According to Morgan Stanley Fund's equity investment manager Chen Xiuzhu, AI remains the most significant investment theme of this era, but the methods for profiting from it are undergoing a shift. She believes that value is migrating upstream along the industrial chain. The marginal value increments for 2026-2027 are likely to be found in storage, optical modules, PCBs, and their upstream materials, which may represent the alpha in AI investment this year.

Chen Xiuzhu pointed out that looking back at the global market over the past year or so, the money the market has paid for AI has largely gone to the chips themselves. The logic is straightforward: this is a computing power arms race, and whoever has GPUs wins. Therefore, the most certain winners are the "sellers of shovels." Consequently, capital has surged, highly concentrated, into computing power and accelerator cards—the most direct and obvious main path. Nvidia has indeed delivered on its performance, and this logic is understood by almost everyone, with all investors positioned similarly.

By 2026, the bank observes that value is migrating upstream along the industrial chain. From GB200 and GB300 to Rubin, which will enter mass production in the second half of this year, and to Rubin Ultra in 2027, the total bill of materials (BOM) value for each generation of rack is rising stepwise, but it is by no means evenly distributed. The performance of AI computing systems is increasingly bottlenecked by data transfer efficiency, and the value of data movement is beginning to approach that of data processing. What the generational upgrades are truly amplifying is the upstream circle surrounding the GPU. Among these, the most valuable and most likely to be recognized belatedly by the market are three segments experiencing both volume and price increases: storage (Memory), the upstream of optical communication, and the upstream of PCBs. These may be the source of alpha in AI investment this year.

Storage: Memory is the single largest driver in this round of value migration. Memory has entered a super-cycle, with bulk NAND/DRAM prices continuing to rise significantly quarter-on-quarter in 2026, marking the largest increase in a decade. New fab capacity from Micron and SK Hynix will not ramp up until 2027 at the earliest, creating a long-term supply gap. Within the structure, high-end HBM is even tighter. Nvidia recently confirmed that Samsung, SK Hynix, and Micron have all passed HBM4 certification for Vera Rubin, further tightening overall supply. From HBM3E to HBM4 and then to HBM4e used in Rubin Ultra, the value of memory per rack is increasing by more than fourfold. HBM capacity is already sold out through the end of 2026. In the GB200 era, memory accounted for only 5–10% of the rack BOM. Due to both increased usage and price, its share rises to about 25–30% in VR200, with the absolute value increasing approximately fourfold compared to GB300. This directly compresses the GPU's share of the BOM from about 65% in GB200 to about 51% in VR200, even though the absolute value of the GPU is still growing.

Upstream Optical Communication: Optics - With each subsequent generation, the interconnect bandwidth inside the rack is rising exponentially. The physical limits of copper interconnects have been reached, necessitating a switch to optics. For instance, co-packaged optics (CPO) begins to scale in the Rubin generation. Currently, the tighter constraints are on optical chips (EML, CW lasers), silicon photonic chips, optical engines, and passive chips. Therefore, companies that can self-supply or secure their upstream supply are "profiting handsomely," while those only doing assembly are "squeezed from both ends."

Upstream PCB: Copper Clad Laminate and Its Raw Materials - The PCB increment comes from two layers: usage (number of layers, area, new modules) and material upgrades. A teardown by SemiAnalysis also confirms the material trend: the PCB value in the VR NVL72 has increased significantly compared to GB200/GB300. The two main drivers are material upgrades and a significant increase in the area/number of layers of high-end PCBs. This also drives the upgrade of CCL from M7 to M8/M9, a comprehensive upgrade of copper foil on main computing boards and network boards to HVLP4, and the need to upgrade glass fiber cloth to reduce dielectric loss. In the PCB chain, the most valuable parts are similarly the upstream copper clad laminate (CCL) and its raw materials. The logic is driven by a dual engine of single-board upgrades and doubled usage: material grades rise from M8 in GB200 to M9/M10 in subsequent scaling; the number of layers keeps doubling. Price increases have already materialized, with Kingboard having raised prices multiple times this year, cumulatively over 40%. Moving one step further upstream, low-dielectric glass fiber cloth (low-dk second-generation electronic cloth) and HVLP3+ high-end copper foil are the tightest bottlenecks.

Chen Xiuzhu stated that in each round of bottlenecks, it is necessary to identify who has pricing power (concentrated supply, able to pass on price increases to their profit statement), who has customer certification (NVIDIA/Google/Meta, etc.), who can expand capacity (who gets the incremental equipment/materials/yield first), and whose profits will not be quickly eroded by competition. Chips/CoWoS/EUV remain the "crown jewels" with the highest barriers and deepest moats, but the marginal value increments for 2026-2027 lie in storage, optical modules, PCBs, and their upstream materials. Value increases with generational upgrades, supply is genuinely scarce, and there is a strong ability to pass on costs—this is the alpha in AI for this year.