Assessing Data Centre Investment Risk - The New Reality

AI infrastructure creating a new financial asset class is not news.

However, this doesn’t detract from investor challenges of adapting at pace to levels of capital that would have been almost unimaginable just a few years ago. Levels of capital that are being deployed into a market that has matured rapidly in scale and sophistication and one drawing in tens of billions of dollars of project finance, infrastructure debt, structured lending and institutional capital markets.

By now the numbers are familiar. Around $70 billion of infrastructure and project finance closed during the first quarter of 2026. Approximately 35% of this volume related to digital infrastructure. The market is on course to exceed $100 billion of data centre financing annually.

Dominant discussions are now centred on liquidity, leverage, refinancing, risk pricing, yield, institutional capital, and capital recycling, reflecting the transition of AI infrastructure into one of the world's largest and fastest-growing infrastructure investment sectors.

Upwards, Inwards, Outwards

At same time, data centre industry investment is not just scaling upward among traditional players but also inward to the AI equipment industry (mostly NVIDIA currently) and outward to the energy transition.

In almost every territory it is access to power, not land or equipment supply chains that is being cited as the major bottleneck, and one which is drawing in new investment considerations around the financing of power generation and storage infrastructure.

As one CEO involved in grid supporting and alternative energy storage at scale said recently: “I spend a lot of time explaining the peculiarities of the energy industry to investors.” 

At the upper end of public cloud AI infrastructure (tech titans), AI factory (LLM builders) and Neoclouds the market has changed to one where the dominant capex need is for GPU and surrounding IT infrastructure. This represents a shift from the traditional development arena where construction and M+E equipment required the greater levels of finance.

Capex allocation and the Nvidia effect

Two-thirds to three-quarters of total data centre Capex is now going to GPUs (Blackwell, Vera Rubin) — meaning most investment ultimately flows to Nvidia rather than construction/infrastructure providers. For providers who invest and design for accelerated server refresh cycles that are in the gift of NVIDIA.

In olden days (2 years ago) hardware typically remained productive for around nine years through successive customer deployments. But AI has disrupted this model. New large language models emerge every few months, hardware generations evolve rapidly and different GPU platforms may be optimised for either model training or inference.

This makes long-term investment planning significantly more difficult, as operators cannot yet predict how AI infrastructure will be repurposed once today's GPUs reach the end of their commercial life.

In terms of financing Cole Crawford, founder of NativX, says: “If you're a Neocloud looking at your financial  tooling — Blackstone, Magnetar, Anchorage, Blue Owl, Digital Bridge — you can get debt if you've got the supply chain. But your lender wants their money back in 12 months, and you need to have made all your money in 24, because a GB300 today is one generation old the moment Vera Rubin ships six months later. With silicon releases every nine months — you're two generations old before your financing term is up.”

With 10–20 year ROI/ROA timelines being modelled against hardware cycles that shift every 1–2 years, operators face a genuine dilemma: optimize for today's efficiency or build in flexibility for tomorrow's unknown architecture.

So what is the solution?

There's no universal right answer — some operators are deliberately deploying prior-generation systems to capture near-term token pricing premiums (since compute value today outweighs long-term efficiency losses); others have paused construction mid-build to redesign around new chip generations. This is becoming a genuine point of investment strategy differentiation, not just an engineering detail.

Pricing Risks of power project on project finance

With lengthening grid connection queues common across Europe the interest in on site power is spiking.

From an investor perspective lenders appear increasingly comfortable with structures for funding combined data centre and power plant projects.

Questions being asked in the finance world are whether markets are charging sufficiently for the additional complexity of Project-on-project risk?

What is the premium for the additional headache of funding a data centre project combined with a power plant project?

And what do combined projects do to loan to cost levels?

The risk analysis for funders evaluating the added complexity of potential power generation risks include:

• Captive power
• Power plant underperformance affecting data centre uptime 
• Dependency between the two assets
• Limited alternative customers for captive power generation
• Potential tenant termination if power supply fails.
• Refinancing risk
• Execution risk
• Construction risk
• Counterparty risk
• Stranded asset risk
• Default risk

Operator and Off taker Risk

Then there are the varied risk profiles of different operators.

