Morgan Stanley: 2026 will be the year of AI technology hardware
```
Author: Dong Jing
Source: Hard AI
Morgan Stanley predicts that 2026 will become a key year for explosive growth in AI technology hardware, mainly driven by strong demand for AI server hardware.
According to news from Hard AI on November 3rd, Morgan Stanley stated in its latest research report that AI server hardware is undergoing a major design upgrade driven by GPUs and ASICs. Nvidia’s soon-to-be-launched GB300, Vera Rubin platform and Kyber architecture, along with AMD’s Helios server rack project, will all deliver greater computing power and rack density.
Morgan Stanley emphasized in the report that with Nvidia's Vera Rubin platform launching in the second half of 2026, rack demand is expected to surge from about 28,000 units in 2025 to at least 60,000 units in 2026. More importantly, in order to support higher power consumption GPUs, the value of power supply solutions could increase more than tenfold by 2027, while liquid cooling will become standard, with per-rack value also continuing to climb. In addition, components such as PCBs/substrates and high-speed interconnects will see major specification upgrades.
This means the entire AI server hardware supply chain is being revalued. The firm maintains a positive rating on a number of AI hardware supply chain companies, believing that the next generation of AI server design upgrades will bring rich returns to related manufacturers.
Explosive Growth in AI Server Rack Demand
Morgan Stanley stated clearly in the report that the driving force for AI hardware growth is shifting from the H100/H200 era to a new cycle driven by NVIDIA’s GB200/300 (Blackwell platform) and subsequent Vera Rubin (VR Series) platforms.
According to Nvidia’s product roadmap, its GPUs’ power and performance are undergoing leaps in upgrade.
From the H100’s 700W TDP to B200’s 1,000W, then GB200’s 1,200W, up to the Vera Rubin (VR200) platform in the second half of 2026 with a GPU max TDP soaring to 2,300W, and the 2027 VR300 (Kyber architecture) even reaching 3,600W.
(Nvidia Product Roadmap)
Morgan Stanley believes that this increase in computing power density directly drives a comprehensive overhaul from server rack design to its components.
Morgan Stanley predicts that, for just Nvidia’s platform, AI server rack demand will leap from roughly 28,000 racks in 2025 to at least 60,000 in 2026, more than doubling. Meanwhile, AMD’s Helios server rack project (based on the MI400 series) is also making good progress, further intensifying the market demand for advanced AI hardware.
The report emphasizes that as rack design upgrades from a single GPU to an integrated rack system, ODM manufacturers with strong integration capabilities and stable delivery records—such as Quanta, Foxconn, Wistron, and Wiwynn—will dominate the GB200/300 rack supply.
Power and Cooling Bottlenecks Ignite Value in Power & Liquid Cooling Solutions
One of the most striking perspectives in the report is that the power and cooling challenges brought by AI hardware upgrades are translating into big business opportunities for power supply and cooling vendors. This is seen as one of the fastest-growing links in value in this upgrade cycle.
On power supply, as total power consumption per rack grows sharply, traditional power architectures can no longer keep up.
The report predicts that power architectures will shift to 800V High Voltage Direct Current (HVDC) solutions. This shift will greatly enhance the value content of power supply solutions.
Morgan Stanley expects that, by 2027, the power solution per Rubin Ultra rack (adopting Kyber architecture) will be worth more than ten times the current GB200 server rack. At the same time, by 2027, the value of power supply per watt of power consumption in AI server racks will also double compared to the current stage.
In cooling, liquid cooling has shifted from an optional solution to a necessity. The report notes that as the GB300 platform’s TDP breaks through 1,400W, liquid cooling is becoming standard. This directly drives up the value of cooling components. Specific data show:
A GB300 (NVL72) rack’s total value in cooling components is about $49,860. By the next-generation Vera Rubin (NVL144) platform, thanks to even higher cooling requirements for compute trays and switch trays, cooling component value per rack rises 17% to $55,710. The value increase of cooling modules for switch trays alone reaches as high as 67%.
Morgan Stanley says this clearly shows that key suppliers in the liquid cooling industry chain—such as cold plate modules, cooling fans, quick connectors (NVQD), etc.—will directly benefit.
Comprehensive Value Chain Upgrade: PCB/Substrates and High-Speed Interconnects Heading Up
Aside from power and cooling, the report also highlights the profound impact of AI platform upgrades on PCBs (Printed Circuit Boards) and various interconnect components. Every GPU iteration brings higher requirements for PCB layer counts, material grades, and size.
According to Morgan Stanley’s sorted Nvidia GPU evolution roadmap:
ABF substrate: Layers rose from H100’s 12 layers (12L) to Blackwell (B200)’s 14 layers, then Vera Rubin (VR200)’s 18 layers, with dimensions correspondingly enlarged.
OAM (OCP Accelerator Module) motherboard: PCB spec went from H100’s 18-layer HDI (High Density Interconnect) up to GB200/300’s 22-layer HDI, and potentially reaches 26-layer HDI on VR200.
CCL (Copper Clad Laminate) materials: Migrating from ultra low-loss (M7) to extreme low-loss (M8) grades to meet higher data transfer rate requirements.
Morgan Stanley believes that these technical detail upgrades mean the PCB manufacturing process will be more complex and higher in value. Heightened demand for high-layer-count HDI boards and advanced CCL materials will provide structural growth opportunities for PCB and upstream materials suppliers with relevant technological strength.
The report clearly indicates that the design upgrades of AI GPU and ASIC servers will strongly support a wave of capacity expansion in the PCB/substrate industry.
Meanwhile, Morgan Stanley says data and power interconnect solutions are also upgrading to meet higher data transfer speed and capacity needs.
This article is from the WeChat Official Account "Hard AI". For more AI front-line information, please visit here
Risk Warning and DisclaimerThe market has risks and investment needs caution. This article does not constitute personal investment advice and does not take into account the individual user's specific investment objectives, financial situation, or needs. Users should consider whether any opinions, viewpoints or conclusions in this article are suitable for their particular situation. Investment based on this is at your own risk. ```