Remarks to the Residual Value Dinner, Delivered by a Retired Bond Insurer the Night Before the Fed

Remarks to the Residual Value Dinner, Delivered by a Retired Bond Insurer the Night Before the Fed

(Transcript, lightly edited for the parts people laughed at.)

Thank you. Sit down, please. I'm told the headline number is $63 billion, so let me start with the number that isn't in the headline.

That number is the guarantee.

Here is the structure for anyone who skipped the press release. A newly formed special purpose vehicle buys Broadcom's custom accelerators and networking gear and leases them to Anthropic. Bank of America, Citi and Morgan Stanley arrange the first lien, reportedly about $43.5 billion, or 69% of the stack. Blackstone writes the second lien. Deliveries start in 2027. And Broadcom provides what the documents politely call residual value support. The first-lien lenders are covered if the hardware loses value, and Broadcom absorbs those losses up to the full amount of that senior debt.

I wrote guarantees for a living, so let me translate. Broadcom sells the silicon today and books the revenue today. In exchange it holds a contingent liability that matures when the silicon is old. Broadcom stock rose about 4% on Tuesday. The market capitalised the premium and left the claim for later, which is how the market has always handled premiums.

My old shop wrapped mortgage paper. We were triple-A, we were "one of the strongest names in the sector," and I believe Moody's used something like that phrase about Broadcom this week, before adding that the risk climbs if several of these deals land in quick succession. Read that sentence twice. The agency isn't worried about this deal. It's worried about the sequel, and the sequel is already in the pipeline: Apollo and Blackstone agreed to finance a $35 billion expansion under an earlier arrangement, and Anthropic filed for what could be a record IPO on Monday.

(pause)

Now the part I'd like the young people in the back to write down. The value of a used aircraft depends on the economy, on fuel prices, on whether the airline across the street goes bust. It does not depend on whether Boeing decides next spring to make last spring's airplane obsolete. A chip's residual value depends on exactly that. The guarantor of the used accelerator is also the sole author of the new one. Broadcom has underwritten a depreciation curve that Broadcom itself draws. Every roadmap slide it presents from here on is, in effect, a claims-adjustment memo. I'm not suggesting anyone is cynical enough to play it that way. I'm suggesting that a conflict of interest does not need cynicism, only a calendar.

And the calendar is already crowded. Look at what happened on the very same Tuesday. Nvidia rose 0.9%. CoreWeave jumped 7.9%, the same CoreWeave whose long paper was yielding mid-teens a fortnight ago. The S&P 500 gained 0.6% to 7,722.60, the Nasdaq added 1.2% to 26,477.81, and the Dow moved 16 points. A majority of S&P stocks fell. The index went up because six or seven names went up. When a market this size is held aloft by the same few balance sheets that are also guaranteeing the financing for each other's customers, you don't have diversification. You have one risk with several tickers.

(sip of water)

Meanwhile, in Seoul. SK Hynix reported record operating profit of 83.3 trillion won, roughly $62.8 billion, nearly sevenfold the year before and ahead of estimates. The stock jumped about 7%. The Kospi was up around 5.6% early Wednesday near 5,950, after a 3.1% gain on Tuesday, in a month where it had been cut by something like a fifth. I say this with respect: a record profit at the memory maker tells you the chips are being bought. It says nothing about whether the buyers can carry them. Every insurer I ever knew posted a wonderful loss ratio the year before the claims arrived.

You'll want me to say something about the Fed, because the Fed is the reason half this room is wearing a tie. Warsh decides at two o'clock Eastern tomorrow. Hold is priced. A December hike is priced. The 10-year sat at 4.99% on Tuesday, up about 84 basis points this year, and Brent slipped 1.2% to $109.41, which is the only reason anyone is calling it calm. I'll tell you what one portfolio manager told Reuters, and I'll paraphrase because I'm a guest: capital has been leaving AI trades for bonds, and flows like that can swamp small moves in Fed expectations. Hear it the other way round. If flows that size can swamp the Fed, then the thing being swamped isn't the Fed. The long end of the curve is now an AI risk gauge with a Treasury label on it. When the guarantees get tested, you'll see it there first, and the central bank will be the last to know.

Let me finish with the thing that kept my old firm alive for forty years and killed it in eighteen months. We priced the guarantee off the history of the asset. The history was good. The history is always good, right up to the point where the asset class becomes large enough to change its own history. Mortgage defaults used to be uncorrelated because nobody had ever lent against everyone's house at once. GPU collateral has the same property today. Nobody has ever had to remarket a few hundred thousand accelerators into a market where the guarantor is simultaneously trying to sell the replacements.

Maybe it works. Perhaps demand grows into every chip, the lease payments keep arriving, and Broadcom's guarantee stays a footnote in a prospectus nobody opens. I hope so. I own some of the paper.

But if you are going to underwrite it, charge for the right risk. Don't charge for the chip's age. Charge for who controls the clock.

Thank you. Enjoy the veal.