The AI Power Bottleneck Is a Delivery Date

AI power scarcity is a problem of location, timing and completion. A generation forecast, a turbine reservation and a commissioned campus describe different stages of the same investment—and carry different risks.

1. Growing demand makes execution more valuable

A large electricity-demand forecast is useful context, but it does not identify an investable bottleneck by itself. Investors need to locate the stage that prevents a particular project from becoming productive: power supply, connection approval, electrical equipment, construction, cooling or final acceptance.

The IEA’s 2026 central projection puts annual data-center electricity use at 950 TWh in 2030, compared with an estimated 485 TWh in 2025. The implied increase is approximately 96%, or 14.4% annually over five years. The growth rates are our calculations from the two endpoints, not additional IEA observations. [iea-ai]

Electricity has to be available at the required site, with the required reliability and commercial terms. A national generation surplus cannot necessarily solve a local connection constraint. Nor can a campus purchase energy in the abstract and assume that the physical system can deliver it at every hour.

For a project developer, the economic consequence is time to revenue. Equipment can be ordered, staff hired and capital committed before a facility can serve paying workloads. Delays therefore affect both the first billing date and the period over which fixed costs must be financed. For a supplier, the same delay can shift shipment timing or customer acceptance rather than eliminate end demand.

2. Identify the dependency that controls the completion date

The IEA describes indicative development ranges of 5–15 years for new grid infrastructure and 1–3 years for data centers. These are broad planning and delivery ranges, not an estimate of how long every campus waits for its connection. Their significance is the mismatch between systems that must ultimately work together. [iea-grid]

Consider an illustrative campus with a building ready in two years and a required grid reinforcement ready in six. Accelerating the building by another three months does not move the power-dependent opening date. The valuable intervention is one that removes the binding dependency, not necessarily the most visible construction activity.

The reverse is also possible. A site can secure electricity while lacking the qualified labor or controls integration needed to commission its cooling system. If that becomes the critical path, more generation equipment does little to improve near-term productive output. Bottlenecks migrate as individual stages are completed.

3. Manufacturing ramps are another clock

GE Vernova’s July 2026 release described a 20 GW annual gas-turbine output milestone for Q3, followed by 24 GW in 2028 and actions toward 30 GW in 2030. The figure preserves that forward-looking status. It is a manufacturing-output path, not a schedule of data-center commissioning. [gev]

The accompanying call distinguished 3 GW of quarterly gas shipments from 20 GW of orders and reservation agreements. Equipment backlog and slot reservations together reached 116 GW; the transcript split that into 53 GW of backlog and 63 GW of reservations. A reservation can be commercially meaningful without being identical to an equipment order. [gev-call]

These categories must not be added to installed capacity. Shipments still need a customer project with fuel supply, site works, construction and commissioning. And an annualized factory rate is not the same as actual shipments over a complete calendar year. Preserving those distinctions prevents a strong demand story from becoming an unsupported operating-capacity claim.

Manufacturing expansion can itself move the constraint upstream. Extra assembly space may not be sufficient if qualified components, skilled labor or testing capacity become binding. The investment test is therefore broader than installed machinery: can the complete production route sustain conforming output at the required delivery rate?

4. Test whether the supplier earns from removing the delay

Physical relevance earns a company a place in the research universe. Financial conversion determines whether that relevance has become an economic advantage. The strongest case combines accepted deliveries with favorable realized pricing, controlled execution costs and cash that can be sustained after reinvestment.

Vertiv’s Q2 2026 sales were $3.274bn, versus $2.638bn a year earlier. GAAP operating profit rose from $442.4m to $637.9m, and operating cash flow reached $1.100bn. The release also noted temporary supply-chain congestion and phased project timing. These disclosures support examining delivery and economic conversion together; they do not identify exact commissioned liquid-cooling megawatts. [vertiv]

Eaton provides a useful scope check. Its Electrical Americas segment reported 18% organic growth, while Electrical Global’s 44% reported growth included 25 percentage points from Boyd Thermal, 18 from organic growth and one from currency. Company or segment growth cannot simply be relabeled growth in AI capacity. [eaton]

The cash question needs equal care. GE Vernova management described customer down payments as funding the output expansion. Advance cash validates commitment and can reduce financing pressure, but it arrives with future performance obligations. A quarter of excellent cash collection can precede, rather than follow, delivery. [gev-call]

For each supplier, ask what has been sold, what has been shipped, what has been accepted and what cash remains after the required investment. Compare operating margins with operating margins and cash-flow periods with matching periods. An EBITDA margin at one supplier and a segment operating margin at another do not form a clean profitability ranking.

5. The strongest counterargument is faster use of existing infrastructure

A scarcity thesis should include the mechanisms that could relieve it. The IEA estimates that technologies and regulatory changes could unlock 1,200–1,600 GW of hosting capacity for advanced-stage projects in global queues. It identifies conditional non-firm connections as one route, with users accepting limits on output or consumption at certain times. These are system-level potential estimates, not committed data-center connections. [iea-grid]

This creates a commercial tradeoff. A flexible workload may accept interruptions to obtain earlier access; a service requiring uninterrupted availability may face a different calculation. Researchers should test how much flexibility is actually available, who bears curtailment costs, and whether backup arrangements erase the apparent saving.

Onsite generation is another option, but it exchanges one set of dependencies for another. The IEA estimates that reliable onsite gas supply can require generation capacity 30–70% above load and notes turbine-supply constraints. Fuel, permitting, reliability design and equipment availability therefore remain part of the route to operation. [iea-ai]

ConstraintAlpha view: Follow the completion date and the contract behind it. Megawatts announced indicate ambition. Energized, accepted and economically productive capacity is stronger evidence. The opportunity lies in identifying which supplier can remove the binding delay—and retain an attractive return for doing so.