Investing.com — Mizuho Securities Asia raised its forecast for TSMC’s monthly CoWoS packaging capacity to 140,000 units by 2026 and 190,000–200,000 units by 2027, up from prior estimates of 120,000 and 170,000–180,000 respectively, as a sharply upgraded outlook for AI-driven server CPU demand forces a broad revision to the firm’s semiconductor supply model.
Taiwan Semiconductor Manufacturing (TW:2330) trades on the Taiwan Stock Exchange under ticker 2330, and its U.S.-listed ADR (NYSE:TSM) is among the most closely watched proxies for AI infrastructure spending. The upward revision to packaging capacity forecasts signals that Mizuho now sees TSMC’s manufacturing footprint as needing to expand faster than previously modeled to meet accelerating orders.
“We foresee stronger demand outlook for AI and server CPU in 2027, including Nvidia Vera CPU, Intel and AMD server CPU, and CSP players’ server CPU (Google, AWS, Microsoft, and Meta),” said Mizuho analyst Kevin Wang. “This has prompted higher capacity requirement for advanced nodes and CoWoS.”
The revision to CoWoS capacity runs in parallel with an aggressive upgrade to the firm’s demand forecasts. Mizuho now projects 630,000 CoWoS units for NVIDIA Corporation (NASDAQ:NVDA) at TSMC in 2026, rising to 1,005,000 in 2027, with the increase attributed to higher demand for Nvidia’s Vera CPU in 2026–27 and stronger production for the Rubin architecture in 2027. The Broadcom Inc (NASDAQ:AVGO) estimate for 2027 was trimmed slightly to 425,000 from 450,000, while the MediaTek figure was nearly doubled to 180,000 from 93,000, driven by stronger Google TPU demand.
On the advanced node side, Mizuho expects TSMC’s N3 monthly capacity to reach 170,000 by 2026 and 200,000 by 2027. N2 capacity is projected at 90,000 by 2026, 150,000 by 2027, and 200,000 by 2028. Looking further out, the A14 node is expected to reach 15,000 units per month by 2027 and 40,000 by 2028, a timeline that underscores how the most advanced silicon generations will be in meaningful production well before the end of the decade.
The macro driver behind all of these revisions is a structural step-up in server CPU demand. Wang is explicit: “We believe server CPU demand for AI applications, including x86 and ARM-based, to grow above 50% YoY in 2027. ARM-based server CPU should be more than double in 2027, compared with 2026.” Mizuho’s unit-level forecasts flesh out that thesis in considerable detail. The firm projects Nvidia Vera CPU production at more than 7 million units in 2027, Advanced Micro Devices Inc (NASDAQ:AMD) Venice CPU at 5 million units, Google CPU at more than 4 million units, AWS CPU at more than 3 million units, Microsoft CPU at roughly 1 million units, and Meta CPU at 100,000–200,000 units.
