What are top managers and allocators saying on recent podcasts and at conferences about semiconductors and the AI compute build-out — and how is positioning shifting?

Managers and allocators describe the current AI compute build-out as a massive, multi-year "super cycle" fueled by unprecedented capital expenditure from hyperscalers and AI labs, estimated to be nearly a trillion dollars [2, 4, 6]. This investment is driving a historic expansion of the semiconductor sector, which has grown from 3% to **17% of the S&P 500** in a decade [11, 14]. Unlike the dot-com era, this boom is viewed as more fundamentally sound due to the staggering profitability of leading companies . The consensus is that the intense supply and demand mismatch for AI hardware will persist for several years, with some Korean semiconductor firms signaling that constraints will extend beyond 2027 [3, 7]. This sustained demand is "decommoditizing" the hardware industry, creating significant pricing power for key suppliers across the infrastructure stack, from energy and data centers to the chips themselves [1, 9].

The primary investment thesis is shifting as the core bottlenecks in the AI build-out evolve. Initially focused on semiconductor availability, the constraint is now migrating to physical infrastructure, specifically **electrical power and data center capacity** [18, 25, 30]. This has led to a strategic repositioning from "follow the GPU" to "follow the gigawatts," with a company's power procurement strategy now seen as a leading indicator of its ability to scale [18, 22, 23]. While TSMC's production capacity remains a fundamental governor of the overall pace , future chokepoints are anticipated further down the supply chain, including a potential cap on the entire build-out from ASML's production of EUV lithography tools . This dynamic landscape is forcing investors to look beyond the primary chip designers to the foundational enablers of compute.

This evolution of bottlenecks has created a new framework for capital allocation, bifurcating the market between "buyers of shortage" and "sellers of shortage" . The buyers—hyperscalers like Microsoft and Amazon—face potential margin compression from their massive CapEx outlays . In contrast, the sellers—suppliers of scarce resources like power and memory—are experiencing expanding profitability and cash flow [18, 26]. Memory is a particularly acute example of this trend, with demand far exceeding forecasts and supply tightness expected to last well beyond 2026 [24, 29]. The market is experiencing unprecedented strain, with the cost of high-bandwidth memory (HBM) more than doubling this year and supply commitments extending out to **2029 and 2030** [21, 28]. This has made memory a core component of AI-focused portfolios .

As the market matures, opportunities are broadening beyond GPU manufacturers to other parts of the semiconductor ecosystem, driven by shifts in compute architecture. The rise of persistent, agent-based AI is creating a new demand profile that heavily utilizes CPUs for serial processing tasks, benefiting companies like Intel and AMD [19, 27]. This could lead to a fundamental change in server design, potentially flipping the standard architecture from a 1:4 CPU-to-GPU ratio to a **4:1 CPU-to-GPU ratio** . Concurrently, some see a technological migration toward specialized Neural Processing Units (NPUs) for AI inference workloads [13, 17]. While this broadening is creating new winners and lifting the entire sector [12, 14], some managers urge caution, citing inflated valuations and the need for disciplined, fundamentals-based investing to avoid being swept up in the market hype .