TSMC to expand CoWoS equipment orders: huge AI chip demand from NVIDIA, Amazon, Meta, Tesla

TSMC is urgently expanding its orders for CoWoS packaging equipment, with Taiwanese factories welcoming huge orders for equipment machines.

2

The unstoppable wave of AI is coming in many forms, with the AI GPUs and AI processors being made mostly at TSMC. TSMC estimates that the server AI processor revenue would more than double in 2024, accounting for around 11-13% of total revenue in 2024. However, server AI processors will account for 50% of annual revenue by 2029.

The server AI processor that TSMC mentions refers to the GPU, AI accelerator, and CPU that perform training and inference functions, but it does NOT include network, edge, or terminal device AI. TSMC emphasized at the North American Technology Forum that its 2.5D advanced packaging CoWoS (Chip-on-Wafer-on-Substrate) is a key driving technology for the AI revolution, which allows for more processor cores and HBM on a single interposer layer.

HBM3E memory is the latest generation (5th generation HBM) and will be produced on CoWoS-S and CoWoS-L, while next-generation HBM4 (6th generation HBM) will see TSMC upgrading its CoWoS packaging technology to meet the upgraded, next-generation performance requirements of HBM4 and HBM4E.

• Read more: NVIDIA's next-gen R100 AI GPU: TSMC 3nm with CoWoS-L packaging, next-gen HBM4 in Q4 2025
• Read more: TSMC to expand advanced packaging capacity at 3 plants: CoWoS, SoIC, SoW for AI GPU demand
• Read more: NVIDIA and AMD have reserved all of TSMC's CoWOS and SoIC advanced packaging for 2024 and 2025

TSMC will see a "grand expansion" of its production lines in the next 3 years, where DigiTimes reports that it's estimated that from Q4 2024 to 2026, Longtan, Zhunan, Zhongke, Nanke, and the recently announced Chiayi AP7 fab will be launched.

Semiconductor equipment industry members said that TSMC not only disclosed the huge revenue contributions that will be created from server AI processors over the coming years, but revised its production capacity plan and released "urgent orders" for CoWoS equipment to the equipment supply chain.

The monthly capacity is expected to hit 36,000 pieces by the end of 2024, which is double the 18,000 pieces in 2023. However, TSMC will be boosting CoWoS throughout the years, with 50,000 pieces expected in 2025, and at least 55,000+ in 2026 and beyond.

DigiTimes reports that equipment industry players further pointed out that all TSMC's AI chip manufacturers are chasing production capacity. Not only NVIDIA, which has heavily invested in 50% of its production capacity, placing orders -- in advance -- for the next 3 years, but also want to catch up with AMD and seize AI business opportunities.

Google, Intel, Meta, Microsoft, AWS, Tesla, and many other major manufacturers are also 'bound' by TSMC's one-stop service. These customers are required to provide the production volume in advance, which also makes TSMC, which is full of orders and advances, get hit with a wall of AI demand, which is prompting this huge production expansion.