The Board's AI Capital Test

The first case is contracted capacity. It is supported by enforceable customer commitments, credible conversion of backlog into revenue and an operating plan for delivery. The board should ask about contract duration, termination rights, customer concentration, pricing, capacity allocation and the cost of bridging supply. A large backlog is relevant evidence, but it is not cash, utilisation or margin. ¹²⁴

The second case is platform defence. A cloud, advertising, software or data business may need capacity to retain customers, protect a product position or prevent a competitor from controlling a critical layer. Its return may appear across the existing franchise rather than as stand-alone AI revenue. The evidence is therefore customer retention, product performance, pricing power and avoided strategic loss—not a generic statement that the market is moving quickly.

The third case is a capability option. A bounded commitment can buy learning, scarce access, talent, data or a future right without assuming the full market materialises. Good options have a maximum cost, an expiry date, observable learning goals and a decision owner. An option with no boundary is simply an open-ended programme with a more attractive name.

The fourth case is speculative exposure. It can be legitimate when the board deliberately accepts concentrated, uncertain return for exceptional upside. SoftBank provides unusually direct language: its chief risk officer describes another planned $30 billion follow-on OpenAI investment, high uncertainty from technology, competition and regulation, and monitoring of financing, market and portfolio risk. The useful lesson is not whether that bet succeeds. It is that concentration and financing belong in the decision record. ⁶

The latest disclosures contain substantial counter-evidence to a blanket capex warning. Alphabet says cloud revenue grew 82% and backlog reached $514 billion. Microsoft reports 84% growth in commercial remaining performance obligation. Amazon says AWS grew 37% and both its AI and chips businesses exceeded $25 billion annual revenue run rates. Supply remains constrained. A board that responds by freezing investment could destroy customer value and strategic position. ¹²⁴

The return question still matters because capital converts into expense and cash demands on different clocks. Accelerators and CPUs are shorter-lived than buildings, networking and grid connections. Alphabet expects higher depreciation and data-centre operating costs, including energy. Microsoft's official commentary separates shorter-lived compute assets from facilities intended to support monetisation for 15 years or more. Amazon's free-cash-flow definition subtracts property and equipment purchases, making the current investment cycle visible even while operating performance is strong. ¹³⁴

A useful board pack therefore puts six measures on one page: contracted demand, delivered capacity, utilisation, unit economics, lifecycle expense and free-cash-flow effect. Add customer and supplier concentration, because a capacity plan tied to a small number of labs, chip vendors, utilities or anchor customers can be diversified in legal form while remaining economically correlated.

The test is not a single hurdle rate applied without context. Contracted capacity may justify staged expansion against signed demand. Platform defence may need a portfolio view. An option should be judged by learning and preserved choice. Speculation should be sized against loss capacity. What matters is that the case is explicit before approval and testable afterwards.

North American hyperscalers can fund capacity from large global platforms, sell across several layers of the stack and reallocate compute between model development, first-party products and cloud customers. That flexibility is economically valuable, but it also makes stand-alone return attribution difficult. Their reported growth is evidence of demand for their businesses; it is not a global proxy for enterprise adoption or for the economics available to a regional operator.

Europe is combining commercial investment with industrial policy. The European Commission says 19 AI Factories and 13 antennas are being established, while an AI Gigafactory programme supported by up to €10 billion in EU and national funding is intended to unlock at least €20 billion of private investment. That creates access, sovereignty and ecosystem value, but the board case must still separate public-policy support, commercial demand and conditions attached to capital. ¹⁰

For lower-income and emerging markets, the constraint set is broader. The World Bank describes four foundations—connectivity, compute, context and competency—and argues that viable markets need coordinated supply, demand and rules. A July working paper covering 85 countries adds an important warning: cheap electricity rarely creates compute-export competitiveness by itself because hardware cost, delivery frictions, regulation and trust can dominate. A proposal for an ‘AI hub’ should therefore prove hardware access, anchor demand, connectivity, serviceability and institutional credibility, not merely an attractive power price. ⁸⁹

A board can approve several individually defensible commitments and still create one correlated exposure. The same frontier lab may be an equity investment, a cloud customer, a supplier dependency and the source of demand supporting new capacity. The same chip vendor can sit behind several cloud contracts. The same grid or financing market can constrain sites that appear geographically separate. Legal diversification is not necessarily economic diversification.

The portfolio record should therefore connect equity, debt, leases, purchase commitments, customer concentration, supplier dependence and energy exposure. Stress the shared assumptions: slower model demand, a pricing reset, a delayed grid connection, export controls, component inflation or one anchor customer changing its plan.

This is especially important for speculative and platform-defence cases, where the return may depend on several linked outcomes rather than one contracted cash flow. Show the board both the stand-alone case and the loss if shared assumptions fail together.

The International Energy Agency projects global data-centre electricity consumption to more than double to about 945 TWh by 2030. It estimates that data centres account for around one-tenth of global electricity-demand growth over the period, but the local effect is far more concentrated. The United States and China dominate absolute growth; Southeast Asian demand is also expected to more than double. The IEA estimates around 20% of planned projects could be at risk of delay unless grid constraints are addressed. ⁷¹²

This changes board timing. A server delivery is not useful capacity until power, networking, cooling, permits and customers are ready. A cheap site is not a low-cost site if energisation is late, curtailment is frequent or water and community conditions are unresolved. Capacity that misses the demand window can turn a contracted case into speculation.

The counterpoint is equally important: efficiency, geographic flexibility and better grid integration can change the outlook. The IEA's scenarios span materially different energy paths depending on adoption, supply chains and efficiency. Energy forecasts should be treated as planning ranges, not a reason to assign false precision to a 10-year return model.

The strongest challenge to this framework is that delay can be more expensive than overcapacity. Cloud demand is growing quickly, major providers remain supply constrained and long-lived sites can support several technology generations. A board that requires perfect attribution before acting may miss customer commitments, lose scarce power or surrender platform position. ¹²³⁴

That is why the test classifies uncertainty instead of pretending to remove it. Contracted demand can support staged delivery. Platform defence can be funded against explicit franchise measures. An option can preserve speed while bounding loss. Speculation can move rapidly when its concentration and downside fit the agreed risk appetite.

The governance objective is not slower investment. It is faster recognition when the evidence no longer supports the original case, with a pre-agreed action that prevents sunk cost, strategic prestige or management momentum from becoming the decision rule.

For every material commitment, record the case, the accountable executive, the geography, the asset mix and the dependency chain. Then answer eight questions: What customer or franchise demand supports it? Which assets are short-lived and which are long-lived? When is usable capacity available? What utilisation and unit-economics range is required? Who pays for power, grid and water? Where are supplier and customer concentrations? What financing or public-policy assumptions are embedded? Which dated finding triggers expansion, reallocation or stop?

Review the record at least quarterly and whenever the model, chip, supplier, customer, site, regulation or funding structure changes. Do not allow a strategic option to become permanent operating spend without renewal. Do not present voluntary policy support as contracted demand. Do not use U.S. cloud growth to stand in for global adoption, or a low power price to stand in for a viable regional market.

The decision is not anti-investment. It is an attempt to preserve the investments that have evidence while making uncertainty visible enough to govern. Boards can move quickly when the case is clear: expand contracted capacity against conversion; fund platform defence against explicit franchise measures; buy options with bounded loss; and accept speculation only within an agreed risk appetite.

AI infrastructure may become one of this decade's most consequential capital cycles. The organisations that benefit will not be those that simply spend the most or wait the longest. They will be those that know which case they are funding, can see when the case changes and have already agreed what to do next.