So You Want to Start a Bank

The bank run0:00

On March 9, 2023, Silicon Valley Bank lost over 40 billion dollars of deposits in a single day. It expected over 100 billion more to leave the next morning. SVB was the bank where startups, and the venture funds behind them, kept their cash. And about 94% of its deposits weren't covered by deposit insurance. So when the money started to move, it moved all at once. Picture a few hundred founders who share the same handful of investors.

A worried call from an investor, a ping from company Slack, and they all check their balances. Then they all move their cash, on the same morning. The Fed's own review called it "highly correlated withdrawals", from a concentrated network of VC investors and tech firms. The simpler term for this is just called a bank run. Long story short, SVB blew up from things related to fast redemptions, and them owning too much low interest bonds, which lost value as rates rose.

It was an old story. And many channels have already covered this problem. I suggest checking out Patrick Boyle's explanation. After SVB blew up, Silicon Valley had a different problem. How and where do they get their money from this time?

The squeeze1:08

This is Evan Conrad, CEO of SF Compute. In his blog post from October 2026, he predicts the end of startups due to difficulties provisioning compute for them. He writes that "The world's compute is sold out." He also says it's "increasingly difficult to get capacity online from folks who are not investment grade". Investment grade is the top tier of credit ratings, the borrowers lenders trust the most. And in his words, "Startups are not investment grade."

So compute sellers ask them for more of the cash up front, and he's seen it go "as high as 100% of the contract paid up front". That was required, he says, because "neither the lender nor the cloud could take on any additional credit risk". The lenders who pay for new GPU clusters, in his words, "underwrite solely on the credit risk of the offtaker". The offtaker is the big customer who signed up to rent the GPUs.

His forecast is that within 6 months, most AI startups won't be able to buy compute at all.

On Carta, 966 startups shut down in 2024, against 769 in 2023. At the same time, the money pooled. Half of the US venture money invested in 2025 went into just 0.05% of the deals.

But Americans also filed a record number of new business applications that same year. So what is really happening is that business venture is squeezed. New companies keep starting, while the money and the compute keep going to a few.

Most of the news talk about the K-shape economy, and how the winners win bigger while losers get left behind. We saw this with stock valuations, small businesses, entry level careers. From the 3 stories earlier, the exact same thing is happening in the startup world, and this trend could be problematic for future innovation. With my experience with finance, what I really think is happening is that cost of capital has now become the bottleneck for future innovation.

That's what you pay to get money, the interest a lender charges, or the slice of your company an investor takes. For us to solve this problem, we must first understand banking.

How a bank earns3:19

Here's the whole business. A bank borrows money cheaply, mostly from people like you. When you leave money at a bank, that's a deposit. For you, it is savings. For the bank, it's a loan from you that you can ask back at any moment. Then the bank lends at a higher rate, to homeowners, farmers and businesses. The gap between what it earns and what it pays is called the spread.

Take all US banks in the second quarter of 2026, and they earned about 5.35% a year on their loans and bonds. At the same time, they paid about 2.03% for their deposits and borrowing. So the spread was about 3.3 points. Banks also earn fees, on cards, accounts and services, about 1.46% of their assets a year. But the spread is their main income. That spread then has to pay for everything.

The staff, the branches, the loans that go bad, and whatever's left is profit.

On its own, it doesn't sound like much of a business. You earn about 5 dollars a year for every 100 dollars you guard, and about 2 of that goes back out to depositors and lenders, so to get more out of that thin spread, you need leverage. Leverage sounds scary, and from this channel you would often hear me advising against it. That said, leverage in banking works very differently.

For normal people, leverage requires you to borrow money from someone else and you pay them back in the future. For banks making a loan, there is no one they have to borrow from first. They just create it out of nothing. Yes, money can be created out of nothing, but not without limits. The Bank of England spelled it out in 2014. When a bank makes you a loan, it types a matching deposit into your account, and that deposit is new money.

It's money because you can spend it. A shop takes a bank transfer just like cash. In their words, "bank lending creates deposits." The same paper names the limits. The loan has to pay more than it costs, against every other bank competing for it, and regulators set rules on top. Then the new money moves. You spend it, it lands at another bank. For the banking system, it never left.

It's still a deposit, so banks as a whole can keep lending against it. So what stops them? The owners' slice. Every loan makes the pile bigger, and the owners' money has to stay above the regulator's line. When the slice gets too thin, the bank has to stop lending, or put in more capital.

