The Capacity Commitments Your Capex Plan Cannot See
Which futures become unavailable when we commit to it?
Executive Summary
AI infrastructure commitments are becoming harder to see at the moment when they are easiest to change.
Annual capital plans still matter, but they no longer define the full exposure. Capacity is now secured through a mixture of owned assets, finance and operating leases, non-cancellable supplier agreements, dedicated cloud capacity, joint ventures, residual-value guarantees, land, power and connectivity arrangements. Each form can be rational. Together, they can remove strategic option value years before the corresponding compute is used.
The public evidence is clearest among hyperscalers and other AI-intensive infrastructure operators. Meta reported approximately $182.88 billion of leases not yet commenced and $237.67 billion of non-cancellable contractual commitments at 31 March 2026; those commitments include third-party cloud capacity and infrastructure, with some lease commencements extending through 2036. [S1] Amazon separately reported $106.347 billion of leases not yet commenced and $103.768 billion of unconditional purchase obligations, alongside operating leases, finance leases and financing obligations. [S3] These totals are not directly comparable, not wholly AI-specific and not debt-equivalent. They are evidence of a more important control issue: economically connected capacity decisions can sit in different approval systems and become visible at different times.
The appropriate response is not a new accounting category or a presumption that the build-out is excessive. Strong demand, contracted backlog and constrained supply can make early commitments value-accretive. External capital and joint ventures can distribute risk. Lease options can preserve flexibility. The decision is therefore conditional.
Enterprises with material AI-capacity commitments should govern a unified capacity-exposure portfolio across financing forms. The portfolio should measure four things that conventional project and capex views often separate: total economic exposure, utilisation risk, practical reversibility and exit rights. The decisive question is not simply how much will be spent this year. It is how much future choice has already been surrendered, under which demand assumptions, and at what cost that choice could be recovered.
The commitment happens before the spend
A technology investment committee approves an AI platform programme. The business case includes accelerator purchases and an annual cloud budget. The numbers fit the capital envelope. The programme is therefore described as funded and controlled.
Elsewhere, procurement signs a multi-year minimum-capacity agreement to protect availability. Real estate advances a dedicated facility lease because the delivery lead time exceeds the programme schedule. Treasury supports a venture structure that reduces immediate owned capex. The energy team secures power on terms designed for a future load profile. Each decision has its own sponsor, model and approval threshold.
Nothing in this sequence is inherently imprudent. The failure appears when the organisation asks a portfolio question and receives five functional answers.
The capital plan shows owned assets. Procurement shows contracted spend. Real estate shows lease liabilities when the relevant accounting events occur. Treasury shows guarantees and partnership exposure. Operations shows capacity and power dependencies. The project register may show only the programme that originally created demand.
By the time these views are reconciled, the economically important decision may already have been made. The organisation has not merely authorised expenditure. It has narrowed the range of futures it can choose without paying to exit.
That distinction matters because AI infrastructure combines assets with very different clocks. Microsoft reported that roughly two-thirds of its fiscal 2026 third-quarter capex related to short-lived assets, primarily GPUs and CPUs, while the remaining spend supported assets intended to monetise over 15 years and beyond; it also reported $4.7 billion of finance leases, primarily for large data-centre sites. [S4] A short-lived compute asset can therefore sit inside a much longer-lived physical and contractual envelope. The hardware may be replaced several times while the site, power, lease or financing arrangement remains.
Annual capex is a flow measure. Capacity exposure is a path-dependency measure.
What the conventional views miss
Capex governance is usually designed to answer whether an asset should be purchased, how it will be funded and whether the expected return clears a threshold. Project governance asks whether an initiative has scope, sponsorship, resources and delivery controls. Procurement asks whether terms are competitive and supply is secured. Treasury considers financing cost, liquidity and risk allocation.
Those are necessary questions. They are not the same question.
