Meta's AI is paying off. The bill is growing faster

Our answer

Yes, the AI spending is paying off, and the place it pays is the ad auction rather than a new business. But in 2026 the bill started growing faster than the thing that pays it. For the first time in this cycle Meta's capital plan no longer fits inside its own cash flow, and most of what it has committed has not yet reached the income statement. The returns are real. They are not yet large enough to carry the bet on their own, and the next two quarters are where that gap either closes or shows.

The returns show up in the price of an ad

The cleanest evidence that the AI works is that advertisers are paying more for each ad while Meta shows more of them. In the second quarter of 2026 revenue rose 28% to $60.80 billion, ad impressions rose 14% and the average price per ad rose 12% ^[1]. The quarter before, revenue had grown 33% ^[5]. Volume and price rising together is the signature of demand, not of a squeeze on a fixed audience.

Management puts the credit on its AI systems. On the July call it said that improved user-understanding models, combined with its GEM ads-ranking model and sequence learning, "generated an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook", and that Advantage+, its automated campaign product, had passed "$75 billion in annual revenue run-rate" ^[3]. Those are the company's own numbers, and the 10-Q is more careful: Meta says it is "unable to quantify the exact impact that each trend had on our advertising revenue" ^[2], and on the call the price gains were also put down to better macro conditions and currency ^[3].

We still read the ranking story as mostly right. Currency and macro can lift what an advertiser pays; they do not produce more clicks and conversions, and clicks and conversions are what management's numbers count. The ad engine is converting compute into revenue.

The bill stopped fitting inside the cash flow

Now the other half. Meta spent $72.22 billion on capital expenditure in 2025 and still had $43.59 billion of free cash flow left over ^[4]. In January it guided 2026 to $115-135 billion ^[4], in April it raised that to $125-145 billion citing "higher component pricing" ^[5], and in July it narrowed it to $130-145 billion ^[1]. Second-quarter capex alone was $31.08 billion ^[1], a little over half of that quarter's revenue.

The cash flow statement shows what that does. Operating cash flow was a healthy $31.86 billion in the quarter, and free cash flow was $784 million, against $8.55 billion a year earlier ^[1]. Meta bought back no stock in the first half ^[1], and in May it sold $25.00 billion of senior notes, taking its notes outstanding to $84.00 billion ^[2].

The first half as a whole still produced $13.17 billion of free cash flow ^[1], so the strain so far is one quarter deep. The plan is what changes the picture. First-half capex was $50.92 billion ^[2], which leaves $79-94 billion for the second half under the July guidance. At the second quarter's pace of operating cash flow, that is a gap of roughly $15-30 billion to be met from cash on hand or borrowing.

None of this is a solvency problem. A company generating more than $30 billion a quarter of operating cash can borrow comfortably. But it changes the nature of the bet. Through 2025 the advertising business paid for the build-out with cash to spare. On the 2026 plan it pays for all it can, and the balance sheet covers the rest.

Most of the cost has not arrived yet

The quarterly income statement understates what has been committed, because a data centre is expensed over years and much of Meta's newest capacity is leased rather than bought.

Depreciation and amortisation was $6,356 million in the second quarter, against $4,342 million a year earlier ^[1], and the servers bought this year have barely started depreciating. Beyond the balance sheet sit leases that have been signed but have not started: $278.99 billion as of 30 June, plus about $68 billion more signed in July that begins in 2027 and 2028 ^[2]. Total non-cancelable contractual commitments were $349.31 billion ^[2].

The joint ventures add a further layer. Meta holds 20% of the venture building its data-centre campus in Louisiana, does not consolidate it, and has given residual value guarantees with a threshold of about $28 billion; its maximum exposure to loss there was $46.03 billion ^[2]. The El Paso venture with BlackRock, announced the day before the second-quarter results, follows the same template: about $14 billion of development costs carried mostly by the partner, a lease back to Meta, and a residual value guarantee of about $13 billion ^[6].

These structures are sensible financing. They are also a way of turning a capital bill into a rental bill, which means the cost reaches the income statement later and lasts longer. Most of the expense of the 2026 build lands in 2027 and beyond.

The promise to watch

The margin pressure is already visible. The Family of Apps segment earned $23.39 billion on $60.37 billion of revenue in the second quarter, against $24.97 billion on $47.15 billion a year earlier ^[1]. That is an operating margin of about 39% against about 53%. The quarter carried $2.40 billion of legal charges and $1.18 billion of severance ^[1], and management said that without them operating income would have risen 9% ^[3]. Even so, a business growing revenue by more than a quarter while its operating profit falls is a business whose costs are growing faster than its sales.

Meta has repeated one promise all year: "We continue to expect to deliver operating income this year that is above 2025 operating income" ^[1]. Operating income was $83.28 billion in 2025 ^[4] and $41.65 billion in the first half of 2026 ^[1], so the second half has to clear about $41.6 billion. Then, on 26 August, Meta said it expects to accrue a legal expense of approximately $10 billion in the third quarter, a charge "not contemplated in the expense range" it gave in July ^[7]. That release did not repeat the operating-income promise; it said only that the other ranges given in July remain unchanged ^[7]. With the accrual, the second half needs roughly $10 billion more than the plan Meta was guiding to in July.

The case for spending anyway

The bull case is not foolish, and Meta makes it well. Its finance chief says the company is "supply constrained", with "numerous ROI positive places that we would put compute toward if we had it", and Zuckerberg says Meta is getting "offers for compute at a significant premium over what we paid for it" ^[3]. Last October he set out the downside as he sees it: the worst case is that Meta has "prebuilt for a couple of years", with "some loss and depreciation", before growing into the capacity ^[8].

That framing is fair as far as it goes. It treats the build-out as one bet. We think it is two. The first is compute for ranking and recommendations, which the ad numbers above already justify. The second is the push toward frontier models and superintelligence, which drives much of the growth in 2026 spending and has produced, so far, no revenue line that can be measured against it.

Our verdict

On balance, we think the core AI spending has earned its keep and the frontier spending has not yet had to. Ad prices rising 12% on rising volume is as good evidence of a return as a business that sells no AI directly can show. But the 2026 plan moves Meta from funding its build-out out of surplus to a second half that, on its own guidance, needs cash reserves or borrowing on top of all its operating cash flow, with several hundred billion dollars of leases still to arrive on the income statement. The returns are real. The bill is simply larger, and it is growing faster.

What would change our view, in either direction:

• Free cash flow in the second half. A return to several billion dollars a quarter would say the first half was the peak of the strain. Another quarter near zero would say it is the new normal.
• The operating-income promise. Beating 2025 despite the $10 billion accrual would show the ad engine absorbing the bill. Missing it, or quietly dropping the line, would be the first time the bill clearly won.
• The 2027 capex number. Meta has given none. A figure growing faster than revenue again would mean the frontier bet is still compounding its cost.
• Price per ad. The whole case rests on it. If the gains fade while depreciation keeps growing by more than 40% a year, the arithmetic turns against the build-out quickly.

---

Cover photo: "Lulea Facebook datacenter building" by Christopher Down, CC BY 4.0, via Wikimedia Commons; cropped, resized and colour-graded.