The rule is a ratio. Capital, divided by everything the bank holds, its loans, its bonds, its cash. Capital is the owners' money, the shares they paid for plus the profits the bank kept. A brand new US bank has to keep that at 8% or more for its first 3 years. Flip that around, and its assets can be at most about 12.5 times its capital. Say the owners put in 100 million dollars, and the regulator asks for 12%.

Then the bank can hold about 830 million dollars of loans and bonds, funded mostly by deposits. That 100 million is the capital, the owners' own money. Everything stacked above it is mostly deposits, money the bank owes its customers, and pays back whenever they ask. Every point of spread is earned on that whole pile, while the owners only put in the bottom slice. That's why a thin spread can still pay the owners well.

It is also why a bank can't take a big loss, because the owners' slice is thin, and a small loss on the whole pile can wipe it out.

Collateral6:42

So a bank wants collateral. It's the thing the lender can take if you don't pay.

Say you walk into a pawn shop with gold worth 100, and you walk out with 70 in cash. The share you got is the loan-to-value. The share the shop holds back, the other 30, is the haircut. Together they always add up to 1.

If you never come back, you've defaulted, and the shop sells your gold. What the shop loses after that sale, as a share of the loan, is the loss given default, or LGD. If the gold sells for more than you borrowed, that loss is zero. So there's the rare case when they don't care when you default on your loans.

Then there's the chance you walk away. The chance of default, times the loss given default, is the expected loss. A lender can't read your mind, and he won't know if you are a nice guy who is a profitable and a cooperating customer. So his next edge at making good bets on you is to know the thing so well that he sells it for close to what it's worth, and that shrinks the loss given default.

And the harder a thing is to sell, the bigger the haircut. That's why a Pokémon card lender, Alt, offers at most 40% of a collection's value.

To see collateral at its best, let's look at pawn shops again. The big US pawn chains file their numbers publicly. At FirstCash, about 75% of pawn loans are against jewellery, mostly gold. For everything else, its store system recalls what similar items recently sold for in its own stores. The chains lend about 25 to 65% of what they expect an item to resell for. Now here's the trick.

Say the shop lends 50 dollars on something that resells for 80. If you default, the shop sells your stuff back at its own counter, and your default turns into a 30 dollar gain for the shop. That's why pawn fees are a bit under half of FirstCash's revenue, and selling the goods brings in most of the rest.

So any collateral comes down to four questions. How much of its value do you lend? Who holds it? Who buys it if the borrower walks away? And how does its value move while you wait?

For gold at a pawn shop, those answers are easy. You lend at a fraction of gold's price, you store it in the shop's own safe, you have a counter to sell it to direct buyers, and its value is mostly not so volatile, and it's liquid. Now ask the four questions about a startup. A normal loan gets repaid out of revenue, and most startups barely have any. There's not much to hold or resell either.

Even a founder's own shares are hard to pledge, because "Not many companies allow share pledging."

So lenders invented venture debt. And how they get repaid is out of the startup's next funding round. Essentially, they are betting if you can reach the next funding round. If you watched the show Shark Tank, you would have some familiarity on this process as well as numbers.

For the startup scene, a typical loan runs a few million dollars, at 10 to 15% interest, plus a small slice of the company's shares on top. So a startup pays way more than what a bank pays its own depositors. There are a lot of reasons why you'd want high interest debt from a venture lender over lower interest debt from a bank, but that would be a topic for another day.

If you'd rather get it from a bank, the best-known bank that lent to startups was Silicon Valley Bank.

Lending to startups today10:01

Wasn't it the bank that blew up in 2023? The business model clearly doesn't work!

Silicon Valley Bank is interesting, as they seemed to be doing what others have not yet dared to during that time, that is providing business loans to unproven business models. Supposedly. Because when you look at who SVB actually lent to, it tells a different story. At the end of 2022, about 56% of SVB's loans went to venture and buyout funds, backed by their investors' promises to invest. Only about 9% went to startups living on investors' money.

And it barely lost anything on its loans. The logical inference is what SVB really wanted was the account and relationship. You see, one of its loan contracts required borrowers to keep their main operating accounts at SVB. And almost half of SVB's deposits paid no interest at all. Essentially, the loans were the bait, the zero interest deposits were the catch, and they haven't lost much money lending to supposedly risky people doing risky things.

So their business model so far has been baiting founders to deposit money in them, pay almost half of it 0% interest, then YOLOing into long term government bonds. It is their bond losses and a quick redemption from customers that blew up Silicon Valley Bank, not because startups are a structurally unviable asset class to be lending to.

