The Forty Billion Dollar Bet

0:01 Edmund Hale — This is The Fathom Report, extended edition, with Edmund Hale.

0:04 Edmund Hale — A word before we start. This edition draws on The Leadsman, the analytical newspaper. Where something is its judgement rather than established fact, we will say so.

0:13 Edmund Hale — The headlines. SpaceX has closed what the paper calls the largest private debt commitment yet for artificial intelligence infrastructure.

0:21 Edmund Hale — Elon Musk has warned that more than two hundred smaller neocloud firms face failures and consolidation.

0:28 Edmund Hale — South Korea reported a surge in chip exports. Japan stationed missiles on an island near Taiwan. And DeepSeek released software that lists Huawei's chips beside Nvidia's.

0:37 Edmund Hale — After three short reports, our editor, Rose, joins me to take the first story apart: a forty billion dollar financing, and the question of what, exactly, it is secured on.

0:48 Edmund Hale — First, South Korea. Last week the reporting turned to customs data on semiconductor exports.

0:53 Edmund Hale — They came to twenty-seven and a half billion dollars in the first twenty days of October, up more than two hundred and twenty-two percent on a year earlier.

1:02 Edmund Hale — Samsung's preliminary operating profit for the third quarter was about one hundred and seven trillion won, roughly eighty billion dollars, up about seven hundred and eighty-three percent.

1:13 Edmund Hale — Revenue rose about one hundred and twenty-seven percent.

1:16 Edmund Hale — The piece credits high-bandwidth memory, the stacked memory that Nvidia's chips require. SK Hynix is Nvidia's main supplier, it says, and Samsung the second.

1:24 Edmund Hale — Its point is about timing. On the Thursday the Kospi fell one point eight percent and Nasdaq contracts nine tenths of a percent, as investors waited for Meta and Tesla results.

1:34 Edmund Hale — The memory makers have been paid; their customers still have to show returns.

1:39 Edmund Hale — Next, Japan, and Keen Sword, the Japanese and American exercise that began on the nineteenth of October and runs to the twenty-ninth.

1:46 Edmund Hale — It is the largest of its series, it says: about forty-five thousand Japanese personnel, twelve thousand American, forty ships and two hundred and sixty aircraft.

1:55 Edmund Hale — Its notable feature, in the account, is geography. Japan placed its Type Twenty-Five anti-ship missiles on Yonaguni Island, about one hundred and ten kilometres from Taiwan.

2:05 Edmund Hale — The Leadsman describes that as a first outside a permanent base, but its own source list marks the claim as resting on a single source.

2:12 Edmund Hale — France, Germany and the Netherlands took part as full participants for the first time, not as observers.

2:18 Edmund Hale — The article stresses that this is an exercise: the launchers are to be removed when it ends, and no permanent basing has been announced.

2:25 Edmund Hale — Its forecast, at moderate confidence, is that Japan will announce permanent basing for at least one anti-ship missile unit by the thirty-first of March twenty twenty-seven. That is a forecast, not a decision.

2:37 Edmund Hale — Then China's DeepSeek.

2:39 Edmund Hale — On the twentieth of October, according to the reporting, it released DeepInfer, free software for running one of its current models.

2:46 Edmund Hale — It lists Nvidia's current chips and Huawei's Ascend nine hundred and fifty processors side by side as supported hardware.

2:53 Edmund Hale — The piece is careful here. Equal billing, it says, describes the documentation, not the performance, and no independent comparison of the two chips was available.

3:02 Edmund Hale — It adds that the first version handles offline batch jobs only, with no memory reuse between requests and no load balancing, so a company must build those parts itself.

3:12 Edmund Hale — Its forecast: by the twentieth of April twenty twenty-seven, at least two more major Chinese laboratories will publish similar software with Ascend support. A forecast to check in April.

3:22 Edmund Hale — Now to the story at the top of the bulletin. With me is our editor, Rose. Rose, welcome.

3:28 Rose Calloway — Thank you, Edmund. A forty billion dollar bet, and I am told the interesting part is not the forty.

3:32 Edmund Hale — The facts first, as the reporting gives them.

3:34 Edmund Hale — On or around the twenty-fifth of October, Elon Musk confirmed that SpaceX had closed about forty billion dollars of callable debt financing to buy Nvidia chips, with deliveries due in the first half of twenty twenty-seven.

3:46 Edmund Hale — About ten billion is bank loans and about thirty billion is investment-grade bonds. Goldman Sachs led the investor discussions, and Apollo Global Management is among those taking part.

3:57 Edmund Hale — The article calls it the largest single debt commitment for artificial intelligence infrastructure by any private entity. Rose, what stood out first?

4:05 Rose Calloway — The structure, not the size. The article itself says the big cloud companies have issued larger investment-grade debt. What is new is the collateral.

4:14 Rose Calloway — The bonds, it says, are backed mainly by Starlink subscription revenue.

4:17 Rose Calloway — It puts that at about four point six million subscribers paying a hundred to a hundred and twenty dollars a month. Here is my own sum, not the paper's: that is about five point five to six point six billion dollars a year.

4:29 Edmund Hale — Which squares with the paper's line that it is in excess of five billion.

4:33 Rose Calloway — It does. But the paper says the subscriber count is an estimate, and that SpaceX has not disclosed the exact figures. So the foundation of the argument is an assessment, which the paper rates as high confidence, not a reported number.

