Memory and photonics rotation gains traction as capex cycle diversifies beyond GPUs

The AI capex boom has been heavily concentrated in GPU players like Nvidia, but chip manufacturing cycles are maturing and capital is rotating into supporting infrastructure. Memory (DRAM and NAND) demand is accelerating as data centers scale inference workloads and require massive storage. Micron's stock has surged more than 100% year-to-date, buoyed by expectations that enterprise-AI cloud deployments will drive DRAM and SSD demand through 2027. Applied Materials and KLA are seeing robust equipment order books as fabs scale capacity, and watch lists are flagging memory rotation as a thematic shift away from GPUs.

Photonics is another focal point. Broadcom's optical interconnect revenue has grown 50% year-over-year as hyperscalers build out high-speed intrachip communication and data-center networking. Marvell Technology and Wyle Electronics are seeing strong demand for photonic components. The market has been systematically rewarding positions in memory and photonics over the past month, with rotations between Applied Materials and KLA reflecting chip-cycle seasonality. This suggests that traders are positioning for a broader multi-year semi supercycle rather than a near-term GPU-only rally.

The bifurcation creates winners and losers. Integrated device manufacturers like Intel face margin pressure as custom silicon from hyperscalers eats share, while fabless and foundry plays (Broadcom, Marvell, Qualcomm) win. Traders are also eyeing smaller memory specialists like SK Hynix peers and newer lithography equipment makers as inflationThe rate at which prices rise across an economy.-hedging plays. Some observers warn that memory capex intensity could overshoot, creating a supply glut post-2027, but near-term demand indicators remain robust.

The technical picture shows a constructive setup. Memory chip futures have broken through key resistance, and semiconductor ETFs reflect broad-based strength. However, valuation dispersion is extreme: GPU names trade at 30-40x earnings while memory trades at 10-15x, suggesting either memory is cheap or GPU is overvalued. Consensus points to rotating into memory as capex cycles broaden.