The bill has arrived. After two years of breathless AI investment announcements, enterprises are opening their monthly cloud statements and discovering something uncomfortable: AI spending does not behave like software spending. It never did, but nobody wanted to say so while the demos were still impressive.
Businesses are projected to spend $680 billion on AI models, agents, and related infrastructure next year. That number is large enough to feel abstract, so consider what it represents in practice: every Fortune 500 company, every mid-market firm with a CIO who attended a conference in 2024, every consultancy that rebranded its analytics practice as “AI-powered,” all of them paying consumption-based fees that fluctuate with usage rather than fixed subscription costs they can budget against in January and forget about.
That is the core of the problem. The SaaS model that enterprise IT departments spent 15 years learning to manage gave them predictability. You negotiated seats, you paid per seat, you knew your number. Pay-as-you-go AI pricing is different in kind, not just degree. Inference costs vary with query complexity, model size, and the number of agents you’ve chained together. A workflow that costs $4,000 in a quiet month can cost $40,000 when a product team decides to run it against a new dataset. CFOs are encountering this for the first time, and a correction looms for OpenAI and its peers as organisations start auditing what they’re actually getting for the spend.
The cloud parallel is instructive, up to a point
Reuters Breakingviews draws the comparison to cloud computing’s maturation, noting that the market can still grow even as CFOs try to rein in costs. That framing is right, but it understates the speed of the current moment. Cloud cost management became a discipline gradually, over nearly a decade, as AWS bills crept up and FinOps teams emerged to contain them. The AI version of that cycle is compressed. Enterprises went from pilots to production deployments in 18 months, and the governance frameworks did not keep pace. The result is that cost reckoning is arriving at the same time as the technology is still proving its value, which makes the conversation harder.
The cloud analogy also has a limit. Cloud costs, once understood, are reasonably controllable: you right-size instances, you reserve capacity, you shut down what you don’t use. AI inference costs are harder to predict because they’re tied to user behaviour and model behaviour simultaneously. A more capable model costs more per token. Users who discover a useful tool use it more. Both effects compound in the same direction.
What the market signals are telling us
Two data points from June 2026 are worth reading together. Databricks chose to raise private capital at a valuation of up to $175 billion rather than pursue an IPO this year. Reuters characterised this as a smart use of a private funding lull while SpaceX and Anthropic occupy the public debut conversation. Read another way, it’s a company with full visibility into enterprise AI spending choosing not to test public market sentiment right now.
Meanwhile, US tech stocks recorded their worst month of 2026, with the sell-off driven explicitly by investor concern over AI valuations. Julie Biel, Chief Market Strategist at Kayne Anderson Rudnick, put it plainly: it is “hard to gauge where the biggest profits from AI will be seen.” That is not a bearish statement about AI as a technology. It is a sober statement about AI as an investment thesis at current valuations. The distinction matters.
Underneath both signals sits a structural concern that Reuters Breakingviews raised in a separate analysis: circular financing, exaggerated profits, understated costs, and hidden liabilities as recurring patterns in the AI boom. The piece invokes John Kenneth Galbraith’s concept of the “bezzle”: the undiscovered embezzlement that exists during boom periods, when asset prices are rising fast enough to conceal losses that only become visible when conditions tighten. Whether the AI market has a bezzle in that sense is an open question. That serious financial analysts are asking it publicly is itself a signal.
Where this leaves enterprise buyers
The practical problem for a CIO or CFO right now is that the governance tooling for AI spending is roughly where cloud governance was in 2013. There are vendors selling AI cost observability platforms, but the category is young and the market is fragmented. Most enterprises are managing AI spend through a combination of manual budget caps, spreadsheet tracking, and periodic “what did we spend last month?” conversations that happen after the fact rather than before it.
The firms that will come through this correction in better shape are the ones treating AI procurement like any other category of enterprise spend: with a sourcing strategy, a cost baseline, a usage policy, and a clear ROI framework for each deployment. That sounds obvious. It is not yet common practice.
A few concrete things that matter here. First, seat-based SaaS contracts are not a safe default any more: many vendors have already moved to consumption pricing or are moving there, and contracts signed on the old model will be renegotiated. Second, the number of models an enterprise runs matters. Every model in production is a cost centre, a security surface, and a governance obligation. Model rationalisation is not a theoretical exercise. Third, the build-versus-buy decision for AI capabilities is more complex than it was for conventional software because the cost structure of building (GPU infrastructure, fine-tuning, evaluation) and buying (consumption fees, vendor lock-in) have both changed significantly in the past 18 months.
The consulting angle is real, and it cuts both ways
For advisory firms, this environment creates genuine demand. Enterprises need help with AI cost governance, model rationalisation, and vendor selection. AI vendors facing a correction need help with pricing strategy, customer success, and positioning in a market where buyers are suddenly more sceptical. That is a real dual mandate, and it is already showing up in the types of engagements firms are being asked to scope.
The risk is that consulting firms oversell the complexity. AI cost management is not a multi-year transformation programme. It is a procurement and governance problem, and most organisations can make meaningful progress in 90 days with the right focus. The temptation to frame it as something larger is understandable commercially. It is also the fastest way to lose credibility with CFOs who are already irritated about the bills they’re receiving.
The $680 billion projection will probably prove directionally correct. Enterprise AI adoption is real and the underlying value is real in enough cases to sustain the market. But the path from here to there runs through a period of cost scrutiny, vendor consolidation, and governance build-out that the AI industry has not yet fully priced in. Public markets, apparently, have started.