HSBC research found that 44% of tasks in 'agent AI' workflows rely on CPUs, which is 3 to 4 times higher than in traditional AI computing. This makes CPUs and memory the biggest bottlenecks limiting data center computing power. The bank significantly raised its global server shipment growth forecast for 2026 to 20% year-on-year and believes that, constrained by supply chain capacity, real market demand could be as high as 60%. This growth momentum is expected to continue until 2028.
Over the past three years, the narrative logic of the AI server market has been highly singular: whoever possesses more GPUs holds the key to achieving Artificial General Intelligence (AGI).
However, as AI applications evolve from simple chatbots to 'Agentic AI,' capable of automating complex processes, the underlying computing power requirements of data centers are undergoing a fundamental 'seismic shift.'
According to Zhui Feng Trading Platform, HSBC Securities pointed out in its latest in-depth industry report that an era of 'computing power rebalancing' has begun, with CPUs regaining their role as the decision-making hub in AI data centers.
From 'Chatting' to 'Working': The Computing Power Earthquake Triggered by Agentic AI
Early AI models primarily relied on large-scale GPUs for parallel computing, often referred to as 'brute-force aesthetics.' In contrast, the workflow of Agentic AI is entirely different, acting as a 'super orchestrator.'
The HSBC report noted that Agentic AI not only needs to generate content but also completes tasks through Retrieval-Augmented Generation (RAG) and accessing external APIs (such as email systems, web searches, and collaborative office software). These tasks involve numerous logical branches, task synchronization, and multithreaded processing.
Comparative data shows that in Agentic AI tasks, the parallel advantages of GPUs are diminished, while the flexibility of CPUs in multitasking makes them shine.
According to the latest research from Cornell University, approximately 44% of the computing power in Agentic AI workflows relies on CPUs, which is 3-4 times higher than the CPU share in traditional AI workflows. CPUs are no longer merely 'assistants' to the system but have become an indispensable 'brain' in the computing engine.
Why do server shipment forecasts continue to soar? Supply falls far short of demand.
This structural change in computing power demand directly translates into robust support for server shipments.
HSBC Securities has significantly raised its baseline forecast for global server shipments in 2026 to a 20% year-on-year increase (previously 4%), far exceeding the market consensus of 10%-15%. Even more striking, HSBC believes that the "underlying demand" in the market could be as high as 60%.
Why hasn't the market fully realized this demand?
The main reason lies in bottlenecks within the supply chain. The report emphasizes that current shortages in CPU and DRAM memory supplies have reached 30%-40%, while key components such as SSDs and power chips also face supply shortfalls of 10%-30%.
This implies that global server shipment growth over the next two years will occur through a constrained, incremental release akin to "squeezing toothpaste." This persistent supply shortage has led to significant backlogs, not only securing the shipment logic for 2026-2027 but also reinforcing HSBC’s confidence that the global server growth cycle could extend to 2028.
Capital Expenditure Surge: 'Extra Fuel' from the Big Beautiful Act
Driving this wave of hardware expansion is not only technological upgrades but also policy incentives.
According to HSBC's estimates, with the implementation of the U.S. Big Beautiful Act (One Big Beautiful Bill Act), companies purchasing data center assets (including servers) can enjoy 100% bonus depreciation and immediate R&D expense deductions. This provides cloud service providers (CSPs) with strong financial incentives for hardware investments.
HSBC predicts that the combined capital expenditure of major cloud service providers (including Google, Microsoft, Amazon, Meta, and Oracle) will reach $743 billion in 2026, representing an 81% year-on-year increase.
The allocation of this capital injection has undergone a structural shift: in 2025, approximately 35% of CSP capital expenditures were directed toward servers; by 2027, this proportion is expected to rise to 59%. This reallocation of resources—from acquiring land and constructing buildings to purchasing computing power equipment—precisely targets benefits for the server supply chain.
Who Stands to Benefit? The 'Rebalancing' Dividend of the Supply Chain
The market style is shifting from the pure "GPU logic" to a more balanced "computing node value chain."
HSBC noted that as the value share of CPUs in racks increases, server ODMs (Original Design Manufacturers) and key component suppliers will become the core beneficiaries of this computing power rebalancing.
In summary, this arms race for computing power has entered the "deep water zone." In the past, it was about "whoever had the graphics card would win," but in the future, it will be about "whoever can manage computing power allocation well will gain efficiency."
The recovery in CPU demand driven by AI not only injects long-term certainty into the server market but also outlines a clear hardware upgrade logic line for investors.
Editor/Lambor
