What are top managers and allocators saying about semiconductors and 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 available supply [1, 3, 9]. The scale of this cycle is projected to be an order of magnitude larger than the cloud era, with some analysts forecasting a **$10 trillion total spend** [2, 12]. This explosive demand has accelerated the semiconductor industry's growth, putting it on a run rate to hit $1 trillion in annual revenue in 2024, four years ahead of previous forecasts [2, 12]. The boom's breadth is also notable, with the semiconductor sector expanding from 3% to 17% of the S&P 500 in a decade, lifting a wide range of companies beyond market leader Nvidia [10, 15]. While some analysts see bubble-like characteristics, others argue that valuations are reasonable given the massive growth projections and that investment is concentrated in highly profitable tech giants [2, 10]. The primary near-term risk cited is a potential slowdown in enterprise AI adoption if companies fail to realize a clear return on their massive investments .

The primary constraint on AI's growth is a series of cascading supply chain bottlenecks, with the most critical chokepoint now **shifting from chip supply to physical infrastructure** . Initially, the main impediments were advanced node semiconductors and high-bandwidth memory (HBM), a shortage expected to persist for years, if not a decade, due to the long, multi-year investment cycles for fabrication plants [3, 5, 6, 8, 11, 13]. However, as chip manufacturing scales, the bottleneck is increasingly becoming the availability of electrical power and data centers [17, 22, 23, 27]. This has led some investors to shift their thesis from "follow the GPU" to "follow the gigawatts," identifying power as the new primary driver of the buildout . This fundamental scarcity is "decommoditizing" the hardware industry, creating significant pricing power for key suppliers of constrained components like memory, power, and advanced packaging [17, 26].

Positioning is shifting in response to an evolution in AI workloads and computing architecture. A significant trend is the migration from cloud-based testing to on-premise enterprise production, driven by corporate needs for data security and sovereignty [6, 8, 20]. Concurrently, the industry is moving beyond generative AI for content creation toward "Agentic AI" for task execution, which is expected to fuel a new, massive demand cycle for both GPUs and CPUs [6, 8, 20]. This shift could cause a **flip in server architecture** from being GPU-heavy (e.g., 1:4 CPU:GPU) to CPU-heavy (e.g., 4:1) to support persistent, agent-based computing . Furthermore, as the market matures, a transition from compute-intensive training to more widespread inference workloads is anticipated, which may open the door for more specialized and efficient chips, potentially increasing competition for Nvidia's current market dominance .

In this environment, value is accruing primarily to the infrastructure and semiconductor layers, a reversal from the cloud era where software and application companies captured the most value . While Nvidia maintains overwhelming dominance, running over 98% of non-Google AI workloads, investment is broadening to include memory suppliers, other semiconductor firms, and companies in emerging markets [4, 15, 21]. Memory, in particular, is seen as a core component of AI coverage, with some analysts predicting the amount of memory required per user will **increase fivefold** [14, 25, 29]. The strategic importance of this buildout has also elevated semiconductors to the level of geopolitical assets, prompting governments to implement industrial policies to secure supply chains and manage dependencies, particularly concerning Taiwan's central role in manufacturing and US-China tech relations [2, 8, 30].