Insight, Energy & Industrial Policy
A 74 MW solar plant in Timor-Leste and a data centre campus in Virginia are ordering from the same factories. Lead times for large power transformers have roughly doubled since 2021. Nobody publishes what small buyers in small markets are paying, or waiting.
Two items in today's briefing are usually read as belonging to different worlds.
The International Energy Agency's 2026 analysis of energy and artificial intelligence estimates that AI-focused data-centre electricity consumption rose about 50% in 2025, and that total data-centre demand could roughly double from 485 TWh in 2025 to 950 TWh by 2030. More striking than the electricity figure is the physical one: the capacity of dedicated AI factories has more than tripled in around eighteen months, creating pressure on transformers, high-bandwidth memory, power electronics, storage and grid connections.
Separately, Timor-Leste is advancing its first utility-scale renewable project: roughly 74 MW of solar with 80 MWh of battery storage near Manatuto, backed by the Asian Development Bank, expected to add at least 134 GWh annually and avoid around 92,700 tonnes of CO<sub>2</sub> a year, displacing a grid currently dominated by imported diesel.
These are not separate stories. Both projects need step-up transformers, switchgear, high-voltage cable, power electronics and a grid connection. They are ordering from the same small set of factories, and factory slots are allocated by order book rather than by need.
Most analysis of AI and energy is about electricity volume. We have argued before that the more important question is which sectors AI makes productive, and that in markets without reliable power the decisive question is who absorbs the volatility. This is a third thing again, and it is more immediate than either.
Before anyone can consume a terawatt-hour, somebody has to install the equipment that moves it. That equipment is scarce in a way that has become severe.
Large power transformers had lead times of roughly 7 to 14 months before the pandemic. They now run beyond 24 months in most major markets, with specialised units approaching 36 to 48 months. The IEA's own 2025 transmission-grid work, based on an industry survey, found lead times had almost doubled on average since 2021, with manufacturers reporting record backlogs. A quarterly industry survey in 2025 put power transformers at an average of 128 weeks and generator step-up units at 144 weeks. Demand for generator step-up transformers rose 274% between 2019 and 2025, with substation transformers up 116% over the same period.
Those are the numbers that decide whether Manatuto is energised on schedule. They are not in any energy-demand model, because energy-demand models forecast electricity rather than the availability of the objects required to deliver it.
Survey figures converted from weeks. All figures are drawn from North American surveys, which is itself part of the argument below. A four-year lead time on a transformer is not a procurement inconvenience. It is longer than most concessional-finance disbursement windows.
Manufacturing slots go to the order book. A buyer's position in that book is a function of volume, repeat business, prepayment capacity and relationship, and a first utility-scale project in a country of one and a half million people scores poorly on all four.
The disadvantages compound in ways that are worth naming individually, because several of them are self-inflicted by the development-finance system rather than by the market.
Order size buys priority and small buyers have none. A hyperscaler ordering across dozens of sites is a strategic account. A single project ordering a handful of units is a scheduling nuisance. This is ordinary commercial behaviour and no manufacturer should be criticised for it.
Prepayment and forward buying are the standard mitigation, and concessional finance forbids them. The way large buyers manage a four-year lead time is to reserve capacity ahead of need, often years before a specific project is confirmed. A project financed by a development bank generally cannot order major equipment before financial close, because disbursement is conditional and procurement rules require a competitive tender against a defined scope. Those rules exist for good reasons, principally to prevent capture and waste. Their effect in the current equipment market is to guarantee that publicly financed projects join the queue at the back, after every privately financed buyer who was free to move earlier.
That is a design problem, not a market failure, and it is fixable by the institutions that created it.
Competitive tender adds months at exactly the wrong point. Tendering, evaluation, challenge periods and contract award consume time during which slots continue to be allocated to others. The integrity value of the process is real. So is the cost, and the cost has risen sharply as lead times have extended, which means the trade-off has changed even if the rules have not.
