跳到正文
Semiconductor Engineering 芯片与封装· The SE Staff·· 5 小时前AI 评分35

芯片行业一周回顾:HBM 经走私流入中国、Oracle 取消腾讯 70 亿美元协议、三星扩产 2nm HBM 基底裸片

Chip Industry Week In Review

AI 导读

据报道,Samsung 和 SK hynix 的 HBM 正通过将低性能服务器处理器与 HBM 捆绑走私的方式流入中国 AI 芯片开发商手中;约 600,000 颗 Nvidia H200 AI 芯片滞留香港等待输往中国大陆的批准。

正文 · 原文

Top Of The Week

HBM from Samsung and SK hynix is reportedly reaching Chinese AI chip developers through smuggling operations that exploit U.S. export-control loopholes. According to The Information, intermediaries package HBM with low-performance server processors, ship the combined components to China, and then remove the memory for resale, including to customers of Huawei.

Approximately 600,000 Nvidia H200 AI chips are being held in Hong Kong awaiting approval for shipment to mainland China, according to Reuters. Beijing has limited imports in an effort to reduce reliance on U.S. technology and encourage adoption of domestically developed processors.

Micron reportedly paused construction of a $4.5B memory-chip plant in Taiwan due to a prolonged labor dispute, according to Bloomberg. Micron has not confirmed the suspension, and the effect on its longer-term Taiwan manufacturing plans remains unclear.

Samsung Electronics is reportedly planning to expand advanced foundry capacity for HBM base dies at two locations.

  • In South Korea, the company reportedly plans a new foundry facility at Giheung, with 2nm and 4nm production lines and an estimated ~US$17B investment, along with a possible advanced packaging expansion.
  • Separately, Samsung reportedly secured orders from AMD and Nvidia for 2nm HBM base dies to be produced at a proposed second foundry in Taylor, Texas, with mass production targeted for 2030. Note: Samsung has not confirmed the reported Giheung investment or Taylor customer orders.

Oracle reportedly canceled a 5-year agreement worth an estimated $7B to provide Tencent access to up to 100,000 advanced AI chips through data centers in Southeast Asia, following pressure from the Trump administration, according to the Financial Times. U.S. officials raised concerns about Chinese companies accessing advanced processors through overseas cloud services, while officials in China’s Guangdong province urged Tencent to use domestically developed chips.

Bosch plans to establish a dedicated semiconductor division in January 2028, expanding its foundry business and opening more manufacturing capacity to external customers. The company is targeting €5B in semiconductor sales by 2030 and plans to expand its semiconductor test center in Penang, Malaysia, by the mid-2030s.

Marvell‘s pending acquisition of Lumentum, initially valued at $115B, is drawing mixed reactions, with analysts seeing opportunities to challenge Broadcom in AI optical interconnects, but also raising concerns about supplier concentration and customer relationships. Questions about the purchase price and regulatory approval have added to investor concerns.

U.S. election results raised concerns about tighter restrictions on data center development, particularly in major markets such as Texas and Ohio, as well as increased congressional oversight of AI, including AI safety, cybersecurity, and industry practices. Restrictions that delay or reduce data center construction could weaken demand for AI accelerators, HBM and networking chips, and increased regulatory scrutiny could create additional uncertainty for AI infrastructure investment. (See chip stock impacts here.)

More acquisitions and deals

  • Onsemi confirmed it plans to acquire Ferric, developer of integrated voltage regulators, for $600M in cash, plus up to $175M in milestone payments.
  • Entegris acquired Advanced Abrasives and announced plans for a new SiC materials manufacturing line in Taiwan. The company also partnered with Element Six to develop diamond-based thermal-management materials for advanced packaging and expanded its deal with Kioxia on BiCS Flash 3D flash memory.
  • Supermicro is initially planning to use 10,000 SambaNova SN50 RDUs for its new AI disaggregated inference systems as part of a strategic collaboration.
  • China’s biggest EDA company, Empyrean, plans to acquire a ~50.4% stake in Easy-Logic for US$33M.
  • Unconfirmed: Mitsui Chemicals reportedly plans to buy JSR, semiconductor materials maker, for at least $8B.
  • Two acquisitions finalized: ADI‘s purchase of Alif Semiconductor, adding its AI-native MCU and fusion processors, and Forge Nano‘s acquisition of HZO for its nanoscale coating platforms.

