A Primary Research Primer on Data Centre and AI Infrastructure for Investment Professionals

The data centre and AI infrastructure sector is the physical layer on which the global AI economy runs. These are the facilities, power systems, cooling infrastructure, and compute hardware that make it possible for hyperscalers, AI labs, and enterprises to train models and serve inference at scale. For most of the last decade, data centres were a slow-moving, yield-oriented corner of real estate. The AI buildout has turned them into one of the most intensely watched infrastructure categories in global markets, combining capital intensity, geopolitical stakes, energy grid constraints, and a rate of technological change that makes the economics genuinely difficult to predict.

This primer covers how the industry is structured, where the money sits, how players across the value chain make their returns, who the relevant experts are, and the questions that produce genuine insight on an expert call. Woozle has run primary research programmes for long/short equity funds, specialist tech-focused pods, and PE deal teams across data centre operators, power infrastructure providers, and AI compute supply chains.

What Is Data Centre and AI Infrastructure?

A data centre is a facility built to house servers, networking equipment, and storage at high density, with guaranteed power, cooling, physical security, and network connectivity. The term covers everything from a 500-square-foot edge node to a multi-gigawatt hyperscale campus consuming as much electricity as a small city. AI infrastructure refers specifically to the compute-dense subset of this market: facilities and supply chains designed to support GPU clusters used for model training and inference.

Data centre equipment and infrastructure spending reached $290 billion in 2024, with Alphabet, Microsoft, Amazon, and Meta investing nearly $200 billion between them. The top three hyperscalers alone plan to invest more than $500 billion in capital expenditures for infrastructure supporting AI deployment in fiscal year 2026. These are numbers that reclassify data centres from a niche infrastructure sub-sector into a first-order macroeconomic force.

The industry breaks into four structural segments. Hyperscalers (AWS, Microsoft Azure, Google Cloud, Meta) build and operate their own facilities at a scale no one else can match. Colocation providers (Equinix, Digital Realty, Vantage, CyrusOne, Iron Mountain) own the facilities and lease power-plus-space to tenants who bring their own hardware. Neoclouds (CoreWeave, Lambda Labs, Nebius, Nscale) are a newer category: they acquire GPU hardware, typically Nvidia H100s, H200s, and GB200s, and rent compute capacity by the hour to AI developers and enterprises who cannot secure sufficient GPU supply from the hyperscalers. Power and cooling infrastructure suppliers (Vertiv, Schneider Electric, Eaton, ABB) sell the critical physical infrastructure that makes any of these facilities function.

The margin concentration within this stack is counterintuitive. The companies with the smallest public profile — the Vertivs and Schneider Electrics of the world — often sit in a more defensible position than the operators above them. Power distribution, cooling systems, and UPS infrastructure represent under 10% of total data centre cost but are the items with the longest lead times and highest switching costs. The colocation operators capturing hyperscale demand are generating EBITDA margins above 50% at the best-positioned players, funded by 10-to-15-year leases signed before a single rack goes live.

Why Are Investors Looking At This?

The structural driver is not subtle. Traditional data centres operate at 10-15 kW per rack. AI workloads demand 40-250 kW per rack. This is not an incremental upgrade to existing infrastructure. It is a full redesign of the facility layer, requiring new power architectures, liquid cooling systems, and site selection criteria centred on grid access rather than real estate convenience. Every Nvidia GB200 NVL72 rack draws approximately 120 kW. The incumbent infrastructure estate was built for a world where racks drew 5 kW. Retrofitting or replacing that estate is a decade-long capital programme, and investors who own the right assets in the right locations for the right duration stand to compound through it.

The return profile varies sharply by segment. Colocation REITs like Equinix and Digital Realty trade on adjusted EBITDA multiples and AFFO yields, reflecting the long-lease, asset-heavy nature of the business. Equinix sustains a 51% EBITDA margin on top of a network effects moat built around interconnection: its facilities host the physical meeting point between cloud providers, carriers, and enterprises, which creates a switching cost that pure wholesale players do not have. Neoclouds trade on revenue multiples and contracted backlog, with investors pricing in whether GPU commodity pricing erodes unit economics before the hyperscaler capex cycle slows. Power infrastructure suppliers trade on order backlogs and book-to-bill ratios, with Vertiv's $8.5 billion backlog and 1.2x book-to-bill serving as a frequently cited leading indicator for the broader sector.

The live debate is a genuine bull/bear standoff. Goldman Sachs's baseline model implies $765 billion in annual AI capex in 2026, growing to $1.6 trillion by 2031. Investors who are long argue that demand is structural, not cyclical: every new AI model generation consumes more compute, inference workloads scale with adoption, and lead times on power and cooling equipment mean supply cannot catch up for years. Those who are cautious point to customer concentration risk — CoreWeave generated 62% of its revenue from Microsoft at IPO — the possibility of model efficiency gains reducing compute intensity, and the precedent of the late 1990s telecom buildout as a reference for infrastructure euphoria ending badly. H100 rental rates have already declined 60-75% from their peak.

The more recent complication is on the supply side of supply: the power grid itself. High-voltage transformer lead times, which ran 24-30 months pre-2020, now stretch to five years. Projects that exist on paper and in press releases are not the same as projects that will come online on schedule. This is the variable repricing the sector in real time.