Dave Girouard - Making Better Loans - [Founder’s Field Guide, EP. 52]
What this covers
Patrick interviews Dave Girouard, co-founder and CEO of Upstart, who came to lending from Apple and eight years at Google. The conversation opens on operating habits rather than finance. Girouard argues that the first step in any decision is to fix how long it will take and who is involved, since a timely decision that is sometimes wrong beats a slow one. He says a CEO who sits in every decision signals a problem, and that leaders should push on timelines only along the critical path. He also puts the quality of the people around you ahead of any plan.
The middle of the episode covers how Upstart works. It began with young, post-college borrowers with a thin file, where education and field of study turned out to predict income stability. He explains how training data and algorithm sophistication depend on each other, and how the model forecasts a loan's cash flow month by month, including early prepayment. Upstart then moved from unsecured loans to auto loans, with homes, small business and heavy equipment as stated next steps. He claims the model can cut about 75% of a bank's losses at constant approvals, or roughly double or triple approvals at constant losses.
The most contested material concerns fairness: Upstart says its models raised approvals and lowered rates for every demographic it measures, and Girouard calls persuading regulators and advocates his hardest task. He also argues lending out-earns payments, and that buy-now-pay-later firms such as Square's Afterpay could strain the card networks. Shorter detours cover crypto and credit reporting, and how public and private investors judge founders differently.
Girouard argues Upstart can beat the 30-year-old FICO score by using continuously learning AI risk models that price credit more accurately and more fairly across demographics, and that lending, not payments or super apps, is the largest profit pool in financial services.
- A three-digit credit score cannot capture the time-based repayment behavior that a monthly cash-flow model can predict
- Centralized models improve with every loan and can be extended from unsecured loans to secured and new asset classes
- Lending's net interest income in the US is comparable to all technology-industry profits, so it is the core of financial services economics
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The FICO credit score was introduced around 1989 by Fair Isaac as a universal three-digit number, a radical improvement over the bespoke loan-officer process that preceded it, but it has since become an industry crutch 30 years later.
“when Fair, Isaac came out with the FICO score in 1989, suddenly there was this universal three digit number”
Leaders should identify the critical path of a project and challenge timelines only on work that actually gates the outcome, since pushing on tasks that do not need to be done for weeks wastes energy and goodwill.
“I think one of the important things you really have to know is what is on the critical path here? There's no sense of us all worrying about something that frankly doesn't need to be done for two or three weeks”
Most banks refuse to lend to anyone below a 680 FICO score, yet many people with 580 to 620 scores are good credits, and even people with 750 or 800 scores subsidize losses elsewhere in the system and pay too much.
“the vast majority of banks that are out there we'll just say, 'Hey, if you're below 680, we just don't take that risk.'”
A single three-digit number cannot capture the subtleties of whether a person will repay a loan, when they will repay it, and at what size or type of loan, leaving substantial room for better modeling.
“a three digit number is never going to capture all the subtlety of a person and whether they would pay back a loan and when, and what size of loan, on what type of loan.”
Having the quality of people around you is the whole game, since one exceptional engineer can be worth ten average ones, and Upstart's success rests on having the right people at the table rather than the right plan.
“having the quality of people around you is the whole game”
A CEO should not be involved in every company decision; if the CEO is in every decision, that signals a problem, and the CEO should only break ties or work on the most important decisions where they have a unique perspective.
“if I'm involved in every decision the company is making, that's a bad sign. I should be breaking ties or helping to work some of the most important decisions”
Upstart reports that its models raised approval rates and lowered interest rates for every demographic group it measures, including race, gender, ethnic origin, and age.
“our algorithms actually increase, improve credit outcomes for every single demographic we can name, race, gender, ethnic origin, et cetera, age”
US net interest income from lending is comparable in scale to all profits in the technology industry, making lending the central profit pool in financial services even though the two figures are not strictly comparable.
“if you looked at the net interest income earned in lending in the US it is comparable to all profits in the technology industry.”
The Federal Reserve found that 30% of Black Americans have FICO scores in the lowest decile, suggesting the existing credit system itself may be unfair and that AI can be a more equitable alternative.
“the fed put out this study that 30% of black Americans have FICO scores in the lowest decile, the lowest 10% of Americans.”
