Cloud Capex Growth to Moderate, Shifting AI Hardware Investment Dynamics

A recent industry report from GF Securities (Hong Kong) indicates that capital expenditure (capex) from the top five US Cloud Service Providers (CSPs) is projected to continue its upward trajectory in the coming years. Forecasts suggest spending will hit $1.19 trillion by 2027, rising to $1.428 trillion and $1.642 trillion in 2028 and 2029, respectively. However, a pivotal trend emerges: the growth rate is expected to decelerate, stepping down from 45% in 2027 to 20% in 2028 and 15% in 2029. This signals that while total investment is expanding, the pace of growth is entering a phase of moderation.

Core Computing Tracks: The Resilient Case for GPUs, CPUs, and Optical Modules

The report highlights stronger structural opportunities within specific segments of the AI investment wave. Analysts maintain a positive outlook on the prospects for Graphics Processing Units (GPUs), Central Processing Units (CPUs), and optical modules (critical for high-speed data center interconnects). These components form the foundational building blocks of AI computing infrastructure, with demand directly tied to CSPs' server deployment and network upgrade cycles. Even amidst a slowing overall capex growth rate, investment priority for these core hardware categories is likely to remain elevated.

Memory Chips: Caution Amid the Boom

In contrast to the optimism for compute hardware, the report adopts a more cautious stance towards the memory chip sector. This prudence stems from evolving industry dynamics. The analysis points out that leading AI platforms, exemplified by Nvidia, are actively pursuing cost-reduction strategies, a key part of which involves scaling back specifications for high-performance memory like High Bandwidth Memory (HBM).

  • Specification Reduction Example: For instance, the planned HBM stack count for a next-generation AI platform has been revised down from 12 layers to 8 layers.
  • Industry-Wide Trend: This move is not isolated. Major tech firms including Meta and Google are also shifting towards 8-layer stacks in their custom chip designs.

These developments suggest that the consumption of high-end memory chips per AI system could be lower than previously anticipated, introducing uncertainty into the demand growth forecast for memory chips, particularly premium HBM.

Investment Implications: Focus on Value, Beware of Divergence

Overall, the report outlines a new phase in AI investment, where CSP spending is transitioning from "rapid expansion" to "quality growth." This shift will inevitably lead to a divergence in how benefits are distributed across the upstream supply chain. Investors may need to sharpen their focus on the indispensable "hardcore" components within the compute chain, such as GPUs, CPUs, and optical modules. For the memory chip segment, a more nuanced approach is warranted, carefully evaluating technology roadmaps and customer concentration to navigate potential demand volatility arising from product specification changes and platform vendors' cost-control measures.