A common view is that NeoClouds carry materially different credit risk than compared to hyperscalers because the scale of free flow generated by hyperscalers is at levels never before seen. The view is that Neoclouds with investment grade off-takers on long term contracts command better pricing but single or dual Neocloud-customer concentration is priced as meaningfully riskier.

Of course, says Lars Ove Skorpen, Senior Partner and Head of Power & Renewable Energy within Investment Banking at Pareto Securities, there is an open question about long term hyperscaler contracts which are likely to have walk-away clauses meaning a 10-year off-take agreement isn't necessarily as durable as its headline term might suggest.

This neatly combines customer concentration, speed to market and power and capex risk.

New Entrants – learning the hard way

2026 marks a shift from announcement-driven froth toward execution-driven differentiation, where operational competence and multi-pronged capital exposure will separate winners from failed projects.

Blanka Thibaud, led site selection for AWS's France and Saudi Arabia's data centre footprint, and is founder of TechHer Capital and President of Via Terra, which is advising hyperscalers and data center operators across Europe on site selection, plus developers on portfolio/use-case strategies targeting 100MW+ hyperscale campuses.

"Capital is abundant, but deployment is staged and cautious. Despite headline mega-funds (Blackstone, BlackRock, DigitalBridge, plus a recent IPO), actual risk-taking remains conservative. Investment follows a curve: low early-stage capital, rising sharply only once a connection agreement is secured. TSOs typically won't grant a connection agreement without a signed operator already in place, meaning early-stage development is funded via smaller million-level tickets (family offices, and increasingly traditional commercial banks testing the space), while the large equity 'pockets' everyone reads about in press releases sit ready to deploy only once the project de-risks," she says.

Blanka believes the wave of new developers may have assumed building a data center "shell" was comparable to standard logistics construction: it isn't, given the cost and complexity of specialized equipment.

As reality sets in, value is shifting toward experienced local developers who can navigate zoning/permitting and political relationships: a less glamorous but increasingly prized skillset.

We are also witnessing a rewriting of regulatory sequencing, which is diverging by country.

Ms Thibaud references recent moves in Germany, which has flipped the traditional order. Here, connection-queue priority now requires proven project maturity (owned land, permits in hand) before power access. France has already begun a similar shift: a July 2025 CRE deliberation authorizes RTE to move to a "first-ready, first-served" model from June 2026, prioritizing projects on land control, financial guarantees and equipment orders in order to curb speculative applications. Industry sources suggest France may also move toward an approved list of contractors/consultants to further filter out speculative applications, though this is not yet confirmed in public record.

Three distinct asset classes, three risk/return profiles

1. Power + land: now requires a locked-in connection agreement (not just theoretical power access), land under option, with permitting/zoning as a seller contingency. Currently seen as the highest-yield, highest-IRR segment, arguably now overvalued: in deals I have reviewed, some landowners are pricing land at up to 10x underlying value.
2. Shell/vertical construction: increasingly managed by operators themselves after being burned by contractors historically, a distinct investment/yield category.
3. Turnkey (fully fitted out): the most capital-intensive category, but yields are now compressed relative to CapEx, making it the least attractive of the three.

How supply chain risk is being addressed 

Supply chain risk is viewed as real but manageable. While some bottlenecks persist in transformers, switchgear, and thermal/heat-rejection equipment, vendors (Schneider Electric, Vertiv, , Legrand, ABB, and power-generation OEMs like GE Vernova, Siemens Energy are actively expanding production capacity.

This is helping the developers plan ahead.

However residual risks such as tariffs, geopolitics, conflict, Strait of Hormuz-related petrochemical supply disruption, and Chinese rare-earth export controls remain.

Conclusion

AI demand is real and evident. For investors and developers the challenge is no longer simply identifying where demand for AI infrastructure will emerge. The differentiator lies in understanding which risks can be engineered away, which can be financed, and which must simply be priced.

Today some data centre financing appears less like conventional commercial real estate and more like the financing of integrated energy, digital and advanced manufacturing infrastructure.

Value has moved beyond construction cost and tenancy to grid access, power strategy, AI IT infrastructure refresh cycles, sovereign AI requirements, supply-chain resilience, regulatory sequencing and operator capability as interdependent factors influencing both project bankability and long-term asset value.

Investors are no longer just underwriting a building; they are underwriting an ecosystem in which energy assets, compute infrastructure and software platforms are increasingly inseparable.

For the data centre industry, this represents both an opportunity and a test.