The bullish compute thesis assumes AI capex can keep growing rapidly and, more importantly, that customers will eventually generate enough economic value from AI to justify all that infrastructure. If monetisation disappoints, hyperscalers could moderate spending surprisingly quickly.

Memory is cyclical and vulnerable to oversupply, but demand is increasingly tied to real hardware requirements. AI accelerators need large amounts of high-bandwidth memory, while servers still need DRAM and storage.

So I see memory’s risk as more about supply, pricing and cycles, whereas compute carries a bigger valuation and AI-ROI assumption.

Both can fall, but if the market starts questioning whether every extra dollar of AI capex produces adequate returns, I would be more worried about compute.