Machine learning models need training data in which rows are historical loan repayment outcomes and columns are applicant attributes, and more sophisticated algorithms can only be used once enough training data exists, creating a co-dependence between data and algorithm.
“if you don't have enough training data, you can only use a fairly simplistic algorithm. Using more sophisticated algorithms requires more training data.”
Lending generates more profit than payments because payments are a commoditized volume game, whereas lending involves risk and unknowns that require bespoke judgment, which keeps competition lower.
“payments in some sense, tend to be somewhat commoditized in the sense of it's a volume game.”
Convincing the world that AI can be fair and more fair than the FICO system is the hardest challenge, and Upstart invests heavily in continuous engagement with the CFPB, lawmakers, and consumer advocates.
“my co-founder is probably talking to the CFPB multiple times a week and not just with the regulator”
Upstart's co-founder argues that 90% of the interest paid in the world is unnecessary, because the obtuseness of risk models drives up the cost of credit, and simpler and smarter models could dramatically reduce it.
“90% of the interest paid in this world is entirely unnecessary. It's because of the obtuseness of risk models.”
Buy now pay later players, especially Afterpay now owned by Square, can build two-sided networks that stress the Visa and MasterCard model by creating payment rails consumers prefer that bypass card networks.
“the buy now pay later players, particularly when you think about one of the largest of them Afterpay, being in the hands of Square now... you have a two-sided network that actually can really put some stresses onto that Visa, MasterCard model.”
US personal lending is roughly a $100 billion annual market and was growing faster than any other consumer lending category before COVID, though it remains the smallest of the major categories.
“Starting from what's known to be the smallest is what we do, personal lending, a hundred billion or so annually”
Credit reporting agencies could move toward a distributed, crypto-style model with reputation and pseudonymity, and crypto may make cross-border financial products easier by offering a more global infrastructure despite differing national rules.
“Credit reporting agencies could go the way of crypto and distributed and there's just all sorts of notions of reputation and pseudonymity”
Crypto records the past immutably and AI predicts the future, so the two are closely related, and immutable, unambiguous historical records would make AI prediction much better, which makes combined products promising.
“crypto is recording the past immutably and AI is about predicting the future. If you have an immutable clear unambiguous record of the past, you can be much better at predicting the future.”
Upstart initially focused on young, post-college borrowers with thin credit files, and found that data such as education and field of study could predict income stability over time.
“we were very interested in young borrowers, post-college, people with little to no credit history, what in the industry they would call a thin file”
Public market investors are not making long-term bets because they can sell at any time, whereas venture investors seek to be early to a perceived insight, so private markets struggled to see Upstart's differentiation until it showed profitable growth.
“The public investors aren't making long-term bets. They're buying something they could decide to sell the next day.”
Upstart's core model predicts the month-by-month cash flow of each loan, including which months a borrower might miss a payment and whether they will prepay early, which is a time-based approach unlike traditional default-based scoring.
“what our model very simply is trying to do is predict the cashflow for every month of a loan”
In the future consumers should have the best available credit offer for any circumstance instantly, with no application process, as a number that continually adjusts to their financial life.
“From the consumer perspective at a moment's notice you have the very best offer of credit you have for any particular circumstance you find yourself in.”
Upstart's models can cut about 75% of a bank's losses while holding approval rates constant, or can roughly double or triple the number of people approved at a constant loss rate, giving banks a choice between conservative and aggressive use.
“we can knock out sometimes in the range of 75% of your losses”
Upstart started with unsecured lending and then taught its model to price secured loans such as auto loans, and plans to extend it to homes, small business cash flows, and heavy equipment, implying near-unlimited addressable lending opportunity.
“we've sort of taught the model new tricks and instead of being backed by an automobile, it might be a home. It might be the cash flows of a small business.”
Upstart never closed more than a single viable lead term sheet beyond its seed round, and Dave attributes this to failing to tell the story from the industry's perspective rather than from the company's inward view.
“We were too inward looking in terms of telling our story. We really needed to tell the story from industry perspective”
The most important step in a decision is to decide in advance how long it will take and who needs to be involved, because a timely decision that is occasionally wrong is almost always better than a slow one.
