For the rotation out of memory into compute to last, I would want three confirmations.

First, TSMC and the hyperscalers need to maintain or raise AI capex guidance. That would support Nvidia, AMD and other logic names even if conventional memory pricing becomes less certain.

Second, compute earnings need to show that demand is translating into margins and cash flow, not merely huge AI spending. That matters increasingly as investors question the returns being generated from AI infrastructure.

Third, memory needs a genuinely weaker fundamental signal. China's DRAM progress alone may create a valuation scare, but it does not necessarily mean near-term oversupply. If HBM remains tight and DRAM contract prices continue rising, money could rotate back quickly.

So my test is simple: strong AI capex + resilient accelerator margins + deteriorating memory pricing/order visibility. If all three appear together, this becomes a durable rotation. Without the third, it may just be profit-taking within an AI trade that still needs both compute and memory.