Specification mismatch narrows the field further. A 50 Hz grid, an unusual voltage class, seismic requirements, island-grid characteristics or a small order of non-standard units reduces the number of factories that will quote at all. The smaller the market, the more likely its requirements sit outside the mainstream production run.
Every figure quoted above comes from North American market surveys. Wood Mackenzie, PwC, industry associations and trade press track lead times and prices for utilities and developers in the United States, because those buyers are numerous, organised and worth surveying.
Nobody publishes equivalent figures for what a small island utility, a rural electrification agency or a development-bank-financed project actually pays and actually waits. The premium may be modest or it may be very large. There is no dataset either way.
This matters more than it sounds. Decisions about project timelines, contingency budgets, disbursement schedules and procurement rules are being made across the development-finance system on the basis of numbers describing a completely different class of buyer. The disadvantage is real, widely acknowledged in private, and entirely unquantified.
That is a measurable gap and closing it is not technically difficult. Lead times, price premia and failed-tender rates across a portfolio of publicly financed energy projects could be assembled from procurement records that already exist. It requires somebody to decide the question is worth answering, which is generally the binding constraint in applied measurement.
The briefing frames the choice facing emerging economies as whether to compete in AI software or in the infrastructure supporting it. That framing quietly assumes the second option means hosting compute, and hosting is the weakest version of it.
A data centre is close to a textbook enclave. Very high capital intensity, very low permanent employment, a large and inflexible electricity draw, and backward linkages into the domestic economy that end when construction does. Forward linkages exist only if local firms can actually buy and use the compute, which requires a domestic software and data sector that in most cases is the thing the country was hoping to build. Hosting delivers electricity sales, some construction, some property tax, and a grid connection that is no longer available to anyone else.
Albert Hirschman's argument about linkages remains the right lens, and it points somewhere different. The value in an industrial-policy sense is in supplying the physical stack: transformers, switchgear, cable, cooling, power electronics, installation and commissioning capability. Those have deep backward linkages into metals, electrical steel, copper and manufacturing skills, and forward linkages into every other electrification project in the region. They are also, at this moment, the scarcest thing in the chain, which is the strongest position a supplier can occupy.
The obvious objection is that heavy electrical manufacturing takes decades to establish and cannot be conjured in response to a shortage. That is correct for transformer cores. It is much less correct for the assembly, installation, testing and commissioning layer, and for regional maintenance capability, which is where a great deal of the constraint actually sits and where several countries could plausibly compete.
Timor-Leste will not manufacture transformers. It will need people who can install, commission and maintain 74 MW of solar and 80 MWh of storage for twenty years, and whether that capability is built locally or imported for the duration is a policy choice being made now, usually by default and inside a procurement document.
Permit early procurement of long-lead items. Reserve manufacturing capacity before financial close, with a separate approval track and a written-off cost if the project does not proceed. The expected cost of occasionally losing a deposit is far lower than the expected cost of a two-year delay across a portfolio.
Aggregate orders across projects and countries. Ten small projects ordering separately are ten nuisances. The same ten aggregated into a framework agreement are an account worth scheduling. No single institution can do this alone, which is precisely the kind of coordination a multilateral is for.
Treat equipment lead time as a portfolio risk with an owner. It is currently handled project by project, discovered late, and absorbed as delay. It behaves like a correlated portfolio exposure and should be monitored as one.
Measure what small buyers actually face. Before designing around the problem, establish its size. The numbers being used at present describe American utilities, and there is no reason to believe they transfer.
The Lab works on this at the intersection of energy access and local manufacturing. The question of what an incumbent industry carries with it when it redeploys, which we set out in The Incumbent's Second Life, applies directly to any country considering entering the electrical-equipment chain.
If you are financing energy infrastructure in a small market and want to know what your projects are actually waiting for, tell us what you need to know.
This is an independent insight piece by Transitions Lab. For the Lab's applied work, see Energy Access & Off-Grid Systems and Local Manufacturing & Supply Chains. See also The Load That Grows When It Is Hot on the same equipment-and-capacity constraint pushed into cooling systems by a load correlated with weather. To discuss a study, see Contact.