SMIC's AI semiconductor yields plummet 30%, sales forecast halved.
SMIC, China's largest foundry (semiconductor contract manufacturing company), is facing a red flag for its plans to produce artificial intelligence (AI) chips. With yields at key processes hovering around 30% due to US sanctions on advanced equipment, global investment bank Morgan Stanley has effectively halved SMIC's AI chip-related revenue forecast. China's semiconductor boom has hit a snag due to technological limitations and external sanctions.
The biggest challenge facing SMIC is the low yield of its AI graphics processing units (GPUs). According to a Morgan Stanley report on the 7th (local time), SMIC's AI GPU yield is projected to remain at 30% by the end of 2025. This is a significant improvement compared to its competitor, Taiwan Semiconductor Manufacturing Company (TSMC), which boasts a yield of over 90% for its 5nm process.
US sanctions are cited as the fundamental cause of this low yield. SMIC is unable to secure cutting-edge extreme ultraviolet (EUV) lithography equipment from ASML in the Netherlands, forcing it to rely on older deep ultraviolet (DUV) equipment. Implementing 7nm-level fine processes using DUV equipment requires the use of "multi-patterning," a technique that involves repeatedly drawing circuits. This process is technically complex and increases production steps, leading to increased costs and lower yields. This is compounded by the lack of precision maintenance technology for Western manufacturers' equipment. In fact, SMIC suffered approximately 6% yield and production losses due to equipment inspection failures in the second quarter of 2025.
◇ Huawei Hit by Direct Hit: AI Chip Production Costs 'Snowballing'
US EUV Equipment Sanctions Hit Directly… Limitations of DUV Multi-Patterning Process Revealed
Huawei's AI Chip Production Costs Soar… Technology Gap with TSMC Widens
SMIC, China's largest semiconductor foundry, is facing disruptions in AI chip production due to US sanctions on advanced equipment. The inability to secure EUV equipment has led to yields dropping to around 30%, causing production costs for customers like Huawei to skyrocket and widening the technological gap with TSMC. Photo = Technetbook View larger image
SMIC, China's largest semiconductor foundry, is facing disruptions in its AI chip production due to US sanctions on advanced equipment. The inability to secure EUV equipment has led to yields dropping to around 30%, causing production costs for customers like Huawei to skyrocket and widening the technological gap with TSMC. Photo = Technetbook
China's largest foundry (semiconductor contract manufacturing) company SMIC's plans for artificial intelligence (AI) chip production are facing a red flag. With yields in key processes remaining at around 30% due to US sanctions on advanced equipment, global investment bank Morgan Stanley has effectively halved its forecast for SMIC's AI chip-related revenue. China's semiconductor growth is now facing challenges due to technological limitations and external sanctions.
The biggest challenge facing SMIC is the low yield of its AI graphics processing units (GPUs). According to a Morgan Stanley report on the 7th (local time), SMIC's AI GPU yield is projected to remain at 30% by the end of 2025. This is a very serious situation, compared to its competitor, Taiwan's TSMC, which boasts a 5nm process yield of over 90%.
The fundamental cause of this low yield is attributed to US sanctions. SMIC, unable to secure cutting-edge extreme ultraviolet (EUV) lithography equipment from ASML in the Netherlands, relies on older deep ultraviolet (DUV) equipment. To achieve 7nm-level fine processes using DUV equipment, it is necessary to use "multi-patterning," a technique that repeatedly draws circuits. This process is technically complex and requires multiple production steps, leading to higher costs and lower yields. Furthermore, the lack of precision maintenance skills for Western manufacturers' equipment has compounded the problem. In fact, SMIC suffered approximately 6% yield and production losses due to equipment inspection failures in the second quarter of 2025.
◇ Huawei Takes a Direct Hit, AI Chip Production Costs 'Snowball'
The production disruption has dealt a direct blow to Huawei, China's largest information and communications technology (IT) company. SMIC planned to produce Huawei's AI chip, the "Ascend 910B," at a monthly rate of 7,000 wafers this year, and transition to the high-performance "Ascend 910C," a chip combining two 910B dies, starting in 2026. However, this structure is bound to further increase yield and cost burdens. The low yield has led to a surge in chip production costs. This year, the production cost of a single Ascend 910B chip is expected to reach 50,000 yuan (approximately 9.74 million won), while the next-generation 910C chip is expected to cost 110,000 yuan (approximately 21.42 million won).
◇ Sales Forecast 'Halved'… Widening Technology Gap
Morgan Stanley has significantly lowered its revenue forecast for SMIC's AI chip division. The 2025 revenue forecast was lowered by nearly 50%: from 146 million yuan (approximately 28.4 billion won) to 58.5 million yuan (approximately 11.3 billion won), from 212 million yuan (approximately 41.3 billion won) to 94 million yuan (approximately 18.3 billion won) in 2026, and from 286.5 million yuan (approximately 55.8 billion won) to 136 million yuan (approximately 26.4 billion won) in 2027.
The report projects that SMIC's 910B chip yield will recover to 70% by 2027, but short-term production disruptions and financial impacts are unavoidable. Accordingly, SMIC is urgently investing an R&D budget of 30 to 75 million dollars (approximately 41.6 to 104.1 billion won) and is staking everything on securing yields by rapidly increasing production facilities, but it is assessed that the technological gap with Samsung Electronics and TSMC still remains.
By looking at the original chart from Morgan Stanley you can easily tell these are just bogus numbers made up by them. They put yield of 910B wafers at 30% and 910C wafers at 5%. If it were based on any real data 910C wafers would have the same yield as 910B because 910C is not a monolithic chip. It's just two 910B dies packaged together.