Produced in partnership with Blue Owl.

In the first installment in this series, I focused on the case for maintaining exposure to US assets while looking for opportunities to diversify. In this installment, I discuss one such opportunity: investing alongside technology companies in digital infrastructure.

AI is not new. Venture capitalists and technology firms have been working towards today’s AI models for decades. From 2015-2022, AI development was all about pattern recognition applied to a specific task. You probably experienced early versions of AI when you opened your phone one day and discovered that it could identify all the photos of the same person and group them together.

The real race for scale began in 2022, when AI made a major advance from “recognition” to “foundational” models. Foundational AI models are large, general-purpose models trained on vast and diverse datasets. Unlike the previous task-specific models, foundational models are “generative” and can be adapted to produce a wide range of outputs: coherent, contextually relevant text, realistic images and video, music, speech and code.

Earlier AI models used millions of parameters (the variables that drive the model) to perform a specific task. Today’s general-purpose AI models employ billions of parameters, leading to a massive increase in the need for data storage, processing and computing power. For example, in 2022, OpenAI’s GPT-3.5 model employed about 2.6 billion petaflops of training computation. In 2025, GPT-4.5 training used 210 billion petaflops.^[1] One petaflop of computation is equivalent to about 10,000 high-end laptops, working flat out, all at the same time.

The Infrastructure Surge

In the generative AI and hyperscaler ecosystems, the modern AI technology stack has three main layers:

• Model layer: This is where the foundational models live. These models serve as core intelligence engines. Companies like OpenAI, Google and Meta sell access to these models or license them for others to build on.
  
  
• Application layer: This includes all the tools, software, and interfaces that end users directly interact with. Companies in the application layer use AI models to deliver products and services: writing assistants, chatbots, image generators, coding copilots or enterprise automation tools. Examples include Salesforce, Adobe and SAP.
  
  
• Infrastructure layer: The infrastructure layer provides the computing power, networking, storage and tools needed to train and deploy AI models. This layer includes cloud providers (e.g. Amazon, Microsoft, Google, Meta), hardware manufacturers (e.g., NVIDIA, Intel) and a host of other specialised companies.

The infrastructure layer is the backbone of AI and cloud. All of the data and computing power resulting from AI models and applications needs somewhere to live, leading to a surge in investment in data centers: specialised facilities dedicated to storing and processing data.

Hyperscalers do not have the infrastructure to support this growing demand, but they are going “all in” to create it. In 2025 alone, hyperscalers like Google, Meta, Microsoft and Amazon are expected to spend over $350 billion on data center construction. [2]  Over the next five years, hyperscalers are projecting a 3.5x increase in data center capacity from around 85GW to over 200GW by 2030.[3]

Why so much? Technology companies are positioning for a future where AI/cloud infrastructure and services dominate IT spending, and they need scale to capture this opportunity. The more scale these companies can achieve (in hardware, data and global reach), the better the economics and the competitive edge.

The Investment Opportunity

The scale of the data center surge creates opportunities for investors to participate, alongside world-leading technology companies that are making transformational advances in their core business. Data center opportunities have attracted strong investor interest, along with rising scrutiny of asset valuations, capital expenditures and future revenue potential.

Is AI a bubble? It’s certainly possible that exuberance about AI’s potential has inflated asset valuations, but timing a correction is notoriously difficult. Former Fed Chair Alan Greenspan originally coined the term “irrational exuberance” in a speech on December 5, 1996, where he called a technology bubble. Technology stocks kept growing significantly until March 2000, and had you sold in December 1996, you would have had more wealth destruction than if you continued investing and lived through the 2000 drawdown. It’s also worth noting that a number of companies that survived the dot.com bubble are still thriving today because they had the capital and the expertise to capture the long-term internet opportunity.

Against this backdrop, opportunities to partner directly with hyperscalers to develop build-to-suit data centers can be an attractive option for investors seeking exposure to the AI theme.

Data centers are complex facilities requiring land, large amounts of power and structures to house servers and other technology. Power is the biggest challenge, with large data centers requiring power measured in gigawatts rather than megawatts. One gigawatt is the average amount of power that San Francisco is drawing at any given moment.

Platforms that can procure power and develop land into a built-to-suit “powered shell” allow the hyperscalers to focus on their core technology expertise, providing significant value to these companies. Investors receive the stable cash flow benefits of traditional infrastructure assets, the potential for tech-like growth, diversification and inflation protection.

• Stable, long-term cash flows: Much like traditional infrastructure assets (toll roads, utilities, etc.) the cash flows derived from data center assets are tied to long term contracts (15-20 years) and are the contractual obligation of companies with strong credit quality. The tenant companies are responsible for the technology inside the data center. Investors receive cash flows from long-life assets (land, structures), not servers or other technology that can degrade over time. In a large data center, we believe hyperscalers are putting in three to five dollars for every dollar of investor capital.

• Long-term capital appreciation potential: Value is created by converting land to stable, operating data centers with long-term leases, tenanted by some of the largest, most credit-worthy tech companies in the world. The long-term value of stable, de-risked digital infrastructure is likely to benefit from the rising tide of data and computing, providing upside exposure to the data center theme without picking winners and losers.
  
  
• Low correlation to traditional assets: Infrastructure assets, including data centers, tend to behave differently from stocks and bonds because the value of long-term contracts is less exposed to shorter-term economic and market cycles. The low correlation of infrastructure to traditional assets provides diversification that can reduce overall portfolio volatility.

• Inflation protection: Lease structures are often “triple-net”, meaning the tenant is responsible for the three main maintenance costs – property taxes, maintenance and building insurance – in addition to the base rent. Rent escalators negotiated at the time the lease is signed provide visibility into net operating income growth and help preserve real returns.

Finally, tech companies need partners with scale, who understand their challenges and have the resources and experience to help them solve these challenges. These businesses need to plan for additional capacity and are seeking partners that can deliver consistently across different geographies. Building a platform to meet these needs requires substantial scale, including teams focused on site selection and power procurement, and expertise in developing these projects in different geographies from construction to community relations. For investors, the benefit of this scale comes from opportunities to partner with world-leading innovators on their most mission-critical projects.

[1] Source: https://ourworldindata.org/grapher/artificial-intelligence-training-computation?country=GPT-1~GPT-3.5~GPT-4

[2] Source: Microsoft, Amazon, Alphabet, and Meta public filings (various dates); RBC: (AI & Hyperscale Recap, October 2025); McKinsey: The Cost of Compute: A $7 Trillion Race to Scale Data Centers (April 2025). Note that capital expenditure estimates reflect projected investments in power infrastructure, data center construction, and IT equipment, based on McKinsey’s demand model that incorporates AI adoption, supply constraints, and regulatory considerations.

[3] Source: McKinsey: The Cost of Compute: A $7 Trillion Race to Scale Data Centers (April 2025)