The AI Gold Rush Just Got a Lot More Physical

AI has a funny way of moving the goalposts. First it was all about the powerful chips that could train and run massive models. Then attention shifted to the memory needed to keep those chips from sitting idle. Now the story is changing again — and this time the real constraints look a lot more like power lines, construction sites, and cooling systems than flashy semiconductors.

The AI Story So Far: From Chips to the Next Layer

The first phase centered on processing power. Specialized accelerators designed for large-scale parallel computation enabled the training and operation of increasingly capable models. This layer delivered rapid performance gains and concentrated significant value in a narrow slice of the supply chain. It was a necessary starting point.

The second phase emerged naturally as clusters scaled. Moving data at the speeds and volumes required by these systems exposed limits in memory bandwidth and capacity. High-bandwidth memory and advanced storage solutions became essential to prevent expensive processors from sitting idle. Demand in the highest-performance segments tightened supply, and producers with relevant technology saw extended order visibility. This shift reflected the physical realities of large-scale systems rather than a change in direction.

Power Just Became the Real Constraint

The current phase is moving further into the supporting infrastructure. Capital expenditure figures and project timelines show that the binding constraints now sit in areas that receive less daily attention but directly determine whether announced plans can be delivered on schedule. Major operators continue to commit very large sums to build capacity. Research from Goldman Sachs and others has modeled cumulative AI-related infrastructure investment in the trillions over the coming years, with annual spending already running in the hundreds of billions and projected to increase further. Much of this spend is no longer driven primarily by the processors or the memory attached to them.

Power has become the clearest and most structural limit. AI workloads require sustained high power density at unprecedented scale. Projections for data center electricity demand have shifted from steady incremental growth to something closer to a step change in several regions. In the United States, some analyses show AI-related demand potentially rising more than thirtyfold by the mid-2030s under aggressive scenarios. Grid interconnection queues in key markets already extend multiple years, and equipment such as large power transformers faces extended lead times. Industry tracking indicates that a meaningful portion of planned 2026 capacity has already moved into later years due to these constraints. These are not marginal issues; they directly affect timelines, costs, and the ability to bring capacity online.

This reality creates sustained demand in power generation, delivery, and behind-the-meter solutions. Projects able to secure reliable supply through grid access, new generation, or co-location arrangements hold a practical advantage. The relevant businesses range from utilities with exposure to large loads to developers of flexible generation and the equipment providers that support high-density delivery. Many of these segments reward operational reliability, long-term contracts, and scale more than pure technological differentiation.

In this area, companies such as Vistra Corp (VST) and Constellation Energy (CEG) have built meaningful exposure to data center power demand through a combination of existing generation assets and strategic restarts or expansions. Both bring scale and the ability to enter long-term supply arrangements that align with the multi-year nature of data center commitments.

Moving All That Data Is Getting Expensive

Networking and data movement represent another growing constraint. Large clusters require extremely high bandwidth and low latency both within facilities and across them. As workloads shift toward distributed inference and continuous operation, the economics of moving data become more important relative to raw compute. Investment is flowing into higher-speed switches, optical components, and the overall networking fabric required to connect thousands of accelerators efficiently. Companies with established positions in these areas are seeing demand tied directly to cluster expansion.

In this layer, Arista Networks (ANET) and Broadcom (AVGO) stand out as high-quality businesses with strong competitive positions in high-speed data center networking. Both have demonstrated the ability to scale with hyperscaler demand while maintaining solid operational discipline and cash flow characteristics.

AI Runs Hot. Cooling It Is Now Serious Business.

Cooling follows as power densities increase. Traditional air-based systems reach practical limits quickly in high-performance environments. Liquid cooling solutions are moving from specialized deployments to broader adoption in new facilities. The companies that can deliver reliable thermal management at the required densities, along with associated services, are positioned to benefit from the same capacity expansion.

Here, Vertiv Holdings (VRT) has emerged as a leading player with strong exposure to data center thermal management and power infrastructure. The company combines technology leadership with the ability to support large-scale deployments and recurring service opportunities.

Someone Has to Actually Build and Run These Things

The physical construction and ongoing operation of the facilities add another layer. Building large, high-density data centers on accelerated timelines requires coordination across multiple disciplines. Once operational, these sites need continuous maintenance, monitoring, and upgrades to maintain high utilization. This creates demand for providers that can handle complex environments across electrical, mechanical, and other trades. Businesses with geographic reach, established customer relationships, and the ability to deliver integrated or self-performed work tend to capture recurring elements of this spend.

In this segment, established players such as EMCOR Group (EME) and Comfort Systems USA (FIX) stand out for their ability to provide integrated electrical and mechanical services with a meaningful recurring maintenance component. Both operate at scale across multiple regions and have demonstrated the operational discipline required to support large, complex facilities over time. In Europe, Spie SA (SPIE.PA) offers a comparable multi-technical services platform with exposure to industrial and infrastructure maintenance that aligns with data center needs.