“the most important thing about a decision is to decide upfront how long you're going to take to make that decision and who needs to be involved in it”
Leaders should let people contribute to decisions but should not try to create consensus, because consensus is nearly impossible to reach when decisions are complex.
“give people a voice but don't try to create consensus because consensus is almost impossible”
Long timelines often hide an unexamined assumption that one task must finish before another starts; teams should ask whether the work can be done in parallel.
“Often there's hidden assumptions in that two months. There's something going on there that they assumed this can't happen until that is done.”
Getting outside parties to move at your pace works best by showing them your reasoning, so that a rational counterpart who understands your needs would ask for the same thing, rather than by berating them.
“if they understand where you're coming from, what you need to do and why, if they were you, they would be asking for the exact same thing you are.”
Entering lending as an outsider can be an asset because knowing what you don't know makes you a fast student who can challenge industry assumptions without being blind to how the industry works.
“if you know what you don't know, you are a fast student and you can challenge a lot of the assumptions, but not be blind to the way the industry works.”
Upstart focused on the heart of lending, the credit decision of who gets a loan and at what price, believing a more accurate default-prediction model would improve outcomes, whereas other online lenders mainly moved existing products onto the internet.
“We really wanted to look right at the heart of lending which is the credit decision, who gets a loan and for what price?”
Upstart discovered that factors like education, field of study, and employer were predictive of income stability, and a 20-year-old co-founder with strong credentials was rejected for loans because he had no credit history, illustrating that the credit score misses creditworthy borrowers.
“This guy went to Yale. He had a perfect SAT score. He was earning a six figure income. He had zero debt. He was getting rejected all around because he didn't have three years of credit”
Upstart avoids hand-judging which schools, employers, or industries are better, instead letting software learn their predictive value from loan performance, aiming for a 360-degree view of the individual.
“we're trying to create a 360 degree view of the individual but we aren't judgmental about the data in any way. We really are letting the software interpret performance.”
The right response to AI bias concerns is to test all outcomes rigorously, and Upstart tests every applicant for bias and reports the data to the Consumer Financial Protection Bureau quarterly on behalf of its bank partners.
“we actually test every single applicant for bias. We provide that data to the CFPB, the preeminent regulator for consumer protection, every quarter”
About 97% of Upstart's revenue comes from three bank-paid fees: a referral fee, a platform origination fee, and a loan servicing fee.
“Almost all of our revenues, something in the range of 97% of our revenue is fees that banks pay us.”
Upstart chose not to pursue a bank charter because most competitors were becoming banks, and it did not see itself fitting the bank model, so it instead partners with banks and leverages their balance sheets.
“it became clear that most of the participants were going to actually pursue bank charters and become banks. And that was a viable path, not one we wanted to pursue”
About 70% of Upstart's loans involve no human intervention at all, a process inspired by Google's low-headcount business model, which Dave cites as the goal of a business that generates a lot of revenue with little human capital.
“Today about 70% of our loans, there's no human intervention at all. It's literally lights-out processing”
Most Upstart borrowers receive an approval in the moment, without documents or phone calls, after getting a rate in seconds, and this surprising ease of access is central to the company's value proposition alongside price.
“they can get a rate in a few seconds and then if they like it, most of them are approved literally in the moment, no documents to upload”
Upstart's centralized AI models get better as every consumer loan and every bank partner contributes data, creating a compounding advantage because the model with the most data and experience is the best one.
“every consumer that gets a loan, every bank that works with us is both contributing to and benefiting from these centralized AI models. So we do get this advantage because the model with more data, more experience is the best one.”
Just as Google showed that search, previously treated as a commodity checkbox, was where the money was, lending is commonly dismissed as a commodity while being where most financial services profits are made.
“Search was considered a commodity and all the sort of portals at the day we're adding all these other features and just thought of Search as a check box.”
A founder-CEO who has achieved some success has authority to change direction quickly, whereas a hired CEO faces a board that wants to see its hire succeed, a founder who is still involved, and an expectation to justify the choice.
“once you've achieved some success and you're on the right path, you have a lot of authority.”
Upstart survived a brutal pivot away from income share agreements, which was a niche product, and a later pivot to serving many banks rather than one, and keeping the whole team aligned through those changes proved the company's resilience.
