Analyst Kazuyoshi Saito: AI chip shares not a bubble

The article opens with the worry that gripped AI and semiconductor shares in the summer of 2026: concerns about overinvestment by US hyperscalers and Nvidia's "circular transactions" in which it provides funding to customer companies. Rather than treating those as decisive, the analysts interviewed argue the demand picture has shifted in a way that makes the fears look misplaced. Saito's confidence rests on a simple multiplication: as AI models become more capable, token consumption rises exponentially, and memory consumption per token also climbs, together pushing semiconductor demand sharply higher. He also points to CoreWeave, a US cloud provider, investing funds amounting to four times its revenue, arguing such expansion cannot be excessive.

Imanaka adds a structural angle: over the past year, the inside of AI data centers has changed. Until now, usage centered on generative AI like OpenAI's ChatGPT, which performs one inference per human instruction. Going forward, autonomous AI agents that make multiple inferences per instruction and handle tasks like summarizing documents, entering required items, and processing accounting on their own are spreading. Because they frequently use RAG to search external knowledge, system loads rise, and Imanaka says each AI chip now needs multiple CPUs to coordinate the overall system. Next-generation data centers will also need large amounts of NAND flash memory for long-term storage, indium phosphide substrates and optical amplifiers for high-speed optical wiring, and DC power distribution and MLCCs for efficiency. Imanaka links the SOX index's rise since April to these new demand sources.

For Japanese companies, the article sees significant benefit: Saito expects semiconductor manufacturing equipment sales to grow 30% across the market in 2026, with new TSMC and other plants starting full-scale operations from 2028, describing a golden age continuing toward 2030. The stakes for investors seem to hinge on whether the shift to autonomous AI agents actually takes hold as described, since that is what would validate the CPU and memory demand forecasts; if agent adoption stalls, the demand multiplier argument would weaken. The article also notes that the era of buying only Nvidia to benefit from semiconductor growth is over, and suggests SOX-linked ETFs or index funds for those unsure about individual stock selection.