Summary
- Meta reported $278.99 billion of operating and finance lease obligations that had not commenced at 30 June 2026, covering data centres, colocation and certain network infrastructure.
- The comparable figure was $182.88 billion at 31 March, implying a quarterly increase of about $96.11 billion, or 52.55%.
- The leases are scheduled to start from the remainder of 2026 through 2036, with terms ranging from more than one year to 30 years.
- Meta separately entered into about $68 billion of additional data-centre leases in July, expected to commence in 2027 and 2028 for terms of 18 to 20 years.
- Uncommenced leases are not current debt, cash already spent, current lease liabilities or proof that the contracted capacity is operational.
- The investment case now depends on delivery, power, utilisation, unit economics and revenue measures that the filing does not provide contract by contract.
A contract can arrive years before its balance-sheet liability
The most important word in Meta’s disclosure is not “billion”. It is “commenced”. A company may sign an agreement for a data-centre building, a colocation hall or network infrastructure before the asset is ready for use. Until the agreed service starts and accounting control tests are met, the future payments can remain an uncommenced lease obligation rather than a lease liability recognised on the balance sheet.
That timing explains why the $278.99 billion figure cannot be treated as if Meta borrowed the amount on 30 June. Nor is it evidence that the company handed over that sum during the quarter. The obligation is an undiscounted contractual payment stack extending across leases with different start dates and durations. When each lease commences, accounting recognition and cash payment schedules will follow their own terms.
The separation is economically useful even though the categories must remain distinct. It shows how much infrastructure ambition sits beyond the assets and liabilities visible today. Meta can reserve future capacity while landlords, developers and equipment suppliers construct it. The company gains a claim on supply; the counterparty gains a long-dated customer commitment that can support financing.
The same separation creates delayed visibility. Investors see a large aggregate, but not a site-by-site bridge between signature, construction, energisation, service availability and lease recognition. The contractual decision is real before the operating result is observable.
The quarterly change is the signal, not a new fleet already online
At 31 March, Meta reported $182.88 billion of uncommenced lease obligations. Three months later the figure was $278.99 billion. The increase of roughly $96.11 billion, or 52.55%, records a rapid expansion in future contracted payments.
It does not mean Meta added $96.11 billion of operating data centres in one quarter. Some contracts may relate to facilities that will not commence for years. The filing says the June-quarter stack begins across a period running from the remainder of 2026 through 2036. Lease terms range from more than one year to 30 years.
Those long horizons are part of the strategy. AI infrastructure requires land, utility connections, substations, cooling systems, fibre, servers and skilled construction labour. Waiting until demand is fully visible can leave a buyer behind competitors in power queues or equipment allocations. Contracting early turns uncertainty about supply into a scheduled option on capacity.
But the option is not free. A long lease fixes obligations before the useful output of future chips, models and products is known. Accelerator performance can improve faster than a building depreciates. Model architectures can alter networking or cooling requirements. A facility well matched to one generation of hardware may need costly adaptation for the next.
The quarter-to-quarter rise therefore measures commitment velocity. It says more about Meta’s willingness to secure a future estate than about capacity already delivering inference or advertising revenue.
July adds acceleration, not a number to paste onto June
Meta says it entered into approximately $68 billion of additional data-centre leases in July. These are expected to commence in 2027 and 2028 and run for 18 to 20 years.
The disclosure is material because it shows that contracting continued after the quarter closed. It also supplies a narrower timing range than the aggregate June figure. Yet the two numbers have different measurement dates. The $278.99 billion is the position at 30 June. The $68 billion describes contracts entered into during July.
Adding them to produce a supposed audited total of $346.99 billion would erase that boundary. The filing does not present such a combined balance. It does not say whether later changes, offsets, currency effects or other classification movements should be included alongside the July additions. The responsible conclusion is directional: the commitment stack was still growing.
The July terms also reveal how far the commercial wager extends. An 18-to-20-year agreement beginning in 2027 or 2028 can persist well beyond the current accelerator generation and today’s product cycle. Meta is not merely placing an equipment order. It is selecting part of its cost base for the 2030s.
That may be rational if the company expects infrastructure scarcity to endure and its products to absorb the capacity. It becomes expensive insurance if supply loosens, external prices fall or demand shifts.
Securing capacity transfers one risk and accepts four others
The clearest benefit is availability. A hyperscaler that controls a long queue of suitable sites is less exposed to a future scramble for power and construction. Contracts can also give developers enough revenue certainty to finance campuses whose upfront costs would otherwise be difficult to bear.
The first accepted risk is delivery. A signed lease does not energise a building. Grid interconnections can slip, turbines and transformers can arrive late, permits can be challenged and cooling designs can change. If a service start is delayed, the economic effect depends on contract protections that Meta does not disclose in aggregate.
The second is specification. Meta must match facilities contracted years ago with evolving server densities, networking topologies and heat loads. A lease can secure square metres without guaranteeing that the space supports the most valuable future workload at the best cost.