“We survived a brutal pivot. I mean, this beautiful idea of an income share agreement, so perfect on the white board, came up against the cruel reality of the world”
Selling to banks is a technology sales cycle in its worst form because banks are among the most regulated and conservative industries, so a company selling into them must raise a lot of money and accept a long adoption timeline.
“Raise a lot of money because you're going to need a lot of time. I mean, selling into banks is sort of a technology selling cycle in its worst possible form”
Consumers have zero tolerance for waiting and paperwork, so solving for speed and low friction matters more to them than a couple of percentage points of interest rate difference.
“Consumers of course, has zero tolerance for waiting, zero tolerance for requesting documents”
The two pillars of the lending value proposition are price and experience, and removing friction and wait time in the experience is tied directly to AI capability.
“There's price and there's experience.”
Most business work consists of making decisions, acting on them, and getting outside parties on board, so a company that does all three faster habitually gets more attempts at success.
“most of what we do in business is number one, making decisions and number two, acting on those decisions and then really as a sort of third thing, a lot of times you have to sort of get others, other companies, other people on board with those decisions.”
Meetings often end with a plan but no completion date; leaders should habitually ask when work will be done and why it cannot be done sooner, not just for every item.
“no one ever said, 'When are we going to be done with those? How fast can we get that done so we can move on to the next thing?'”
Upstart grows when its funnel converts more of the people who request a rate into actual loans, and better models let it offer lower rates to more people while removing friction.
“Our whole business tends to grow when our funnel gets more efficient. That means for every person who requests a rate, how many convert into loans?”
In almost every application, if you could predict the future you could make things better, and Dave believes this logic applies broadly across fields such as autonomous driving and language translation as well as cache memory placement.
“In almost every application, if you could predict the future, you could make it better.”
Dave argues that Upstart's AI business is defensible because it is pro-consumer and helps people enter the banking system, so the regulatory and public trust it builds is part of its competitive moat.
“I think we've built a great moat. I mean, I think ultimately our business and our AI is very defensible in terms of it's very pro-consumer.”
Unsecured personal loans are taking market share from secured products like home equity loans and credit cards because they can be approved in a day rather than two months and are priced more competitively.
“unsecured lending is taking a lot of market share away from secured lending. People are using unsecured loans to redo their kitchen, instead of trying to get a home equity loan, which can take you two months.”
Upstart went public in December 2020 as a traditional IPO after private markets repeatedly showed skepticism toward fintech lending, and the company chose public markets to prove its differentiation through growth and profit.
“We went public in December 2020. So mid COVID”
Upstart is not a lender; it markets to consumers, runs them through its models, and refers them to one of its 25-plus bank partners that originate and fund the loans.
“We're not a lender ourselves. We're not a bank. We sit in the middle as the technology provider”
Renaissance Technologies works to shave off a few basis points with trading algorithms, but lending has roughly two orders of magnitude more inefficiency, making it a far larger opportunity.
“Renaissance Technologies coming up with these trading algorithms... to kind of shave off three basis points. And we kind of think it's laughable because we're working at something that has two orders of magnitude more inefficiencies.”
No credit model is perfect, but Upstart does not need to be perfect, only better than the status quo, which Dave believes it is without question.
“No system is perfect. No model is perfect. We just have to be better.”
Larry Page at Google routinely pushed teams to compress aggressive timelines, asking why a two-month plan could not be a week, which forced teams to examine their assumptions about what must finish first.
“Larry would say, 'Well, why can't you get it done in a week.' No matter what you brought to him, he could one-up you and just challenge you”
Upstart's success came less from its game plan or perfect decisions than from having the right people around the table, so hiring, people development, and mission drive company success.
“having the quality of people around you is the whole game.”
Upstart's original model was income share agreements in which young people repay a fraction of future earnings, and a co-founder built a model to predict lifetime earnings as the foundation of the company.
“you would pay it back as a fraction of your future earnings. And income share agreements”
AI can predict where information should be stored in memory so that it is retrieved optimally, an application Dave found surprising given his computer engineering background in caching.
“you can actually use AI to predict where you should store a piece of information so that it will be optimally retrieved”
Financial services and healthcare were largely avoided by internet entrepreneurs for the first 20 years of the internet because of regulation, but the sector has changed substantially in the last decade.