Reports and analysis

Earnings

Revenue growth/decline year over year:

Quick Links To More News

In-Depth — Auto, Security and AI newsletter, shift left becoming a system-level problem, and who verifies AI.

Global — Europe boosts chip-design funding; Asia expands materials and test capacity; U.S. advances AI-driven chip design R&D.

New Technologies — RISC-V automotive MCUs, high-bandwidth compute, high-density testing, and ceramic data storage.

Security — 3D-stacked HW shield against side channel attacks; AMD vulnerabilities, compromised Android devices, secure-element certification.

Trending Video — Chiplets: shared resources, a mix of tools, and multiple dies add new challenges for design teams.

Research — AI chip inspection; cryogenic transistors; edge AI efficiency; ferroelectric memory.

Quantum Hardware — New QPUs; error-correction platforms; CMOS-based qubit readout; reliability and error suppression.

Workforce | Education — Apprenticeship funding; semiconductor doctoral fellowships; AI’s impact on engineering careers.

Events and Webinars — ACM/IEEE ICCAD, Electronica & SEMICON Europe. Webinars on shift left, verification, and debug of high-power motor controllers.


Global

Americas

  • UC San Diego and Qualcomm launched a $750K AI-driven chip-design research initiative, funded by a $500K Qualcomm gift and $250K in UCSD matching funds, focused on AI methods for SoC design, processor architectures, hardware accelerators, memory systems, and interconnects.
  • CSIS examines Albany NanoTech as a model for semiconductor technology clusters, highlighting shared R&D infrastructure, industry partnerships, and workforce development supporting sub-2nm technologies, advanced packaging, photonics, and quantum systems.
  • The OCP Foundation is opening nominations for project leads across all 12 top-level OCP Projects, ranging from hardware management to the AI computing continuum.

Europe

  • The European Semiconductor Industry Association (ESIA) called for dedicated semiconductor funding under the EU’s 2028–2034 budget, including support for R&D, pilot production, advanced packaging, and industrial deployment. The association urged policymakers to coordinate national and EU funding, preserve the Chips Joint Undertaking’s cross-sector role, and ease restrictions on intellectual property transfers within multinational companies.
  • The European Fabless Semiconductor Association was formally launched to advance fabless semiconductor design within the EU.
  • The EU launched a European Chip Design Platform (EuroCDP) with €525M to stimulate new chip design companies.
  • Synopsys and the EU Chips JU signed a deal to expand access to its EDA software for semiconductor startups and research teams through the EuroCDP.

Asia

  • South Korea reaffirmed plans to spend ₩21.3T (US$16B) on AI and major initiatives covering semiconductors, physical AI, and data centers, in its proposed 2027 budget, nearly doubling this year’s allocation.
  • Panasonic Industry plans to invest about ¥92B ($583M) to expand AI-related materials production and R&D, including circuit-board materials, capacitors and semiconductor packaging materials.
  • Two workers were killed and five injured in an explosion and fire at Panasonic Industrial Devices Materials‘ factory in Hsinchu, Taiwan. The facility produces PCB and semiconductor packaging materials. The cause and production impact have not been determined yet.
  • South Korea’s Techwing plans to invest approximately ₩52.6B ($39M) in a new semiconductor equipment factory in Cheonan to expand production of memory test handlers and HBM cube probers.

New Technologies

Automotive and physical AI

  • Infineon introduced its first automotive MCU platform based on the RISC-V architecture, adding hardware-based physical protection and crypto agility, with support planned for post-quantum security requirements.
  • Infineon and Qualcomm are collaborating on a microcontroller-based subsystem for Snapdragon Ride reference platforms used in central car computers for automated driving and digital cockpit applications.
  • STMicro announced a new microcontroller portfolio covering industrial, automotive, and physical AI applications.

