The Data Center Buildout Has a Toilet Paper Problem

In 2020 the United States did not run out of toilet paper. Production never fell. People who normally bought one pack bought two, entirely reasonably, and the shelves emptied anyway. Then the empty shelves convinced everybody else to buy three.

Nobody behaved stupidly. Ordinary caution, repeated by millions of people who could not see each other's cupboards, did all of it.

That pattern has a name, and MIT has been demonstrating it for decades. Supply chain people call it the bullwhip effect, because a small movement at the customer end produces a violent one at the factory end. Prof. John Sterman has a name for the move that drives it, ordering more than you need because deliveries are slipping. He calls it phantom ordering, and the paper is titled "I'm not hoarding, I'm just stocking up before the hoarders get here."

The teaching tool is a board game about beer, created by Jay Forrester at MIT and studied extensively by Prof. John Sterman. Four players form a supply chain. A retailer, a wholesaler, a distributor and a brewery. Real customer demand holds steady, steps up once from four cases a week to eight, and then never moves again. Nobody is told that. The one thing players cannot do is see each other.

Somebody orders a little extra to cover a delivery running late. Their supplier runs out. Now the people waiting are getting a fraction of what they asked for, so they ask for more, to be safe. Their suppliers read the bigger orders as growth and build for it. Sterman and Gokhan Dogan ran a version where none of that can pay off. No competition between customers for supply, no capacity limit, no randomness, and every player told that real demand is constant. Twenty percent of them hoarded anyway, raising their targets and their orders when deliveries faltered, which produced a demand bubble and cost them money. You can play it yourself, and most people who do order too much on their first try.

Sterman and Dogan are careful about one thing that matters here. Ordering ahead can be perfectly rational when several customers are competing for limited supply, which is exactly the turbine market. Their experiment stripped that competition out to see what was left underneath. What was left was still 20 percent.

I played it in Prof. Nelson Repenning's class at MIT. Nothing else in the curriculum lands the same way, because everyone leaves knowing they would do it again.

Now look at the turbine order book

The headline number in power equipment is 116 GW, up from 100 GW in a single quarter, with at least 125 GW expected by year end. It is read as proof that demand for power stays extraordinary for years, so paying up today is safe.

That figure is two different things, and GE Vernova reports them apart. Equipment backlog went from 44 to 53 GW. Slot reservations went from 56 to 63 GW. A backlog entry is a signed order. A slot reservation is a held place in the build queue, taken by a buyer who is not certain they need it and is certain they cannot get it later if they wait.

Reservations are the larger half, and they were still the larger half at the end of the quarter. The company signed 20 GW of new orders and reservations in those three months. It booked 12.1 GW of gas turbine orders and shipped 3.3 GW. The combined book grew by 16 GW anyway, because the reservation side grew alongside the order side.

That is the difference between what people need and what people ask for when they are afraid of being last. It is also, precisely, what the players do in the game.

I wrote in April that the turbine, not the interconnection queue, now gates 2028 gas. That still holds. The question here is different. If most of the constraint is made of options, what is the constraint measuring?

Most of this book is not data centers

One more number from the second quarter earnings call, and it complicates the title of this piece. Of the roughly 100 customers holding those 116 GW across 26 countries, Strazik said about 80 percent are traditional customers and 20 percent are data centers.

So the turbine book is not a data center order book. It is mostly utilities and industrial buyers.

That does not weaken the argument. It is the argument. In the beer game the retailer is the only player who can see the customer. Everyone upstream is reading order flow instead of demand, and the further upstream they sit the harder they swing. A utility watching data center announcements land in its territory and reserving turbine slots against them is playing the wholesaler's position exactly. The load belongs to somebody else, and the utility has less visibility into when and how much of it will actually arrive.

The evidence that cuts the other way

GE Vernova's management reads this the other way and says so plainly. Scott Strazik, on the second quarter call:

In the second half of the year, we expect to convert many of these SRAs into orders, driving continued growth in our backlog while achieving an important inflection point with gigawatts and backlog greater than SRAs.

That is a public forecast that the reservation half stops being the larger half within six months, made by the person with the best view of the book. It deserves a straight answer rather than a footnote.

Here is the straight answer. It is a prediction, and predictions can be checked. Backlog gained on reservations last quarter, growing 9 GW against 7. But it started 12 GW behind and is still 10 GW behind. Crossing over by year end means closing that remaining gap across two quarters, 5 GW a quarter against the 2 GW just achieved.