“Financial services along with maybe healthcare are two of the most difficult industries to penetrate.”
Visa and MasterCard's network effect is nearly unbreakable because any challenger must have both merchant presence and consumer presence, which has made it very hard for anyone to compete.
“the strong network effect of Visa and MasterCard has been around a long time and just unbreakable, frankly.”
FICO scores remain central to consumer credit, including mortgage eligibility for sales to Fannie Mae and Freddie Mac, which embeds the score into how credit works.
“the government uses it to decide which mortgages can be sold to the government sponsored entities, Fannie Mae and Freddie Mac”
Instead of a three-digit score, Upstart's models use roughly 12, 20, 40, or 60 variables about each applicant, combined with repayment data, to predict outcomes.
“you go from a three digit number to maybe 12 different variables or maybe 20 or 40 or 60”
Banks were formed to take deposits and make loans, and lending is still their main activity, so a company that does not want to make loans probably does not want to be a bank.
“Banks were formed as an entity to take deposits and make loans and unsurprisingly, it's still the main thing they do”
Half or more of Americans are left out of the banking credit system or get terrible rates, and Upstart's approach to measuring risk on a continuum can address this gap.
“because half or more of Americans are completely left out of the system”
Despite the obvious benefit of better risk models, the lending industry widely views lending as an ultimate commodity dependent on macroeconomic conditions, making it hard to convince people that a better lending mousetrap is possible.
“just enormous disbelief that it's possible to build a better mouse trap in lending, it's just viewed as the ultimate commodity subject to macro economic conditions”
Girouard describes himself as a natural company builder but not a natural fundraiser, which created incentives to reach cash-flow positive to avoid repeated fundraising rounds.
“I was a natural company builder. I was not a natural fundraiser.”
The US auto lending market is roughly five to six times the size of personal lending, at about $700 billion in originations, and mortgages are another order of magnitude larger.
“something like auto where we're moving into, five or six times, 700 billion in originations.”
Insurance is the other side of the coin from lending because it is an entirely risk-based industry, but its business models present challenges that make it a harder expansion target.
“insurance is, of course, it's like the other side of the coin for lending. It's an entirely risk-based industry.”
The kindest thing anyone did for Girouard was his father, who worked 50 hours a week and took extra jobs during vacations, so the family could feel secure despite having little money.
“my dad would literally work those extra two weeks”
Whether merchants pay more for buy now pay later than they pay for Visa and MasterCard, and whether BNPL has pricing power or faces commoditization, remains an open question.
“are merchants actually paying more for these services than they're actually paying for visa, MasterCard”
One exceptional software engineer can be worth roughly ten times an average one, but taken to an extreme this view becomes dysfunctional.
“one exceptional software engineer really can be worth 10X what an average software engineer is. And you can take this to an extreme where it is dysfunctional”
After eight years at Google and building Google Cloud, Girouard began to doubt whether his success came from himself or from Google's momentum, which motivated him to prove he could build something outside Google.
“Maybe anybody could have showed up at Google and built this. The company had so much mojo in those days”
Many fintechs such as Revolut, Robinhood, Square, and SoFi are pursuing super apps, while Upstart deliberately remains focused on lending, which Dave sees as a contrarian and uphill position.
“There's a whole move toward the super app, the one app to rule them all.”
Steve Jobs could berate vendors and demand unreasonable timelines because Apple had leverage that most companies lack, so the same tactics are not reliably available to others.
“Apple can get away or Steve jobs could get away with things that others probably can't.”
Upstart's business grew and became profitable as its AI models matured, and public markets were willing to understand and value that growth, which is unusual in fintech.
“The AI models matured. They were just at this hockey stick moment of how they worked. And that led to just a lot of growth and profitable growth, which in the FinTech world is pretty unique.”
Upstart's three founders came together around the same idea despite large differences in age and background, and the company's survival through a brutal pivot is credited to keeping the founding team intact.
“we still have the three founders day-to-day active as well as a lot of the exec team has been with us since the beginning”
The AI risk models are most exciting when they enable breakthrough consumer experiences, such as approving someone instantly for a loan that would otherwise take days, because the underlying problem is risk.
