware to offer a full-stack quantum platform, while continuing to support Classiq's integrations with other vendors' hardware.

Why it matters: Intel is positioning quantum computing as a long-term growth pillar alongside CPUs and GPUs. Software that makes quantum machines usable matters as much as the hardware.

  • Silicon spin qubit: A qubit that stores quantum information in the spin of an electron inside a silicon chip made with conventional processes.

9. 2027 server CPU shipments forecast to rise 45.7%, with AMD set to top Intel in revenue

AMD headquarters in Santa Clara, California
AMD headquarters in Santa Clara, California · Photo: Coolcaesar / Wikimedia Commons, CC BY-SA 4.0

DIGITIMES Research forecasts global server CPU shipments of 52.6 million units in 2027, up 45.7% year on year, driven by agentic AI, expanding AI servers, cloud giants' in-house CPUs and replacement of general-purpose servers.

Intel is expected to keep the top spot by units with 39%, with AMD closing in at 30.4%. By revenue, AMD is projected to reach $31.2 billion, ahead of Intel's $28.7 billion. Separately, TrendForce said on October 22 that 2027 server CPU shipments would grow by more than 40%.

Why it matters: AI is lifting demand not only for GPUs but for the CPUs that orchestrate them, and the long Intel-dominated server CPU market is being reshaped.

10. ASE lifts 2026 capex again to $11.5B; co-packaged optics enter volume production

ASE Group headquarters in Kaohsiung, Taiwan
ASE Group headquarters in Kaohsiung, Taiwan · Photo: 人人生來平等 / Wikimedia Commons, CC BY-SA 4.0

ASE, the world's largest chip packaging and testing company, raised its 2026 capital spending plan from $10.5 billion to $11.5 billion at its October 29 earnings call. The extra $1 billion will be split between buildings and equipment, with about 60% going to its leading-edge advanced packaging and test (LEAP) business. Spending in 2027 will be higher still.

ASE raised its 2026 LEAP revenue target from $3.7 billion to more than $4 billion and expects it to double again in 2027. Co-packaged optics (CPO) entered volume production in the fourth quarter, with more than $500 million in CPO revenue expected in 2027.

Why it matters: AI chips cannot be built without advanced packaging that combines multiple dies. Rising packaging investment shows this step is still a bottleneck.

  • CPO (co-packaged optics): Placing optical components inside the same package as the chip to cut power use and signal loss.

11. Samsung Electro-Mechanics Q3 operating profit up 165%; MLCC long-term deals with about 50 customers

Label on a box of 20,000 Samsung Electro-Mechanics MLCCs
Label on a box of 20,000 Samsung Electro-Mechanics MLCCs · Photo: Solomon203 / Wikimedia Commons, CC BY-SA 4.0

Samsung Electro-Mechanics posted record quarterly revenue of ₩3.7633 trillion and operating profit of ₩691 billion, up 165% from a year earlier. The component division, which makes MLCCs, grew revenue 36% to ₩1.883 trillion, and the package solution division, which makes FC-BGA substrates, rose 58% to ₩939 billion.

On the earnings call, the company said it has signed or is negotiating long-term MLCC supply agreements with about 50 customers. With capacity short of demand and raw material costs rising, it is negotiating prices with customers.

Why it matters: AI servers use far more MLCCs and larger, more complex substrates than ordinary servers. As customers rush to lock in supply, component makers are gaining pricing power.

  • MLCC (multilayer ceramic capacitor): A tiny component that stores and releases charge to keep circuit voltage stable.
  • FC-BGA: A substrate that connects high-performance chips to the motherboard.

12. TI and NXP see industrial chip recovery, with industrial sales up 40% and 39%

Texas Instruments sign in Dallas
Texas Instruments sign in Dallas · Photo: Texas Instruments / Wikimedia Commons, CC BY 3.0

Texas Instruments' third-quarter revenue rose 28% to $6.06 billion, with growth across industrial, automotive and data center markets; on the earnings call it said industrial revenue grew more than 40%. It guided fourth-quarter revenue to $5.80 billion to $6.30 billion.

NXP reported third-quarter revenue of $3.79 billion on October 27, up 19%. Industrial and IoT revenue rose 39% to $802 million, which the company attributed to AI-enabled processors, data centers and factory automation.

Why it matters: Industrial chips suffered a long inventory correction in 2023-2024. Data center power equipment and factory automation are now driving a clear rebound.

  • Analog chip: A chip that handles continuous signals such as voltage, current or temperature, used in power management and sensing.

13. Silicon Motion Q3 revenue reaches $560M, up 131% year on year

A Silicon Motion SM2258XT SSD controller on an SSD board
A Silicon Motion SM2258XT SSD controller on an SSD board · Photo: Raimond Spekking / Wikimedia Commons, CC BY-SA 4.0

SSD controller maker Silicon Motion reported third-quarter revenue of $560.2 million, up 131% from a year earlier and 24% from the prior quarter. Gross margin was 53.6% and operating margin 29.0%.

The company said its 4nm PCIe Gen 6 MonTitan controllers for enterprise and AI SSDs are beginning to scale. It guided fourth-quarter revenue to $588 million to $616 million.

Why it matters: AI data centers consume large volumes of high-performance SSDs, and that demand is spreading from memory chips to the controllers that manage them.

