Guide · for directors

AI oversight for boards: twelve questions to ask management

Boards are approving the largest technology investments most companies have ever made, often with little more than a strategy deck to go on. These twelve questions come from what the biggest AI spenders disclose in their SEC filings: where the money goes, what has been promised, who the company depends on and whether the story matches the numbers.

How to use this guide

The questions are grouped in four parts: spending, commitments, dependence and disclosure. Each comes with why it matters, an example of what public companies’ filings show (computed from current SEC data, so the figures stay up to date) and what a good answer from management sounds like. None of it requires technical knowledge of AI; it asks the questions a board already knows how to ask of any large investment, applied to one that moves unusually fast.

The examples are drawn from the largest AI spenders because their filings are the most detailed record of how AI investment behaves at scale. The same questions apply to a company buying AI services, building products on models from others, or supplying the build-out.

Spending: is the investment sized to what it can earn?

Most AI spending is capital spending, and capital spending is easy to approve and hard to reverse. These questions test whether the plan is tied to a return.

01What share of our operating cash flow is going to AI, and what would we do if that cash flow fell?
Why it matters.
Spending funded from cash flow depends only on the business; spending that needs borrowing also depends on lenders. A board should know which kind it is approving.
What the filings show.
The five largest buyers spent 97% of their operating cash flow on capex in Q2 2026, against about a third in 2023. Amazon and Oracle now spend more than they generate. See is AI a bubble?.
A good answer.
A specific share, a ceiling the board has agreed, and a plan for what gets cut first if cash flow drops.
02What revenue or savings does each major AI investment have to produce, and by when?
Why it matters.
Without a target set in advance, every outcome can be described as progress. A target turns AI spending into something that can be reviewed.
What the filings show.
The hyperscalers’ capex rose 87% on a year earlier while their core cloud revenue rose 38%. Our capex payback calculator shows how much new revenue a year of spending needs.
A good answer.
A number, a date and an owner for each large commitment, reported back to the board on a schedule.
03How long are we depreciating AI hardware over, and is that how long it will stay useful?
Why it matters.
Depreciation life is an accounting choice that moves reported profit. Longer lives flatter earnings; if the hardware becomes obsolete sooner, the write-down arrives later and larger.
What the filings show.
Most hyperscalers lengthened server lives between 2022 and 2025. Amazon has since shortened its server life from 6 to 5 years. The useful-life calculator shows what one year of difference does to profit.
A good answer.
A life justified by how long comparable hardware has actually stayed in productive use, reviewed as new generations arrive.

Commitments: what have we promised beyond the budget?

The largest AI obligations often never pass through the capex line. Leases, cloud contracts and supply agreements commit a company for years.

04What leases, cloud contracts and purchase commitments have we signed that are not yet on the balance sheet?
Why it matters.
Commitments signed today become costs later, whether or not the demand they were signed for arrives.
What the filings show.
Five companies on our map disclose $1.11T of data-center leases signed but not yet started; Meta reports hundreds of billions of dollars of non-cancelable commitments, mostly for third-party cloud capacity.
A good answer.
A single schedule of all multi-year AI commitments, by year, that the board sees alongside the capital budget.
05If demand for what we are building came in at half the plan, what could we cancel, and what would it cost?
Why it matters.
Flexibility has a value that is easy to give away in a hurry to secure capacity. Knowing the exit terms before signing is cheaper than discovering them later.
What the filings show.
Nvidia’s filings show what happens when demand for a product falls after supply was committed: a $4.5 billion charge for excess H20 inventory and purchase obligations.
A good answer.
Exit terms for each major commitment, and a downside case the board has reviewed.
06Do we have the power, and the grid connection, that our AI plans assume?
Why it matters.
Chips arrive in months; power can take years. Plans that assume capacity before the power is secured are plans with an unpriced delay.
What the filings show.
The largest buyers now sign 20-year power contracts: Microsoft for a restarted Constellation nuclear plant, Meta with Vistra for 2,609 megawatts. See AI data center.
A good answer.
A dated power plan for each site, with contracted rather than assumed supply.

Dependence: who are we relying on?

AI concentrates dependence on a few suppliers and, for many companies, a few customers. Concentration is not a problem until one of them changes terms.