Test, measurement

  • Keysight released XR3, a low-profile 14-bit oscilloscope designed to increase measurement capacity in rack-based test systems.
  • Siemens introduced an automated component model extraction solution developed with Bosch that cuts frequency-response-function measurement time by up to 93% and produces ISO 20270-compliant models for virtual NVH development.

AI hardware & memory

  • Qualcomm completed hardware validation of its high-bandwidth compute (HBC) silicon, exercising the full memory read/write path and near-memory computations.
  • Cerabyte completed S3 integration testing of its ceramic data storage within established object-storage workflows, in collaboration with CERN openlab.

Design


Research

MCL, PVA TePla and ASE researchers developed an AI-based image-enhancement method for detecting flip-chip interconnect defects. By improving scanning acoustic microscopy images, the method increased detection of cracks and delamination in extreme-low-k dielectric layers from 5.2% to 100% in tests involving 19 defects across three packages.

TU Munich, Applied Materials et al. evaluated cryogenic 2nm GAA transistors and superconducting interconnects for energy-efficient AI computing. Using a calibrated chip-design flow, they reported a 3x performance increase and 7.7x improvement in energy efficiency for a modeled AI accelerator’s compute fabric, excluding memory, data movement, and refrigeration power.

ETH Zurich-led researchers benchmarked 10 commercial sub-watt edge AI devices, comparing inference performance and energy consumption across convolutional neural networks and small vision transformers. The study found that memory traffic, rather than computational workload, was the primary factor determining latency on the tested neural-network accelerators, highlighting the importance of data movement in low-power AI hardware design.

HUST and PolyU researchers developed annealing-free ferroelectric transistors with ultrathin oxide-semiconductor channels for NVM and neuromorphic computing. The devices achieved a memory window of up to 20.6V, endurance of 10⁶ cycles, and stable 6-bit operation across 64 states. A neuromorphic network evaluation achieved 93.02% recognition accuracy.


Security

Government security news

Alerts and attacks

  • AMD issued security bulletins covering a newly discovered EDK II PXE driver vulnerability that can allow pre-boot code execution and a contention-based side-channel technique for exploiting known vulnerabilities in AMD CPUs.
  • Dell issued BIOS updates for a critical Qualcomm chipset vulnerability affecting five Latitude, Inspiron, and XPS systems. The flaw can cause memory corruption when a malformed RSA key is imported into Qualcomm’s Keystore.
  • Google banned Videostrong Technology from making certified Android devices after its security researchers determined that the company was responsible for creating an XCPE backdoor that enrolled 1.3 million set-top boxes into a criminal botnet.
  • CISA’s alerts for the week are here.

Security offering and deals

  • Vietnam has its first-ever homegrown security chip after government-backed Viettel Group successfully integrated a domestically developed cryptographic algorithm into silicon.
  • u-blox announced that it will integrate Kudelski Labs’ embedded security IP into its future chipsets.
  • Chinese tech giant Tencent released Hy4, its powerful new open-source AI model. Amid growing concerns about agentic hacking, the company said it limited the model’s cybersecurity training data to defensive material only.

Security technical papers

NC State and Princeton researchers proposed a 3D-stacked hardware shield to protect post-quantum cryptography accelerators from side-channel attacks. The design stacks a guard layer above the cryptographic hardware, using decoy-key activity and physical shielding to obscure power leakage from NTT computations.


Trending Video

Accelerating Plug-And-Play Prototyping In Chiplet-Based Designs: Purna Mohanty, CEO of SignatureIP, talks about the impact of chiplets and shared memory resources, the challenges of version control across multiple flavors of I/O, and the challenges engineering teams face when hardware-accelerated tools are running at different speeds than the actual design.


Quantum Hardware

New launches

  • Universal Quantum launched its iQPU quantum processor, combining a large trapped-ion compute surface, ultra-low-noise integrated control electronics, and chip-to-chip interfaces designed to scale quantum computing systems. The processor is manufactured on a commercial 200mm semiconductor production line.
  • IBM, ORNL, and AMD opened a new co-design platform for quantum error correction at IBM’s quantum facility in New York, co-locating the 120-qubit IBM Quantum Nighthawk with ORNL’s Umbra HPC system.

