Data Centers: Ain’t No Mountain High Enough

The major hyperscalers are supercharging the data center industry with a staggering amount of capital expenditure. Amazon, Microsoft, Google and Meta are expected to invest an additional $290 billion by 2027, with a large portion allocated to IT and hardware infrastructure. The barriers to entry in data center development remain solid too. The typical costs of the building shell, mechanical, electrical and cooling and fit-out can easily range between $10 and $12 million per megawatt (MW) in traditional prime markets.²

The risk of speculative build is currently low due to the extremely high costs to build hyperscale data centers and the specialized use cases by their tenants. The key consideration for investors is the bifurcation of counterparty risk exposures. Not all technology companies or neocloud providers have the low leverage ratios and positive cash flow of the four main hyperscalers.

Financing options, such as project finance and asset-secured loans, are generally available due to the long-term revenue visibility from contracted revenues. While exit strategies are still developing, the acquisition of Aligned Data Centers in October 2025 for about $40 billion, by a consortium including infrastructure funds, Microsoft and NVIDIA, provides valuable insights into potential exit avenues.

Figure 2: Capital expenditure by technology companies, $ billions

Figure 3: Private infrastructure investment volumes, $ billions

The race to create a frontier large language model continues to stimulate innovation in efficiency and performance. In a fast-evolving landscape, competitive positions shift quickly and new alliances are formed between chipmakers and technology providers. OpenAI and NVIDIA dominated the AI landscape prior to the November 2025 release by Google of its frontier model Gemini 3. Unlike most other frontier models trained on Nvidia’s GPUs, Gemini 3 uses proprietary and currently more cost-effective TPUs³ for AI training.

Current demand for data center capacity is predominantly driven by the compute and power intensive requirements of AI training. However, the subsequent phase will involve the practical application of AI through inference, such as generating AI-powered responses. Ongoing optimization of inference workloads has the potential to reduce per-token costs and accelerate the adoption of AI models. Additionally, smaller language models, some of which already demonstrate impressive performance, can substantially decrease the expense associated with inference.

A broader transition towards AI inference is anticipated to generate robust demand for data center colocation services and cloud on-ramps. The emergence of more “intelligent” models will concurrently broaden the scope of applications for both consumers and enterprises. Conversely, significant risks to AI adoption include regulatory concerns surrounding AI security, privacy and model governance. Policymakers and regulatory bodies are expected to prioritize the protection of private data and the prevention of misinformation and fraudulent activities.

Figure 4: Adoption of AI models, share, %

Figure 5: Power demand for AI training vs. AI inference, gigawatts (GW)

Current AI investment, while unprecedented in its scale, is underpinned by robust fundamentals rather than speculative trends. The investment thesis for data centers is reinforced by significant unmet demand and constrained supply. This is evidenced by record-low vacancy rates in critical markets, high utilization levels and extended project development timelines. In North America, despite a substantial increase in inventory, the primary market vacancy rate has declined to merely 1.6%, with pre-leasing maintaining a strong rate of 74.3% of capacity currently under construction.⁶

Figure 6: Global data center supply vs. demand, GW

The supply bottlenecks are not going to ease soon. The competition for power grid connection is intense in many parts of the world from new power load to manufacturing and electric transport. On-site generation—whether renewables coupled with battery energy storage or other power sources—is still far from being mainstream. Environmental and social factors are rising in focus, causing local public scrutiny and project cancellations due to data centers’ heavy use of power, land and water resources.

In the current data center landscape, investors are advised to focus on several key strategic considerations:

1. Portfolio diversification: Construct diverse data center portfolios, balancing traditional colocation services with specialized AI training and inference capabilities.
2. Rent renewal assessment: Rigorously assess rent renewal assumptions for data centers, particularly those in tertiary locations and tailored for specific AI training applications.
3. Tenant financial scrutiny: Conduct thorough scrutiny of the profitability and financial health of AI market participants, as these entities are critical drivers of demand.
4. Infrastructure readiness: Prioritize data center projects that have advanced power supply solutions and chip deliveries in the pipeline.
5. Exit strategy stress resting: Stress-test exit assumptions and allow for longer realization or continuation vehicles.

---