Amid the global wave of investments in cloud AI servers, Taiwan Semiconductor Manufacturing (TSM) is seizing unprecedented growth opportunities as a leading foundry.
According to a recent UBS report, every 1GW server project could generate $1–2 billion in revenue for TSMC, equivalent to 1.0–1.5% of its projected 2025 sales. With OpenAI and multiple hyperscale cloud providers announcing server deployment plans in the tens of GW range, TSMC's revenue growth potential far exceeds market expectations.
**Varying Capacity Demands Across AI Platforms** The report highlights that TSMC's revenue from Nvidia's next-gen AI GPU platforms will grow progressively. For every 1GW server deployment, TSMC stands to earn around $1.1 billion from the Blackwell Ultra/Rubin platform, rising to $1.4–1.9 billion for the Rubin Ultra/Feynman platform.
This growth stems from multiple factors: process node migration, increased GPU units per rack, adoption of advanced packaging like CoWoS, and a potential transition to panel-level packaging by 2028.
From a capacity perspective, Nvidia’s Blackwell Ultra requires ~3.5k wafers/month (kwpm), including 2.3k kwpm for xPU chips, 0.5k kwpm for CPUs, and 0.7k kwpm for networking chips. CoWoS packaging demands an additional 2.3k kwpm.
Broadcom’s ASIC solutions show efficiency advantages, with workload-specific designs potentially driving higher chip demand than GPUs. For instance, Broadcom’s TPU v7p for Google requires ~4.9k kwpm of N3 capacity per 1GW—far exceeding Nvidia’s 2–4k kwpm.
In terms of revenue share per rack, ASIC solutions contribute 10–11% to TSMC’s total revenue, versus 4–6% for GPU solutions. Specifically, Google’s TPU v7p could deliver $1.895 billion per 1GW for TSMC.
AMD’s AI GPU demand is the lowest among major players. Its upcoming MI455, using a chiplet design, needs just ~1.0k kwpm of N2 capacity per 1GW, translating to ~$746 million in revenue for TSMC.
**Massive Opportunities per GW Deployment** UBS breaks down that each 1GW server build requires TSMC to supply 2–5k kwpm of advanced-node (N3/N2) capacity and 3–6k kwpm of CoWoS packaging capacity.
From a capex angle, every 1GW project drives $1–2 billion in wafer fab equipment (WFE) investment for logic chip production, covering both front-end and CoWoS capacity. This means each AI infrastructure expansion directly fuels TSMC’s capacity needs and capital expenditure growth.
**OpenAI Deals Hold $34.4 Billion Revenue Potential** UBS estimates TSMC’s potential revenue from OpenAI’s announced deals—totaling 26GW across Nvidia (10GW), AMD (6GW), and Google (10GW)—at $34.4 billion. Breakdown: - Nvidia: $11 billion - AMD: $4.5 billion - Google: $18.9 billion
At a 40–45% operating margin, these projects could contribute ~$14.6 billion in operating profit, underscoring how the AI infrastructure wave is turbocharging TSMC’s earnings.