The Announcement Is Not the Asset
The governance task is not to predict one correct capacity number; it is to preserve the right to learn before every layer becomes irreversible.
The capex number is the least useful unit of governance
AI infrastructure is often presented to boards as a single, dramatic capital number. That framing is convenient for headlines and weak for decisions. It compresses assets with radically different economic lives, supply constraints and exit paths into one total. A GPU fleet that may be technologically superseded within a few years is not the same commitment as a powered data-centre site expected to operate for decades. A finance lease is not the same as a cash purchase. Reserved third-party cloud capacity is not the same as owned equipment. Yet all can appear in the same conversation about “AI capex”.
The scale makes this distortion consequential. Alphabet said it expected 2026 capital expenditure of $175–185 billion after spending $91.4 billion in 2025. Microsoft reported $34.9 billion of quarterly capital expenditure, with roughly half directed to short-lived assets and the balance to long-lived assets, including $11.1 billion of finance leases. Meta raised its 2026 capital-expenditure range to $125–145 billion, citing higher component prices and additional data-centre costs. Amazon reported that a $66.1 billion year-on-year increase in purchases of property and equipment had pushed trailing-twelve-month free cash flow to an outflow, while attributing the increase primarily to AI investment.
Those figures are not comparable measures of “AI conviction”. They differ in scope, accounting treatment, business mix and timing. The governance question is therefore not whether the headline is large. It is whether each commitment layer has evidence appropriate to its reversibility, lead time and economic exposure.
Replace the budget with a commitment ladder
A useful board framework separates the programme into four layers.
Short-lived compute
Accelerators, CPUs, networking equipment and associated systems are the most visible layer. Their commercial life can be shorter than their accounting life because model architectures, chip performance and price-performance improve quickly. They can often be redeployed across workloads, but only if software, networking and operational systems make that substitution practical.
This layer should be governed by utilisation, queue depth, workload priority, serving cost and refresh economics. Procurement volume should follow credible demand signals rather than a general belief that AI usage will rise. Microsoft’s distinction between short-lived compute and long-lived sites is useful precisely because it exposes the different evidence each requires.
Long-lived sites, power and cooling
Land, power connections, buildings and cooling systems have longer lead times and are harder to reverse. Their value depends not only on aggregate demand but also on location, power availability, network topology, permitting and the ability to support future hardware generations. These commitments can preserve strategic access to scarce capacity, yet they can also lock the enterprise into a geography or design before demand is sufficiently understood.
The control point is not quarterly utilisation alone. It is staged commitment: options on land, contracted power, modular construction, equipment-ready shells and explicit gates for fitting out additional capacity. The relevant evidence includes time-to-power, construction milestones, contracted customer demand, regional concentration and the cost of delaying versus accelerating.
Financing structures
Finance leases, operating leases, purchase commitments and supplier financing can alter the timing and presentation of cash and capital expenditure without changing the underlying economic obligation. Microsoft’s disclosure illustrates the issue: total capital expenditure and cash paid for property and equipment diverged substantially because finance leases and payment timing were material.
Boards should review the obligation schedule, not only the capex line. The questions are simple but demanding: What has been irrevocably committed? When does cash leave? What termination, renewal or substitution rights exist? Which costs sit outside the headline measure? A governance system that compares companies or business units without normalising these structures will reward presentation rather than commitment quality.
Third-party capacity
Cloud contracts, reserved instances, colocation and dedicated hosting can preserve flexibility or merely move rigidity off the balance sheet. The distinction depends on minimum-spend terms, duration, portability, data-egress costs and whether capacity can be reassigned. Outsourcing is not inherently more reversible than ownership.
This layer should be tested against total committed cost, utilisation floors, concentration risk and the operational cost of switching. It is often the fastest way to meet uncertain demand, but the option has value only when contracts retain genuine choices.
Match evidence to irreversibility
The commitment ladder becomes practical when every approval is paired with an evidence threshold. The threshold should rise with irreversibility and lead time.
For short-lived compute, require workload-level demand, observed utilisation and a credible plan for redeployment. For sites and power, require capacity scenarios, milestone gates and a quantified value of securing scarce inputs early. For financing, require a complete obligation view across cash purchases, leases and commitments. For third-party capacity, require contract-level tests of minimum spend, portability and exit.
This is not a demand for false precision. Public disclosures do not consistently isolate AI spending from broader technical infrastructure, nor do they align asset lives, lease treatment or customer commitments. Internal data will also be imperfect. The answer is to state which uncertainty matters at each gate and decide what evidence would change the next commitment.
The governance task is not to predict one correct capacity number; it is to preserve the right to learn before every layer becomes irreversible.
That principle changes portfolio reviews. A programme can be strategically attractive and still fail a near-term commitment gate. Another can have modest immediate economics but deserve an option payment because power or land scarcity would make later entry impossible. The comparison is not between “invest” and “do not invest”. It is between commitment structures with different learning rights.
