The forecast is not an order.

Two sheets of paper lie side by side on a desk at a company that builds large power transformers. The company, the desk and both documents are invented. The sheet on the left is an industry outlook: a curve that climbs steeply to the right, a headline number in the trillions, and the word could in its first sentence. The sheet on the right is an order form for twelve transformers, with a delivery window, a 20 percent deposit, a cancellation clause and a signature line that nobody has signed. Both sheets describe the same market. They answer different questions.

The outlook asks how large the market might become. The order form asks who owes what if the project it serves slips. A supplier deciding whether to add a production line needs both answers, and the trouble starts when the first is read as if it were the second. My argument is that a case for new capacity should keep projected demand and contracted customer commitment on separate lines, because they carry different information. Projected demand describes a possible market. A commitment allocates part of the risk of serving it: it says which customer will pay, when, and what happens if that customer changes its mind.

This is not an argument that forecasts are useless, or that a company should build only against signed orders. Forecasts are how anyone sees an opportunity before it arrives, and capacity built strictly against firm orders often arrives too late. Nor is it a claim that the current build-out of data centers is a bubble. The claim is narrower. However large a forecast is, it cannot tell a particular supplier how much to build, which customers will pay, or who absorbs a delay. Those answers come from commitments, and the case for new capacity should show how much of it rests on each kind of evidence.

What a forecast is for

A market forecast is a model of possible futures. Its job is to show where an opportunity might lie, to test whether today's capacity could cope with it, and to compare one future with another. A good forecast states its scope, its horizon and its assumptions, and it usually comes with a range.

The data-center build-out is a useful case because its forecasts are so large. On March 27, McKinsey published an analysis of how industrial companies might win a share of that build-out. By the firm's estimates, global spending on data centers could reach $7 trillion by 2030.¹ Each word in that sentence does work. Global means the total is spread across every region and every buyer. By 2030 means it accumulates over several years, not one. Could means it is an estimate of what may happen, not a tally of what has been signed. And the money does not all go to one kind of supplier. In the firm's earlier and more detailed analysis, published in April 2025, about 60 percent of the AI-related portion was expected to go to the makers of chips and computing hardware, and about 25 percent, or $1.3 trillion, to what the authors call energizers: power generation and transmission, cooling, and electrical equipment.²

The same 2025 analysis shows why a single number needs its range. It modeled three paths for AI-related capacity. In the central one, on which its headline figures rest, AI data centers would require $5.2 trillion of capital spending by 2030. If demand were constrained, the figure fell to $3.7 trillion; if demand accelerated, it rose to $7.9 trillion.² The distance between the low and high cases is larger than the low case itself. That is not a flaw in the analysis. It is an honest statement of how uncertain the future is, and the authors drew the obvious conclusion: overinvesting risks stranded assets, while underinvesting means falling behind.²

The International Energy Agency handles uncertainty the same way, for a different quantity. Its April 2025 report Energy and AI estimated that data centers used about 415 terawatt-hours of electricity in 2024, roughly 1.5 percent of the world's consumption, and projected about 945 terawatt-hours by 2030 in its Base Case, alongside other cases in which adoption, efficiency and infrastructure turn out differently.³ Electricity is not money, and the two sets of projections should not be stitched together into one. What they share is a method: several futures, each labeled, none presented as a promise.

From interest to cash

Now follow the supplier's decision. Its one production line builds 120 large transformers a year and is fully booked. Customers are asking for more. Should it build a second line?

"Customers are asking" covers several different things, and they carry very different weight. In rough order of strength, the stages run something like this. Interest: a developer mentions that its next campus will need transformers. A request for a quotation: the developer asks for a price and a delivery date, which costs it an afternoon. A reservation: the customer holds a production slot, perhaps for a small fee, and can usually let it go. A firm order: a signed contract with a deposit, a delivery schedule and a penalty for cancelling. Cash: the deposit, and later the balance, actually received. These are illustrative commercial stages, not a description of how McKinsey or anyone else builds a forecast. What matters is that each step up moves some of the risk from the supplier to the customer.

Parts of the market are already climbing this ladder. McKinsey reports that in North America the lead times for medium-voltage switchgear and transformers can now reach 80 and 50 weeks respectively, and that the largest cloud operators are reserving production slots earlier and paying premiums for schedule certainty.¹ A premium for certainty is a customer agreeing, in cash, to carry part of the risk. The supplier's question is how much of its proposed new line is backed by commitments of that kind, and how much by the outlook alone.

What a signature settles, and what it does not

A signed order is stronger evidence than a forecast, but its strength lives in the terms. Four of them decide who carries the risk.

Cancellation rights. An order that can be cancelled without penalty until shipment is closer to a reservation than to a sale. The penalty schedule tells the supplier how much it recovers at each stage if the customer walks away, and therefore how much of the production line the order really underwrites.

Deposits. A deposit is the part of the commitment that has already changed hands. It covers some of the supplier's costs if the order collapses, and its size is a rough measure of how much of the downside the customer has accepted.