OK, startup founders lost one of the means to access capital, but lately they want something different from capital. Compute. How do you lend them something like that? When you try to provision a GPU, you learn the compute prices aren't stable at all.

Silicon Data's index of H100 rental prices fell 45% from its start in September 2024 to February 2025, and later climbed about 50% off a low. So the price can swing by about half in either direction in a short time. Even Amazon shortened how long it expects some of its servers to last, because the technology moves so fast. And as far as anyone found, no GPU-backed loan has been publicly foreclosed.

So nobody has seen who buys the chips after a default.

Lenders found a way around the chip. They lend against the customer, the offtaker from Conrad's post. CoreWeave's loan this March shows how it works. 8.5 billion dollars, "secured by HPC infrastructure and an associated customer contract". It cost 2.25 points over the benchmark rate, and Moody's rated it investment grade. So cheap credit goes to borrowers with a customer lenders already trust, and the K-shape shows up again, this time in credit.

Conrad's own company, SF Compute, runs a market for GPU hours today. A startup can buy GPU time ahead, in blocks of 8 GPUs, priced by the hour. If it doesn't use all of it, it can resell the rest on the same market. SF Compute says its buyers get back about 25% of what they spend that way. It is like ticket scalping. You hold the seat, and you resell it before departure.

But every hour is paid up front, all of it, before delivery. Nothing is bought on credit, and no lender stands behind the buyer yet. So a startup still has to find the whole bill in cash, the same 100% up front that Conrad warned about.

Erebor13:12

How do we fix this? How do we avoid a K-shaped economy that makes it hard for small startups to get the capital and compute they need to compete? There are some attempts.

Erebor Bank was the first new bank the current Comptroller of the Currency conditionally approved, in October 2025. It targets "technology companies and ultra-high-net-worth individuals that utilize virtual currencies". Palmer Luckey is a director and principal shareholder. The first question was why anyone would start a bank for this, instead of running a VC fund with their own money all over again. The answer so far came down to three things.

A lender gets paid first. It gets better leverage. And it gets paid in interest along the way, instead of waiting for an exit.

The leverage gap is big. A VC fund under the usual SEC exemption can borrow at most 15% of its commitments, and only for 120 days. A bank held to Erebor's 12% floor can carry about 8 times its capital in assets, funded by deposits. But a lender's upside is capped. It earns its interest and a small slice of shares, not the full moonshot you get to brag about on the All-In Podcast.

So it only works if almost every loan comes back. And the deposits are the prize. In 2010, bankers interviewed by researchers said the real motivator was "the chance to secure the start-up's deposit accounts". It seems that in the startup ecosystem, trust and relationships often matter more than raw returns.

Now let's look at Erebor's call reports and see what it looks like so far. Hopefully, we don't find the same patterns as Silicon Valley Bank. Its deposits reached about 4.06 billion dollars by June 2026. And in June, 89% of Erebor's assets sat in cash. Loans as a share of deposits were just 1.9% in June. That looked suspiciously low. I guess it's because holding cash is safe, and they really want to protect their reputation after the last bank failure spooked the market.

And holding so much cash barely earns. In its first half year, Erebor lost 16.3 million dollars. Erebor hasn't shown yet what it will lend against, and the hours a startup buys on SF Compute aren't collateral for any lender yet. So we haven't seen progress yet on the cost-of-capital crisis for small startups.

Hours as collateral15:30

Conrad's plan is to make those hours the collateral. In his words, "If a lender gets a GPU hour as collateral, the credit risk of the customer matters less", and the lender has "a real path to selling the collateral". And it's still a plan. In late October, Conrad wrote that to test it, "we need capital we can use to backstop a pilot on it". They've built the market, he says, "but it takes capital to do this, and we don't have that yet".

The hard part is time. "A GPU-hour is consumed the hour it exists." So a lender that seized a defaulter's hours would have to sell the hours still to come, before they pass. So time and volatility make it hard to put a price on this collateral.

Is all hope lost? What if we leave the hours alone and lock the price instead? The exchange CME has filed futures on GPU rental prices. One contract covers 730 GPU-hours, and it settles in money against the month's average rental price.

Say you lock in today's index price, about 2.9 dollars an hour, on one contract. If the month's average comes in 60 cents higher, you receive that on every hour, 438 dollars. If it comes in lower, you pay the gap instead. So a lab that still rents real GPUs at the market price has locked what its compute costs. And the GPU owner on the other side has locked its revenue.

In SF Compute's words, "Long-term contracts are required to get supercomputers financed." We don't know if this ends up inviting a bunch of degen gamblers to speculate on GPU rental prices. But at least this trick locks in the revenue, and that's what lenders want to see.

