No AI Bubble, But Memory and Power Will Be the First Bottlenecks

I noticed an interesting detail: when ADATA Chairman Chen Libai said "let's discuss the AI bubble after 2030," he was essentially viewing AI from the perspective of a memory manufacturer. This reminds me of last year in the lab, when I ran a large protein complex prediction with AlphaFold2. The 128GB VRAM was maxed out, the server's power consumption spiked to 800W, and the data center AC went into alarm mode. Since then, every time I optimize a model, I subconsciously ask myself: is the memory and power budget sufficient for this task?