The Case for Essential AI Infrastructure

Executive Summary 

Artificial intelligence is catalyzing one of the largest capital expenditure cycles in modern history. In 2024 alone, global data center spending reached an estimated $455 billion, up more than 50% year over year, and forecasts suggest that annual outlays could surpass $1 trillion by 2029. The bulk of this spending is flowing into AI factories—data centers packed with GPUs and accelerators. While these factories are essential to the current AI race, they are also capital-intensive, short-lived, and heavily exposed to hyperscaler return on investment. 

By contrast, the essential layer of AI infrastructure—factory infrastructure, power distribution, cooling, semiconductor manufacturing infrastructure, and supporting energy systems—remains underfunded relative to its strategic importance. These assets are long-lived, policy-supported, and capable of generating durable cash flows across multiple industries. 

This white paper argues that essential infrastructure is the more attractive risk-adjusted investment: it captures upside from AI growth while retaining value even if hyperscaler ROI proves elusive.  

1. Framing the Investment Lens: Price, Value, and Durability

Howard Marks reminds us that “price is what you pay, value is what you get”. Value is derived from an asset’s earning power over time, not from investor sentiment in the short run. Applying this framework to AI requires distinguishing between: 

• Speculative, short-lived assets: GPUs and AI factories with 2–3 year refresh cycles and uncertain future utilization. 

• Durable, foundational assets: power systems, cooling equipment, transmission networks, and semiconductor factory infrastructure with decades-long utility. 

We define essential AI infrastructure as encompassing: 

• Factory infrastructure: energy systems, cooling, thermal management, interconnect networks, site construction. 

• Equipment: industrial cooling plants, liquid immersion systems, backup power, high-voltage distribution. 

• Semiconductor factory infrastructure: fabs, cleanroom systems, process tool installations, specialty manufacturing facilities. 

• Grid and enabling systems: substations, transmission upgrades, renewable energy integration, storage. 

These categories constitute the enduring backbone of AI, cloud computing, and digital industry.  

2. The Capital Surge into AI Factories

• Historic scale: Global data center CapEx reached ~$455B in 2024 (Dell’Oro Group) and is projected to grow ~21% annually, topping $1T by 2029. 

• Hyperscaler dominance: Meta, Microsoft, Alphabet, Amazon, and others are responsible for the majority of spending, with CapEx ratios approaching 50% of operating income. 

• GDP impact: In the U.S., as much as half of GDP growth in 2024–25 has been linked to AI-related data center construction (Plain English podcast). 

• Composition: Roughly 60% of new facility costs are GPUs and accelerators, while power and cooling account for much of the remainder. 

• Financing opacity: Hyperscalers increasingly rely on special purpose vehicles (SPVs) and partnerships with private credit providers to shift costs off balance sheets, echoing pre-2008 financing practices. 

This surge rivals historic infrastructure cycles—railroads, telegraph, and fiber optic buildouts—but differs in one crucial respect: unlike steel rails or fiber, GPUs depreciate rapidly and must be replaced every few years.  

3. Why AI Factories May Not Provide Durable Value

AI factories are essential in the short run but vulnerable as long-term investments: 

1. Depreciation treadmill: GPUs have a useful life of 2–3 years, forcing constant reinvestment. 
2. Overcapacity risk: As more hyperscalers and private investors build factories, utilization rates may fall, undermining returns. 
3. Energy and regulatory constraints: By 2030, global data centers may require ~200 GW of power, straining grids and raising costs. Local backlash against water use and land conversion is growing. 
4. Macroeconomic exposure: Factories are highly sensitive to interest rates, supply chain shocks, and tech cycles. 
5. Capital crowding-out: The concentration of investment into AI factories has starved other industrial sectors of capital, repeating dynamics observed in the telecom buildout of the 1990s. 

In short, factories are necessary but fragile. Their earnings power is tied to the uncertain ROI of hyperscalers.  

4. The Case for Essential AI Infrastructure

Unlike AI factories, essential infrastructure offers durability and diversification: 

• Long-lived assets: Substations, cooling plants, semiconductor equipment, and transmission lines operate for decades, not years. 

• Multi-sector demand: These assets serve AI, cloud, semiconductor manufacturing, telecom, and industrial IoT alike. 

• Policy alignment: Infrastructure is directly supported by the U.S. CHIPS Act, Inflation Reduction Act, and global energy transition programs. 

• Scarcity value: Power, water, and cooling capacity are increasingly constrained, giving infrastructure owners pricing power. 

• Valuation advantage: While markets chase Nvidia and hyperscaler stocks, infrastructure assets remain underrecognized, creating opportunities for disciplined investors. 

Examples: 

• Cooling providers like Carrier have seen unprecedented demand as liquid cooling becomes mandatory for high-density AI workloads. 

• Semiconductor equipment and fab infrastructure providers benefit from the CHIPS Act and multi-year national security imperatives. 

• Utilities and transmission developers gain from power demand growth regardless of which AI platforms dominate.  
  
  

5. Risk-Adjusted Return Comparison

A stylized comparison illustrates the asymmetry: 

Essential infrastructure provides the qualities Marks associates with true value: steady earning power, resilience across cycles, and attractive entry valuations relative to hype-driven assets. 

 6. Investor Implications

1. 1. If AI succeeds, hyperscaler CapEx accelerates, creating more demand for power, cooling, and manufacturing equipment. 
   2. If AI underperforms, factories may be stranded, but infrastructure retains cross-sector utility. 

1. 1. Essential infrastructure should be treated as a core allocation—a defensive-growth position that benefits across scenarios. 

1. 1. Public markets: utilities, semiconductor equipment firms, infrastructure REITs. 
   2. Private markets: direct investments in cooling, transmission, and fab infrastructure projects. 

1. 1. Grid regulation, tariffs, and semiconductor incentives are major drivers of investment outcomes. 

 7. Conclusion

The story of AI investment is not just about chips and factories. It is about the underlying systems that make them possible. While hyperscalers race to outspend each other, investors rooted in evidence-based principles should ask: where does durable value accrue? 

The answer is clear. Essential AI infrastructure—the power, cooling, manufacturing, and grid systems that undergird the AI economy—is the foundational layer. It generates durable, multi-sector cash flows, benefits from policy alignment, and offers attractive valuations relative to speculative assets. 

In an era defined by transformative technology, essential infrastructure represents the kind of investment that endures. It is the backbone of AI—and the more resilient place to allocate capital for long-term value.