AI chip race expands: UMC enters silicon photonics, Korea tightens talent controls

UMC is pushing into silicon photonics in Singapore to address connectivity bottlenecks in AI data centers. The company is migrating production from 8-inch to 12-inch wafers to reduce optical loss and improve efficiency, and has already assigned more than 100 employees to the Singapore operation. Silith Technology will be its first major customer. UMC has outlined an expansion roadmap: advanced packaging services next year, an open optical-chip platform by 2028, and potential thin-film lithium niobate chiplet deployment. A partnership with IMEC will open the platform to additional clients from 2027. This move places UMC in direct competition with established players including STMicroelectronics, Broadcom, Nvidia, and Foxconn Industrial Internet.

South Korean prosecutors intensified their investigation into DDR5 chip pricing on July 15, raiding Rambus, Montage, and Renesas over suspected price-fixing of DDR5 interface chips used by Samsung, SK Hynix, and Micron. The three companies under investigation control over 90% of DRAM output. The probe follows US searches and subpoenas and is accompanied by a separate lawsuit accusing the three memory makers of inflating prices.

Government control over frontier AI models is tightening on both sides of the US-China divide. Washington held pre-launch talks with OpenAI regarding GPT-5.6 and temporarily restricted Anthropic's Fable 5 and Mythos 5. A June 2 executive order created a voluntary framework allowing federal agencies up to 30 days of early model access. Beijing is taking a different approach, considering stricter launch reviews, delayed releases, foreign-user limits, and tighter export controls. This shift in strategy follows the traction gained by Chinese models including DeepSeek, Moonshot AI's Kimi K2.6, and Z.ai in US corporate workflows.

Talent mobility and intellectual property are becoming contested territory. A South Korean court on July 9 barred two former Samsung Electronics NAND designers from working for SK Hynix or its affiliates until April 30, 2027, citing their access to protected next-generation NAND technology. The injunction covers 18 months—shorter than Samsung's requested two years—and imposes a KRW5 million (approx. US$3,353) daily penalty for violations. In a separate case, SK Hynix failed to block a former employee from joining Samsung in an HBM matter.

Hanmi Semiconductor and Hanwha Semitech are locked in a patent battle in Seoul over three TC-bonder patents covering flux inspection, light wavelengths, and squeegee mechanisms. Both companies are competing for SK Hynix's HBM4 orders; Hanmi has already secured a KRW44.2 billion HBM4 TC-bonder contract, while Hanwha reportedly received a similarly sized order. Hanwha is seeking KRW1 billion in damages and destruction of Hanmi's Griffin and Dragon equipment; Hanmi denies infringement. The next hearing is set for September 10.

MediaTek has secured the 5G modem slot across all four Google Pixel 11 models, a victory that ends Samsung's long-standing grip on the Pixel supply chain and beat Qualcomm in Google's evaluation. The win strengthens MediaTek's US expansion after it entered the Apple Watch supply chain in 2025 and could open further opportunities in smartphones, PCs, vehicles, and AI glasses, though Apple's in-house connectivity chip strategy may limit future growth.

US pressure on Korean memory suppliers continues to mount. According to DIGITIMES analyst Luke Lin, Washington is pressuring Samsung Electronics and SK Hynix to build memory fabs locally, even as Korea plans more than US$500 billion（約80兆円） in domestic semiconductor investment. Intel's EUR5 billion (approx. US$5.72 billion（約9200億円）) Fab 34 expansion in Ireland could lift capacity to 20,000–22,000 wafers, or up to 30,000 with higher utilization, but Intel's recovery ultimately depends on capacity, yield, and 14A funding from 2027.

The article frames AI competition as no longer confined to chip design and manufacturing, but extending into every layer of the technology stack—connectivity (silicon photonics), memory supply chains, intellectual property disputes, talent mobility, and now model governance itself. UMC's expansion into silicon photonics in Singapore is one example: the company is moving to 12-inch wafers to reduce optical loss and serve a fast-growing connectivity market, signaling that data-center bottlenecks are driving investment in optical interconnect alongside traditional semiconductors.

Government intervention is reshaping the landscape. The US and China are both treating frontier AI model releases as national-security matters, moving beyond passive observation to active pre-launch coordination (Washington's 30-day early-access framework) and potential delayed releases and export restrictions (Beijing's strategy). This stands in contrast to the free competition that has historically governed chip supply; now model access timing is a policy tool. Simultaneously, memory-market regulation is tightening: prosecutors in South Korea are investigating three suppliers that control over 90% of DRAM output for suspected price-fixing, and talent restrictions (the April 30, 2027 bar on former Samsung engineers at SK Hynix) are being weaponized in disputes over next-generation technology access.