Who Carries the Risk When AI Spending Fails to Pay Back

The laziest AI argument right now is the bubble argument. It sounds decisive and asks almost nothing of management. Prices are high, spending is huge and some promises will disappoint, so the label sticks and the conversation ends. As strategy it gives an executive nothing to do. Railways and telecoms both overbuilt, yet the railways still reshaped economies and the telecom bust left useful fiber behind. Cloud computing absorbed years of capital before it became clear which companies had built operating advantage and which had merely rented hope.

For executives, a narrower and less comfortable question is where the payback actually sits. AI spending has moved out of the software budget and into the capital structure, where it appears as data centers, power contracts, GPUs, long leases, vendor guarantees, prepayments, equity issuance and debt. A founder or board that only asks whether a model is "useful" is looking at the wrong account. The strategic work begins when you can name the asset being created, the cash flow you expect from it, who absorbs the downside, and what evidence would justify another round of spending.

For AI capital spending, the question that matters is who owns the payback.

What the market is already signaling

Signs of this are visible across the stack. In May 2026 S&P Global Ratings estimated that the five large cloud providers it rates could spend about $750 billion on capital expenditures in 2026, equal to 38 percent of revenue. Earlier in the year it had already described hyperscalers as "hyperspending" after guidance moved well above expectations. Financial newspapers ask the same thing more bluntly: who pays for AI? Usage-based pricing, token rationing, cheaper model alternatives and pressure on cloud free cash flow all suggest an industry moving from subsidized experimentation toward economic sorting. The chip supplier may get paid first and the cloud provider may spend first, while the enterprise buyer may learn the true unit cost only after a workflow has come to depend on the tool.

Neoclouds sharpen the ownership problem. CoreWeave's growth has been built on expensive infrastructure, large customer contracts and heavy borrowing. TeraWulf's pivot from bitcoin mining to AI data centers shows another version of the structure, in which a smaller infrastructure company can borrow against a lease because a giant technology firm helps backstop the deal. Iren's reported AI-cloud deals take a third route, using customer prepayments to reduce near-term funding risk while the company races to add capacity. Each arrangement settles who keeps flexibility if demand changes. A contract gives one party revenue visibility and the other a fixed obligation. Guarantees support the buildout and can park reputational risk on a larger balance sheet, and prepayments, which look like proof of demand, may still hide whether end customers will keep paying at the margin the buildout needs.

Split AI spending into four accounts

Before approving the next large AI commitment, executives should break the single "AI spend" line into four separate accounts of capital expenditure.

The usual failure is letting one account pass for another. Signed infrastructure contracts get read as operating proof, vendor roadmaps as customer willingness to pay, rising share prices as unit economics, and model benchmarks as workflow adoption. A data-center announcement becomes strategy only when the company can explain how the capacity will turn into advantage, and a press release does nothing to answer that.

Capacity owners and payback renters

The Tobias Carlisle-style capital-expenditure question is still worth asking. Value investors are trained to notice what is already in the price and who earns the return on incremental capital, and operators should run the same question inside their own companies. Some firms may be genuine capacity owners: they control scarce infrastructure, have credible customers, can finance the buildout, and can redeploy assets if one demand pool weakens. Others may be payback renters, buying access to expensive capacity because the market rewards visible AI ambition while the underlying workflow has yet to earn the right to consume that capacity at scale.

Size does not decide which one a company is. A large company rents payback when it commits capital without evidence, and a smaller firm can own payback if a narrow workflow produces real cash, defensible data and faster learning than competitors. What matters is whether the AI spend adds strategic degrees of freedom or quietly uses them up, and boards need better vocabulary for telling the two apart. Careful capital spending on AI is mostly a matter of sequencing: buy options before obligations, fund learning before scale, turn vendor enthusiasm into contract terms, tie infrastructure commitments to usage evidence, and keep exit paths open until the workflow has proved it deserves permanence.

Put the CFO ahead of the roadmap

AI strategy usually arrives as a technology roadmap with a finance appendix. For large commitments the order should flip. The chief financial officer's job is to put the work on a payback schedule, which need not slow it down. Before the next approval of any large AI commitment, finance should ask five questions:

1. What minimum revenue, cost reduction, risk reduction or learning milestone justifies the next spend?
2. Which obligations stay fixed if usage disappoints?
3. Which assumptions depend on lower model, chip, power or inference prices?
4. Which counterparties are being treated as proof of demand, and are they end customers or intermediaries?
5. What decision would make us stop, shrink or renegotiate before sunk-cost logic takes over?

The questions are uncomfortable because they ask finance to underwrite a roadmap that has usually been presented as inevitable. A company that cannot underwrite its own AI roadmap has joined a financing chain and is hoping the next participant pays more.

What this means for enterprise buyers

Most readers will not build data centers. They will buy the AI products, cloud services, agent platforms and consulting projects that sit on top of the buildout, and the payback table applies to them too. If a vendor's economics depend on expensive infrastructure, the buyer should ask how that cost will eventually surface. It might come as a shift to usage pricing, throttling of low-margin customers, premium models reserved for higher tiers, or contract minimums that turn experimentation into obligation. Capacity in the buyer's region, with the right privacy and latency profile, may also be scarce just when the workflow becomes important.

A procurement team that treats AI like ordinary software-as-a-service can easily sign a contract whose real economics behave more like a capacity reservation. That arrangement can work if someone prices the option, names the utilization assumption and plans for a cheaper substitute. It goes wrong when nobody takes on those jobs.

A one-page payback table for your largest commitments

This week, take the company's three largest AI commitments, including vendor contracts if you are not building infrastructure, and put each through a one-page payback table:

1. Name the asset: physical capacity, workflow capability, data advantage, customer product, risk control or learning option.
2. Name the obligation: capital expenditure, lease, minimum usage, prepaid credits, headcount, integration work, governance burden or vendor lock-in.
3. Name the payback owner: business unit, vendor, customer, hyperscaler, financing partner or shareholder.
4. Name the proof milestone: live usage, cash yield, retention lift, error reduction, cycle-time compression or pricing power.
5. Name the stop rule: what would force cancellation, renegotiation, downsizing or a move to a cheaper architecture.

Companies that do this will still spend on AI, sometimes heavily, but they will know what kind of spending each commitment is. The table separates strategic infrastructure from theme exposure and customer demand from financing choreography. It also shows which commitments are learning options that can still be dropped and which have already hardened into permanent obligations.