When Compute Becomes a Financial Product: What's Actually Different From 2008

Credit means that a certain confidence is given, and a certain trust reposed. Is that trust justified? And is that confidence wise? These are the cardinal questions.

— Walter Bagehot, Lombard Street (1873)

What the episode is about

A MacroMicro feature from 2026-09-02, hosted by Vivianna, with the question right there in the title: if AI compute gets securitised, does that set up the next subprime crisis?

The starting point is two things happening at once. Nvidia and Wall Street launching a compute financing programme on the order of five hundred billion dollars, and the Chicago exchange listing the world’s first compute futures. She reads both as the same signal: compute is moving from physical data centre infrastructure into something you can trade, pledge as collateral, and lever up.

The first time I heard the words “compute futures,” my reaction was probably the same as everyone else’s — that sounds extremely 2008. What makes the episode worth an hour is that it doesn’t stop at the gut reaction. It lays 2008 and today side by side, item by item, and the conclusion it lands on is more useful than either “yes it’s the same” or “no it isn’t.”

The main points

One. What securitisation really does is turn people who couldn’t lend into invisible lenders. Package future compute leases, hardware, and data centre cash flows as the underlying asset and sell it on. The upside is obvious: the tech giants get funded fast, and banks lend more willingly because there’s collateral. The downside is the same fact seen from behind — investors with no direct connection to that loan are now wired into the same line. Mortgage-backed securities, and the collateralised debt obligations sliced out of them, did exactly this.

Two. The dangerous property of a credit cycle isn’t that it rises, it’s that it self-reinforces. Asset prices rise, collateral looks more valuable, financing terms loosen, more money comes in, prices rise again. The line she keeps returning to: precisely because it self-reinforces, it can invert overnight rather than gently cooling off.

Three. The giants are stuck in a textbook game. Under the shared premise that the AI market keeps expanding, nobody dares slow capex unilaterally — the one who blinks loses share and competitiveness outright. So everyone is forced to keep the arms race running.

Four. But that equilibrium flips, and it flips fast. Once the market starts doubting monetisation, the premise changes: now the company that announces capital discipline first, or even cuts capex, is the one more likely to protect its share price from being hunted. That was my favourite part of the episode — the identical action gets the opposite verdict either side of the premise flip. And once any one of them fires that first shot, the rest follow within a very short window.

Five. Four structural differences. On the lender side, 2008 was the scarce-reserves era, where any tremor in liquidity could break the system; today reserves run around three trillion dollars and money market funds sit near eight trillion, and the lenders aren’t only banks anymore — plenty of non-bank institutions are in the game. On the borrower side, a lot of the 2008 participants had no income and no assets; today’s borrowers are companies with enormous balance sheets and real operating moats.

Six. The collateral difference is the counter-intuitive one. Houses depreciate slowly and meet a hard need, which sounds like ideal collateral — but it’s exactly that belief in long life and permanent demand that grew the illusion that prices only go up. GPUs have short lives, forcing the giants to squeeze out cash flow and recover capital fast, which sounds more dangerous. Yet because the asset depreciates quickly, lenders are forced to stay cautious about the collateral. The weakness becomes the brake.

Seven. The deepest one: whether the ceiling is fixed. Real estate’s addressable market is roughly fixed — population and housing need can’t compound severalfold in three to five years, so an overflow of credit can only produce too many buildings, oversupply, and a price collapse. AI runs the other way: borrow more, spend more, and you improve the odds of a technical breakthrough; even falling prices per model call plausibly widen the application space and create revenue. Her conclusion is that this isn’t a straight replay of 2008 but a capital-driven bet on productivity — conditional on the technology actually continuing to break through, which is why it all comes back to real deployment and real monetisation progress.

In her own words: history rhymes, it doesn’t repeat.

Going further

”Everyone’s saying bubble — should I get out of what I’m holding?”

That was my own first thought, so let’s start there.

