Rotation from the AI/semiconductor sector to other sectors has actually been going on for weeks.

But the biggest recent news is that Moonshot’s Ambiance released Kimi K3, which is being praised for maintaining top-tier intelligence output with fewer resource overheads.
This has raised a concern: if large-model costs can be reduced further, will it lead to a decline in demand for data centers and memory?
I believe the appearance of K3 and the trend it represents will only further accelerate the adoption of AI, because the cost of large models is gradually becoming affordable. Cost optimization will free up more profit margins, and also save more capital that can be allocated to improving terminal UI/UX and optimizing inference.
Training models only account for a small portion of data center spending; model inference is the real big head. And as Token prices fall and intelligence levels continue to improve, AI can handle more and more work, which will also drive more inference demand.
So there are no signs of any slowdown in memory demand, and there’s at least two or three years to wait for Chinese competitors such as CXMT (Changxin Memory Technology) to catch up. A 50% pullback in the memory sector is the result of profit-taking by the trading crowd plus leveraged positions being liquidated.
I tend to think the issue isn’t big.