For two years, the boardroom conversation about AI has been almost entirely one-sided. The question was always some version of the same thing: how fast can we move before we fall behind? Budgets were approved on the strength of momentum, and momentum passed for strategy. That phase is ending. The question being asked now, in rooms that do not leak to the press, is the one nobody wanted to ask first: what is this actually costing, and what is it returning?
The shift is not subtle. It is reshaping how companies think about capital, about risk, and about the shape of their own futures.
From an R&D line item to the core balance sheet
The first thing that changed is the scale and the location of AI spending. A few years ago, AI was a research project — a line item in the technology budget, interesting but contained. That is no longer true anywhere.
AI has become capital expenditure. Data centers, chips, networking, power contracts, hiring — these are not experiments; they are assets being acquired and built. In the largest technology companies, this spending now runs to the tens of billions of dollars a year, and it is increasingly financed with debt. The bond market has noticed: a growing share of new investment-grade and high-yield issuance is now coming from companies funding their AI buildout. When a company starts borrowing heavily to buy infrastructure that has not yet shown a clear return, the balance sheet has entered the conversation in a way that a research budget never could.
None of this is, by itself, wrong. Every transformative technology required a period of heavy upfront investment. Railroads, electricity, the internet — all were built on borrowed money before they returned it. The question is not whether companies are spending. It is whether the spending is being made with discipline, and whether the returns are arriving on anything like the promised schedule.
The return question is getting harder to dodge
For a long time, the answer to “what is AI returning?” was essentially: it is early. That answer worked for two years. It is starting to wear thin.
The problem is not that AI is useless. It demonstrably is not — the productivity gains in software development, customer service and content production are real and measurable. The problem is the gap between the gains and the outlay. A company can save real money with AI in its operations and still be losing the race, if its competitors are deploying more efficiently and if the infrastructure it bought is depreciating faster than expected.
There is also the subtle matter of what the spending is actually for. Some of the biggest AI expenditures are not about products at all; they are about staying in a game where the cost of entry keeps rising. That kind of spending — defensive, mandatory, hard to point to a single revenue line — is the hardest to justify in a downturn. Boards are beginning to notice which of their AI projects pay for themselves and which are simply premiums on staying relevant.
The market has stopped rewarding the story
The discipline is not coming only from inside companies. The capital markets have started to do their own filtering.
For a stretch, any company that mentioned AI saw its shares treated favorably. That indiscriminate reward is over. Investors are increasingly distinguishing between companies that can show earnings quality, free cash flow and credible monetization, and companies that are merely spending to keep up. The penalty for the latter is growing, and it shows up in widening valuation spreads and in the price of their bonds.
This is healthy. It is the mechanism by which capital flows to the most productive uses of the technology rather than the loudest claims. But it also creates a new kind of pressure inside companies. Executives who were comfortable saying “we are investing heavily in AI” now have to answer the follow-up: “and what does that investment produce?” That is a harder question to deflect, and it is changing how proposals get approved.
What disciplined AI investment actually looks like
So what separates the companies that will come out of this phase strong from the ones that will come out wounded? The pattern, from the evidence so far, comes down to three things.
The first is specificity. The companies doing best are not funding general-purpose AI initiatives; they are funding a small number of sharply defined problems — a customer-service function, a supply-chain forecasting system, an internal productivity platform — where the ROI can be measured in months, not in a vague promise of transformation.
The second is reuse. The cheapest AI dollar is the one spent once and used a hundred times. Firms that build a capability once and spread it across the organization get a very different economics from firms that fund a hundred one-off pilots. The gap between those two approaches is not small; it is often the difference between a profitable investment and a write-off.
The third is the willingness to stop. Healthy organizations kill AI projects that fail, on schedule, without embarrassment. The unhealthy ones keep funding losers because canceling an AI project feels like admitting the hype was wrong. In this environment, the ability to say no is becoming one of the most valuable management skills.
The risk nobody is pricing yet
There is one risk that still gets too little attention, and it is the one that could hurt the most. If a meaningful share of the enormous AI infrastructure now being built turns out to be overbuilt — if demand grows more slowly than supply, or the technology improves faster than the assets can be amortized — then the write-downs will be enormous. Whole balance sheets have been built around the assumption that AI demand keeps growing at this pace.
That assumption is reasonable today and has been for a while. But the history of every infrastructure boom — railroads, fiber optics, telecom towers — is that the overbuild always arrived, and it always arrived when nobody expected it. The companies that manage their AI spending as if that could happen, rather than as if it cannot, will be the ones left standing if it does.
None of this means the AI investment wave is a bubble about to burst. It means it is a real investment cycle, with real winners and real losers, and the winners will be decided by execution and discipline, not by who spoke the most confidently about the future.
What boards should actually be doing
So what should a board actually be doing in this environment? The evidence points to a few practical moves.
First, ask for the unit economics. Not the vision — the unit economics. What does one customer interaction cost with AI versus without? What is the payback period on the infrastructure? If the answers are not on the table, the project is not ready for approval. Second, tie AI spending to a business outcome, not to a technology budget. The proposals that survive should be the ones that can name the metric they will move and the date they will move it by. Third, stress-test the downside. Ask what happens if the revenue assumption is wrong by half, or if the technology improves so fast that this year’s investment is obsolete in two. Companies that run those scenarios do not enjoy them, but they survive them.
None of this is heroic or headline-worthy. It is the unglamorous work of managing a large investment through a period of genuine uncertainty. But it is exactly the kind of work that separates the companies that treat AI as a tool from the companies that were treated by it.
The boardroom question has changed. It used to be “are we doing enough?” Now it is “is this working, and how would we know?” The companies that can answer that second question honestly — with numbers, not narratives — are the ones that will be around for the third act.