4:44 Edmund Hale — Why does the collateral matter so much?

4:46 Rose Calloway — Because, the analysis argues, the companies that rent out chips, the so-called neoclouds, have generally been unable to borrow this way.

4:53 Rose Calloway — Their customers are on short contracts and are often young firms of uncertain longevity, so the revenue is neither recurring nor contracted far ahead.

5:01 Rose Calloway — SpaceX, it says, solved the problem not by being a better artificial intelligence company but by being a better credit. The satellite revenue existed before the chips and is separate from any bet on them.

5:10 Edmund Hale — Then I would put the case more strongly than the paper does, and that is my own reading. If the lender is repaid from broadband bills, the chips are almost incidental. The deal is a bet on Starlink more than on artificial intelligence.

5:23 Rose Calloway — I would disagree a little. The money buys chips, and the paper reads the callable structure as SpaceX expecting chip prices to stay high through the first half of twenty twenty-seven, while keeping room to refinance.

5:34 Rose Calloway — So the chip market is in the structure.

5:36 Rose Calloway — Still, I concede the direction of your point. The lenders are underwriting the satellite business first.

5:41 Rose Calloway — Here is another of my own sums. Forty billion dollars against five and a half to six and a half billion a year is a little over six years of the whole subscription revenue, up to about seven and a quarter.

5:51 Rose Calloway — Only the thirty billion of bonds is described as backed by Starlink, and thirty billion over the same range is about four and a half to five and a half years.

5:59 Rose Calloway — And revenue is not profit. The article gives no bond term and no interest cost, so I cannot tell whether that revenue covers the debt.

6:06 Edmund Hale — The piece says an investment-grade rating requires projected revenue to more than cover debt service across the term. Does it show that?

6:14 Rose Calloway — It asserts it. It gives no term, no coupon and no rating agency view, so the claim is plausible and unverified. I would want the bond documents before leaning on it.

6:22 Edmund Hale — Now the warning. Musk said more than two hundred neocloud firms face widespread failures and consolidation.

6:28 Rose Calloway — The logic is about the cost of capital. If a borrower of SpaceX's standing can raise cheap money, the cost of computing falls for SpaceX and, by extension, for Amazon, Microsoft and Google.

6:38 Rose Calloway — Small chip renters, it says, cannot match their prices.

6:40 Rose Calloway — Note whose voice that is. The warning comes from someone who has just borrowed on terms the sector cannot get.

6:46 Rose Calloway — The article says as much, calling him a market participant. That is not a reason to dismiss it, but it is a reason to weigh it.

6:52 Rose Calloway — My harder question is this. The paper says chip supply is normalising, and that weakens the neoclouds, whose pitch was access to chips when nobody else could get them.

7:00 Rose Calloway — Yet the same article describes a buyer locking in forty billion dollars of chips for delivery next year. If chips were easy to get, why lock in?

7:07 Edmund Hale — Because easier supply and unlimited supply are different things.

7:11 Rose Calloway — I concede that. But the evidence the article offers for normalising is a record quarter of revenue at a chip manufacturer.

7:17 Rose Calloway — That is evidence of production, not of price, and the scarcity premium is the thing that matters to the renters.

7:22 Edmund Hale — The paper then makes a dated forecast, which is not yet an outcome.

7:26 Edmund Hale — At least three neocloud firms with more than five hundred million dollars of chip-backed debt will file for bankruptcy protection or announce merger terms by the thirtieth of June twenty twenty-seven. Moderate confidence.

7:39 Rose Calloway — Two observations. First, my own arithmetic: three firms out of more than two hundred is under one and a half percent of the sector.

7:45 Rose Calloway — That is a far lower bar than the headline warning of widespread failure, though it counts only firms with large debts.

7:51 Rose Calloway — Second, merger terms count. Mergers happen in healthy industries as well as failing ones, so the forecast is easier to meet than the warning suggests.

7:58 Rose Calloway — The piece names one way to be wrong: a demand surge that lifts how much the renters can sell. That, it says, would delay consolidation but not prevent it.

8:06 Edmund Hale — And the wider worry, concentration?

8:08 Rose Calloway — The reasoning is that if three to five groups control accessible computing, rules written for the model layer do not reach the infrastructure.

8:15 Rose Calloway — Antitrust built on market share and consumer harm may also miss harm to new entrants, researchers and governments.

8:20 Rose Calloway — It calls the deal not a competition event but a directional signal. I think that restraint is right.

8:25 Edmund Hale — If you had to put the piece in a sentence?

8:28 Rose Calloway — The argument is that to accumulate computing at scale you need investment-grade credit, and that needs recurring revenue which can survive a downturn in artificial intelligence.

8:37 Rose Calloway — I would add that the revenue is estimated, the repayment terms are unseen, and the sector warning comes from an interested voice.

8:44 Rose Calloway — Its argument about structure holds. Its timetable for casualties is a forecast.

8:48 Edmund Hale — Rose, thank you.

8:50 Rose Calloway — Thank you, Edmund. Mind the collateral.

8:52 Edmund Hale — That was The Fathom Report, extended edition. Our thanks to Rose, and to The Leadsman, whose analysis we drew on throughout. This is Fathom FM.