Upstart — What I Think
Upstart asked the one question the credit industry had quietly decided was settled: what if the FICO score is just wrong? Not biased, not outdated. Wrong, as in leaving real money on the table by judging people on too few, too crude variables. A three-digit number built decades ago decides who gets a loan and at what rate for the entire country, and Upstart’s founding bet was that with modern machine learning and a thousand variables instead of a handful, you could see creditworthiness the bureau was blind to. Find the people FICO unfairly rejects, lend to them at a fair rate, and pocket the difference between the real risk and the perceived risk. That’s the whole company, and it’s a genuinely good idea.
What Upstart understood that the banks didn’t: the credit score is a compression algorithm, and like all compression it throws information away. Education, employment history, the school you attended, the field you work in, signals a human loan officer might weigh but a FICO score flattens to nothing. Upstart’s thesis is that a model trained on enough of that texture can approve more people with the same loss rate, which is the holy grail of lending: expand the pool without expanding the risk. When it works, it’s not magic, it’s just better-priced information.
Where the thesis hits its hard wall, and this is the lesson of the whole company: a lender that doesn’t hold its own loans is at the mercy of someone else’s mood. Upstart’s model was to originate loans with its fancy AI and then sell them to banks and institutional buyers. That works beautifully when capital is cheap and buyers are hungry. It works terribly the instant rates spike and funding dries up, because suddenly the brilliant underwriting engine has no one to sell to, and origination volume, the thing your stock is priced on, falls off a cliff through no fault of the model itself. Upstart got caught being a marketplace that thought it was a technology.
The truth: the underwriting insight is probably right and genuinely valuable. But “I have a better model” is not a business if you don’t control the balance sheet that funds the loans. The AI was never the fragile part. The capital structure was. A lending company is a funding company wearing an algorithm.
Favorite & worst CEO
One founder-CEO, so: on its leadership, Dave Girouard, co-founder and CEO, formerly the man who built Google’s enterprise apps business. I respect the core conviction enormously, that the credit score is a lazy, lossy proxy and that machine learning can underwrite better than a fifty-year-old formula. That’s a real, contrarian, technically-grounded thesis and he’s stuck to it through brutal cycles. My honest critique of the era is the same as my critique of the model: building a superior underwriting brain on top of a funding model you don’t control means the thing you’re best at can be neutered by macro conditions you can’t influence. The engineering vision was sound. The dependence on someone else’s appetite for your loans is the strategic flaw the whole story keeps circling back to.
Part of “What I Think About the Top 50 Fintech Companies of All Time.” I’m Prajjwal Chittori. prajjwalchittori.com.