Will compute-backed securities repeat the 2008 MBS fiasco? The honest analysis of the risks — stranded assets, ratings agencies, and how hedging changes the game.
Whenever a new "asset-backed" instrument appears, the comparison to mortgage-backed securities (MBS) isn't far behind — and for good reason. The 2008 financial crisis showed what happens when financial innovation meets corrupted incentives. So will compute-backed securities (CBS) end the same way? Here's the honest bull and bear case.
The fear is straightforward. If you securitise 10 years of GPU cash flows, and then a genuinely new computing architecture arrives — a post-transformer design, an optical or photonic chip, a fundamentally different way to compute — those GPUs become stranded assets. The revenue disappears, and the security's value collapses.
As one Wall Street observer put it, financial assets want predictable depreciation, and exponential technologies don't offer that. They come in nested S-curves: everything runs out, and something new arrives. AI is evolving monthly, which makes 10-year cash flow projections inherently speculative.
There's also the human risk. Lou Ranieri — the inventor of the mortgage-backed security — gave a sobering answer when asked about the 2008 collapse: "It's not the instrument that was broken. It's the ratings agencies getting corrupted." If the Dyson swarm is an obvious good investment but the ratings agencies get corrupted again, CBS could produce "another Big Short collapse."
Three arguments stand out:
1. Compute is fundamentally productive. The best a house can do is shelter you. A GPU cluster earns revenue every hour of every day — training models, serving inference, running workloads for paying customers. The underlying asset produces real, measurable, growing cash flow. That's a very different foundation than subprime mortgages.
2. No corrupt policy motive (yet). The MBS market was lubricated by an explicit policy goal: the American dream of home ownership, which pressured ratings agencies to look the other way. There's no direct analog here yet. The closest pressure point is the US-China AI race, which could push policymakers to lubricate private capital markets for compute — a risk worth watching, but not a foregone conclusion.
3. Hedging changes the game. This is the most important difference. In a sophisticated market, investors can hedge against both directions of the risk: an algorithmic breakthrough that tanks GPU values, or a Taiwan invasion that sends compute prices through the roof. Options and futures on compute give sophisticated actors the tools to price and manage these risks — which is precisely why many argue compute derivatives are a prerequisite for the $7 trillion AI infrastructure buildout to become investable at all.
The bull case gets a concrete boost from real-world behaviour. CoreWeave — one of the largest AI cloud providers — has clients under contract to 2029 for A100 chips, a GPU introduced in 2020. Six-year-old silicon, still under multi-year revenue contracts.
Why? Because once hardware is paid for, the marginal cost is just electricity turned into intelligence. And that conversion is improving: a modern 10-billion-parameter model outperforms GPT-4, and dozens fit on a single chip where GPT-4 needed 16. Old chips don't stop earning — they just get more efficient at what they do.
The honest answer: CBS carries real structural risk, and the MBS comparison is not frivolous. But the differences — productive underlying assets, revenue-linked returns, and the availability of hedging instruments — are material. The outcome will largely depend on one thing: whether the institutions rating and underwriting these securities maintain their discipline, or whether the US-China AI race corrupts the process the way housing policy did.
As with any new asset class, early participants get the best pricing — and take the most model risk. Do your own research.
Not financial advice — for educational purposes only.