“a model that allows you to approve somebody in a moment for a loan that would otherwise have taken days, that's what it's all about.”
Being a public-company CEO requires repeatedly telling the company story to analysts and investors, which feels unusual to Dave, but is part of the job of bringing the market along.
“It feels like a little unusual just to keep repeating the story, telling the story, answering the same questions over and over again.”
A meaningful improvement in lending should be measured in percentage points rather than basis points, because the market is so large that even large relative gains translate into very big opportunity.
“if you can actually make a significant improvement there not 10 basis points, but a hundred percentage points”
Dave's desire to start a company was partly prompted by a colleague's email, sent to an outsider, saying he had done well at Google but would not survive outside it, which made Dave want to test whether he could succeed on his own.
“he's done pretty well here at Google but I don't think he'd survive on the outside.”
Upstart's co-founder speaks with the CFPB multiple times a week and engages lawmakers and consumer advocacy groups, which Dave describes as an ongoing cost of building trust in AI lending.
“my co-founder is probably talking to the CFPB multiple times a week and not just with the regulator, we're talking about lawmakers”
Partner banks on Upstart's platform set their own lending parameters, including who they approve, loan types, and return targets, via adjustable settings Dave calls nerd knobs, while Upstart supplies the underlying technology.
“They actually get to have all the knobs and the dials of the types of loans they want, who they want to approve, what their return targets are.”
Upstart is not a crypto company, and the strong talent attraction of crypto makes this a potential liability for recruiting, though the company remains a crypto believer that will evaluate developments over time.
“The attraction of crypto to a lot of the talent in the world is so strong today that we do feel like it can be a bit of a liability”
Before about 2016, online lending was out of favor after some public online lenders went public and then performed poorly, so companies like Upstart operated almost invisibly while payments firms like Stripe and Square attracted attention and capital.
“prior to say 2016, the idea of online lending, there were some companies that went public and then suddenly they didn't do so well.”
A 2003 Google recruiter said the company was earning seven or eight million dollars a day in revenue even though employees were away skiing at Lake Tahoe, which Dave cites as a model for low human-capital intensity.
“they're still making like seven or $8 million a day in revenue.”
Dave believes DoorDash's founder-led structure gives it major advantages, while Uber is less advantaged despite having CEO Dara Khosrowshahi, who he considers an excellent non-founder CEO.
“I think DoorDash, for example, the founder led has incredible advantages, Uber a little less so. I think Dara is an amazing CEO but he's not the founder”
Jamie Dimon reportedly said he would be happy if his loan officers were playing golf rather than making loans, which Dave reads as a sign that large bankers are nervous about the economy and reluctant to lend.
“I'd be pretty happy if our loan officers are playing a lot of golf right now and kind of making loans.”
A model improvement can be used to lower losses or to increase reach, with the typical bank landing in the middle of these trade-offs.
“Banks will land somewhere in the middle, but the ultimate value prop is you can have a more inclusive lending program”
Upstart's median borrower is in their late twenties, younger than the American average, and tends to be highly educated and online, with personal loans as the core product and auto refinancing recently added.
“The median age of our borrowers, it's probably the late twenties. It's still trends toward younger, relative to the American demographic.”
Upstart exported many Google practices but picked and adjusted them, since Google did not do everything right.
“Google didn't do everything right. So there was a lot of picking and choosing”
Upstart plans to expand its lending capabilities beyond the US and into more or all forms of lending, guided by a combination of ambition and realism about what can be accomplished.
“can we take this outside of the US when and how? Can we take it to more or even all flavors of lending, when and how?”
Upstart generates around 40,000 to 50,000 repayments or delinquencies each month that serve as training data, growing its data advantage as loans accumulate over six or seven years.
“Every month, there's 40 or 50,000 repayments or delinquencies that happen and that's just a bit of training data.”
The rapid growth in GPU rental demand for AI training, reportedly doubling every three to four months, signals an impending explosion of predictive models being applied everywhere.
“I saw this insane demand curve for GPU, basically renting GPU's to train artificial intelligence models and it was doubling like every three or four months”
The host framed the speed discussion as four stages: triaging which decisions deserve time, making decisions, executing them, and aligning external parties.
“I'd love to just pick apart each of those, I'll call it four stages”