  • SSD controller: The chip inside an SSD that manages reading and writing data to NAND flash.

14. Fire hits Taiwan power-chip wafer maker Episil; less than 10% of capacity affected

6-inch silicon carbide (SiC) wafers in a carrier box
6-inch silicon carbide (SiC) wafers in a carrier box · Photo: FDominec / Wikimedia Commons, CC BY-SA 4.0

A fire broke out on the morning of October 28 in a cleanroom at Episil-Precision's Innovation plant in Hsinchu Science Park. All 92 people inside were evacuated without injury, and the blaze was put out in about three hours. About 3,700 square meters burned, with damage initially estimated at NT$130 million.

On October 29 the company said the affected area accounts for less than 10% of total capacity and output, and that it has activated backup production at other plants. It said it is fully insured. Fire officials are investigating the cause.

Why it matters: Episil makes epitaxial wafers used as the base for power semiconductors. The company says the impact is limited, but with few suppliers of this material, customers are watching the recovery.

  • Epitaxial wafer: A wafer with an extra thin crystal layer grown on top, used to improve the performance of power chips and other devices.

15. Taiwan's LandMark to boost silicon photonics laser capacity 2.5-fold in 2027

A 300mm silicon photonics wafer
A 300mm silicon photonics wafer · Photo: Ehsanshahoseini / Wikimedia Commons, CC BY-SA 4.0

Taiwanese optical materials maker LandMark Optoelectronics plans to lift its 2027 capacity for CW lasers used in silicon photonics to 2.5 times the 2026 level, and hopes to keep a similar pace of expansion in 2028. It described current capacity as very tight.

Its board approved equipment spending of up to NT$2.27 billion, and the company signed a four-year CW laser supply contract covering 2027-2030 with a US customer.

Why it matters: Silicon photonics and CPO chips rely on external lasers for their light source. As optical links in AI data centers multiply, these lasers could become a new bottleneck.

  • Silicon photonics: Building optical pathways on silicon chips so data moves as light rather than electricity.
  • CW laser: A laser that emits light continuously, used as the light source for silicon photonics chips.

16. LB Semicon expands power-chip packaging capacity on a 10-year Qualcomm deal

Qualcomm logo
Qualcomm logo · Image: Original: Qualcomm / Wikimedia Commons, public domain

Korean chip packaging company LB Semicon will invest about ₩30 billion to expand capacity under a 10-year supply agreement with Qualcomm. By 2028 it plans to raise 200mm-equivalent capacity at its Pyeongtaek headquarters from 26,000 to 36,000 wafers a month and 300mm capacity from 3,000 to 5,000 wafers a month.

The agreement took effect in February 2026 and renews automatically every two years. It does not fix volumes in advance; LB Semicon supplies Qualcomm's orders within its available capacity. The company began volume shipments of Qualcomm power management ICs (PMICs) in August.

Why it matters: It shows a Korean back-end firm long dependent on display driver chips moving into a global chipmaker's supply chain to diversify.

  • PMIC (power management IC): A chip that delivers the right voltage and current to each part of a device.

17. Taiwan's Kinsus explores aluminum nitride ceramic as AI chip substrates grow larger

A BGA-packaged chip mounted on a circuit board
A BGA-packaged chip mounted on a circuit board · Photo: F5onr / Wikimedia Commons, CC BY-SA 4.0

Lin Ding-hao, R&D vice president at Taiwanese substrate maker Kinsus, discussed materials for advanced packaging substrates at the TPCA Show in October. As AI chip packages grow, keeping substrates flat, dimensionally stable and resistant to warping becomes a central challenge.

He noted that glass cores, a leading next-generation candidate, have attractive material properties but are brittle and hard to make reliable. Kinsus is therefore researching aluminum nitride (AlN) ceramic, which conducts heat well, as a substrate material.

Why it matters: AI chips put more dies and memory on a single substrate, so substrates keep getting bigger. Warping under heat causes defects, sparking a race among glass, ceramic and other new materials.

  • Warpage: Bending of a substrate caused by chips and substrate expanding at different rates (CTE) when heated.
  • Glass core substrate: A next-generation substrate whose core layer is glass instead of organic material, making it flatter and less prone to warping.

18. Nvidia reportedly weighing microchannel cooling for Rubin Ultra

An Nvidia DGX GB200 server rack
An Nvidia DGX GB200 server rack · Photo: Pokiiri / Wikimedia Commons, CC BY-SA 4.0

Reports continue to suggest that Nvidia's next-generation Rubin Ultra AI chip may use microchannel cold plates (MCCP) that run coolant directly above the die. With power per chip possibly reaching 2,300W, conventional cold plates may struggle, and Taiwanese cooling firm AVC has been named as a partner.

According to SemiAnalysis, the Rubin GPU's cold plate already uses a microchannel design with channel pitch narrowed from 150 microns to 100 microns (0.1mm). Nvidia has not officially confirmed the final cooling design for Rubin Ultra.

Why it matters: As AI chip power keeps climbing, cooling is becoming part of package design rather than just a facility issue.

  • Microchannel cooling: Running coolant through hair-thin channels to remove large amounts of heat from a small area.

Sources

Compiled by Bits & Bites from public reporting. Not investment advice.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *