The World Is Building AI Infrastructure. The Economics Are Still Catching Up.
Artificial intelligence has moved from a strategic aspiration to a budgetary commitment. Across the United States, Europe, the Middle East and parts of Asia, companies are allocating capital to AI infrastructure with a sense of urgency that reflects more than enthusiasm for new technology. AI is increasingly framed as a prerequisite for relevance, not a discretionary investment, and that framing has begun to shape corporate behaviour at scale.
The spending that follows is substantial. Cloud capacity is being expanded, specialised chips are being secured years in advance, and long-term power and real-estate commitments are being made in parallel. Governments are being drawn into the process through industrial policy, energy planning and incentives aimed at attracting or retaining AI infrastructure within national borders.
Publicly, the logic is straightforward. In a race to build foundational capability, hesitation is assumed to be more dangerous than excess. Privately, however, the way this expansion is being financed suggests that confidence in when the returns arrive is more conditional than the rhetoric implies. Also, analysts have dubbed 2026 the year of AI Payback. Enterprises are moving away from pilot programs toward strictly measuring hard ROI. Software companies failing to show productivity gains are seeing renewal rates drop.
A buildout that assumes time
The scale of the AI buildout is now large enough that investors and analysts have begun comparing it to earlier infrastructure cycles rather than to software adoption curves. Capital expenditure estimates tied to AI run into the hundreds of billions of dollars over the coming years, while current revenues attributable directly to AI remain much smaller.
This imbalance sits at the heart of the argument put forward in Praetorian Capital’s widely circulated AI Addendum. After speaking with participants across the AI ecosystem, from data-centre operators to suppliers and financiers, the firm described a recurring admission: while belief in AI’s long-term importance is widespread, clarity on the financial math required to justify today’s spending levels is not.
The point is not that the investment is irrational, but that it is being made ahead of proof. The industry is building capacity on the assumption that demand, pricing power and monetisation models will mature in time to support it. That assumption may well prove correct, but it introduces a temporal risk that is rarely acknowledged in public narratives.
That risk is already influencing how companies choose to carry the financial burden.
Where the balance sheets show restraint
Rather than funding all AI-related infrastructure directly on their corporate balance sheets, a growing number of large technology firms are relying on special-purpose vehicles, joint ventures and project-level financing structures. These tools are standard in capital-intensive sectors such as energy or transport, but their increasing prominence in technology marks a shift in financial posture.
Financial Times reporting has highlighted how tens of billions of dollars of AI data-centre debt are now being raised through structures that reduce its immediate visibility at the parent-company level. The stated rationale is flexibility and risk management, but the implication is harder to ignore: companies are preparing for the possibility that returns may take longer to materialise than the strategic narrative suggests.
A frequently cited example is Oracle, which has raised roughly $18 billion in debt to support cloud and AI infrastructure expansion. Oracle remains strongly cash-generative, yet portions of that borrowing are tied to specific projects rather than carried as conventional corporate debt. This does not remove risk, but it isolates it, aligning repayment more closely with the performance of individual assets.
For companies with robust cash flows, this choice is telling. It reflects a desire to participate fully in the AI buildout while limiting the consequences if the path to monetisation proves uneven.
The enterprise side is still negotiating value
While infrastructure investment assumes scale and utilisation, enterprise adoption continues to move more cautiously, shaped by organisational realities rather than technological capability.
Kurt Muehmel, Head of AI Strategy at Dataiku, said the challenge many companies face is not whether AI can perform tasks, but whether its impact can be measured in ways that justify reliance on it.
“Value becomes murky when AI is positioned as a broad, general productivity improvement,” Muehmel said. “It is difficult to quantify and link to measurable business outcomes.”
According to him, confidence only begins to form when AI is applied to well-defined processes and tied to outcomes such as reduced cycle times, lower costs, revenue gains or risk mitigation. Achieving that level of clarity requires more than deploying models; it involves redesigning workflows, assigning accountability and accepting new dependencies, all of which introduce internal friction.
