AI Capex Reset Hits Memory Expectations

Hyperscalers are now reassessing the economics of generative AI data centers, viewing current spending as carrying very high variable costs and inadequate returns on investment. This reassessment is leading to reductions in AI‑focused infrastructure capex, after a rapid buildout that pushed annual AI capex into the hundreds of billions of dollars. The shift puts the durability of the recent AI infrastructure supercycle under scrutiny across equity markets.

Memory semiconductor names tied closely to AI data centers sit at the center of this adjustment. Micron Technology (MU), SK hynix (000660.KS), Samsung Electronics’ memory segment (005930.KS), and Western Digital (WDC) are all major suppliers of DRAM, NAND, HBM, and storage hardware into hyperscale data centers. Their earnings and valuations are highly sensitive to bit demand, pricing, and utilization levels that had been underpinned by aggressive AI server and storage expansion.

Historical capex booms in sectors such as telecom networking, PCs, and solar show that when an end market proves less economic than expected, upstream suppliers can experience severe equity drawdowns in the 70-90% range. In those episodes, high operating and financial leverage, combined with valuations built on extrapolated growth, amplified downside once capex was cut. The current AI data center cycle fits several of these conditions, making a sharp repricing in AI-levered memory and storage stocks a plausible, though not guaranteed, outcome.

For Micron, SK hynix, and Samsung’s memory operations, aggressive investment in advanced DRAM, NAND, and HBM capacity ties future profitability closely to sustained AI data center demand. Western Digital (WDC), while more focused on NAND and HDD storage, is also exposed through cloud and AI storage backends. If hyperscaler AI capex remains under pressure, these companies could see significant earnings volatility and valuation resets, particularly if non-AI workloads and other end markets fail to absorb the capacity built on AI-centric assumptions.