07Which single supplier, model provider or cloud would hurt us most if it changed prices or terms?
Why it matters.
A product built on one model provider or one chip supplier inherits that supplier’s pricing power. The question is what switching would cost and how long it would take.
What the filings show.
Nvidia names TSMC and Samsung as its foundries and SK Hynix, Micron and Samsung as its memory suppliers, and it has committed $279 billion to suppliers to secure capacity. Even the largest buyer of all depends on a few suppliers. See Nvidia’s customers and suppliers.
A good answer.
A named list of critical dependencies, the cost and time to switch each, and who owns that risk.
08How concentrated is the AI revenue we expect, and how creditworthy are those customers?
Why it matters.
A backlog is only as strong as the customers behind it. When a few customers are themselves funded by investors, their funding becomes your risk.
What the filings show.
Oracle’s remaining performance obligations are $664B, 9.3 times its annual revenue, with about 13% expected within a year. CoreWeave reports that Microsoft supplied 67% of its 2025 revenue. See RPO.
A good answer.
Concentration by customer, the share of contracted revenue due within a year, and what the company would do if the largest customer slowed.
09Are we financing any of our own customers, or being financed by any of our suppliers?
Why it matters.
Vendor financing is normal, but revenue a seller has helped finance is lower-quality revenue, and support from a supplier usually comes with strings.
What the filings show.
Our money-loop map lists 13 equity investments, credit lines, guarantees and customer warrants between AI suppliers and their customers, from Amazon’s facility for Anthropic to AMD’s warrants to OpenAI and Meta. See vendor financing.
A good answer.
A full list of financing given or received in connection with AI purchases or sales, with the board’s view of each.

Disclosure: does what we say match what we report?

Investors, lenders and regulators increasingly compare what companies say about AI with what their numbers show. Boards should make the comparison first.

10Where in our reported numbers will shareholders be able to see AI working?
Why it matters.
If AI is central to the strategy, it should eventually be visible in revenue, margins or costs. If it can’t be seen anywhere, investors will draw their own conclusions.
What the filings show.
Alphabet reports Google Cloud’s operating income and, since 2026, revenue from selling TPU systems; Microsoft discloses its revenue from OpenAI because OpenAI is a related party. Companies that want credit for AI increasingly report it.
A good answer.
A specific line, segment or metric where results will show, agreed before the claims are made.
11Are our public statements about AI specific and measurable, or general?
Why it matters.
Specific claims can be checked against results; general enthusiasm can’t, and it invites skepticism when results lag.
What the filings show.
On every company page, we count how often a company’s earnings releases mention AI and how many of those sentences carry a number. Some of the largest companies talk about AI far more often than they quantify it.
A good answer.
Fewer, more specific statements, each tied to a reported number, and a review of the language before it is published.
12Who on the board can read the AI parts of our filings and our suppliers’ filings critically?
Why it matters.
AI changes cost structures, financing and risk in ways that traditional technology oversight was not designed for. A board needs at least one member, or an adviser, who can test management’s answers.
What the filings show.
Everything in this guide comes from public filings, which are available for every competitor, supplier and customer. The same reading can be done for any company: Nvidia, Microsoft or any of the others on the map.
A good answer.
A named director or adviser responsible for AI oversight, with time on the agenda at least twice a year.

What to do with the answers

Most boards will find that management can answer some of these questions precisely and others not at all. The gaps are the useful part: a commitment no one has scheduled, a dependency no one owns, a claim no number supports. Put the unanswered questions on the next agenda with an owner and a date, and ask for the AI commitments schedule and the payback targets to come back to the board at least twice a year.

If it would help to work through these with your board or investment committee, we run briefings built on the same filings research: where your company sits in the AI stack, how exposed it is to the capex cycle and its financing, and the signals worth monitoring. See work with us, or read the AI stack map and is AI a bubble? for the market context behind these questions.

Ask about a board briefing

Free · quarterly · from SEC filings

After each Big Tech earnings season we update the map from the new filings and publish what changed: who spent, who got paid, who kept it and what the filings reveal about the loop. Free, in Rob May’s Investing in AI newsletter.

Form not showing? Subscribe on Substack. No spam; unsubscribe any time.

Questions about AI oversight for boards

What questions should a board ask about AI?

Start with money and commitments: what share of cash flow AI is taking, what each investment must produce and by when, what has been committed beyond the budget, and how long AI hardware is depreciated over. Then dependence (critical suppliers, concentrated customers, financing) and disclosure (where AI will show in reported results).

Who is this guide for?

Directors, audit and risk committee members, investment committees and executives preparing for board discussions about AI spending. It applies to companies that build AI infrastructure, buy it, or build products on top of it.

Is this legal or governance advice?

No. It is a set of questions drawn from what public companies disclose in their SEC filings. Directors should take their own legal and governance advice on their duties.

Why use public filings for a board guide?

Because they show how the largest AI spenders actually behave, with numbers they are accountable for, and because the same reading can be applied to any competitor, supplier or customer. Every example here links to its source.

How often are the examples updated?

The figures are computed from the same SEC data as the rest of the site, so they update when the data is refreshed after each earnings season.

Can you brief our board on this?

Yes. We run board and investment-committee briefings on AI spending, financing and oversight, with pre-reading and a written summary. Use the inquiry form on the work-with-us page.

Research and education, not investment advice: nothing here is a recommendation to buy or sell any security. Figures come from companies’ SEC filings as reported; companies named were chosen because they show how the stack works, not as a view on their shares.