A capacity-exposure portfolio asks how the pieces behave together when the forecast changes. It treats a financing form as one attribute of a commitment rather than as the boundary of the decision.
| Form | What the normal control sees | What the portfolio must also see |
|---|---|---|
| Owned build | Capital cost, depreciation, delivery | Redeployment value, site specificity, power dependency, hardware-vintage mismatch |
| Lease | Term, payment profile, accounting treatment | Commencement lag, extension logic, break rights, subletting or transferability |
| Capacity contract | Unit price, committed volume, supplier performance | Minimum spend, resale rights, repricing, demand-shortfall exposure |
| Joint venture | Ownership, funding, governance rights | Occupancy dependency, guarantees, residual-value risk, practical exit route |
| Power or site agreement | Availability, price, delivery schedule | Constraint on location, utilisation timing, stranded enabling capacity |
The table does not imply equivalence. A residual-value guarantee is not the same as debt. A capacity contract is not the same as an owned data centre. A joint venture may genuinely share risk. The point is that all five can reduce the organisation’s ability to defer, resize, relocate or cancel capacity.
Meta’s El Paso transaction illustrates the mixture. Meta and BlackRock announced an 80/20 venture with approximately $14 billion of development costs, $12.5 billion of debt financing, leases that can extend to a potential 20-year term and residual-value guarantees with an aggregate threshold of approximately $13 billion that declines over time. [S2] The structure may accelerate delivery and bring in external capital. It also means that the economic assessment cannot stop at Meta’s 20 per cent equity interest or at the initial four-year lease term. The relevant exposure includes occupancy, extension choices, guarantees, site-specific infrastructure and the conditions under which value can be transferred or recovered.
The same control problem appears in less elaborate forms. Amazon’s disclosed commitments span operating lease liabilities, finance lease liabilities, financing obligations, uncommenced leases and unconditional purchase obligations. [S3] Some purchase obligations include energy, property and equipment, software and content, so the total is not a pure AI measure. That limitation is important. Yet it reinforces the governance point: published categories aggregate different purposes, while internal decisions may fragment a single capacity strategy.
The mechanism is option loss
The pressure begins with a real operating condition: demand is growing, supply is constrained, lead times are long or the cost of being late is high. Management responds by securing scarce inputs before revenue or utilisation is fully observable.
The organisation then uses several structures because no single structure solves every constraint. Owned build provides control. Leases accelerate access. Supplier contracts reserve capacity. Ventures bring capital and specialist capability. Power and land arrangements secure the enabling environment.
Fragmentation follows. Different functions approve different pieces, on different dates, against different baselines. Accounting recognition may begin after commercial commitment. Project reporting may start after land, power or capacity has been reserved. A venture may appear to reduce balance-sheet intensity while preserving a strong operating dependency.
The consequence is not automatically financial distress. It is reduced manoeuvrability.
Once capacity is tied to a location, power profile, supplier, hardware generation or minimum payment, later choices inherit earlier assumptions. A new model architecture may reduce compute per task. Unit prices may fall. A preferred accelerator may change. Power delivery may slip. Demand may remain strong but move to another geography or service tier. None of these developments makes the original decision irrational. They change the value of the options that were retained.
This is why cancellation language alone is an incomplete test. A contract can be cancellable in law but effectively irreversible in operation because replacement capacity is unavailable, sunk enabling work is specific to the site or exit would impair a linked commercial commitment. Conversely, a long contract can preserve practical flexibility through transfer rights, volume bands, repricing mechanisms, resale provisions or staged commencement.
The governance object is therefore not duration by itself. It is capacity optionality: the economically usable ability to defer, resize, redirect, reprice, transfer or terminate exposure under plausible downside conditions.
A financing structure is flexible only to the extent that the enterprise can exercise that flexibility when its demand, technology or location assumptions change.
A worked capacity decision
Consider an illustrative AI-intensive enterprise planning for 120 normalised units of compute demand three years ahead. This is a composite decision sequence, not a sourced company case.
The portfolio is assembled as follows:
- Forty units remain variable cloud consumption and can be reduced with limited notice.
- Thirty units are reserved under a minimum-spend agreement to secure price and availability.
- Twenty units are dedicated capacity in a colocation facility with a seven-year lease.
- Thirty units depend on a jointly financed site, with the enterprise as anchor customer and a residual-value support obligation.
The forecast is approved because each component addresses a different risk. Variable cloud preserves speed. Reserved capacity reduces scarcity risk. Colocation supports data and latency requirements. The venture provides long-duration scale without full owned funding.
