Beyond the Hype: A Framework for Evaluating AI Infrastructure Investment

The AI infrastructure investment market is producing a familiar pathology: an abundance of capital chasing a scarcity of genuinely credible projects. Announced investment figures in the hundreds of billions are now commonplace. Less common is the rigorous, independent evaluation that distinguishes infrastructure that will perform — technically, commercially, and operationally — from infrastructure that will not.

This piece sets out a framework for that evaluation. It is grounded in 25 years of direct experience designing and implementing digital infrastructure at scale across three continents, and in the specific demands that frontier AI workloads impose on the facilities built to run them. It is not a theoretical framework. It is the framework we apply when advising clients who have real capital at risk.

Why Standard Due Diligence Is Insufficient

Infrastructure investors who have underwritten conventional data centre projects understand how to evaluate a facilities business. They know how to read an annualised power usage effectiveness figure, assess a lease structure, and stress-test an operator's track record. What the AI infrastructure moment demands is something different: an evaluation framework calibrated to requirements that have no direct precedent in the history of commercial computing.

The difference is not merely quantitative — more power, more cooling — though the quantitative differences are dramatic. It is qualitative. A facility designed for enterprise cloud workloads and a facility designed for frontier AI training are not the same class of asset in the same way that a general cargo vessel and a liquefied natural gas tanker are not the same class of vessel. They share superficial characteristics — they float, they carry cargo — but their engineering, their operational requirements, and the risk profiles of investing in them are fundamentally different.

Investors who apply standard data centre due diligence to AI infrastructure projects will consistently underestimate technical risk, overestimate operational feasibility, and — ultimately — overpay for assets that cannot deliver the utilisation and return profiles they were underwritten to produce.

The Five Dimensions of Evaluation

A credible evaluation of an AI infrastructure project must address five dimensions. Each is necessary; none is sufficient on its own. The failure modes in each dimension are distinct, and advisers who over-index on one — typically power, because it is the most visible — while neglecting others will miss material risks.

Dimension One: Power — The Threshold Criterion

Power is not one factor among many. It is the threshold criterion. A project without a credible, costed, and timetabled route to reliable power at the required scale is not a viable AI infrastructure investment, regardless of its other characteristics. This principle sounds obvious, but the number of projects currently seeking investment on the basis of unresolved or aspirational power strategies is significant.

The evaluation questions are specific: Does the project have a grid connection agreement, or is it in the interconnection queue? If in the queue, at what position, and what is the realistic timeline? What is the utility's track record of delivering connections on the stated timeline in that market? Is there a credible behind-the-meter generation alternative if grid delivery slips? What is the energy cost structure over the full underwriting horizon, and what assumptions are embedded in the operating cost model?

In markets where the interconnection queue stretches five to seven years — Northern Virginia being the most prominent example — a project that has not yet secured a grid connection agreement is not early-stage; it is pre-feasibility. The distinction matters for underwriting.

Dimension Two: Thermal Architecture

The thermal engineering of an AI data centre is categorically different from that of a conventional facility. Current-generation GPU clusters draw 80–200kW per rack; leading hyperscalers are designing for configurations beyond 300kW. Air cooling — the dominant technology in conventional data centres — becomes physically impractical at sustained rack densities above approximately 30–40kW. Above that threshold, liquid cooling is not a premium option; it is an engineering necessity.

The evaluation framework must distinguish between facilities with genuine high-density liquid cooling capability and facilities that have been marketed as AI-ready on the basis of air-cooled infrastructure with selective liquid cooling additions. The questions to ask: What is the maximum sustained rack density the facility can support? What proportion of the total white space has liquid cooling infrastructure in place, versus planned or optional? What is the facility's mechanical design basis, and was it originally designed for high-density AI workloads or adapted from a conventional design?

Retrofitting liquid cooling into a facility designed for air cooling is possible but costly, disruptive, and technically constrained by structural factors — floor loading, under-floor space, ceiling heights — that are difficult and expensive to change. A facility that requires extensive retrofitting to support the workloads it has been underwritten to host is a facility whose pro forma economics should be revisited.

Dimension Three: Network Fabric

The network requirements of large-scale AI training are unlike those of any other data centre workload. A large language model training job distributes computation across thousands of GPU accelerators simultaneously, generating high-volume, low-latency east-west traffic — communication between compute nodes within the facility — that is fundamentally different from the north-south traffic patterns (client to server) for which most data centre network designs are optimised.

