Insuring the Asset, Not the Hype: Bernie Margulies on De-Risking GPU Finance
Our guest on Where Capital Meets Compute is Bernie Margulies, founder of American Compute, which builds risk management and credit backstops for GPU finance. The pitch is simple: a lot of the projects raising money to buy GPUs are not the most creditworthy, they are young companies without decades of financial history, and that makes them hard to underwrite. American Compute structures insurance products that transfer that risk out of the deal so lenders, lessors, and equity investors can get comfortable enough to fund it.
Bernie did not come from insurance. He started in investing, spent his early career in Singapore, got pulled into the startup scene in San Francisco, and eventually fell down what he calls a rabbit hole of insurance, applying a financial mechanism to a brand new asset class. The company’s most well known product is residual value insurance for GPUs, where American Compute is arguably the only structurer in the market. Below is a breakdown of the conversation with Kony, organized by theme.
The real risk is not obsolescence, it is margin
Everyone obsesses over technology obsolescence, the fear that a new chip makes the old one worthless. Bernie argues that is the smaller risk. The bigger one is NVIDIA’s margin. NVIDIA runs a 70 to 75 percent gross margin on its hardware today, which historically sat closer to 40 percent. When you finance a GPU project, you are not really paying for silicon cost, you are paying NVIDIA’s margin.
You’re literally just paying NVIDIA seventy-five percent. You’re not financing hardware, you’re financing NVIDIA.
The danger is not that the technology ages. It is that NVIDIA, if it ever faced real competition, could cut prices in half, still print money, and vaporize the residual value of every GPU already financed in the field. That is why so much of American Compute’s qualitative work is spent studying silicon margin profiles rather than chasing spec sheets.
The data: model the distribution, discount the mania
On the quantitative side, American Compute collects secondary market transaction data on enterprise IT equipment, GPUs, CPUs, memory, storage, and switches, working with resellers and remarketers. What comes back is not a clean number. It is a messy distribution of quotes, and from that spread they model expected losses and future resale outcomes.
The critical step is stripping out the hype. Memory prices are spiking right now, partly real and partly sellers squeezing margin, so American Compute references pre-mania price levels from 2023 and 2024 rather than today’s skewed distribution. The underwriting philosophy is deliberately pessimistic: underwrite the bottom of the distribution, not the middle.
When will they insure a brand new chip
Not on day one. Bernie waits a few months to see whether the market actually likes an asset, a concept he frames as infant mortality. Some model years are just worse than others, and there is no way to know which until the market votes.
His second filter is what he calls evergreen: hardware anyone can use, in any data center, for any workload. Air cooled beats liquid cooled because more sites can take it. General purpose beats specialized. Counterintuitively, the generic asset holds residual value better than the exotic one.
You don’t want to be super specialized. The more generic it is, the better. Anybody can use you, anybody can put you into their workloads. That holds a better residual.
He is also candid about a growing problem: fully configured racks and superpods are hard to resell as units, so American Compute values them by their components, which doubles as a worst case salvage scenario, the equivalent of selling a boat for its steel.
There is only one risk in town
Pressed on which risks are insurable, Bernie is blunt. Lenders are not really scared of the asset. They are scared of credit, of the AI bubble popping and the off taker running out of money. And that, he says, is precisely the risk you cannot sell insurance against.
There’s only one risk that anybody really cares about. It’s credit risk. And that, I would argue, is very hard to insure.
So American Compute attacks it synthetically, from other angles. Residual value insurance shifts the focus from the borrower’s credit to the asset’s worth, the same logic a mortgage uses when it lends against a house rather than a borrower’s income. Individual off taker default, SLA breaches, and delivery timelines can each be insured in isolation, even if the market wide bubble cannot.
The product lenders actually want: off-take interruption
The newer product, and now the more popular one, is off-take interruption cover. Lenders told Bernie the same thing repeatedly: the residual value backstop is fine, but repossessing and reselling hardware is a nightmare they never want to reach. What they really want is assurance that the revenue keeps coming, because as long as the off take holds, the loan performs.
American Compute structures it as a 90 day coverage window. If the off taker leaves, for default, an SLA dispute, or simply going elsewhere, the policy covers the interruption and buys time to find a replacement, which Bernie believes is genuinely findable. His sharpest reframe is casting the whole thing as a familiar financial object.
One way to think about this product is it’s like an artificial debt service reserve account. If you have a three-month reserve account, you now also have 90 days from this bond. That’s six months to figure it out when stuff hits the fan.
Pricing runs 200 to 400 basis points on the loan amount, collected monthly alongside the lender’s own repayments. That is deliberate: if American Compute misses its premium, it is because the lender missed a payment too, so both sides are motivated to keep the deal healthy. It is a one year bond, and they simply decline to renew year two if the deal looks too messy.
The skill of the business is saying no
Insuring the projects that big, well capitalized borrowers do not bother insuring means American Compute is structurally exposed to adverse selection. Bernie’s answer is that managing it is the entire craft.
The best underwriters are able to say no. The ones that blow up don’t.
Smaller projects carry more perceived risk, but he argues the risk is often overblown, and the job is picking the one deal out of several where the fear is misplaced. He keeps exposure conservative in other ways too: no ten year policies, no promising an 80 percent residual three years out, only the lower end of the distribution.
How the structure actually works, and whether AI is a bubble
Bernie does not issue the policies himself. American Compute acts as a program administrator, designing the policy form, pricing, and underwriting, while a rated, regulated carrier fronts the paper and reinsurers sit behind it, syndicating the risk. The reason is credibility: a credit committee trusts a carrier with a 30 year AM Best rating far more than a policy signed by, in his words, a young man named Bernie. Reinsurance also solves the capital problem, letting a young company borrow balance sheets instead of parking hundreds of millions in collateral.
Asked the closing question, is AI in a bubble, he does not dodge. He is AI pilled and uses the tools every day, but he expects a correction, likely not next year, possibly within five.
I’m seeing a lot of unqualified operators getting access to financing when they really should not. Their business model is spending VC dollars as quickly as possible.
His hope is that the industry underwrites conservatively and lends to the right people on the right terms, because too many on-demand bets and too few off take contracts make the whole system fragile. It is a fitting note for a show about where capital meets compute: the constraint is not enthusiasm for AI, it is discipline in financing it.
Listen to the full episode of Where Capital Meets Compute for the complete conversation.