Every time I think "it's time to focus on applications," new hardware hype emerges?

The past three years of the US stock market's AI bull run can be seen as a process of continuously seeking out system bottlenecks, which can be divided into five thematic phases: First Phase: GPU Era (2023) Whoever owns computing power, owns the future. The representative company is Nvidia. Second Phase: Data Center Era (First half of 2024) There aren't enough server rooms, not enough power supply, not enough cooling systems, and network bandwidth is also insufficient. Capital begins to move from GPUs to the entire data center industry chain. The representative company is Vertiv. For the first time, the market forms a key realization: AI is an infrastructure movement. Third Phase: Power Era (First half of 2024) Data centers can be built, but there is one resource that cannot be quickly increased: electricity. The representative company is Vistra. For the first time, the AI trend expands from the tech sector to the industrial sector. Fourth Phase: HBM Era (2025) When the bulk of new demand comes from inference, the data simply cannot be delivered to the GPU quickly enough, shifting the AI bottleneck from computing to data movement. The representative companies are the three major memory manufacturers. Fifth Phase: High-speed Interconnect Era (2026) Data transfer once again becomes the bottleneck; the market is no longer trading individual chips, but entire system architectures. The representative company is Broadcom. What’s frustrating is that every time people think hardware speculation has run its course and it's time to speculate on applications, the market circles back to new hardware. Why is today’s AI industry trend so different from that of the current Internet? In this Wednesday’s live broadcast, I will talk about this surge. You are welcome to book your spot (UTC+8):