July 2, Nomura recently stated in a deep-dive semiconductor report that cloud providers will still find it difficult to stop expanding in 2027. AI model iteration, growth in inference demand, expansion of data center construction plans, and tight supply in memory and advanced packaging will all make cloud providers continue to lock down chip, packaging, substrate, memory, and server resources.

Nomura's logic is that AI capital expenditure is not a short-term choice for a single company, but a competitive pressure among large cloud platforms. As long as companies like Microsoft, Google, Amazon, and Meta are still competing for AI models, enterprise customers, and inference traffic, they will find it difficult to proactively slow down infrastructure construction. Even if costs rise, stopping could mean losing the competitive position of the platform.

The report specifically mentions that although TSMC is expanding CoWoS advanced packaging capacity, small substrate suppliers may become a new bottleneck. In other words, the bottleneck is not only in GPUs, but also in advanced packaging, ABF/substrates, HBM, server assembly, and power infrastructure.

Nomura is therefore bullish on supply chain companies such as TSMC, ASE, Aspeed, MediaTek, GlobalWafers, KYEC, Elite Material, and Zhen Ding. Its judgment contrasts with concerns about "AI overheating": the real problem is not the disappearance of demand, but that the supply chain is still insufficient; as long as bottlenecks exist, cloud providers will continue to pay for scarce capacity.#Gate股票转仓功能上线