Good morning, Anissa here!

The AI boom requires an enormous amount of new power plants, chips and data centers. But it also requires something that’s less visible: new financial plumbing that makes it possible to lend against this costly hardware.

Miami fintech startup Barkr is one of the firms building the pipes, specifically for valuing AI chips being used as loan collateral, and has already attracted the attention of Nvidia. I recently met its CEO and co-founder, Thomas Galbraith, while moderating a panel in San Francisco about how lenders are underwriting the AI infrastructure build-out.

Galbraith sits in a very specific part of the complicated web of asset-backed financing.

His firm provides insurance-backed valuations of hard assets, like cars, art work and commercial equipment, which now includes AI chips. Barkr doesn’t lend the money itself, but gives lenders a number they can trust—and take to the bank.

This work takes backbone: If Barkr gets the price wrong and a borrower defaults, it pays the difference.

Founded in 2023, Barkr got its start calculating the value of assets such as collectible cars and private jets so that lenders could better price their loans. About six months ago, a lender came to Galbraith with a new challenge. “We‘re going to start lending against GPUs,” he recalled them saying. “And we can’t find anyone who can price them, let alone provide a guarantee on the price.”

While Nvidia graphics processing units have emerged as a new asset class over the past few years, asset-backed financings for the chips are not yet ubiquitous. That’s because lenders and borrowers often disagree on the residual value of the GPUs, and there’s no independent party that can iron out the discrepancy. That means the terms of GPU-backed financing aren’t favorable compared to deals involving traditional assets.

Asset-backed lending has always involved two risks: borrowers‘ ability to pay, and whether an asset’s value will hold up. Lenders have spent decades getting good at underwriting borrowers. Underwriting assets, particularly novel ones like GPUs, remains genuinely hard. Ask five firms to value the same GPU cluster and you will get five different answers, Galbraith said.

Barkr has customized several AI models to value assets being used as loan collateral, then wraps the valuation in a contractual guarantee backed by Munich Re and other insurers. For a lender, that changes the calculus entirely, he said. A trusted price means less risk for the lender and a lower interest rate for the buyer.

In the past six months, Barkr has aided transactions involving about $200 million worth of GPUs deals, and it expects to handle another $300 million in GPU deals before the end of the year. (That’s a relatively small number compared to Nvidia’s sales, and most GPUs are bought by world’s biggest companies, largely using their balance sheets.)

For comparison, three years ago, investment firm Upper90 gave cloud provider Crusoe $200 million in debt, which was backed by 20,000 GPUs. Similar deals have happened since but have not proliferated.

Making GPUs Boring

Barkr, which has 14 employees and raised $3.5 million in venture capital, treats a GPU cluster the way an appraiser treats commercial equipment, by asking what another firm would pay for the hardware if a borrower defaulted on their payments.

“We make them boring assets,” he said. “When you're looking at the insurance sector and how they think about assets and how they think about risk, the more uniform, the more boring, the more steady that it looks, the easier it is for the insurance industry to adopt.”

While the prices AI developers pay to rent GPUs can be volatile (and are currently going up), GPU resale prices are a lot more stable, he says.

GPUs depreciate, but they do so in a predictable way, which is what insurers and lenders want to see, he said. Barkr has found that Nvidia H100 chips, which came out in 2023 for around $30,000, are reselling for around $25,000 to $30,000 apiece, below a peak of more than $40,000.

Galbraith has been obsessed with valuation problems since the early 2000s, when he worked at a global insurance company analyzing the value of art and other luxury items for high-net-worth clients. The firm's very name is a nod to that obsession. It’s a reference to Bob Barker, the late host of The Price is Right, a gameshow where contestants guess the retail price of consumer goods.

Nvidia's Quiet Endorsement

After our event panel, Galbraith drove to Santa Clara for a meeting at Nvidia's headquarters. There, he said he met with Nico Caprez, vice president of AI infrastructure, and others on the Nvidia startups investment team.

Nvidia appears to understand that making its chips easier to finance is a huge opportunity. Nvidia employees have already introduced Galbraith to lenders who want to back GPU deals, he said.

Nvidia has two reasons to care. First, it wants an independent third party validating that GPUs hold their value longer than some skeptics believe. Second, it wants to see the capital markets open up so that companies beyond the investment-grade hyperscalers can access large amounts of GPUs at similar borrowing terms.

“We're solving a very specific problem that they are also very aware of,” Galbraith said of Nvidia.

In other news…

NextEra Energy will buy Dominion Energy in an all-stock transaction valued at $66.8 billion in a bid to create the world’s largest utility provider with a combined market capitalization of nearly $250 billion.

GridCare—which my colleague Ann wrote about in a prior AI Infrastructure column—announced an oversubscribed, $64 million Series raise that brought in additional investors including Sutter Hill Ventures and John Doerr. GridCare is helping AI data centers find power by running advanced AI simulations to find underutilized pockets of power on the grid, which runs at a surprisingly low 30% capacity much of the year.

Brent Mayo, head of data center capital markets at advisory firm Newmark, has left the firm and told people he is joining investment firm DigitalBridge.

Upcoming Events

Wednesday, May 20 — Inside the AI Infrastructure Boom

Wednesday, September 23 — AI Agenda Live SF 2026

Tuesday, October 27 — The Information’s 2026 WTF Summit