Quacks Like a Bond

Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.

Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.

89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Start flowing free

Would you lend against a used GPU?

An actual loan, I mean. Your money, secured by racks of two-year-old H100s and the rental contracts hanging off them. Take a minute before you answer, because the market already has: one tally of the filings puts GPU-collateralized loans past $20 billion, and Lambda bundled $500 million of its chips into the first GPU-backed bond anyone has sold. The lender list reads like a private credit conference: Blackstone, PIMCO, Carlyle, Macquarie, JPMorgan.

I read deals the way I read code: skip the README, open the imports. For a decade the AI trade imported software's standard library, ARR multiples and retention cohorts and gross margin. The documents crossing my desk this month import from a different floor of the building: amortization schedules, residual value guaranties, take-or-pay clauses, loan-to-value ratios. That's the mortgage-and-aircraft toolkit, structured finance, and this August the whole trade switched dialects in public.

Nvidia had quite an August.

On the 10th, Nvidia signed MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build "AI compute infrastructure financing platforms": a plan to mobilize over $500 billion of third-party capital, with the compute itself pledged as collateral and the bonds issued out of special purpose entities. Jensen Huang said it plainly, and I would frame the sentence if I ran a credit fund: "We began by building chips; today, we are helping create a new class of productive, investable infrastructure: AI factories."

A week later, an 8-K. Nvidia is providing residual value guaranties, capped at $105 billion, on leases OpenAI will sign at a campus in Pike County, Ohio, a decommissioned uranium enrichment site that SB Energy will build and own, rented to OpenAI for 20 years. If OpenAI stops paying, Nvidia covers the shortfall. July's reporting had the number near $250 billion. Investors flinched, and it filed at $105 billion.

Read that structure again. The chip vendor is guaranteeing the resale value of the buildings that exist to house its chips. Aircraft manufacturers invented that move to keep planes selling. (Five weeks ago I wrote about where the hyperscalers' lease obligations hide. That was the liability side. This is the asset side: what all that debt believes it's secured by.)

Structured finance, for engineers, in one paragraph.

Take a stream of future cash flows. Wrap it in a shell company so it can live or die on its own. Borrow against it, slice the debt by seniority, and sell each slice to the buyer with the matching risk budget: the senior tranche to an insurer, the mezzanine to somebody braver, the equity, the first-loss piece, to whoever believes hardest. It's how America finances houses, cars, airplanes, and pipelines. It is also, famously, how it financed houses in 2006. The tool is neutral. It's now pointed at compute.

Wall Street got here before the press releases did. Data center securitizations grew from $2.4 billion in 2020 to $15.5 billion last year and are on record pace again in 2026, and Goldman counts nearly half a trillion dollars of AI-related debt issued this year alone. Then in late July the SEC's structured finance office, asked by Latham & Watkins, determined these deals aren't technically "asset-backed securities" under the statute, so the ABS disclosure and risk-retention rules don't bind. The new object failed the interface check, so they cast it to any and shipped.

The best-informed people disagree about the collateral.

Every structure above compiles down to one variable: what the chips, and the sheds full of them, are worth in year five. Watch the insiders argue about that number in GAAP. Amazon cut its server depreciation schedule to five years and ate a $700 million hit. Meta stretched its own to five and a half and booked a $2.9 billion gain. Microsoft, Alphabet, and CoreWeave hold at six. Same hardware, three answers, all audited.

The rental market votes too. An H100 fetched $7 to $10 an hour in early 2024 and under $4 by late 2025, with some providers under $2.

Aircraft finance works because a 737 flies for twenty-five years and a used one prices to the dollar: appraisers, blue books, a whole priesthood. The residual value of an H100 in 2031 is an argument. Twenty billion in loans and a $105 billion guarantee are already priced off somebody's side of it.

Credit changes the question you're paid to answer.

Software equity is a call option. The job is sizing the upside, and the downside rounds to "you lose the check." Credit is the opposite trade: capped upside, a coupon, and a single question, what do you recover when it goes wrong. Different math, and frankly a different species of investor. My diligence training is cohort curves and sales efficiency. None of it prices a refinancing wall, and the people who can price one work at Apollo, not on Sand Hill Road.

Software was the anomaly.

The grown-up version of the bull case deserves airtime, because it's good: this is what maturity looks like. Railroads, power grids, telephone networks, aviation. Every real infrastructure buildout eventually graduated from adventurous equity to boring collateralized debt, and cheaper capital per flop honestly does mean more AI per dollar. Software spent forty years as the exception, an industrial revolution that never needed a mortgage. Maybe AI is just industry again, and this is the paperwork of growing up.

Maybe. But maturity financing assumes a mature asset, one whose value curve both sides can agree on. The last time a dominant equipment vendor stood behind its own customers' paper, it was Lucent in 1999, lending telecoms billions to buy Lucent gear. The gear was real. The demand even showed up, eventually. Lucent still nearly died of its own receivables.

So, a prediction, in writing, where you can hold me to it. Nvidia's Ohio guarantee switches on when the campus goes ready for service, expected 2028. Before that happens, a GPU-backed structure somewhere, a neocloud loan or an ABS tranche, will restructure, and the fight at the center of it will be a depreciation schedule: a number four accountants signed and no market ever agreed to. Credit people like to say equity tourists call these things two years early. They're usually right. I'm writing it down anyway, and when the first one goes, skip the model benchmarks and read the appraisal.

— SWEdonym