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

Managers and allocators describe the current AI compute buildout as an investment "super cycle" where demand far outstrips the available supply of semiconductors, data centers, and electricity [1, 20]. The scale of this cycle is projected to be an order of magnitude larger than the cloud era, with some analysts estimating **$10 trillion in total capital expenditure** [9, 11]. This spending has propelled the semiconductor industry to a $1 trillion annual revenue run rate years ahead of schedule and is seen as the primary driver of current market and economic growth [9, 25]. While some compare the boom's magnitude to the dot-com era, they note it is more fundamentally sound due to the staggering profitability of the leading infrastructure companies . However, others caution that the capex is more widespread than in the dot-com period, suggesting a future correction could have a broader macroeconomic impact . The primary near-term risk cited is a potential slowdown in enterprise adoption if companies fail to realize a clear and timely return on their massive investments .

The primary limiting factor on AI's growth is a series of cascading supply chain bottlenecks that are expected to persist for years [2, 13, 15]. While initial constraints centered on GPUs, where NVIDIA maintains over **98% market share** for non-Google AI workloads , the shortages have broadened significantly. Severe constraints are now reported in high-bandwidth memory (HBM), advanced packaging (CoWOS), and other critical components, with some supply commitments extending years into the future [2, 14, 18, 19]. Some industry executives and analysts believe this supply-demand mismatch will last for at least a decade [3, 7]. More recently, the primary bottleneck is shifting from chip manufacturing to physical infrastructure, specifically **electrical power and data center availability**, which now constrains even the largest hyperscalers [14, 17, 21]. This has led to a decommoditization of the hardware industry, creating significant pricing power for key suppliers .

Demand drivers are evolving from cloud-based experimentation to on-premise enterprise production, motivated by data security and sovereignty needs [3, 7, 19]. A more profound shift is underway from generative AI for content creation to "Agentic AI" for autonomous task execution [3, 19]. This paradigm of persistent, agent-based computing is expected to dramatically increase the computational footprint of every user and is driving a new architectural debate . Some allocators predict this will cause a fundamental flip in server design from being GPU-heavy to **CPU-heavy**, potentially shifting the balance of power among semiconductor firms [14, 29]. This new architecture, combined with more advanced "reasoning" models, is also expected to drive a fivefold increase in the amount of memory required per user [22, 26, 28].

In response to these dynamics, investor positioning is shifting. Value is currently accruing to the infrastructure and semiconductor layers, a reversal from the cloud era where software captured the most value . This has driven the semiconductor sector's weight in the S&P 500 from 3% to **17%** over the last decade [5, 10]. Capital allocation frameworks are moving beyond a singular focus on NVIDIA to a broader "sellers of shortage vs. buyers of shortage" thesis . This view favors component suppliers like memory and power companies, which are seen as having expanding profitability, over hyperscalers who face margin compression from their massive capex . The investment mantra is evolving from "follow the GPU" to "follow the gigawatts," identifying electrical power as the next key area for investment . While momentum remains strong, the rally's narrowness is a noted concern .