Chip memory crunch widens gap between AI winners and losers

The artificial intelligence capex boom is hitting a structural bottleneck: memory chips. High-bandwidth memory (HBM) and advanced DRAM are in acute shortage as hyperscalers race to build out AI infrastructure. Companies with secured supply contracts from Samsung, SK Hynix, and Micron are winning; those without are facing project delays and margin pressure.

Broadcom, which supplies memory interconnect and networking components, has outperformed Nvidia in recent weeks as investors increasingly focus on the bottleneck in the supply chain beyond just GPUs. Western Digital, a memory-storage play, has also rallied, up roughly 3x Nvidia in one month according to social-media mentions in the batch. The divergence signals that memory is the constraint, not compute power. Data-center customers are dual-sourcing and negotiating hard to secure HBM supply.

For equities, this is a rotation story within semiconductor and infrastructure plays. Companies exposed to memory, thermal management, and power delivery are gaining relative strength. Companies dependent on merchant allocations of HBM from TSMC and Samsung are facing pushback. This also benefits memory manufacturers themselves; Micron and SK Hynix are seeing pricing power and demand strength. However, it creates headwinds for smaller AI infrastructure plays that lack direct relationships with memory suppliers.

The constraint is likely to persist through 2026. TSMC and Samsung are increasing HBM capacity, but lead times stretch to 12-18 months for new fabs. Some observers view this as a healthy correction within the AI trade; others worry it signals that the AI capex cycle may face a supply-chain wall that slows deployment velocity and returns on invested capital.