Samsung Foundry Counterattack Begins; Eyeing Google Chip Orders

*Editor's Note: 'AI PRISM' (Personalized Report & Insight Summarizing Media) is an "AI-based customized news recommendation and summary service" developed with support from the Korea Press Foundation. It selects and provides six customized news items by reader type.

1. Anthropic Builds Own Servers, Cutting Token Production Costs First

- Key Summary: Anthropic has signed at least 12 letters of intent for data center leases with U.S. developers, embarking on building its own servers for the first time. Until now, Anthropic developed Claude by connecting to servers owned by AWS and Google, but it will now shift to a structure in which it directly purchases equipment such as AI chips to build servers and pays only rent to site owners. The direct background is that as users of Claude Cowork (AI agent) and Claude Code (coding tool) surged, the company had to rely on hyperscalers (large-scale data center operators) whenever server capacity ran short, and consumer complaints grew over usage restrictions. The move also aims to resolve the structural problem in which a shortage of AI infrastructure drives up server lease prices, leading to higher token costs. To reduce financial burden, a plan is also under discussion in which Google would guarantee lease payments on the condition that Anthropic uses Google TPUs (Tensor Processing Units).

2. Samsung Foundry Likely to Win Google's Next-Generation TPU Order

- Key Summary: Google is reviewing a plan to entrust production of the memory I/O die, a key component of its 10th-generation TPU next-generation AI chip, to the 2-nanometer process of Samsung Electronics' foundry division. The development comes as the production capacity of TSMC, the world's No. 1 foundry, has reached saturation, drawing attention to Samsung Electronics' comprehensive semiconductor solutions. Some analysts say Google is using foundry orders as leverage to secure HBM (high-bandwidth memory), which is in short supply. The two companies have a history of jointly producing Tensor chips for Pixel phones from 2021 to 2024, and have maintained a cooperative relationship, with Samsung supplying more than 60% of the HBM going into Google TPUs last year. The 10th-generation TPU targets mass production as early as 2028, and concrete negotiation results, such as an actual contract, are expected to take shape only late this year or early next year.

3. TSMC Bottleneck Reshapes AI Supply Chain; Full-Scale Cooperation with Google on 2nm and HBM

- Key Summary: As demand for AI chips has exploded, the production capacity of TSMC, which holds a 73% share of the foundry market, has reached a critical point, and TSMC Chairman C.C. Wei publicly acknowledged at the shareholders' meeting on the 9th of this month that "customer demand is so high that there are limits to handling it." Accordingly, big tech firms with their own chip design capabilities have begun adopting multi-foundry strategies, and Samsung Electronics is expanding its customer base by winning an order for Tesla's next-generation autonomous driving chips worth about 25 trillion won, followed by securing Nvidia's language processing unit (LPU) "Grok3." Google is reviewing a dual configuration in which the core computing chip is entrusted to TSMC's 1.4-nanometer process while the I/O die responsible for memory connections uses Samsung Electronics' 2-nanometer process. Korea Investment & Securities forecasts that Samsung Electronics' non-memory business will turn to profit after running a deficit this year, posting operating profit of 154 billion won in the fourth quarter, and that quarterly operating profit will rise to 731 billion won in the fourth quarter of 2027.

4. 'K-Moonshot' to Solve AI Challenges Faces Personnel Vacancy as AI Scientist PD Resigns

- Key Summary: In K-Moonshot, the inter-ministerial innovation-challenge R&D program promoted by the Ministry of Science and ICT, Lee Min-hyung, CEO of Asteromorph, who served as PD of the AI Scientist mission, expressed an intention to resign, creating a personnel vacancy even before the launch. K-Moonshot aims to solve 12 national missions—including new drugs, brain-computer interfaces (BCI), small modular reactors (SMR), and humanoids—on an AI basis by 2035, and the PD is not a principal investigator (PI) who conducts research directly but an overall coordinator responsible for establishing roadmaps for each mission and planning projects and managing milestones. Lee runs a startup aiming to develop an "AI scientist" that generates and verifies scientific hypotheses, and recently attracted seed investment worth 42 billion won; the resignation is reportedly linked to restrictions on concurrent positions and for-profit activities that would apply upon the transition to special research fellow status in July. The Ministry of Science and ICT is reviewing a plan to minimize the vacancy by having the head of the National Science AI Research Center concurrently serve in the AI Scientist mission, rather than immediately selecting a successor PD.