And a future doesn't ask for the whole bill up front. You post a margin, and settle the difference every day. None of this trades yet, so if you're a degen gambler, you should calm down. The CFTC, the US futures regulator, extended its review to November 9, so mark that on your calendar.

The bottleneck17:32

Before anything gets better, it's getting worse. Lenders have been repricing the companies that borrow to build compute. In June 2026, CoreWeave sold unsecured bonds at face value, paying about 9.6% a year. Unsecured means nothing specific stands behind them, no chips and no customer contract. By mid-October they yielded about 15%.

A credit default swap is insurance against a borrower not paying. CoreWeave's cost about 11% of the debt a year, past the level that's generally a sign of distress. Oracle's costs about 3% a year, almost 4 times what it cost a year earlier. Meta's hit a record in October, but at only about 1.2% a year, it's still small. As one market strategist told Reuters, "Nobody is worried about Meta or Amazon or Google getting the financing they need."

So this is one more example of the 2 trends we've been pointing out. A K-shape, where the big guys get capital on much better terms than the small guys. And a shortage of compute for the small guys, because lenders who are choosy end up slowing down the data center financing, and everyone walks home with less compute.

This video was made to explain an unintuitive bottleneck in the market. Everyone talks about GPUs, power and memory. The bottleneck I keep seeing is the cost of capital. Conrad saw it coming back in March 2024. He wrote that commoditization "king-makes the player with the cheapest cost of capital". And the new, small players "didn't have the credit to get a cheap loan". And so far, I've only found 2 people working on a solution.

Thankfully, it doesn't have to be just them. The business of lending isn't that complicated. As long as you have a way to manage the collateral's liquidity, its volatility and its trading volume, you can find a way to lend people money against it.

If you have a good idea here, it's still an open problem waiting for contributions. You can follow Evan on Twitter. The link is in the description. Alright, that's all for this video. I will see you guys in the next one.

Sources

1. Evan Conrad, "The end of startups: the coming credit crunch" (Oct 22, 2026)
2. Evan Conrad, "On Commoditization, Pricing, and Assets" (2024)
3. Evan Conrad on X (Oct 26, 2026)
4. Federal Reserve, review of the supervision and regulation of Silicon Valley Bank (2023)
5. SVB Financial Group, Q4 2022 earnings release
6. Board of Governors of the Federal Reserve System, Distributional Financial Accounts
7. FDIC, Quarterly Banking Profile, Q2 2026
8. McLeay, Radia & Thomas, "Money creation in the modern economy", Bank of England Quarterly Bulletin 2014 Q1
9. OCC, Comptroller's Licensing Manual: Charters
10. FirstCash Holdings, FY2025 Form 10-K
11. Alt, help centre: Alt Lending
12. Ibrahim, "Debt as Venture Capital", University of Illinois Law Review 2010(4)
13. 17 CFR 275.203(l)-1, the SEC definition of a venture capital fund
14. Silicon Data, Silicon Index (H100)
15. Silicon Data, "The Practitioner's Guide to the Compute Future (Part 1)", on the CME contracts
16. The Register, Amazon Q4 FY2024 (shorter server lives)
17. CoreWeave, $8.5 billion investment-grade GPU-backed financing (2026)
18. SF Compute, realized discount
19. SF Compute, "San Francisco Compute sells $245M worth of compute"
20. amcompute, "GPU and Neocloud Defaults" tracker
21. OCC, conditional approval of Erebor Bank (news release, 2025)
22. OCC, conditional approval letter for Erebor Bank, N.A. (2025)
23. FDIC, approval of deposit insurance for Erebor Bank (2025)
24. FDIC, call reports for Erebor Bank
25. CFTC letter to CME on the NYMEX compute contracts (Sep 21, 2026)
26. PitchBook-NVCA Venture Monitor, Q4 2025
27. US Census Bureau, Business Formation Statistics
28. Carta data via TechCrunch, "2025 will likely be another brutal year of failed startups, data suggests"
29. Bloomberg via Financial Planning, "Wall Street's private banks are vying for AI's paper money elite" (2026)
30. Reuters via Insurance Journal, "AI credit worries send default insurance to record high" (2026)
31. CNBC, "Stock investors are ready to move on from the AI rout. Credit investors are not" (2026)
32. Bloomberg via Yahoo Finance, "Meta Credit Risk Gauge Hits Fresh High on AI Spending Plans" (2026)
33. Reuters via The Star, "Steep bill for AI buildout punishes Meta, other tech favorites" (2026)