Not Every AI Infrastructure Business Is a Good Business

From a business quality perspective, the more durable positions in this environment tend to share certain characteristics. Scale offers advantages in procurement, talent, and the ability to support larger programs. Long-term customer relationships and contractual structures can provide visibility and some insulation from short-term fluctuations in spending. Strong cash generation and balance sheet flexibility matter because supporting infrastructure at this scale requires sustained investment through uneven project cycles. Businesses that combine these traits with exposure to the actual bottlenecks — power delivery, connectivity, thermal management, or facility operations — are structurally better placed than those tied purely to the fastest-moving parts of the technology stack.

What a Disciplined Investor Should Actually Focus On

A smart, long-term investor approaches this phase with the same discipline applied to any other opportunity. The goal is not to own “AI exposure” at any price, but to own high-quality businesses that can compound value through the full cycle of this buildout.

Focus first on competitive advantages that are likely to endure. Scale, deep customer relationships, recurring revenue streams (such as maintenance and service contracts), and the ability to self-perform or integrate complex work create stickiness that pure technology plays often lack. These characteristics tend to produce more predictable cash flows and better resilience when spending patterns shift.

Second, demand evidence of strong free cash flow generation and balance sheet strength. Infrastructure at this scale requires ongoing investment. Companies that can fund growth internally or maintain flexibility through cycles are in a stronger position than those reliant on continuous external capital.

Third, insist on a margin of safety in valuation. Many parts of the AI-related universe already trade at levels that assume near-perfect execution and continued rapid expansion. A disciplined approach means being willing to wait for better entry points rather than chasing momentum. The businesses worth owning are often those solving real operational problems for the operators actually spending the money, not those riding the loudest narrative.

Finally, stay selective and long-term oriented. This buildout will not unfold in a straight line. There will be periods of faster and slower spending, technological shifts, and execution surprises. Investors who focus on durable business models and reasonable prices, rather than trying to time every phase of the story, are more likely to capture the compounding that comes from owning quality compounders through the full cycle.

The Risks Nobody Really Wants to Talk About

Risks deserve direct acknowledgment. Major capital expenditure programs can adjust based on realized returns, improvements in model efficiency, or shifts in strategic priorities. Technological advances in algorithms or chip design can ease pressure on memory or power over time. Execution risk on large infrastructure projects remains material; permitting timelines, grid upgrades, and supply chain realities move more slowly than semiconductor development cycles. Customer concentration among a small number of large operators means any change in their plans transmits quickly. Valuations across much of the related space already reflect substantial optimism about continued rapid expansion. Areas that benefited from earlier phases now trade at levels that leave limited margin for slower progress or unexpected delays.

Bringing It All Together: What Actually Matters Now

The broader reality is that the AI infrastructure buildout is substantial and multi-year, yet it is becoming more industrial and more dependent on physical systems than the early chapters suggested. The initial focus on processing power and then on high-performance memory addressed necessary layers. The current phase centers on whether the supporting systems — power, connectivity, cooling, and operational capacity — can scale at the required pace and cost. This does not diminish earlier parts of the story; it adds the constraints that will determine how far and how fast overall capacity can grow.

Companies able to deliver reliable solutions in these foundational areas, particularly those with recurring service components or long-term contractual visibility, are likely to exhibit more durable demand characteristics than those exposed purely to the highest-velocity segments of the technology curve. Entry price and underlying business quality become more important as the narrative matures and additional capital flows into the space.

A more detailed version with the underlying models, company-level analysis, and full framework is available here: https://investingsprints.com/blogs/researches/

Resources

• Goldman Sachs research on the assumptions shaping the scale of the AI build-out
• Deloitte analysis on U.S. data center power demand and infrastructure needs
• S&P Global / 451 Research 2026 trends in data center services and infrastructure
• Dell’Oro Group forecasts on data center infrastructure spending and capacity in 2026
• Industry tracking on hyperscaler capital expenditure levels and grid interconnection data from regional transmission organizations (various 2025–2026 reports)

The infrastructure layer is where execution will be tested most directly. That is where the next phase of this story will be decided. A disciplined investor’s job is not to predict every twist in the narrative, but to identify the businesses that can deliver real value through the full buildout — and to own them only when the price offers a genuine margin of safety. That approach has worked across many cycles before this one, and it remains the clearest path forward here.

Consolidated Valuation & Financial Metrics Comparison

 

Looking at the gap between Current Price and DCF Buy Price (50% MOS), the clearest margin of safety right now appears in: ANET, AVGO, FIX.

Disclaimer: This content is provided for educational and entertainment purposes only. It is not investment advice, financial advice, or a recommendation to buy, sell, or hold any security. I am not a fiduciary and I am not a licensed or professional financial advisor. Always do your own research and consult with a qualified financial professional before making any investment decisions. Past performance is not indicative of future results, and all investments carry risk of loss.