Citigroup raises TSMC CoWoS capacity forecast: AI demand continues to surge, NVIDIA iterations and cloud vendor ASICs become key drivers

With high demand for AI, Citigroup raises CoWoS capacity expectations, with NVIDIA's product iteration and cloud vendors' ASIC layout becoming the main drivers.

According to the Chasing Wind Trading Desk, Citigroup's latest research report has raised TSMC's CoWoS advanced packaging technology capacity expectations from 800,000 units in 2026 to 870,000 units, reflecting the continued strong demand in the artificial intelligence sector, as well as growth opportunities brought about by larger chip sizes, the ramp-up of ASIC accelerators, and application diversification.

Citigroup analysts pointed out that although some downstream ODM manufacturers have provided lackluster guidance, supply chain leaders represented by Hon Hai still maintain strong momentum and an optimistic outlook. The order outlook for NVIDIA is positive, and its wafer revenue at TSMC is expected to achieve over 50% year-on-year growth in 2026.

The research report emphasizes that the ASIC development plans actively promoted by cloud service providers (CSPs) will become the second engine driving TSMC's steady growth. The demand for advanced packaging is expanding to more application areas such as server CPUs, further supporting optimistic expectations for CoWoS capacity.

AI infrastructure becoming increasingly complex, industry barriers continuously rising

As AI chips move towards more advanced process nodes, the requirements for data transmission speeds are increasing, and system design complexity is rising, leading to continuously climbing industry barriers. The Citigroup report indicates that by 2027/2028, the power consumption of a single AI system cabinet may reach 800-900 kilowatts, placing higher demands on cooling and power systems.

With the increasing importance of scalability and high-speed serial/parallel interfaces (SerDes I/O), more network switch chips and server CPUs will also adopt advanced packaging technology, further driving CoWoS demand. Citigroup believes that the increasing complexity of chip and system design will allow leading suppliers in the AI supply chain to continue enjoying better growth prospects.

"Despite facing challenges from product iterations in the upstream and downstream supply chains, based on recent comments from cloud service providers, we see that AI demand is still surging," said Citigroup analyst Laura Chen in the report, "We are not concerned about mismatches in the upstream and downstream supply chains."

NVIDIA's AI chips remain strong, next-generation products accelerating iteration

According to Citigroup research, NVIDIA's GB200 is still the main configuration for AI data centers, with the GB300 expected to gradually ramp up in the fourth quarter of 2025. More notably, Citigroup expects NVIDIA's next-generation system, Vera Rubin, to be officially launched at the 2026 GTC conference and put into use in the second half of 2026.

The Vera Rubin system will continue to be based on the current GB200/300 system design and Oberon architecture, while introducing N3 process GPUs, higher memory density (including LPDDR and HBM4), dual CX9 connectors, and higher power, which will support the long-term growth of TSMC, Jingyuan Electronics, and ASE in the AI chip business.

Citigroup pointed out that the delivery cycle from chip to system may take 9 months, and although there are challenges from product iterations, the upstream supply chain is prepared for the smooth ramp-up of the GB300

Cloud Vendor ASIC Accelerators Become the Second Growth Engine

Citigroup's report specifically points out that among all cloud service providers, Google and Amazon AWS are the most active in developing their own ecosystems, while Meta is also increasing its investment. Citigroup expects the shipment of ASIC chips to reach 400,000 to 500,000 units by 2026.

In Google's TPU supply chain, Broadcom remains the main supplier, but Google's internal team is also collaborating with MediaTek. Analysts expect MediaTek's $1 billion ASIC revenue target to be achieved on schedule, considering there will be at least 200,000 chip shipments by 2026.

For AWS, the N3 process Trainium 3 is expected to achieve larger-scale mass production in the second half of 2026. Although the supply chain has discussed the N2 process Trainium 4, Citigroup believes this is more likely to be a product for 2027 or 2028.

Microsoft is relatively slow in developing its own AI ASICs and still mainly relies on NVIDIA and AMD's GPU solutions, but Citigroup has observed that Microsoft has resumed activities related to Maia 300, expecting small-scale mass production next year.

Demand for Advanced Packaging Expands, Diversified Applications Provide New Growth Points

Citigroup's research emphasizes that the application of advanced packaging technology is expanding from AI accelerators to more fields. With the increasing demand for scaling and high-speed data transmission, network switching chips and server CPUs are also beginning to adopt advanced packaging technologies such as CoWoS, providing more growth opportunities for Taiwan Semiconductor.

The report also points out that as system complexity increases and data transmission requirements rise, AI infrastructure construction faces higher barriers, allowing industry-leading suppliers to gain competitive advantages and enjoy better growth prospects.

"The growth in demand for TSMC's CoWoS mainly comes from larger chip sizes, the ramp-up of ASIC accelerators in the second half of 2026, and the expansion into other applications such as server CPUs," Citigroup analysts summarized in the report. These factors collectively support Citigroup's upward revision of TSMC's CoWoS capacity expectations