Under the AI infrastructure boom, Taiwan Semiconductor is making a fortune! UBS: $1 billion to $2 billion in revenue per GW!

Under the global investment wave in cloud AI servers, TSMC, as a major foundry, is experiencing unprecedented growth opportunities.

According to the latest research report from UBS, each 1GW server project will bring TSMC an income opportunity of USD 1-2 billion, equivalent to 1.0-1.5% of its expected sales in 2025. With OpenAI and several large-scale cloud service providers announcing plans for tens of GW-level server construction, TSMC's revenue growth potential will far exceed market expectations.

Significant Differences in Capacity Demand Among Different AI Platforms

The analysis in the report shows that TSMC's actual total revenue from NVIDIA's new generation AI GPU platform will gradually increase. For each 1GW server construction, TSMC will earn approximately USD 1.1 billion from the Blackwell Ultra/Rubin platform, which will increase to USD 1.4-1.9 billion by the time of the Rubin Ultra/Feynman platform.

This growth comes from multiple factors: process technology migration, an increase in the number of GPU units per rack, the application of advanced packaging technologies such as CoWoS, and a potential transition to panel-level packaging technology by 2028.

In terms of capacity consumption, NVIDIA's Blackwell Ultra requires approximately 3.5 thousand wafers per month, with xPU chips consuming 2.3 thousand wafers per month, CPUs consuming 0.5 thousand wafers per month, and network chips consuming 0.7 thousand wafers per month. The CoWoS packaging capacity demand is 2.3 thousand wafers per month. Broadcom's ASIC solution has a clear efficiency advantage, as ASIC chips, with their higher efficiency for specific workloads, may generate a greater demand for chip units than GPUs. Broadcom's TPU v7p designed for Google requires approximately 4.9 thousand wafers per month of N3 capacity for each 1GW, significantly higher than NVIDIA's 2-4 thousand wafers per month.

In terms of the proportion of revenue to rack value, the ASIC solution brings TSMC a higher actual total revenue amount, reaching 10-11%, while the GPU solution accounts for 4-6%. Specifically for the Google TPU v7p project, TSMC's revenue opportunity for each 1GW can reach USD 1.895 billion.

AMD's AI GPU capacity demand is the lowest among major manufacturers. The upcoming MI455, which adopts a chiplet design, requires only about 1.0 thousand wafers per month of N2 capacity for each 1GW, providing TSMC with a revenue opportunity of approximately USD 746 million.

Huge Business Opportunities from Each GW Server Construction

The UBS report breaks down AI server investments, indicating that for each 1GW of server construction, TSMC needs to provide approximately 2-5 thousand wafers per month (kwpm) of advanced process capacity (N3 or N2), as well as 3-6 thousand wafers per month of CoWoS advanced packaging capacity. From the perspective of capital expenditure, each 1GW project will drive an investment of USD 1-2 billion in wafer fab equipment (WFE) for logic chip production, including front-end capacity and CoWoS packaging capacity. This means that every expansion of AI infrastructure will directly translate into TSMC's capacity demand and capital expenditure growth.

OpenAI Deal Implies USD 34.4 Billion Revenue Potential

UBS has conducted a detailed assessment of the deals announced by OpenAI. The total scale of OpenAI's deals with NVIDIA, AMD, and Google reaches 26GW, including 10GW from NVIDIA, 6GW from AMD, and 10GW from Google.

Based on these deals, TSMC's potential revenue reaches USD 34.4 billion, with USD 11 billion from NVIDIA, USD 4.5 billion from AMD, and USD 18.9 billion from Google. With an operating profit margin of 40-45%, these projects will contribute approximately USD 14.6 billion in operating profit to TSMC, fully reflecting the driving effect of the AI infrastructure construction wave on TSMC's performance