Use a five-part decision record
Every material AI-infrastructure approval should contain five elements.
The demand claim
State which workloads, customers or products require the capacity, over what period, and with what confidence. Separate contracted demand from pipeline, internal forecasts and strategic aspiration. Backlog can be relevant, but it is not automatically equivalent to utilisation, margin or cash conversion.
Alphabet’s disclosures connect infrastructure investment to Cloud demand, first-party products and frontier-model development. That breadth is strategically coherent, but it also means a single demand metric cannot govern the whole portfolio. Each demand pool needs its own observable signal.
The unit-economics claim
Show how incremental capacity changes revenue, gross margin, cost to serve or research throughput. Include energy, networking, software and operations, not only hardware acquisition. Alphabet reported substantial reductions in Gemini serving unit costs during 2025, demonstrating why efficiency must be reviewed alongside capacity. A lower unit cost can absorb demand growth without a proportional capacity increase; greater usage can also overwhelm efficiency gains.
The commitment map
List cash purchases, lease obligations, long-term supply agreements, third-party reservations and construction commitments on one timeline. Record the earliest exit, substitution or deferral point for each. This prevents a portfolio from appearing flexible merely because obligations sit in different accounting categories.
The stress case
Test at least three conditions: demand arrives later than planned; component economics improve faster than expected; and scarce inputs become harder to secure. The stress case should show which commitments remain productive, which can be redeployed and which become stranded. It should also test the opposite risk: that excessive caution leaves the enterprise unable to serve demand.
The next gate
Every approval should specify the next decision date, evidence required and maximum additional exposure. A gate without a bounded next commitment is a status meeting. A gate with explicit evidence and exposure is a control.
Govern scarcity without using it as a blank cheque
The strongest challenge to staged investment is real: AI infrastructure is constrained by power, sites, equipment and supply chains. Early overbuilding can create advantage if late entrants cannot secure capacity. Microsoft reported demand exceeding supply across workloads, while Alphabet described a tight supply environment. A framework that treats reversibility as the only virtue would miss the option value of early commitment.
The answer is to price scarcity explicitly. Boards should distinguish between buying a strategic option and funding a fully built asset. Securing land, interconnection rights or a supplier allocation may be rational before demand is certain. But the proposal should state the option premium, expiry, follow-on commitment and scenario in which the option is abandoned. Scarcity is then a measurable reason for action, not a narrative exemption from discipline.
Likewise, high growth does not settle the returns question. Amazon’s AWS growth and AI-related investment can support a strong strategic case while free-cash-flow pressure remains visible. Meta can expect operating income growth while expanding capital plans. These facts can coexist. Governance should track whether capacity produces the expected operational and financial evidence over time, not infer success or failure from one aggregate outcome.
Build a board dashboard around commitment quality
A compact dashboard should report six measures for each layer:
— Utilisation and demand coverage, distinguishing contracted, forecast and speculative demand. — Unit cost and contribution economics for the workloads using the capacity. — Irrevocably committed exposure over the next 12, 24 and 60 months. — Time to the next meaningful learning event. — Exit, substitution and redeployment rights. — Concentration across suppliers, regions, power sources and anchor customers.
The dashboard should reconcile to finance but not be limited by financial-statement categories. Its purpose is to show economic exposure before accounting presentation obscures it. Portfolio reviews should then allocate new commitment to the opportunities with the strongest combination of evidence, strategic access and retained options.
This also changes accountability. Technology leaders own workload and architecture assumptions. Finance owns the obligation map and cash consequences. Procurement owns contractual optionality and concentration. Business sponsors own the demand case. The board or investment committee owns the portfolio trade-off. No single function can certify commitment quality alone.
The decision standard
AI infrastructure is neither a conventional IT refresh nor a single moonshot. It is a portfolio of interdependent commitments with different clocks. The useful standard is therefore not “Are we spending enough?” or “Is capex too high?” It is: Have we matched the duration and irreversibility of each commitment to evidence strong enough to justify it?
That standard preserves strategic ambition. It supports early moves where scarcity has demonstrable option value, rapid scaling where demand and economics are visible, and restraint where accounting categories conceal rigid obligations. Most importantly, it turns the discussion from the size of the bet to the design of the commitment.
Sources
- Alphabet — 2025 Q4 Earnings Call — 4 February 2026 — https://abc.xyz/investor/events/event-details/2026/2025-Q4-Earnings-Call-2026-Dr_C033hS6/default.aspx
- Microsoft — Fiscal Year 2026 First Quarter Earnings Conference Call — 29 October 2025 — https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q1
- Meta — First Quarter 2026 Results — 29 April 2026 — https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-First-Quarter-2026-Results/default.aspx
- Amazon — Second Quarter 2026 Results — 30 July 2026 — https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/