Now assume realised demand is 75 units, not 120. The variable cloud can fall from 40 to 15. The other 80 units remain economically committed unless the enterprise can transfer, resell, reprice or terminate them. The portfolio therefore carries 20 units more committed capacity than realised demand even after the most flexible component has been reduced.
The usual response is to revisit utilisation plans: accelerate use cases, move workloads, renegotiate contracts or sell spare capacity. Those actions may work. They also reveal that the organisation is managing a commitment created earlier, not making a fresh capacity choice.
A robust approval would have tested this downside before signing. It would have asked which 45 units could be removed, at what cost, within what period and by whose authority. It would also have tested the opposite case: demand of 150 units. Optionality is not merely an escape route. It is the ability to scale in either direction without allowing one financing form to dictate the entire portfolio.
The worked figure exposes a common error. Diversifying financing forms does not necessarily diversify economic risk. Four contracts can still depend on the same demand forecast, power region, model economics or hardware roadmap.
The strongest case against a new portfolio layer
The sceptical argument is substantial.
The largest disclosed commitments belong to companies with extraordinary demand visibility, technical capability and balance-sheet capacity. They are not passive buyers caught by unfamiliar contracts. They operate infrastructure at scale, can negotiate sophisticated terms and may have customer demand already reserved against future capacity.
Microsoft’s management explicitly linked accelerated capex to demand, capacity constraints and the opportunity to convert installed capacity into revenue. [S4] Meta presents capital partnerships as a way to increase speed and flexibility. [S2] Orange and Morrison’s proposed 50/50 venture combines existing data-centre assets, infrastructure investment expertise, equity and debt to target 400 MW of capacity through a €3 billion programme. [S5] These structures can match asset duration with revenue duration, share development risk and preserve capital for other uses.
A unified portfolio layer could also become bureaucratic. Specialist teams may already exchange the necessary information. A central committee that re-performs finance, procurement, architecture and real-estate work could delay capacity when delay is itself the larger risk. If the enterprise has genuinely consolidated economic exposure, common scenarios and exit rights, the proposed governance change adds little.
This challenge should narrow the conclusion.
The case is not that leases, ventures or long-term contracts are disguised mistakes. Nor is it that every company needs a new committee. The case is that leaders need evidence that portfolio optionality is already governed across those forms. Where that evidence exists, the operating model may be sufficient. Where it does not, the apparent sophistication of individual transactions can conceal a basic control gap.
The sceptic therefore improves the decision test: do not ask whether the enterprise has a capacity governance process. Ask whether one accountable forum can show the aggregate exposure, common demand assumptions, concentration, and executable exit rights across every material form.
The capacity-exposure portfolio
A useful portfolio view should be decision-grade, not an inventory of contracts. It needs enough common structure to compare unlike forms without pretending they are identical.
Total economic exposure
Record the payments, minimum commitments, guarantees, contingent obligations and enabling investments that remain relevant under the same capacity thesis. Separate recognised liabilities from future contractual exposure and from scenario-dependent support. Do not collapse them into one debt-like number.
The purpose is to prevent a low-capex structure from being interpreted as low commitment. It is equally important to prevent a large disclosed total from being treated as fully irreversible when contracts contain usable options.
Reversibility
For each material exposure, identify what can actually be deferred, resized, repriced, transferred or cancelled. State the notice period, decision owner, direct exit cost and operational consequence.
The word option should not be accepted at face value. Meta’s announced El Paso leases have an initial four-year term and four extension options, which may provide flexibility. [S2] The portfolio question is how those choices interact with occupancy needs, guarantees, site investment and replacement capacity. Public disclosure cannot answer that. Internal governance should.
Utilisation and vintage risk
Link every capacity tranche to demand drivers, expected utilisation and the relevant technology vintage. Microsoft’s distinction between short-lived compute assets and 15-year-plus infrastructure is the right kind of separation. [S4] An enterprise should know whether a long-lived site depends on one hardware generation, whether later generations can be installed economically and which demand assumptions are shared across tranches.
A portfolio can be fully utilised in aggregate while still destroying value if the wrong capacity is available in the wrong place, at the wrong efficiency or under the wrong commercial terms.