Analogy is a powerful intuitive tool — powerful enough that it routinely skips the verification step. “This is exactly 2008” sounds like a conclusion, but it’s a hypothesis, and it’s the particular kind of hypothesis that stops you thinking, because it’s so smooth and so vivid.

What the episode does is pry that hypothesis open. It never says “not alike.” It lists four specific places where things differ, and each one is concrete enough to be falsified. That’s the whole difference: a judgement about resemblance can’t be checked, a list of differences can.

So I rewrote those four differences into four things I’d actually go look at: whether financing terms are tightening or loosening, whether the borrower base is spreading outward from cash-rich companies, whether depreciation assumptions on the collateral are being quietly stretched, and whether the ceiling is still expanding. If the first three start moving the wrong way, “this time is different” begins to expire. I won’t claim four is enough — but they’re checkable, and “feels bubbly” isn’t.

“If nobody dares stop first, how would I ever know when it turns?”

I sat with this part for a while.

What the game theory section is really saying is that the current sprint rests not on facts but on shared expectations. Everyone believes the market keeps expanding, so nobody can stop. Which means the thing that has to change for it to reverse isn’t the facts either — it’s the expectation. That’s exactly why it can invert overnight instead of cooling gradually.

There’s a practical corollary for anyone watching prices: the price isn’t what moves first. Price is what happens after expectations turn, not the cause. The earlier signals are the ones tied directly to expectations — a change of language in one company’s capex guidance, an earnings call where “accelerating investment” becomes “investing with more discipline,” lenders starting to demand a spread for the risk, lease terms getting shorter or stricter. None of it is exciting and none of it leads the news, but it’s the trail expectations leave as they move.

My own past mistake ran the opposite direction: I believed too early that I’d spotted the turn, jumped out of an equilibrium that hadn’t broken yet, and missed a long stretch for nothing. So what I do now is write down what I’d need to see before it counts, and write it before I go look, rather than after. Not necessarily right, but at least it blocks my habit of constructing the story afterwards.

”So what am I supposed to do with this episode?”

Honestly, the line between what it can and can’t support is pretty clear.

It supports structural judgements: what this financing chain looks like, how risk propagates, which conditions make it sturdier than 2008, which conditions breaking would make it fragile. It doesn’t support timing judgements — nobody in that episode tells you when it turns, because that isn’t within the reach of this kind of analysis in the first place.

That distinction feels worth practising. The same good source is genuinely helpful on the questions it’s built for and turns into self-issued permission on the questions it isn’t. After enough financial content, the common failure isn’t consuming something false — it’s taking something true and using it to answer a question it was never answering.

Worth a look

• MacroMicro, 2026-09-02 feature: “If AI compute gets securitised, are we heading for another subprime crisis?”
• CME Group’s product pages on compute futures
• Federal Reserve data on bank reserves and money market fund assets (H.4.1 and the FRED database), if you want to check the three-trillion and eight-trillion figures yourself
• Walter Bagehot, Lombard Street (1873, public domain) — on how credit gets built and how it vanishes in a night. Written a hundred and fifty years ago and it doesn’t read old at all

The one thing to take away

One idea: “is it similar” isn’t a judgement — “where is it different” is.

When you catch yourself saying “this is exactly like last time,” that sentence usually isn’t the result of analysis; it’s the place analysis stopped. What’s good about this episode isn’t that it declares AI not-subprime. It’s that it forces itself to name four specific differences — you can only falsify a list, and only what can be falsified counts as a judgement.

Here’s something I’ve tried that you’re welcome to steal, and it works nowhere near investing too. Pick something you recently concluded with “same as last time” — a colleague missing another deadline, a relative raising that topic again, a friend restarting the same plan. Write down three specific ways this time differs from last time (not “it feels different” — timing, people, circumstances, stakes, things you could verify). Then circle one of them and write beside it: “if even this one matches, I’ll admit my analogy holds.”

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.