Samsung’s GAA Blocked by TSMC Wall, Technological Barriers Must Be Overcome First

OpenAI Adopts Process Stability–Focused Technological Pragmatism
Widening Gap in CoWoS-Based Packaging Competitiveness
Urgent Need to Restore Trust Through Yield Stabilization and Low-Cost Orders

OpenAI’s proprietary semiconductor development project, “Tigris,” which is spearheading the artificial intelligence (AI) era, has revealed concrete production guidelines, delivering a stark defeat for Samsung Electronics. OpenAI ultimately selected Taiwan’s TSMC as its sole manufacturing partner, setting aside Samsung’s ambitious turnkey solution encompassing both memory and foundry services, thereby reaffirming the solidity of its silicon alliance. This outcome symbolically underscores that Samsung’s foundry strategy has yet to provide a definitive technological answer in the global AI market.

OpenAI’s Strategic Move to Reduce Dependence on Nvidia

According to industry sources on the 9th, ahead of TSMC’s March earnings announcement scheduled for the 10th, Samsung Electronics’ foundry business has once again come under scrutiny for its marginalization. In particular, following Broadcom’s disclosure late last month that TSMC’s advanced process capacity had reached its limits—effectively formalizing production concentration among big tech clients including OpenAI—the market is revisiting Samsung’s failed turnkey strategy as a painful reality.

OpenAI has partnered with U.S. semiconductor firm Broadcom to develop customized AI chips (ASICs) in an effort to reduce its dependence on Nvidia, the leader in AI chips. Mass production is targeted to begin next year. At present, the chip development process appears to be progressing smoothly. Reuters noted that “the process of sending a chip design to a fabrication plant is called ‘taping out,’” adding that “taping out typically costs tens of millions of dollars, and without paying expedited fees, it takes approximately six months before actual chip production begins.”

During the bidding process in February, Samsung emphasized that it was the optimal partner capable of handling both supply and foundry manufacturing in a single integrated solution. Following the February 4 meeting at Samsung’s Seocho headquarters between Samsung Chairman Lee Jae-yong, OpenAI CEO Sam Altman, and SoftBank Chairman Masayoshi Son, expectations were high that OpenAI might allocate part of its ASIC production to Samsung’s foundry. However, Samsung’s shrinking foothold in the foundry market has led to quarterly losses amounting to trillions of won, constraining its ability to pursue aggressive investment, including delaying the operational timeline of its Taylor, Texas plant from this year to next.

Defeat Despite Samsung’s ‘Turnkey’ Proposal Amid Packaging Gap

Ultimately, CEO Altman chose TSMC to manufacture the ASICs. Rather than Samsung’s integrated solution, OpenAI opted for a division-of-labor model that leverages best-in-class capabilities in each domain. OpenAI sought to avoid the risk that uncertainties in process stability could determine the success or failure of the multi-billion-dollar Tigris project. It therefore adopted a pragmatic approach, selecting TSMC’s 3-nanometer and 2-nanometer processes backed by extensive validated data.

As the key determinant of semiconductor performance shifts from front-end processes to back-end packaging, Samsung’s weaknesses have become more pronounced. TSMC’s Chip-on-Wafer-on-Substrate (CoWoS) technology has effectively become the standard for AI chip production. OpenAI’s Tigris project, which requires advanced packaging capabilities for ultra-large-scale computation, faced a prevailing assessment that Samsung had yet to establish sufficient trust in this domain compared to TSMC.

Paradoxically, the fact that TSMC produces Nvidia’s chips most effectively also reinforced OpenAI’s decision. The compatibility of design assets and supply chain stability derived from using the same production lines as Nvidia represent advantages that OpenAI could not forgo. Consequently, Samsung’s foundry business has found itself unable to position as a viable alternative even for customers seeking to reduce dependence on Nvidia.

The AI semiconductor market has now entered an era of customized chips optimized for specific algorithms. OpenAI’s Tigris stands at the apex of this transition. TSMC has outperformed Samsung in its ability to precisely implement highly detailed design requirements into manufacturing processes with flexibility and sophistication. This has fueled criticism that Samsung’s longstanding mass-production, general-purpose approach is misaligned with the demands of the customized AI era.

The Elusive Dream of Foundry Supremacy

Under these circumstances, Samsung’s strategic options are becoming increasingly clear. Rather than focusing solely on securing mega-scale clients, the company must accumulate process stability and yield in the 2-nanometer and 3-nanometer nodes for its Gate-All-Around (GAA) technology, while gradually rebuilding trust through low-cost orders. Although Samsung’s GAA process is recognized for its potential as a next-generation technology, questions surrounding its mass-production reliability persist.

This is evidenced by Qualcomm, long regarded as a pivotal client capable of driving Samsung’s foundry revival, effectively finalizing a complete shift back to TSMC. According to Taiwan-based IT outlet DigiTimes, Samsung’s second-generation 2-nanometer process, SF2P, has failed to meet the mass-production yield thresholds required by global fabless companies. As a result, Qualcomm has internally decided to allocate the entire production volume of its next-generation Snapdragon flagship processor to TSMC’s 2-nanometer lines.

Samsung’s SF2P process was designed as an ambitious effort to maximize power efficiency and performance based on its experience with GAA at the 3-nanometer node. However, the advantages anticipated at the design stage have not been fully realized in actual mass-production environments. As processes advance into ultra-fine nodes, process variables and physical constraints increase sharply, leading to a significant rise in yield management complexity. Based on currently observed performance and stability metrics, the market has formed a consensus that entrusting large-scale volumes to Samsung remains burdensome.

Qualcomm had previously shifted to TSMC without hesitation when Samsung struggled with heat generation and yield issues in its 4-nanometer process. In the current 2-nanometer competition, Qualcomm once again prioritized stability over risk. It judged that TSMC’s proven mass-production reliability and supply capabilities offer a more secure path to safeguarding corporate interests than Samsung’s technological potential.

Through its monopolization of orders from OpenAI and Qualcomm, TSMC has once again demonstrated unrivaled dominance in the 2-nanometer market. As global core clients concentrate in Taiwan, TSMC is securing opportunities to further enhance process maturity through exclusive access to large-scale production data. In contrast, Samsung is facing difficulties in stabilizing its processes while also encountering limitations in accumulating mass-production experience. This widening gap is likely to raise entry barriers even further for Samsung in the future 1-nanometer race. Failure to overcome the formidable barrier of yield could bring Samsung’s foundry ambitions to an abrupt halt before they fully materialize.