Don't Give Up on AI Chips Yet, Says JoAnne Feeney

The market is conflating lower unit compute with lower aggregate compute. In practice, efficiency gains tend to shift spend from brute-force training into broader inference deployment, which is better for the ecosystem than for any single hardware node: networking, power, packaging, and cloud capacity can keep compounding even if GPU counts per model fall. The nearest-term beneficiaries are the infrastructure toll collectors with pricing power and recurring pull-through, while the losers are names whose earnings rely on a sustained shortage premium rather than durable end-demand.

Memory is the fragile link because it behaves more like a commodity than an enabling platform. If hyperscalers conclude they can extract similar model performance with fewer accelerators or slower refresh cycles, DRAM and NAND bit demand can normalize faster than consensus expects, compressing margins before headline AI adoption slows. That makes Micron-style exposure more vulnerable to a 1-3 month multiple reset than the broader AI hardware complex, especially if supply additions arrive just as spending growth moderates.

Contrarian view: the selloff may be overdone in the short run because efficiency often raises usage by lowering cost per query, which is a demand stimulus, not a demand killer. The real reversal risk is not model quality but capex discipline — if hyperscaler budgets flatten while utilization improves, hardware revenue growth can decelerate even as AI adoption remains strong. Watch for the next cloud capex guide and GPU lead-time commentary; those are the tells that separate temporary de-rating from a genuine demand inflection.