Dependency and concentration
Capacity is constrained by more than compute. Power, cooling, network connectivity, land, permitting, supplier solvency, construction sequence and customer location can turn formally separate commitments into one concentrated risk.
The portfolio should therefore show common failure points. Three contracts with different counterparties may still depend on the same power region. Two geographically separate sites may still rely on one accelerator supply chain. A variable cloud service may be operationally difficult to move because data gravity and platform architecture have removed practical portability.
Decision rights and gates
Each commitment should have a named point at which option value materially changes. That may be land acquisition, power reservation, contract signature, lease commencement, equipment order, financing close or the expiry of a transfer right.
Governance should be strongest before that point, not merely when expenditure appears in the annual plan.
A staged model can be simple:
- Thesis gate: confirm the demand case, strategic need and conditions that would invalidate it.
- Exposure gate: consolidate every linked financing and contractual form under common scenarios.
- Option gate: specify minimum cancellation, transfer, repricing and resizing rights, or explicitly price their absence.
- Concentration gate: test shared power, supplier, site, technology and counterparty dependencies.
- Reforecast gate: revisit the portfolio when utilisation, unit economics, delivery timing or technology assumptions cross defined thresholds.
The output is not another project scorecard. It is a statement of how much future choice remains.
Where the argument applies
The evidence is strongest for hyperscalers, telecom operators, AI laboratories, cloud and colocation providers, and enterprises making material private-cloud, reserved-capacity or data-centre commitments.
It does not automatically apply to an ordinary adopter buying variable cloud services without minimum commitments. A company running short experiments through consumption-based services may have cost-management and architecture problems, but not the long-duration exposure described here. Nor does the argument apply with equal force where treasury and portfolio governance already consolidate leases, contracts, guarantees, power and ventures under one investment view.
There are also limits to what public evidence can establish. Disclosures show scale, duration and financing form. They do not reveal transaction-level utilisation assumptions, cancellation economics, pricing resets or internal committee visibility. The amounts reported by Meta and Amazon combine categories with different risks, and Amazon’s obligations are not all AI-specific. [S1] [S3] Public sources cannot tell us how much apparent exposure is practically irreversible.
That uncertainty is not a reason to ignore the signal. It defines the management question.
Strong demand may justify the commitments. Capacity scarcity may make delay more expensive than overcommitment. External capital may improve risk allocation. The thesis weakens materially when practical exit options are strong or when one enterprise-wide process already governs aggregate exposure. Those are conditions to test, not caveats to append after approval.
The question that changes the decision
The capex question is: Can we afford to build or buy this capacity?
The portfolio question is: Which futures become unavailable when we commit to it?
That second question does not oppose investment. It distinguishes investment conviction from accidental lock-in. It forces leaders to see a lease, a reserved-capacity contract, a venture, a guarantee and a power agreement as parts of one operating choice when they depend on the same demand thesis.
AI infrastructure may prove highly productive. The organisations that benefit most will not necessarily be those that committed least. They will be those that knew exactly where commitment became irreversible, which options they retained and what evidence would justify exercising them.
Sources
- Meta / US Securities and Exchange Commission — Form 10-Q for the quarter ended 31 March 2026 — 2026 — https://www.sec.gov/Archives/edgar/data/1326801/000162828026028526/meta-20260331.htm
- Meta Investor Relations — Meta Announces New Strategic Venture with BlackRock to Develop Data Center in El Paso — 28 July 2026 — https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-New-Strategic-Venture-with-BlackRock-to-Develop-Data-Center-in-El-Paso/default.aspx
- Amazon / US Securities and Exchange Commission — Commitments and Contingencies for the quarter ended 31 March 2026 — 2026 — https://www.sec.gov/Archives/edgar/data/1018724/000101872426000014/R11.htm
- Microsoft Investor Relations — Fiscal Year 2026 Third Quarter Earnings Conference Call — 29 April 2026 — https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3
- Orange Newsroom — Orange and Morrison announce plans to create a data centre joint venture to strengthen Europe’s digital sovereignty — 27 July 2026 — https://newsroom.orange.com/orange-et-morrison-annoncent-un-projet-de-creation-de-coentreprise-de-data-centers-pour-renforcer-la-souverainete-numerique-de-leurope-486247/?lang=eng