He may well be right. He is also the seller, and what he is forecasting is that the places he has sold will turn into purchases.

Two queues, and only one of them was free

Texas has made it expensive to hold a place in line for grid connection, which tells you somebody there concluded it had become too cheap.

Under ERCOT's Batch Zero process, approved by the PUCT in June 2026, a large electricity user posts security at a default of $50,000 per megawatt of the capacity it requests. For a 1,000 MW site that is $50 million. The rule pairs that security with a separate interconnection fee charged on the same per megawatt basis, and that one is not refundable at all.

Pull out later and 20 percent of the balance comes back within 60 days, while the other 80 percent goes to the transmission company and is applied as an offset to its rate base at its next rate proceeding. Miss a milestone in the phased energization schedule by six months and the same treatment applies, and ERCOT has to reallocate the capacity. An applicant also has to disclose whether it is pursuing a substantially similar request elsewhere, triggered by things like a delay of a year or more or a 20 percent change in the capacity asked for, which closes the obvious workaround of applying at several sites and picking later.

The level is the interesting part. That default was proposed at $100,000 per megawatt in March 2026 and cut to $50,000 before it landed. A regulator built a mechanism to price the option of holding a place, then halved the price of it under pressure from the people holding places.

ERCOT reported 137 new applications totaling roughly 140,000 MW of requested load out to 2036, still being processed, nearly all filed under the old rules where reserving cost very little.

A reserved turbine slot was never free, though. GE Vernova's chief financial officer put working capital at a $6.4 billion cash benefit in the quarter, driven mainly by higher down payments on increased orders and slot reservations at Power. Strazik said customer down payments are funding the expansion to 30 GW of annual output by 2030. What a slot costs per gigawatt is not disclosed, because those terms sit in private contracts. That the sums are material is not in doubt.

Exercising costs more than holding, too. First half orders were priced more than 20 percent above fourth quarter 2025 equipment orders, which the company puts down to higher-priced reservations converting into backlog, and it guided the second half to the higher end of a 10 to 20 point range.

So the position is expensive to hold and more expensive to take up, and the reservation book grew anyway. Texas had to price its queue because standing in it cost almost nothing. The turbine market has been charging all along and filled up regardless.

Phantom ordering does not need a cheap option. It needs a frightened buyer, paying real money, who would tell you quite correctly that they are being prudent.

What the game suggests that the order book does not

The correction may not be one clean glut. If the same behavior dominates here, the pattern is not a single overshoot that settles. It is oversupply, then a collapse in new orders, then another shortage later, priced by whoever held a position through the quiet part.

The damage lands furthest from the customer. In the game the brewery swings hardest, because it is furthest from real demand. Here that is not the turbine maker. It is the foundries making large castings and forgings, the high voltage transformer plants, and the specialist trades. That is where the swing arrives latest and largest, and where the balance sheets are thinnest.

This market has run something close to it once. ERCOT battery storage revenue per megawatt fell almost 37 percent in 2025 while installed capacity grew about 80 percent over the same year. Everyone answered the same signal, the supply landed together, and the return that justified building thinned. That is an analogy from an adjacent asset rather than proof about turbines, but it is the same market and the same buyers.

What the game changes about a negotiation

In the beer game the swings collapse the moment players can see real demand instead of the order stream. That is the whole result, and it maps onto three things here.

In an uncertain market, a cancellation right can beat a modest discount. A turbine bought at the top of a shortage is paid for across the twenty years that follow, including the years after the shortage ends. A few percent off the price does not undo that. The right to not take delivery does.

Orders and held places are different things, and the supplier reports them separately. So the share of a book that is firm is a fair question with a published answer.

The real demand signal is electricity being consumed, not capacity in a queue. ERCOT publishes it monthly. In June 2026, 8,926 MW had been approved to energize while 3,966 MW showed up in observed consumption, roughly 44 percent. That measure sums each customer's own monthly maximum, so it flatters the total.

The test Strazik has already set

GE Vernova publishes both halves every quarter, and its chief executive has attached a date to his version. If backlog passes slot reservations by the end of 2026, the held places were real demand waiting to be confirmed and this piece is wrong. If reservations are still the larger half going into 2027, the book kept filling faster than buyers turned held places into equipment.

That is a clean test. It has a number, a direction and a deadline, and none of the three are mine.

