Nomura and Mizuho analysts warn of impending AI semiconductor packaging crisis: CoWoS capacity bottleneck will trigger 'epic shortage' and unavoidable 2027 price hikes.
Wall Street and top-tier Asian brokerages are signaling simultaneously that the semiconductor investment cycle remains far from its peak, with supply-demand mismatches set to drive a wave of price increases. Nomura Securities and Mizuho Securities have issued back-to-back reports challenging the narrative that the semiconductor cycle has peaked. Despite recent pullbacks in AI semiconductor stocks, both institutions point to the rise of agentic AI and sustained capital expenditure from cloud service giants as catalysts for an unprecedented supply crunch spanning the second half of 2026 through 2027.
Nomura warned bluntly that an "epic" component supply gap is approaching, with shortages no longer confined to advanced process nodes and high-bandwidth memory but spreading throughout the hardware supply chain. Beyond tight capacity for advanced packaging, wafer-level substrates, printed circuit boards, copper-clad laminates, IC substrates, high-end capacitors, power management ICs, and optical components now face acute shortages. NVIDIA's Rubin architecture and Amazon AWS's Trainium 3 will enter mass production in the second half of 2026, and since greenfield capacity typically requires two years to build, severe supply constraints will persist through 2027. This broad-based component shortage will constrain both AI server shipment growth and squeeze supply for non-AI sectors including consumer electronics and automotive.
Nomura's proprietary tracking data shows global data center projects have grown from 240 to 280, with approximately 50 gigawatt-scale hyperscale projects underway. New computing power deployments are expected to reach 32 gigawatts in 2027, driven by capital expenditure from Microsoft, Google, Meta, AWS, and emerging providers like CoreWeave. Additionally, the Chinese government has drafted a national AI computing power network plan with intended investments of $295 billion over five years. Nomura has raised its global server market expectations significantly, forecasting AI server revenue growth of 78% in 2026 and 76% in 2027.
In advanced packaging, TSMC Chairman C.C. Wei stated the company will strive to meet all demand and forgo no business opportunities. However, bottlenecks in wafer-level substrates and other sub-components may limit actual output below planned targets. Mizuho Securities raised its forecast for TSMC's average monthly CoWoS packaging capacity to 140,000 wafers in 2026 and 190,000 to 200,000 wafers in 2027, driven by strong demand for AI and server CPUs including NVIDIA's Vera CPU, server CPUs from Intel and AMD, and custom chips from major cloud service providers.
Nomura predicts intense competition for capacity. NVIDIA will occupy approximately 55% of TSMC's CoWoS capacity in 2027, while Google's TPU share will jump from 23% in 2026 to 27%, making it the fastest-growing AI logic chip. AMD and AWS will face severely squeezed capacity. Mizuho forecasts NVIDIA's CoWoS demand at TSMC will reach 1.005 million wafers in 2027, up significantly from 630,000 in 2026. MediaTek's demand, driven by Google TPU requirements, will nearly double to 180,000 wafers from the previously forecast 93,000; Broadcom's 2027 forecast was slightly lowered to 425,000 wafers.
Intel's EMIB-T interconnect technology has emerged as a potential threat to TSMC's advanced packaging dominance. Mizuho noted that Intel's EMIB-T yield has exceeded 95%, with technology adoption discussions underway with MediaTek, Ampere Computing, AWS, and Tesla. Google's next-generation TPU v9 reportedly plans to collaborate with MediaTek and adopt Intel's EMIB-T packaging. To maintain its lead, TSMC is accelerating mass production of SoIC and CoPoS technologies.
The explosion of agentic AI is unexpectedly driving demand for traditional server CPUs. NVIDIA CEO Jensen Huang noted that the future AI economy is token-based and its Vera CPU is designed for low-latency, high-bandwidth agentic AI applications. AMD CEO Lisa Su stated that agentic AI requires CPUs for orchestration and data movement, forecasting that total addressable market for server CPUs will exceed $120 billion by 2030. Arm CEO Rene Haas is similarly bullish on this hundred-billion-dollar opportunity.
Mizuho analyst Kevin Wang provided aggressive quantitative forecasts: server CPU demand for AI applications will grow over 50% year-over-year in 2027, with ARM-based server CPU demand more than doubling compared to 2026. Specific production projections for 2027 include NVIDIA's Vera CPU exceeding 7 million units, AMD's Venice CPU reaching 5 million units, Google's custom CPU surpassing 4 million units, AWS's custom CPU exceeding 3 million units, Microsoft's custom CPU around 1 million units, and Meta's custom CPU approximately 100,000 to 200,000 units.
The surge in CPU demand directly benefits outsourced semiconductor assembly and test providers and packaging equipment suppliers. Mizuho expects ASE Technology Holding's average monthly CoWoS capacity to reach 20,000 wafers in 2026 and 40,000 to 45,000 wafers in 2027; Amkor's average monthly capacity will expand to 20,000 to 25,000 wafers by end of 2027. On the equipment side, TSMC and ASE collectively ordered over 130 thermocompression bonders from Shibaura in 2026, with TSMC planning to order 50 to 80 fluxless bonder tools in the second half of 2026 through the first half of 2027.
Both institutions have taken bullish action. Nomura reiterated "Buy" ratings on nine Asian AI technology companies including TSMC, ASE Technology Holding, and MediaTek, raising target prices across the board and advising investors to buy on market weakness. Mizuho maintained its "Buy" rating on TSMC with a target price of NT$3,000, issued a "Buy" rating on ASMPT with a target price of HK$280 (approximately $35.70), and a "Buy" rating on ASML with a target price of €2,000.