A Board Briefing on AI Capital Expenditure

Summary. Boards are being asked to approve AI infrastructure budgets that are large, long-lived and hard to compare with earlier technology spending. This dossier offers a short decision framework in the form of five questions directors can put to management before approving the money. It is an illustrative analytical scenario, designed to structure a conversation rather than to prescribe an outcome.

The material draws on general public knowledge about infrastructure and technology investment. It does not describe any specific company.

The signal

Spending on compute, data centres and the power to run them has grown into one of the largest categories of corporate capital allocation. The pressure to commit is real: competitors are investing, vendors present urgent timelines and the technology is advancing quickly.

Urgency is exactly when governance matters most. The risk is not that the board says yes or no; it is that the decision is made on a narrative rather than on explicit assumptions that can later be checked.

Question one: what business outcome does this buy?

Ask for the specific products, cost reductions or revenue lines the investment supports, and the metric that will show whether it worked. A request framed as capability, such as being ready for AI, is hard to test. A request tied to named workloads, expected usage and a measurable result is much easier to review.

Directors can also ask what the company would do if it spent nothing. The honest answer defines the true option value of the project.

Question two: build, rent or partner?

The same capability can be owned, leased from a cloud provider, accessed through a model vendor or shared with a partner. Each choice shifts risk between capital cost, operating cost, control and flexibility.

• Owning gives control and potentially lower unit cost at high utilization, but it ties up capital and exposes the company to hardware that ages quickly.
• Renting preserves flexibility but can create dependency and variable bills.
• Partnering can share cost and expertise, at the price of shared governance.

Management should show the comparison, not only the preferred option.

Question three: what are the physical constraints?

AI infrastructure depends on things that cannot be purchased instantly: grid connections, electrical equipment, cooling, land and skilled construction labor. Lead times for some items stretch to years. A budget that assumes delivery on an optimistic timeline can leave expensive hardware idle or force costly workarounds.

Directors should ask which dependencies sit outside the company's control, what the schedule looks like if each slips, and who carries the cost of delay.

Question four: how does the economics change if assumptions move?

Any large AI plan rests on a handful of assumptions: utilization rates, hardware prices, energy costs, model efficiency and demand. Ask management to show sensitivity, not just a base case. In particular, test the effect of lower utilization than planned and of faster hardware obsolescence.

Efficiency gains in software can reduce the compute needed for a given result. That is good for operating cost and potentially bad for a business case built on scarce capacity.

Question five: how will we stage, review and, if needed, stop?

Large commitments are rarely all-or-nothing. Staged funding with defined checkpoints lets the board release capital as evidence accumulates. Agree beforehand what results would justify the next tranche and what results would justify pausing.

Equally important is ownership. A single accountable executive, a regular report to the board and an independent view on progress reduce the chance that sunk costs drive later decisions.

A useful discipline is to require a one-page summary of these five answers with every major request. If management cannot complete the page, the proposal is probably not ready for a vote, and sending it back is a legitimate and inexpensive outcome.

Scenarios

Base case. The company approves a staged budget tied to named workloads, rents some capacity and builds only where utilization is clearly high. Results arrive unevenly, and the board adjusts tranches accordingly.

Upside case. Demand grows faster than planned and early commitments to power and equipment prove prescient, giving the company a cost advantage.

Downside case. Demand disappoints or technology shifts, leaving capacity underused. A staged structure limits the loss; a single large commitment magnifies it.

What to watch

• Whether management can state the business metric the investment is meant to move.
• Utilization data from the first phase compared with the plan.
• Changes in power, equipment and construction lead times that affect the schedule.
• Movement in the unit cost of renting comparable capacity.
• Evidence that efficiency gains are reducing the compute a given workload needs.

This dossier is an illustrative analytical scenario built from general public knowledge. It is analysis, not a recommendation.