Memory, optical connectivity and PCBs rise together: How is AI demand reshaping three supply chains?
Long-context inference is increasing demand for memory, upgrades to optical architecture are raising the value of connectors, and PCB manufacturers are beginning to pass price increases through to gross margins; earnings are materializing at different rates across the three supply chains.
AI infrastructure upgrades are simultaneously increasing memory capacity requirements, optical connection density and the value of server PCBs, but the timing of new supply, capacity ramp-ups and cost pass-through determine when each segment benefits.
Why could the memory shortage last longer?
Long-context inference increases cache requirements: the report estimates that a single 1k-token request generates approximately 0.5 GB of KV cache, while 128k tokens require approximately 64 GB. Caching therefore extends from GPU buffers to a multilayer system that includes HBM, DDR, pooled memory and SSDs; the report forecasts that servers will account for 37% of downstream NAND demand in 2026.
Demand is growing faster than new capacity is coming online. The report forecasts that both DRAM and NAND bit shipments will grow by approximately 20% in 2026, while most new capacity will come online after 2027; Micron’s Singapore NAND plant is expected to begin output in the second half of 2028, while YMTC’s third-phase Wuhan facility could begin mass production earlier, in the second half of 2026.
SK hynix expects supply shortages to persist until 2030 and quarterly earnings to maintain their upward momentum at least until 2027. The report also says products such as HBM4 can be customized to customer requirements, making them less commodity-like and helping reduce cyclical volatility.
How do optical architecture upgrades increase connectivity value?
High-speed optical modules depend not only on chip performance but also on greater precision in the coordination of their internal optical components. The report estimates that an FAU connecting optical fiber to a chip is worth approximately 1.5 US dollars per channel in a pluggable architecture, rising to 9.4 US dollars with an upgrade to NPO; with CPO, the use of an easier-to-maintain DFAU raises the value to 21.9 US dollars.
Interfaces connecting optical fibers to one another are also evolving toward greater density. North American cabling company PH disclosed MPO orders exceeding 500 million US dollars; an MPO order announced by Zhishang Technology on 30 September was worth approximately 3.5 billion yuan, with a North American data center cabling company as the customer.
The realization of earnings is affected by orders and capacity ramp-ups. The report says that, excluding the impact of equity incentives, third-quarter net profit attributable to shareholders at Shijia Photons and Zhishang Technology rose 61% and 141% quarter on quarter, respectively; a low base, rising order volumes and capacity ramp-ups together amplified the increases.
Why are PCB price increases beginning to show up in profits?
Upstream copper-clad laminate prices continue to rise, and PCB manufacturers’ ability to pass those increases downstream is key to realizing earnings. Kingboard Laminates raised prices for the eighth time during the year, increasing selling prices for thin-cloth laminates and prepregs by 15%; Panasonic also notified customers that prices for some copper-clad laminates would rise by 200% starting in 2027.
Gross margins at some midstream manufacturers have already improved: Shennan Circuits’ quarterly gross margins in the first three quarters of 2026 were 29.2%, 32.4% and 34.9%, respectively. WUS Printed Circuit forecasts that net profit attributable to shareholders for the first three quarters will rise by between 144% and 168% year on year, and its Thai subsidiary has moved from ramping up capacity to operating at scale.
Demand is also changing. The report says the Rubin platform uses a cable-free architecture, with multilayer PCBs handling more high-speed transmission, and the value of PCBs per server rises by more than twofold compared with the previous generation; the server CPU-to-GPU ratio is shifting from 1:8 toward 1:2 and even 1:1, also increasing demand for motherboard PCBs.