As a result, many AI initiatives remain in extended pilot phases. They are technically functional but organisationally unresolved, leaving a gap between the capacity being built upstream and the revenue being realised downstream.
Responsibility spreads, caution follows
That gap is reinforced by the way responsibility for AI systems is distributed across the ecosystem.
Tim Pfaelzer, General Manager and Senior Vice President for EMEA at Veeam Software, described an environment in which accountability is layered rather than centralised.
“At a high level, I see three layers,” he said. “Model owners, cloud providers, and the enterprise layer - the company that deploys and governs the use of AI in real workflows.”
In practice, enterprises deploying AI often carry the greatest exposure. They are the ones accountable to regulators, customers and courts, even when the technology itself is developed and hosted elsewhere. This reality shapes behaviour, encouraging caution and extensive governance before systems are allowed to operate at scale.
Pfaelzer noted that requirements for traceability, contractual safeguards and rollback mechanisms are increasingly non-negotiable. These controls are necessary, but they also slow deployment and delay revenue, while infrastructure costs are incurred immediately.
Why capital keeps flowing
Despite these constraints, capital continues to flow into AI infrastructure at speed, a dynamic that cannot be explained by near-term economics alone.
Jessica Constantinidis, Innovation Officer for EMEA at ServiceNow, argued that the current phase of AI development rewards participation over restraint, particularly among the largest players.
“What you’re watching right now is a phase where the biggest players are collaborating because it’s expensive, they’re still learning and they’re trying to avoid direct cost,” she said.
This collaboration spreads risk across partnerships and reinforces concentration within the ecosystem. It also creates a dynamic in which stepping back carries strategic and reputational consequences, even if financial clarity is lacking.
Constantinidis expects this phase to result in a small number of dominant global AI infrastructure and platform providers. Until that consolidation stabilises, companies are incentivised to keep investing, not because the returns are fully understood, but because absence from the buildout is perceived as a long-term liability.
How credit markets are reading the cycle
In credit markets, the AI boom is being watched with interest rather than alarm. Banks and bond investors have been willing to finance data centres and related infrastructure, particularly when cash flows are ring-fenced and supported by long-term contracts.
At the same time, the growing use of project finance and off-balance-sheet structures has drawn attention to how risk is being distributed. Credit analysts are increasingly focused on utilisation assumptions, power costs and the durability of demand, especially as hardware cycles shorten and refinancing risks emerge.
For now, abundant liquidity and investor appetite have allowed the buildout to continue. The question raised by Praetorian Capital’s analysis is not whether funding exists today, but how sentiment shifts if expectations around growth or pricing begin to soften.
Who feels the pressure first
If the AI buildout encounters resistance, the effects are unlikely to be evenly distributed.
Credit markets would likely react first, reassessing risk premiums and refinancing terms for infrastructure projects. Enterprises would follow, tightening budgets and slowing deployments if promised efficiencies fail to materialise. Governments, particularly those that have tied industrial policy or energy planning to AI investment, would then face pressure to justify subsidies and commitments.
Labour markets could amplify the impact. Constantinidis has warned that premature assumptions about workforce displacement could weaken demand and stress financial systems if job losses outpace the creation of new income streams.
“The real risk is the illusion some companies will have that they no longer need people at scale,” she said. “Once masses lose jobs, they lose income and insurance. Then they can’t pay loans.”
In such a scenario, infrastructure does not fail because it exists, but because the economic context supporting it deteriorates.
A familiar phase in a long cycle
None of this suggests that AI is a passing trend or that the buildout is destined to collapse. It suggests that the industry is navigating a familiar phase of a capital-intensive cycle, one in which investment runs ahead of clarity and financial structures are used to manage uncertainty.
Special-purpose vehicles do not eliminate risk. They shape when and where it appears.
What balance sheets, financing choices and investor behaviour collectively signal today is not doubt about AI’s long-term importance, but realism about the path to returns.
The technology is advancing rapidly, and the infrastructure is being built accordingly, yet the people funding that expansion are preparing for a longer, more uneven journey than the public narrative implies.
That tension between strategic urgency and financial caution is not a contradiction. It is the defining feature of the AI moment.