$3.7B gone: SoftBank's turbine deposits hit AI's gas ceiling

SoftBank just burned $3.7 billion on power equipment it no longer wants — 52% of its $7.1 billion quarterly loss. Thirty cents of every dollar it prepaid for AI data-center power is gone, and the deposits were non-refundable.

What actually happened

On Tuesday, SoftBank reported a net loss of ¥1.12 trillion (about $7.1 billion) for the July–September quarter — its first quarterly loss in almost two years, where analysts had expected a ¥129 billion profit. The single biggest item in it was a ¥597.2 billion (about $3.7 billion) charge on power-generation equipment for its US AI data centers.

The charge splits into two parts, and the split matters:

• ¥457 billion ($2.85 billion) — advance payments to turbine and equipment makers that cannot be refunded. SoftBank paid to reserve manufacturing slots, then changed its buildout plans. The suppliers keep the money.
• ¥140.2 billion ($0.88 billion) — contract changes and settlements with suppliers.

The vehicle behind it is Silver Bands 7, co-owned by SoftBank and Vision Fund 2. Over the past 18 months it paid a cumulative $9.48 billion for power equipment — reserving factory capacity and buying gear for US AI data centers, led by the 8-gigawatt PORTS-Pike campus in Pike County, Ohio. Do the division: $2.85 billion of $9.48 billion is 30.1% of everything prepaid, vaporized. The remaining $6.6 billion is still assigned to projects, per SoftBank's earnings presentation.

Why SoftBank over-ordered in the first place

This was not a clerical error. It was the rational move in a broken market, and that is what makes it interesting.

Starting about two years ago, SoftBank put down deposits on more than 13 gigawatts of turbines — against a 9.2-gigawatt plan announced in February (a $33 billion gas plant in Ohio, part of the US–Japan investment framework). That is roughly 40–50% more turbines than it needed. Turbine order queues now stretch into the early 2030s, and SoftBank's CEO Masayoshi Son bet that with a big enough war chest he could corner a slice of the market: if the AI world wanted power, it would have to come to him.

"SB Energy did not have a choice except to overbook turbines in the future if they wanted some turbines right away," Peter Gardett, CEO of Noreva, told TechCrunch. And his line on the loss itself is the one to remember: "This charge is just that price discovery finally happening."

That is the right frame. The $3.7 billion is not a fraud or a demand collapse. It is what it costs to discover, after the fact, what a place in the turbine queue was actually worth.

The same week, the same queue, the opposite trade

Here is the part that stopped me. Three days before SoftBank's earnings, Anthropic paid $3.4 billion to buy gas turbines — 1.7 gigawatts, the largest single power-equipment purchase by an AI lab, which I wrote about on Monday. SoftBank's write-off on turbine deposits is 1.09x Anthropic's entire purchase. Same turbine queue, same week, opposite sides of the same trade: one lab buying in, one investor writing off what it already bought.

The market noticed. SoftBank shares fell as much as 11% in Tokyo the day after the results. GE Vernova — the turbine maker — fell as much as 9.9% on November 10, then said SoftBank's disclosure has no financial impact on it. The suppliers, it turns out, are the ones who got paid either way.

Elon Musk, who posted on November 6 that "SoftBank is defaulting on their turbine orders," was directionally right about the walking away and wrong about the default. CFO Yoshimitsu Goto said at the briefing: "This is simply not correct at all. There is no such default." SoftBank also says it is not downsizing PORTS-Pike, and management insists data-center power supply is still far short of demand.

So is this a cap on gas, or a cap on over-ordering?

TechCrunch's headline says the loss "may have put a cap on AI's natural gas ambitions," and I think the honest version is narrower: it puts a cap on the strategy of buying your way to the front of the queue with non-refundable cash. The gas itself is still riding high — AI's power demand is too large for mass cancellations, and SoftBank's own losses are "unlikely to have a domino effect," as TechCrunch puts it.

But the alternatives are getting a real hearing now. Brent Nelson, senior managing director at Ascend Analytics, told TechCrunch's Equity podcast that most data centers would be best served by a grid connection — a smaller one than they would normally request — with large batteries to shave peak loads or ride out curtailment requests from the utility. Gas looked like the safe, proven bet. It still carries winter curtailment risk when supplies dip and demand spikes.

The number I will be watching next is the $6.6 billion still assigned to projects. If SoftBank's remaining prepaid equipment gets absorbed into PORTS-Pike and the other campuses, this stays a one-off procurement lesson. If more of it gets written down, then the ceiling is lower than anyone thought — and every other builder that prepaid for scarce inputs has the same hidden exposure sitting on its balance sheet.

One more thing worth saying plainly: this is a procurement-sequencing loss, not an AI-demand signal. Do not read it as "the AI boom is slowing." Read it as "the AI boom's power bill just got its first real invoice, and the fine print said non-refundable."

Sources: TechCrunch, "SoftBank's billion-dollar loss may have put a cap on AI's natural gas ambitions" (November 12, 2026); Bloomberg and Wall Street Journal reporting on SoftBank's November 10 results (¥1.12 trillion net loss, ¥597.2 billion equipment charge); CNBC reporting on the $9.48 billion cumulative payments via Silver Bands 7; SoftBank earnings presentation figures as reported by CNBC and AI Stock Wire; US Commerce Department fact sheet, February 17, 2026 (9.2 GW, $33 billion Portsmouth/Pike County project). All ratios in this article are my own arithmetic from those reported figures.