Fig. 1: ORNL’s Umbra HPC system. Source: Oak Ridge National Lab

Advances

  • Researchers from Quantum Motion, UCL, and others demonstrated a superconducting circuit element fabricated using a commercial CMOS process. The titanium-nitride device functions as a compact superinductor and was integrated with a silicon quantum dot to enable sensitive RF charge detection, potentially supporting more compact readout hardware for semiconductor spin qubits.
  • University of Regensburg-led researchers demonstrated a real-time feedback technique that reduced superconducting-qubit operation errors by approximately 54%, from 0.14% to 0.065%. The approach identifies changes in qubit resonance frequency associated with material defects and adjusts control signals to improve reliability.
  • Alice & Bob reported extended bit-flip lifetimes in superconducting cat qubits, averaging 44 minutes on a single-qubit test device and reaching 189 to 252 minutes on selected qubits in its 12-qubit Helium 2 processor. The preliminary results demonstrate strong suppression of bit-flip errors during idle operation, with performance during error-correction operations still to be validated.

Workforce, Education

The U.S. Department of Labor opened applications for a $40M apprenticeship incentive fund that includes semiconductor fabrication among several high-growth sectors. Registered apprenticeship sponsors can receive up to $6K per eligible new apprentice.

UCLA named the first 10 doctoral fellows in its Semiconductor Hub, with research spanning 3D ICs, advanced packaging, silicon photonics, compute-in-memory, and AI hardware. Fellows will also intern with one of the Hub’s industry partners, including Synopsys, GlobalFoundries, and others.

In a new Semiconductor Engineering roundtable, industry experts discuss how AI is changing expectations for new engineering graduates. Demand for electrical engineers remains strong, but AI agents are taking over some traditional junior-level tasks, raising expectations for practical experience, AI fluency, and the ability to tackle higher-level engineering work earlier in their careers.

U.S. community colleges are expanding semiconductor technician training programs as the industry faces a potential shortage of 157,000 workers by 2030, according to the Associated Press. More than 200 colleges and universities are coordinating with semiconductor manufacturers, including TSMC, Intel, Samsung, and SK hynix, to develop hands-on training programs supporting new U.S. fabs.


Events and Webinars

Upcoming webinars are here, including:

Find upcoming chip industry events here, including:

Upcoming Events

SWTest Asia 2026
Nov. 5–6 | Hsinchu, Taiwan

ACM/IEEE ICCAD 2026
Nov. 8–12 | San Jose, CA

GSA: 2026 WISH Conference
Nov. 10 | Santa Clara, CA

Electronic Specialty Gases & Systems Conference (ESGS)
Nov. 10–12 | Phoenix, AZ

Electronica 2026
Nov. 10–13 | Munich, Germany

SEMICON Europa
Nov. 10–13 | Munich, Germany

SC26: International Conference for HPC Networking, Storage & Analysis
Nov. 15–20 | Chicago, IL

Hardwear.io Security Trainings and Conference NL 2026
Nov. 16–20 | Amsterdam, Netherlands

DVCon Europe
Nov. 17–18 | Munich, Germany

2026 SIA Awards Dinner
Nov. 19 | San Jose, CA

2026 MRS Fall Meeting & Exhibit
Nov. 29–Dec. 4 | Boston, MA

SEMIEXPO Vietnam 2026
Dec. 3–4 | Ho Chi Minh City, Vietnam

SEMICON Japan 2026
Dec. 9–11 | Tokyo, Japan

IEEE Semiconductor Interface Specialists Conference
Dec. 9–12 | San Diego, CA

2026 GSA Awards Celebration
Dec. 10 | Santa Clara, CA

IEDM 2026: International Electron Devices Meeting
Dec. 12–16 | San Francisco, CA

来源: Semiconductor Engineering 芯片与封装 · semiengineering.com