Meta's quarterly report for the three months ended June 30 shows two numbers about leases, a few pages apart. On the balance sheet, the company owes $28.65 billion in operating lease liabilities: $2.43 billion due within a year and $26.23 billion after that. In the notes, a single sentence adds that Meta has also signed leases that have not yet begun, mostly for data centers and colocation space, worth about $278.99 billion. They start between now and 2036 and run for as long as 30 years.1 The second number is nearly ten times the first, and it does not appear on the balance sheet at all.
Meta is not unusual. On August 17, coverage of a Wall Street Journal analysis of the footnotes of nine companies (Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX and AMD) put their off-balance-sheet commitments at about $3 trillion, mostly tied to AI. Roughly $1.2 trillion was leases not yet commenced and $1.9 trillion was commitments to buy chips and other hardware, against about $600 billion of capital spending over the past year.23 Alphabet alone reported $811 billion of purchase and contractual commitments at the end of June, up from $332 billion three months earlier.4
The useful distinction here is between an obligation that has been booked and one that has been signed. The balance sheet answers the question "what has started?" The notes answer a different one: "what has already been decided?" For a company in the middle of a buildout, the second answer is larger and says more about where its cash will go. That is not a claim that anything is hidden, or improper, or that these companies have promised more than they can pay. The rules require the disclosure, and the numbers are printed. The claim is only that a reader who stops at the balance sheet is reading last year's decisions and missing this year's.
The balance sheet carries the leases that have started.
The notes carry what has been promised: nearly ten times as much in leases alone.
Why the rule waits for the keys
Accounting has a good reason for the gap. A lease puts a liability on the balance sheet when it commences, which is the day the tenant gets the use of the building. Before that there is no asset in hand, only a contract, so the obligation goes into the notes. Meta's sentence is written exactly that way: these are leases "in addition to the lease liabilities that are included on our balance sheet."1 Purchase commitments follow the same logic. The liability arrives with the goods.
For most companies, most of the time, the gap is small, because leases are signed a few months before the keys change hands. A buildout stretches it. A data center campus can be agreed years before its first rack is powered, and while it is being built the promise is real, specific and, in most cases, binding, but it sits outside the statement most readers look at.
We have been here before
Off-balance-sheet leases were the subject of a long fight in accounting. In 2005, after Enron, the staff of the Securities and Exchange Commission estimated that American public companies had about $1.25 trillion of non-cancelable operating lease obligations that appeared only in the notes, an undiscounted figure.6 In 2016 the International Accounting Standards Board counted US$2.86 trillion of such commitments disclosed by listed companies around the world, "almost US$3 trillion," and issued a standard, effective in 2019, that brought leases onto the balance sheet. The American standard setter was finishing a similar rule at the same time.7
The reform fixed the leases that have started. It was never meant to reach the ones that have not, and that is where a buildout lives. The irony is hard to miss: the world's listed companies once kept almost $3 trillion of leases in their footnotes, and nine companies now keep about the same amount there, much of it in a form the reform did not cover.
Three kinds of promise
A signed obligation is not one thing. Meta's filing shows three kinds, and they turn into cash in different ways.
The first is a lease that has not begun: a fixed schedule of payments that will start on a known date. The second is a non-cancelable purchase commitment, $349.31 billion in Meta's case, for goods and services the company has agreed to take.1 The third is a guarantee. For a Louisiana data center venture in which it holds a 20 percent interest, Meta has provided residual value guarantees with a threshold of about $28 billion. It judges a payment not probable, so it has recorded no liability, and it puts its maximum exposure to loss from the venture at $46.03 billion.1 A lease costs money if things go to plan. A guarantee costs money only if they don't.
The question that matters for each is not its size but its behavior in a bad year. Can it be delayed, cut or walked away from, and at what price? A reader cannot answer that from a single total, which is why adding all three into one headline number, as the coverage did, says more about scale than about risk.
A balance sheet that grows by itself
Timing creates an odd effect, which the model in Figure 2 shows. Suppose a company signs every lease it will need in year zero, and they start in tranches over the next five years. Nothing new is decided after the first day. Yet the balance sheet shows lease liabilities climbing year after year, as each tranche begins, and a reader who watches only the booked line sees a company taking on obligations in year three. It took them on in year zero. The decision and the recording are years apart.
The same gap inside a company
The pattern is not confined to giants. Any technology organization signs promises that its budget shows only as they bill: a cloud agreement that commits a minimum spend over three years, a multi-year software subscription, colocation space reserved for next spring. The budget records this year's run rate. The contract records the decision. When a board reviews technology spending by looking at the year's budget, it is looking at the booked column and approving the signed one without seeing it.
The remedy is clerical, not clever: keep a register of commitments beside the budget, with three columns. What is booked. What is signed but not started, with its start date. What is contingent, such as a guarantee or a minimum-spend shortfall, with the event that would trigger it.
Objections, considered
Nothing here is hidden. Correct, and that is the point. The figures are disclosed in plain sentences. The problem is where readers look and what they add up, and a ratio such as debt to assets, computed from the balance sheet alone, leaves the larger column out.
The promises are matched by customers' promises. For sellers of capacity, partly so. Alphabet reported $519.5 billion of revenue backlog at the end of June, most of it at Google Cloud.5 But a commitment to pay and a commitment to be paid are not mirror images: they differ in who can walk away, on what terms and when. The match has to be shown contract by contract, not assumed from two totals.
Booking signed leases now would be worse. Probably. Recording liabilities for buildings that do not yet exist would distort the balance sheet in the other direction. The answer is not a new rule. It is reading both columns.
The balance sheet tells you what has started; the notes tell you what has been decided.
Before the next signature
- List what has not started. What have we signed whose payments, deliveries or leases begin later, and when?
- Trace each promise to cash. When does it become a payment, and can it be delayed, reduced or cancelled, at what cost?
- Separate the contingent. Which promises cost money only if something goes wrong, and what is that something?
- Review at signing. Does the board see a commitment when it is made, or only when it reaches the balance sheet?
Back to Meta's two numbers: $28.65 billion on the balance sheet, $278.99 billion in the notes. The first is what the company has started. The second is what it has already decided, and over the next ten years it will move, tranche by tranche, from the notes to the balance sheet. It is only news to readers who wait for it to arrive.