Data Centers: Derivative AI Exposure¹

Narratives, for better or worse, tend to drive capital flows—as we saw with prior narrative cycles—autonomous vehicles, cannabis, cryptocurrencies, etc.

Alas, real adoption curves are very hard to predict. To put some numbers behind the 'AI' narrative - VC investment in AI startups eclipsed $290 billion from 2019-2024² with Silicon Valley darlings raising droves of capital to bet on what are highly uncertain business plans.There's also a valuation premium for AI startups vs. non-AI startups.

AI technology is more power hungry than traditional workloads, requiring substantial processing power and energy consumption. For example, CapEx on compute is soaring, with Alphabet spending $12 billion in Q1 of 2024 alone, and Microsoft announcing a path to spend $14 billion+ per quarter in the near future.³ Additionally, significant value has accrued for NVIDIA, the GPU pioneer.

Goldman Sachs projects that the increase to annual U.S. power consumption resulting from AI alone will be about 200 terawatt-hours by 2030, about 4x the annual power consumption of New York City.⁴

US On-Grid Electricity Demand (TWH)

Source: McKinsey data center demand model (2023); S&P Global Power & Renewable Market Outlook (2024)

When a new narrative cycle emerges, there's a gravitational pull to invest directly in a subset of upstart companies seemingly well-positioned to benefit from the 'hype'—or platform shift—as some would call it. Many investors gravitate towards what they view as the purest exposure form-factor with upside asymmetry, and this approach may well be effective for some. However, in our view, attempting to 'pick winners' seems reasonably fallible and exacerbated by the fact that it remains ambiguous where value will accrue across the sector, as both upstarts and incumbents vie for technological supremacy and market share.

During the dot-com bubble, 7,000-10,000 internet-related companies received formal funding, with over 5,000 either acquired or shut down during 2000-2003.⁵ A study conducted by Robert Hendershot analyzed 441 U.S.-based firms categorized as “internet-specific, direct-to-consumer and received their initial venture backing from 1995-2000” who raised over $21 billion combined from private and IPO investors.⁶ Of these 441 companies, just 22 (5%) produced an end-of-2001 liquid value greater than 1.5x the total amount of equity-public and private capital invested in the firm.⁷ While the dot-com bubble is not a perfect parallel, it does provide an analogous situation where extreme consensus-driven hype-narratives have been dangerous.

It is our view, then, that focusing on 'picks-and-shovels' exposures—well-positioned to capture accelerating demand—may offer superior long-term risk/return. One such exposure in which we have identified actionable opportunities of interest, is data centers. This is a space typically dominated by scaled players and publicly traded REITs (e.g., Blackstone, Brookfield, Digital Realty, Equinix, etc.). Each has trade-offs, but investors are generally choosing between (i) stabilized data centers and (ii) data center developments.

Compute-intensive use cases, including adoption of AI training, cloud computing, streaming services, and autonomous vehicles, are driving substantial demand for capacity. In the U.S., data center power consumption is expected to more than double from 17,000 megawatts in 2022 to 35,000 megawatts in 2030, a ~9% CAGR.⁸

Both (i) and (ii) are likely to benefit from a combination of surging demand, large-scale capital inflows, and relative supply scarcity (i.e., exponentially increasing demand and constricted supply).

(i) Data center developments, in particular, appear to exhibit an attractive risk/return profile. Typically, before a data center build-out occurs, capacity is contracted out in some form. Historically, new developments had often assumed 'lease-up' risk. Contractual revenue, further, is backed by some of America's leading companies often providing line-of-sight to cash flow prior to material capital outlays.

(ii) Stabilized data center investment outcomes are likely to be much more sensitive to market multiples. Two key risks of note: (a) obsolescence and (b) capital (CapEx) intensity. While stabilized data centers may be a match for investors with a relatively low return hurdle, and with decade+ underwriting horizons, it's less appealing for firms like ours.

For developments, basis is critical—i.e., margin of safety relative to comparable stabilized multiples. And for stabilized exposures, investors are making a bet that multiple expansion will persist, offset by interim yield. Both exposure types offer different risk/return.

Asset, market, and manager selection will ultimately be key outcome drivers.

Data center business models can be broadly grouped into two categories:⁹

Cloud / Hyperscaler, where a large company with data service needs (typically cloud vendors or telcos)owns and operates data infrastructure for their own purposes.

Co-Location, which involves a third-party company that providesphysical space,cooling equipment, network capacity, and power totenants, who bring their own IT equipment.

Outsourced “co-location” data center demand, in particular, is anticipated to benefit from the need for rapid scale from technology companies. Hyperscaler customers (e.g., Amazon, Google, Meta, Microsoft, and Oracle), who are both tenants and competitors for co-location operators, have used outsourced co-location services to rapidly and efficiently scale across disparate geographies.¹⁰ Smaller enterprise customers also rely on co-location operators for their specialized services. The co-location model is expected to continue to take share from legacy enterprise operators, who will not be able to remain competitive with the operating leverage offered by co-locators, and hyperscalers, who are typically supply constrained.¹¹

The co-location model also benefits from high switching costs; once a tenant has occupied a 3rd party data center, they are very unlikely to move. The cost of equipment required to absorb 10 megawatts of compute space is currently estimated at approximately$1 billion.¹² In addition, the technical expertise, freight capacity, and insurance coverage needed to relocate several billions of dollars of delicate, high-end machines, not to the inherent business disruption caused by equipment downtime of this magnitude, creates significant “stickiness” for existing clients, creating a compelling narrative for the future of co-location—auspiciously positioned to capture outsized value.

Explosive demand for data centers coupled with supply-barriers are creating a market imbalance. We see data center developments as an attractive way to gain derivative AI exposure with a relatively high margin of safety.

History shows that prescience is in short-supply—tread carefully when trying to play the AI boom.