Meta’s $347B AI Leases Help Hide a $3 Trillion Bill

Big Tech’s AI bill is roughly $3 trillion bigger than the capex headlines suggest. Much of it sits in future promises that look harmless—until demand, margins or power costs turn.

The Wall Street Journal’s latest analysis puts the AI infrastructure commitments sitting beyond the obvious headline numbers at roughly $3 trillion. That does not mean $3 trillion of secret debt. It does mean investors, founders and operators who only look at quarterly capital expenditure are staring at the small end of the elephant.

The core story: AI has become a promises business

On August 17, 2026, the Wall Street Journal laid out the uncomfortable bit of the AI spending story: the hyperscalers are not just buying chips, servers and data centres with cash already flowing through their income statements. They are signing enormous future commitments for power, equipment, cloud capacity and leases that have not yet begun.

That matters because an expense is easy to cheer when it is described as “investment.” A long-term obligation is harder to cheer when demand slows, margins get squeezed or a better technology makes last year’s data centre look like a very expensive Blockbuster store.

Meta offers the cleanest example. As of June 30, it disclosed roughly $278.99 billion in leases that had not yet commenced. In July, it added another $68 billion in data-centre leases expected to begin in 2027 and 2028. Add those figures and you get about $347 billion in future lease obligations before a good portion of the assets are even operating.

Meta also reported $349.31 billion in non-cancelable contractual commitments at June 30, mostly tied to third-party cloud capacity, servers, network infrastructure, data centres and Reality Labs hardware. Do not lazily add every big number in a filing together; categories can overlap. But do not ignore them either. The direction is obvious: the AI race has moved from software competition to industrial-scale procurement.

Alphabet is playing the same game at a different scale. Its June-quarter filing reported $811 billion in material purchase commitments and other contractual obligations, including technical infrastructure, inventory and energy arrangements for data-centre use. Of that, $200.7 billion was short-term. Its energy commitments can run for as long as 26 years, through 2054.

That is not a company testing AI with a few clever engineers and a corporate credit card. That is a company laying concrete before it knows precisely what every future tenant will pay.

Why the balance sheet looks cleaner than the real bet

Here is the important distinction: off-balance-sheet does not automatically mean dodgy. Accounting rules have a timing logic. A lease generally hits the balance sheet when it starts. Purchase commitments are often disclosed in notes rather than booked as full liabilities because the product or service has not yet been delivered.

No conspiracy required.

But accounting treatment and economic reality are not the same thing. If a company has promised to take power, lease a facility, buy capacity or guarantee financing for a data-centre build, management has made a commercial bet. The cash may not have left yet. The risk has.

This is precisely why the phrase “Big Tech has fortress balance sheets” has become less useful. Yes, Alphabet, Meta, Amazon and Microsoft are genuine cash machines. They are not 1999-era minnows with a sock puppet and no revenue.

But a fortress can still make commitments that turn it into a very expensive warehouse if its assumptions prove wrong.

The present AI arms race rests on several assumptions happening together:

- Demand for AI products keeps rising hard enough to absorb all this compute. - Customers tolerate prices high enough to cover chips, depreciation, power, networking, cooling, financing and staff. - New models do not become so efficient that yesterday’s planned capacity is suddenly excessive. - Governments and communities keep allowing rapid data-centre expansion. - The companies building capacity remain willing and able to honour their commitments.

That is a hell of a parlay.

The second-order problem: the real constraint is no longer chips

For years, the obvious bottleneck was Nvidia chips. It is still a bottleneck, but it is no longer the whole game.

AI now needs land, transmission lines, substations, natural gas, renewables, water, construction crews, networking gear, cooling systems and finance. You cannot download any of those things from an app store.

This changes who has power in the market. The winning AI businesses may not simply be the ones with the cleverest model. They may be the ones with the best access to electricity, the most credible infrastructure partners and the balance sheet to sign commitments that scare everybody else away.

It also creates a political problem. Bloomberg reported in May that Maryland’s consumer advocate alleged residents could pay an extra $1.6 billion over a decade for grid projects driven primarily by data-centre demand outside the state. Whether that exact allocation survives regulatory challenge is almost beside the point. Voters do not care about the elegance of a transmission-cost formula when their power bill goes up.

The moment ordinary households believe AI is making their electricity more expensive, data centres stop being symbols of progress and become extremely large, noisy political targets.

That should worry founders who believe infrastructure is someone else’s issue. It is not. Regulation, energy pricing and community backlash can become product constraints faster than a competitor can copy your feature set.

The contrarian view: this may be rational — and still dangerous

The lazy take is that every large AI commitment proves there is a bubble. That is rubbish.

The internet was real. Mobile was real. Cloud computing was real. All of them created absurd valuations, failed businesses and investors who confused a genuine technology shift with a guarantee that every price was sensible.

AI can be the most important commercial technology in decades and still produce horrific returns for buyers who pay too much for the wrong infrastructure, at the wrong time, with the wrong contracts.

In fact, the scale of these commitments may be rational for the largest players. If demand keeps compounding and AI becomes basic infrastructure for search, advertising, cloud software, coding, customer support, entertainment and enterprise operations, being short of compute could cost more than overbuilding it.

The overlooked risk is not necessarily that Meta or Alphabet run out of money. It is that the return on each additional dollar gets worse.

That is how capital cycles usually hurt people. The first assets are scarce and highly profitable. Then everyone builds. The best customers get served. Prices fall. Suppliers win for longer than expected. Asset owners discover that “strategic capacity” is another name for a lower return on capital.

For investors, that means owning the story is not the same as owning a good price. For operators, it means renting AI capacity from a giant with a $3 trillion pile of future promises may be far safer than trying to mimic their spending spree yourself.

What this means for you

If you are a founder, stop saying you have an AI strategy unless you can answer one boring question: what is the unit economics after inference, human review, support, security and customer acquisition?

Do not build a business whose only advantage is access to a model API. Models improve, prices move and platforms change the rules. Build proprietary workflow, customer trust, distribution, data rights or a genuinely painful operational problem solved end to end.

If you are an operator, treat AI spend like headcount or inventory: measure it. For every deployment, track time saved, error rates, revenue gained, retention changed and the fully loaded cost per useful outcome. “Our team uses AI heaps” is not a metric. It is a confession that nobody is managing it.

If you are an investor or saver, read the commitments footnote, not just the quarterly capex headline. Look for future leases, purchase obligations, guarantees, power deals and debt raised to fund infrastructure. Then ask the simple question Wall Street often avoids: what revenue has to materialise for this commitment to earn an acceptable return?

And if the answer is vague, do not call it conviction. Call it what it is: hope with a ticker symbol.

AI is going to make fortunes. It will also punish anyone who mistakes a giant future promise for a current profit. Those are not the same thing, mate.

Sources

• Why Big Tech’s AI Spending Is $3 Trillion Higher Than It Seems — The Wall Street Journal
• Meta Platforms Q2 2026 Form 10-Q — lease obligations and commitments
• Alphabet Q2 2026 Form 10-Q — contractual commitments and energy agreements
• AI Push to Add $1.6 Billion to Maryland Power Bills, State Says — Bloomberg