Generative AI has moved quickly from experimentation into early production use in many enterprises. However, very few can confidently forecast what it’s going to cost them in six months.

For a technology that has consumed so much board-level attention and capital, that reflects a lack of certainty, and one that some technology leaders may privately recognize as true of their own organizations.

The spend is real and the direction is clear, but the number at the end of the year can remain genuinely uncertain.

To capture a glimmer of the confidence driving the infrastructure race, Amazon’s CEO has indicated it expects to spend heavily on IT infrastructure to support AI, with an estimated $200 billion in AI capital spending, arguing it is “not going to be conservative” in how it invests in the tech.

In practice, what makes AI different from the infrastructure investments that came before it is not the scale of the commitment but the nature of the consumption.

Cloud computing was unpredictable when it arrived too, but it eventually settled into patterns that finance teams could learn to model. AI hasn’t settled in the same way yet, and much of the reason comes down to how it is being used.

A great deal of enterprise AI use remains exploratory, which is part of what makes forecasting harder. And unlike cloud, which stayed largely within technical teams for years before spreading, AI is moving across the whole organisation almost immediately. That changes everything about how you try to govern it.

The limits of financial visibility

On the surface, some forms of AI appear to offer what earlier infrastructure lacked: clean, granular, real-time data and what it costs. But across the rapidly growing landscape of technology providers leveraging AI in some way, many do not.

In some cases, token-based pricing is precise in a way that early cloud billing never was, and for finance teams accustomed to working with far less, it can feel like a step in the right direction for solving the visibility problem.

We unfortunately still have a long way to go, since simply understanding what was spent last month tells you very little about what will be spent next quarter, particularly once adoption moves beyond the teams who originally shaped the business case.

One must consider that teams across legal, HR, and customer operations are not thinking about token economics (tokenomics). They’re only thinking about whether the tool works.

Cost exposure builds not through any single decision but through dozens of small expansions, each reason in isolation, none of them reflected in a comprehensive forecast. By the time anyone joins the dots, the demand curve has already moved.

Extending the disciplines that already exist

The organizations who are doing a better job managing AI spend have tenured experience managing consumption-based technology. IT asset management (ITAM) teams for example often have more experience dealing with more fixed constructs like users or seats, which makes the consumption-based nature of AI far more challenging.

FinOps teams on the other hand have grounded experience in managing consumption that originated in public cloud. FinOps teams may therefore better positioned to deal with the new tsunami of AI consumption and spending, ensuring that it is governed as adoption scales.

FinOps has also been broadening its scope beyond the initial roots in public cloud, with AI cost management now sitting firmly within that remit for many, a shift reflected in how the FinOps Foundation is increasingly incorporating AI into its guidance. Part of that expansion is about forecasting demand that behaves differently from conventional workloads.

There is also growing interest in whether AI itself can support FinOps practices, particularly in anomaly detection, optimization and, over time, forecasting, as consumption patterns become harder to model.

The challenge is applying FinOps practices early enough so that governance shapes how AI scales, rather than scrambling to restore control once spend has already outpaced oversight.

The compounding difficulty of legacy environments

For organizations whose technology estates were built around consistency, extending governance into AI is harder than it sounds.

AI-first organizations design with cost in mind from the beginning, treating inference the way they would any other product input, with economic constraints shaping architecture decisions before commitments are made.

Retrofitting AI into legacy infrastructure means something different. Existing commercial commitments and operating models do not adapt quickly to a consumption model that is inherently variable, and that friction has a direct bearing on cost.

The difficulty is often that AI is being introduced into environments built around very different assumptions about how demand behaves, and that is part of what makes forecasting harder.

The challenge is not simply new spend, but expenditure ballooning in environments where oversight and control are already difficult to maintain.

Organizations navigating this will tend to run controlled experiments before broad rollout and are deliberate about how adoption spreads. In practice, that is often about containing unmanaged adoption early, before usage patterns, costs and dependencies become harder to unwind.

That same exposure increasingly carries beyond internal governance. As AI appears more often in customer procurement conversations, questions that were once largely internal are starting to be probed externally too.

For organizations whose governance has not kept pace, those questions can force a level of clarity they may not yet be prepared to provide.

From activity metrics to business outcomes

Beyond governance and cost control, there remains a harder question, which is whether AI investment is producing meaningful business value. Most leadership teams are not yet in a position to answer that with confidence, and the metrics currently reaching the board are not making it easier.

Model usage, inference volumes and compute consumed describe activity without explaining value. It is easy to build a compelling board update from consumption data without addressing whether any of it is moving the business.

What gets closer to an answer is understanding whether individual inferences are delivering something a customer would pay for, or something that meaningfully reduces cost or risk.

Incremental business outcome per pound or dollar of AI spend is a harder measure to produce, but it is closer to the economics that matter because it requires a clearer position on what AI is actually delivering.

That is precisely where many organizations are still finding the work harder than it looks, particularly as AI deployment moves ahead of the models used to understand cost and value.

That disconnect matters more as the market expands, because where those economics remain unclear, cost exposure can build in ways that are harder to recognize early and harder to contain later.

For many enterprises, the challenge ahead is scaling AI without allowing spend to outrun the value it is meant to create.

The best cloud storage: tested, reviewed and rated by experts.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Governing board member of the FinOps Foundation and serves as SVP of Cloud & FinOps at Flexera.