Record Semiconductor Revenue: Don't Rush to Claim AI Implementation Yet

When Omdia's data came out, many people's first reaction was that AI has won big again. Global semiconductor revenue in Q2 2026 hit $425 billion, up 31.4% quarter-over-quarter, setting a new record. The numbers are indeed impressive, but let me throw some cold water on this: it looks more like upstream capital expenditure and storage prices are jointly boosting the market, which doesn't mean downstream AI implementation has succeeded.

Look at the structure to understand. Reports say AI demand is reshaping the industry, and Omdia's previous forecasts also placed computing and data storage, DRAM, and NAND at the core. In other words, money is mainly flowing into data centers, VRAM, memory, and accelerator cards. In contrast, general semiconductors only saw a slight increase of just over 2% quarter-over-quarter. This contrast is key. End-device manufacturers are stockpiling, model companies are burning through GPUs, cloud providers are expanding capacity, but the number of people truly willing to pay long-term for AI features is still far from reaching stable repeat purchases.

I've recently been integrating APIs and reviewing client POCs, and the feeling is very direct. The flashier the model demos, the easier things get stuck during the implementation phase. Costs, hallucinations, permissions, data cleaning—none of these are cheap. Record upstream revenue indicates real demand for compute power; but it also shows that the bubble is being cashed out early on the hardware side. This direction is overheated, while actual implementation is still in its early stages.