The third is concentration. Long commitments create dependence on developers, landlords, utilities and cloud-capacity providers. Meta reports $349.31 billion of non-cancelable contractual commitments across several categories, mostly third-party cloud capacity and investments in technical infrastructure, data centres and Reality Labs hardware. This figure is not identical to the lease total, but it confirms that outside counterparties are important to the buildout.
The fourth is demand. Capacity earns a return only when products use it productively. An empty hall can be technically available and economically unproductive. A busy hall can also destroy value if its power, financing and service costs exceed the revenue or cost savings generated by the workloads inside it.
Capex, commitments and debt answer different questions
Meta’s second-quarter capital expenditure, including principal payments on finance leases, was $31.08 billion. Its full-year outlook is $130 billion to $145 billion. Those figures measure current-period investment and expected annual outlay. They should not be added to, or substituted for, undiscounted payments under leases that have not begun.
Long-term debt was $83.66 billion at 30 June. Cash, cash equivalents and marketable securities were $90.26 billion. Those balance-sheet figures show current financing resources and obligations. They do not cancel the future lease stack, and the lease stack should not be relabelled debt simply because both require future cash.
The distinctions help readers ask better questions. Capital expenditure reveals how quickly Meta is deploying owned equipment and finance-leased assets now. Uncommenced leases show capacity contracted for later. Non-cancelable commitments widen the lens to cloud, servers, network equipment, data centres and hardware. Debt shows borrowed principal already recognised.
Cash flow supplies another boundary. Meta generated only $784 million of free cash flow in the quarter after a very large investment outlay. One quarter is not a funding forecast, but it shows why timing matters. Future lease commencements will layer payments and liabilities onto a business that is already converting a large share of operating cash into infrastructure.
No single headline number measures the whole programme. Combining them would exaggerate exposure in some places and hide timing in others.
The missing unit is useful compute, not contracted dollars
The filing does not state how many megawatts, accelerators or data-centre halls sit behind the $278.99 billion. It does not identify every country, supplier, price escalation clause or workload. It does not give expected utilisation or revenue per contracted unit.
Without those denominators, the total proves scale but not efficiency. A company could sign a larger commitment because it secured more capacity, paid a higher price, accepted longer terms or moved from ownership to leasing. Each explanation has a different implication for returns.
The operating scorecard should begin with commencement. How much contracted capacity starts service each quarter, and how much slips? It should then measure usable power, installed accelerators, network availability and time to production workloads.
Utilisation must be defined tightly. Booking a GPU is not the same as completing a useful inference, training or ranking task. Meta should connect infrastructure consumption to accepted work, latency, reliability and cost per result. Improvements in model efficiency should reduce the compute required for a unit of product value, even if total demand keeps growing.
The financial scorecard needs annual lease payments, capitalised assets, depreciation, external cloud spend and revenue or savings attributable to the capacity. A return calculation requires both the cost of reserved supply and the economic output it enables.
The commitment stack is also a forecast of organisational confidence
Meta serves 3.60 billion daily active people across its family of products, giving it an enormous base over which infrastructure can be used. Advertising can fund projects that have not yet produced direct AI revenue. That scale distinguishes Meta from a start-up signing capacity against a narrow customer book.
It does not eliminate forecasting error. The company is making overlapping bets: that engagement and advertising remain strong, that new AI products create value, that future models need the contracted infrastructure and that Meta can finance the intervening years. Weakness in one part can make the others more burdensome.
Long contracts may also shape product decisions. Once capacity is committed, managers have an incentive to find workloads that consume it. This can accelerate useful deployment, but it can also confuse utilisation with value. Filling servers is not a business objective if the resulting product does not retain users, improve advertising or generate revenue.
The filing therefore documents more than an accounting footnote. It records a governance decision to lock in infrastructure across several technology cycles. The board and investors will need evidence that project selection, supplier exposure and demand forecasts are being revised as conditions change.
The next disclosure should turn the queue into a bridge
Readers need a reconciliation from one quarter to the next: new contracts, commencements, cancellations, remeasurements and other changes. Meta’s aggregate figures allow the size of the rise to be calculated, but do not explain every movement.
A maturity view would show when undiscounted obligations become annual cash demands. A capacity view would connect start dates to megawatts, server readiness and expected workloads. A supplier view would expose concentration without revealing commercially sensitive site details. A return view would compare total infrastructure cost with incremental revenue, gross profit or measurable cost avoidance.
The July disclosure creates an early test. Contracts expected to commence in 2027 and 2028 should later appear as identifiable additions to operating capacity and lease liabilities. Delays, redesigns or unused supply should be visible rather than absorbed into a larger aggregate.
Meta has secured an infrastructure queue of extraordinary scale. That may be the price of ensuring that power and buildings are available when models need them. It may also be a costly way to discover that contracted supply and productive demand mature on different calendars. The difference will not be settled by the size of the commitment. It will be settled by what starts, what runs and what each completed unit of work earns.