Except a firm order is not the end of the story

Prof. John Sterman raised the obvious hole in that test, in a public comment on this piece:

In past episodes, including the dot.com boom, even firm orders evaporated quickly once the bubble burst. People ordering turbines now may believe that if they end up not needing them they can sell their place in line to others, but that same logic failed in the dot.com, housing bubble leading to the Great Recession, and other speculative bubbles and crashes because nearly everyone thinks the same way and then finds that there is no demand for the units they want to shed. As a cautionary tale, during the dot.com bust the book-to-bill ratio of JDS Uniphase went from about 2 to negative for at least one quarter, cancellations outweighed new orders.

He is right, and it breaks the symmetry this brief was resting on. I treated backlog as the solid half and reservations as the soft half. Backlog is only the solid half while nobody is trying to leave. The resale argument fails for the same reason the queue filled in the first place, because everyone is reasoning identically and the exit is crowded at exactly the moment it is needed.

So the crossover is necessary and not sufficient. Backlog passing reservations would show the options were exercised. It would not show the orders will be honoured.

The good news is that cancellations are visible in the same disclosures. Backlog has to satisfy a simple identity: opening plus orders booked minus shipped equals closing. Last quarter that was 44 plus 12.1 minus 3.3, which is 52.8 against the 53 reported. It closes, so cancellations were not material. When they stop being immaterial, that identity visibly fails, and the size of the failure is the size of the cancellations.

One caveat on how precise that is. These figures are reported in whole gigawatts, so a quarter with a few hundred megawatts of cancellations would appear to close just as neatly. The identity catches material cancellations, not small ones. That is the level this argument turns on anyway, but it is worth saying before somebody runs the arithmetic next quarter and expects more from it than it can give.

The test is therefore two sided. Does backlog pass reservations, which is Strazik's claim, and does the backlog roll forward keep closing, which is Sterman's warning. Both are checkable every quarter from the same three documents.

Methodology

This is an argument about structure, not a forecast. It projects no load, price or delivery volume.

Turbine figures are from GE Vernova's second quarter 2026 disclosures, all dated 22 July 2026. The combined 100 to 116 GW figure and the 125 GW year-end expectation are in the press release. The 44 to 53 GW backlog, the 56 to 63 GW slot reservations, the 20 GW signed, the customer mix and the quoted forecast are from the earnings call. The 12.1 GW of orders and 3.3 GW shipped are from the investor presentation.

Some widely reported figures are deliberately not used. Several outlets broke the 20 GW into 18 GW of reservations and 2 GW of orders, and reported 10 GW converting during the quarter. Those numbers may well be right, but they appear in none of the three primary documents, so nothing here rests on them.

ERCOT figures are from the June 2026 monthly operational overview published 17 July 2026: 8,926 MW approved to energize against 3,966 MW of observed non-simultaneous monthly peak. That measure sums each load's own monthly maximum regardless of timing, so it runs above true coincident demand and the 44 percent is an upper bound. The 137 submissions totaling roughly 140,000 MW by 2036 are from the March 2026 TAC report.

Batch Zero provisions are from PGRR145 as filed 4 March 2026, approved by the PUCT on 18 June 2026. The filing itself sets security at $100,000 per MW. The $50,000 used here is from the PUCT framework in Project 58481, which cut it in March 2026, and it is a default rather than a flat rate. The rule adds a separate interconnection fee that is not refundable.

Storage figures are from the ERCOT Independent Market Monitor's 2025 State of the Market report: revenue per unit of capacity down almost 37 percent after normalizing for the larger fleet, total net revenue up about 8 percent, installed capacity up almost 80 percent.

No origin date is given for the Beer Game because the record does not support one, and the demand step from four cases a week to eight is the current version's pattern. Both are from Ignacio J. Martinez-Moyano, History of the Beer Game, System Dynamics Review (2024), doi 10.1002/sdr.1767. The 20 percent figure, the experimental conditions and the phantom ordering framing are from Sterman, J. and G. Dogan, "I'm not hoarding, I'm just stocking up before the hoarders get here." Behavioral causes of phantom ordering in supply chains, Journal of Operations Management, 39-40 (2015), pages 6 to 22, doi 10.1016/j.jom.2015.07.002. Those results are established for the game. Applying them to gas turbines is my own reading and is not tested.

Prof. Sterman commented publicly on this piece after it was published. His comment is quoted above and credited, including the JDS Uniphase figure, and the two sided test came out of it.

All analysis by Tafel Power from public sources.