UPSTART

Dave Girouard spent a decade running Google's enterprise division. By 2012, he was ready to start something of his own — and the problem he kept coming back to was credit.

Specifically, the absurdity of FICO scores. A three-digit number invented in 1989 was still the primary way American banks decided who deserved a loan.

It didn't factor in your education, your career trajectory, your actual income potential — just whether you'd paid your bills on time in the past.

Girouard co-founded Upstart with Anna Counselman, who'd also worked at Google, and Paul Gu, a Thiel Fellow who had dropped out of Yale to build things. The original idea was actually income sharing agreements — a way for recent graduates to fund themselves by selling a slice of their future income.

That model hit regulatory walls fast. So they pivoted to personal loans, but kept the core thesis: that machine learning could assess creditworthiness better than a decades-old scoring model.

They raised early funding, built a model that ingested hundreds of variables — employment history, education, area of study, earning trajectory — and started making loans. The early results were compelling enough to get banks interested.

Instead of lending their own capital, Upstart began licensing its AI model to banks and credit unions, taking a fee per loan originated. That shift changed everything.

Upstart is not really a lender. That's the thing most people get wrong about it.

It doesn't hold loans on its balance sheet the way a bank does. It builds and licenses an AI-powered credit model, then connects borrowers to banks and credit unions who actually fund the loans.

Here's how it works: someone applies for a personal loan on Upstart's platform. Upstart's model runs hundreds of data points through its algorithm and spits out a credit decision — often in seconds, often without a hard credit pull first.

If approved, one of Upstart's bank partners funds the loan. Upstart gets a fee for originating the loan and a fee for servicing it.

The bank takes the credit risk, not Upstart. Clean business model on paper.

The catch is that this model is extremely sensitive to what banks are willing to fund. When rates are low and credit is cheap, banks love originating more loans.

When rates spike and credit tightens, banks pull back — and Upstart's fee revenue collapses with them. In 2022 and 2023, that's exactly what happened.

Upstart ended up having to fund some loans itself to keep the platform moving, which is exactly the balance sheet risk it was supposed to avoid. The model works beautifully in a low-rate, risk-on environment.

The question is whether it's durable in any other kind.

Upstart's flagship is its personal loan platform — unsecured personal loans typically ranging from $1,000 to $50,000, aimed at people who might not qualify for the best rates through traditional bank underwriting. The whole pitch is that borrowers with thin credit files but strong income potential can get better rates here than they would at a conventional bank.

They've also pushed hard into auto lending, which is a much bigger market. Upstart Auto Retail is a dealer-facing product — it lets car dealerships offer AI-underwritten financing directly on the lot.

This is a logical expansion because auto lending dwarfs personal lending in volume, but it's also more complex and more competitive.

More recently, Upstart has moved into home equity lending and small business loans. Both are early stage and both face the same headwind: banks need to be willing to fund the originations, and in a tight credit environment, new loan categories are the first things partner banks pull back from.

The core tech — the underwriting model — is what Upstart keeps betting will win across all of these verticals.

Upstart's growth story has two very different chapters. The first was quiet and methodical: spend years convincing small and mid-size banks that their AI underwriting was better than FICO.

They had academic studies backing them up — lower default rates at the same approval levels, or more approvals at the same default rate. Banks are notoriously slow to change, but one by one, credit unions and regional banks signed on as partners.

The second chapter was the pandemic trade. When COVID hit and stimulus money flooded the economy, consumer credit was artificially pristine — defaults were low, spending was high, and everyone wanted a personal loan.

Upstart's origination volumes exploded. Revenue went from $164 million in 2020 to $849 million in 2021.

The stock, which IPO'd at $20 in December 2020, hit $401 in October 2021. Wall Street had decided Upstart was going to eat the entire consumer lending market.

The counterintuitive move that fueled the hype was the pitch to banks: 'Let us underwrite your loans better than you can.' Bankers, who typically hate ceding control of anything, were actually buying it — because the data backed it up in benign conditions. What nobody fully stress-tested was whether the model held up when the macroeconomic environment turned ugly.

Interest rates. That's it.

That's the whole challenge.

When the Federal Reserve started hiking rates aggressively in 2022, Upstart's business model cracked fast. Banks that had been eager to originate Upstart-underwritten loans suddenly had better uses for their capital.

Loan demand from consumers dropped too, because borrowing got expensive. Upstart's loan originations fell off a cliff — from a peak of 180,000+ loans per quarter down to a fraction of that.

To keep the platform from going totally dark, Upstart did the thing it had promised investors it would never do: it started holding loans on its own balance sheet. By mid-2022, it was sitting on hundreds of millions in loans it couldn't sell.

That's credit risk. That's exactly what a fee-based AI platform isn't supposed to carry.

The deeper problem is model risk. Upstart's AI was trained predominantly on a period of historically low defaults.

When defaults started rising in 2022 and 2023, the model's predictive edge became harder to demonstrate. Critics asked an uncomfortable question: did the model actually work, or did it just look great because everyone was paying their bills back during stimulus season?

The company has spent the last two years trying to answer that question without having a definitive answer yet.

Upstart's early investors included Google Ventures, which made sense given the founder DNA — Girouard was a Google exec, and the company was built on the premise that machine learning could disrupt a legacy industry. Third Point Ventures also backed the company early.

These weren't huge rounds by Silicon Valley standards — Upstart raised about $160 million in total before going public.

The real capital came from the IPO in December 2020, raising $240 million at a $1.8 billion valuation. Then the stock went on its historic run.

At its October 2021 peak, the company was worth roughly $32 billion — a company doing under a billion in revenue, priced like it was going to own global consumer credit.

The institutional hype during that period was real. Cathie Wood's ARK Invest was a major buyer on the way up — and a major seller on the way down, which told its own story.

The post-2021 collapse wiped out most of that paper wealth. By late 2023, the market cap was back down around $2-3 billion.

Investors who bought at the peak are still waiting to get back to even.