What are top managers and allocators saying about semiconductors, AI compute buildout, and how is positioning shifting in 2026?

Managers and allocators describe the AI compute buildout as a supercycle an order of magnitude larger than the cloud era, with a projected **$10 trillion** in total capital expenditure [2, 14]. This has accelerated the semiconductor industry's growth, putting it on a run rate to hit $1 trillion in annual revenue by 2026, four years ahead of previous forecasts [2, 14]. The consensus view is that demand for AI hardware vastly outstrips supply, creating an era of scarcity that is expected to persist for several years, and potentially a decade [1, 3, 5, 6, 17, 22]. While initial bottlenecks were concentrated in advanced semiconductors, high-bandwidth memory (HBM), and advanced packaging [3, 7, 10], the primary constraint is now shifting to physical infrastructure, specifically electrical power and data center availability [9, 19, 27]. This has led some investors to shift their focus from "follow the GPU" to "follow the gigawatts," identifying power as the new atomic unit of AI growth [19, 29]. The entire buildout is seen as inflationary due to rising costs for components, energy, and labor .

The nature of AI demand is evolving, driving architectural shifts in compute infrastructure. A significant trend is the migration from cloud-based testing to on-premise enterprise production, driven by corporate needs for data security and control [1, 6, 7, 24]. Concurrently, the industry is transitioning from generative AI for content creation to "Agentic AI," where autonomous systems actively perform complex tasks [1, 7, 12]. This paradigm shift is expected to dramatically increase the computational footprint of every user and is fueling a massive new demand for CPUs to handle serial processing tasks, in addition to GPUs [6, 12]. Some analysts predict this could fundamentally alter data center design, potentially flipping server ratios from being GPU-heavy (e.g., 1:4 CPU:GPU) to **CPU-dominant (e.g., 4:1)**, which would create new winners and losers in the semiconductor market [12, 19]. This evolution is also projected to quintuple the amount of memory required per user [20, 25].

In response to these dynamics, investor positioning has become heavily concentrated in the semiconductor sector. According to Goldman Sachs' prime brokerage data, nearly **20% of their clients** are weighted towards semiconductors . Unlike the cloud era where value accrued to software firms, value in the current AI cycle is being captured by the infrastructure and semiconductor layers . NVIDIA maintains a dominant market position, running over 98% of non-Google AI workloads, and is defending its moat with an accelerated annual product roadmap . While valuations are high, they are often considered reasonable given the massive growth projections . The primary near-term risk cited is a potential slowdown in enterprise adoption if companies fail to realize a clear and timely return on investment . The long-term vision extends beyond the current digital buildout to future, even larger markets for "Personal AI" running on local devices and "Physical AI" integrating with robotics and the industrial economy [6, 7, 24].