$Micron Tech(MU.US)The AI bubble will eventually burst, but I think we are still far from true madness.

There is currently a severe shortage on the supply side: GPUs, power, optical modules, and data center land are all in short supply. 800G hasn't been fully deployed yet, 1.6T is just starting, and CoWoS and PIC capacity is still expanding. The application layer hasn't truly exploded either; many companies are still stuck in POC and small-scale deployments, and it will take time to truly change business models.

Extending the timeline a bit further, if the bubble really is going to burst, I tend to believe it will peak and start declining from late 2027 to the first half of 2028, with more significant declines in 2029 and 2030, affecting the entire AI hardware supply chain. But this decline isn't the death of the industry; it's an adjustment caused by declining returns on capital, combined with excessive valuations and changes in the financing environment. The internet era was the same—it wasn't just too much fiber laid, but also extreme valuation frenzy, a flood of low-quality companies going public, rising interest rates, investors starting to focus on profitability, and tightening financing conditions. These factors combined to create the burst.

However, this decline is not the end, but paves the way for the next phase. When infrastructure bottoms out, it means the underlying costs of computing power, storage, and networking will become very cheap. I believe that when infrastructure costs are low enough, the AI application layer will see a real explosion. The biggest opportunities in the next 5 to 10 years will ultimately shift from selling picks and shovels to selling applications.

The difference between this AI cycle and the internet is that infrastructure and the application layer are likely to overlap for a long time, rather than waiting for hardware to bottom out before the application layer takes over. GPUs are still rising now, while AI search, AI programming, and AI customer service are already making money, AI Agents are entering enterprises, and robots are slowly being deployed. Hardware hasn't peaked yet, but applications are already emerging. These two tracks are likely to run in parallel.

My judgment is that technological revolutions follow a fixed script: infrastructure boom, declining returns on capital, infrastructure valuation reset, and application layer harvesting. This was true for the railway era, the electricity era, the internet era, and to some extent, the cloud computing era. History always repeats itself, and this time is likely no exception. The biggest difference between AI and the internet is that this time, infrastructure and application layers will develop in parallel for a long time, rather than a linear handoff.