The fabric connecting AI compute nodes must deliver aggregate bandwidth in the tens of terabits per second with sub-microsecond latency. This is achievable, but it requires specific infrastructure: InfiniBand HDR or NDR interconnects, or equivalently capable Ethernet fabric, deployed at scale and configured for the traffic patterns of distributed training workloads. The capital cost of this network infrastructure is substantial and is a component that is frequently underestimated in project pro formas.

The evaluation questions: What interconnect technology is deployed or planned? At what scale — rack count, total bandwidth, switching architecture? Has the facility successfully run distributed training workloads at the planned capacity, or is the network design untested at that scale? Who designed the network, and what is their specific experience with AI training fabric?

Dimension Four: Commercial Viability and Demand Certainty

AI infrastructure is capital-intensive and operationally complex. The economics only work if utilisation is high and sustained over a long horizon. This makes demand certainty — the degree to which the facility has contracted, creditworthy off-take — a central evaluation criterion.

The current market presents a particular challenge in this respect. Announced AI infrastructure demand from hyperscalers and large technology companies far exceeds the capacity actually contracted on long-term agreements. A project that has a letter of intent from a large technology company, or that is underwritten on the assumption of securing hyperscaler tenancy, is not the same as a project with a signed, long-term power purchase or colocation agreement. The gap between the two categories is where many AI infrastructure projects currently reside.

The evaluation framework must be rigorous about the nature and enforceability of demand commitments. It must also assess the counterfactual: if the anchor tenant does not proceed, or proceeds at lower scale than planned, what is the realistic alternative demand? In markets where hyperscaler demand is the only viable source of revenue at the scale required to service the capital structure, the absence of a committed anchor tenant is not a risk to be modelled — it is an underwriting question to be resolved before capital is deployed.

Dimension Five: Operational Governance and Technical Management

AI infrastructure is not self-operating. Running a high-density, liquid-cooled data centre with specialised GPU infrastructure at high utilisation requires a level of operational capability that is genuinely scarce. The skills required — expertise in high-density thermal management, GPU cluster operations, distributed training environment management, and the security and compliance obligations of an AI infrastructure provider — do not exist in the same depth or abundance as conventional data centre operational skills.

The evaluation framework must assess whether the project sponsor has the operational capability to run the asset it is proposing to build, or whether it is relying on third parties whose own capabilities have not been independently assessed. It must also assess the governance architecture: who makes technical decisions, on what basis, and with what accountability? What is the escalation path when something goes wrong — as it will — with complex infrastructure of this kind?

A facility with credible power, robust thermal design, appropriate network fabric, and contracted demand will still disappoint its investors if it is operated incompetently. Operational due diligence is not a secondary consideration; it is the difference between an asset that performs and one that does not.

The Due Diligence Checklist

The following checklist represents the minimum scope of independent technical and commercial due diligence that AI infrastructure investors should commission before deploying capital. It is not exhaustive — specific projects will generate specific questions — but it covers the material risk areas in each of the five dimensions.

A Note on Jurisdiction and Geography

The five dimensions above apply universally. But their relative weighting and the specific questions they generate vary materially by jurisdiction. An AI infrastructure project in the Gulf Cooperation Council faces different power, regulatory, and talent challenges than a comparable project in the UK or the US Pacific Northwest. The risk that is most acute in Loudoun County, Virginia — grid interconnection timeline — is a different order of magnitude from the equivalent risk in a Gulf market where sovereign demand provides both power anchor and planning certainty.

This is not an argument for geographic arbitrage as an investment strategy — every market has its specific risks — but it is an argument for evaluation frameworks calibrated to the specific context of each project rather than applied generically. An adviser with direct experience in the market in question is not a luxury in this evaluation; it is a material input to the quality of the analysis.

The Cost of Getting It Wrong

AI infrastructure projects at the scale currently being underwritten represent some of the largest single infrastructure investments in history. A campus-scale AI data centre may represent $2–5 billion of capital over its development phases. At that scale, the cost of technical errors made at the evaluation stage — errors that could have been identified with rigorous independent due diligence — is not marginal. It is the difference between a successful investment and a failed one.

The market is in an early phase. There is abundant capital, significant excitement, and real demand from real AI workloads. But there are also projects that will not perform as underwritten, facilities that will not be operated as designed, and demand commitments that will not be honoured. The investors who distinguish between these categories — systematically, rigorously, and with the benefit of specific technical expertise — will outperform those who do not.

That is what independent advisory exists to do.