Delivery dependencies. A transformer for a data center that cannot get a grid connection is a transformer nobody can install. The IEA warned that unless transmission investment speeds up, around one in five planned data-center projects could be at risk of delay.³ An order can be entirely firm and still slip because something else in the project did. The contract should say what happens then: storage charges, revised payment dates, or the supplier's right to sell the slot to someone else.

Counterparty strength. A commitment is only as good as the balance sheet behind it. A large customer with cash on hand and a small developer relying on financing it has not yet raised can sign identical forms. The second signature transfers less risk, because the developer may be unable to pay what it owes.

Even the best contract does not guarantee that the customer pays on time or that the equipment is commissioned on schedule. What a contract does is name, in advance, who loses what if things go wrong. That is the information a forecast cannot supply, because a forecast has no counterparty.

Build it all, or build it in stages

The supplier has two broad choices. It can build the whole second line now, adding 120 units a year once the line is ready. Or it can build half now and decide on the other half a year later, once it sees which reservations have turned into orders. The staged route has a cost: the second half arrives later, and in a strong market some customers will go elsewhere. The full route has a cost too: if demand comes in low, the new line sits partly idle and the supplier carries it.

Economists have a precise way of describing this trade. Avinash Dixit and Robert Pindyck showed that when an investment is hard to reverse and its payoff is uncertain, the ability to wait for information has a value of its own, and a rule that ignores that value will approve investments too readily.⁴ Their framework does not say that waiting always wins; waiting forfeits whatever the investment would have earned in the meantime. It says that the cost of waiting and the cost of committing both belong in the comparison.

The worksheet below makes the comparison concrete for the invented supplier. Pick a demand scenario and set how many units of the new capacity are backed by firm orders. The worksheet keeps three outcomes apart instead of netting them into one figure: capacity that sits idle, sales that are delayed because capacity is not ready, and capacity that rests on the forecast alone, which is the part exposed if interest never turns into orders.

At the starting values, the central scenario with 45 units of firm orders, full expansion leaves 30 units a year idle and delays nothing, while staged expansion leaves nothing idle and delays 30 units. In the high scenario the delays grow for the staged line; in the low one, the idle capacity grows for the full line. Neither route wins in every future, which is the point. What the worksheet does show plainly is the third row. The full line places 75 units of capacity on the forecast alone, and the staged line places 15 there. That is the speculative part of each plan, and it is the part someone has to agree to carry.

The recommendation that follows is modest. Match the next irreversible commitment to the strength of the evidence and to the party who bears the downside. If firm orders cover most of the full line, build it. If they cover the first half and the forecast covers the rest, build the first half, prepare the second (buy the land, place the long-lead orders that can be resold or cancelled cheaply), and decide in advance which level of firm orders will trigger it.

Objections, considered

Building only against firm orders means building too late. Often that is true. With transformer lead times approaching a year, a supplier that waits for every order to be signed may watch its customers buy elsewhere. Capacity built ahead of demand to win a position can be the right call. Nothing here forbids it. The argument asks that such capacity be named as a bet on the forecast, sized, and approved as a bet, rather than presented as backed by demand that has not yet committed.

Capacity teaches. A second line may be worth building partly for what the company learns by running it: better yields, new designs, a trained workforce. That is a real benefit. It belongs in the case as its own line, with its own estimate, not folded into a demand number.

Orders can be weaker than they look. Several orders may depend on the same customer's financing or the same grid connection, so that one setback cancels many at once. Orders can also be contingent on permits or on the customer's own customers. This objection strengthens the argument rather than weakening it. If a commitment's value lies in its terms and its counterparty, then the case for capacity should examine both, not just count signatures.

This is a bubble call in disguise. It is not. Nothing here implies that data-center announcements are double counted or that today's demand is imaginary; claims like those would need evidence of their own. A forecast can be entirely right about a market and still leave a particular factory empty, because the market's growth can go to a competitor, arrive later than planned, or be served by a different technology.

A forecast describes the opportunity; a commitment tells you who carries the risk.

Before you approve the new line

• How much is committed? The share of the proposed capacity covered by firm orders with deposits, and the shares covered by reservations, quotations and the forecast alone.
• What would cancel it? For each large commitment, the terms under which the customer can walk away, and what it would owe at each stage.
• Who else has to deliver? The grid connections, permits and other equipment each order depends on, and what the contract says if one of them slips.
• What can be staged? The part of the investment that can wait for more evidence, what waiting would cost in delayed sales, and the order level that triggers the next stage.
• Whose balance sheet carries the rest? The capacity resting on the forecast alone, stated in units and in dollars, with a name next to it.

Back to the two sheets on the desk. The outlook is a good document; it is the reason anyone is thinking about a second line at all. But it has no signature line, because nobody stands behind a forecast the way a customer stands behind an order. The order form has one, and it is still blank. Until it is signed, and until the supplier knows what that signature commits the customer to, the second line is a bet on the outlook. It may be a good bet, as long as everyone approving it knows that is what it is.