Upstart Holdings, Inc. (UPST)
Upstart Holdings is an artificial-intelligence company for credit underwriting, founded in 2012 and taken public in 2020. The company’s platform uses machine learning to assess creditworthiness and set terms for consumer loans—personal loans, auto loans, and other retail credit. Rather than relying on traditional credit scores, Upstart’s models consider a wider range of signals about a borrower’s repayment likelihood. The ticker UPST trades on NASDAQ.
The origin and early thesis
Upstart was founded by Dave Girouard (former president of Google’s enterprise division), Paul Gu, and Anna Counselman, with the thesis that machine learning could improve credit decisions for consumers, lenders, and investors. Traditional credit underwriting relied heavily on credit bureaus—FICO scores, payment history, and debt-to-income ratio. These signals work on average but exclude many creditworthy borrowers who lack strong credit histories, like young people, immigrants, or those recovering from past financial stress. Upstart’s claim was that alternative data and machine-learning models could identify credit risk more accurately than traditional methods, benefiting lenders who could approve more borrowers and reduce losses, and consumers who could access credit at better terms.
The company initially focused on personal loans—unsecured credit. Banks and online lenders adopted Upstart’s platform to originate and underwrite loans. Upstart earns revenue by charging per-loan fees (often called “funding fees”) and sometimes by taking a cut of the interest spread. The business model separates the underwriting (which Upstart does) from the funding of loans (which banks and investors do), so Upstart does not hold significant credit risk on its balance sheet.
Expansion into auto lending and the 2020s cycle
In the late 2010s and early 2020s, Upstart began pivoting into auto lending, which is a larger market than personal loans. Auto lending is secured credit (the lender can repossess the car if the borrower defaults), so it is lower-risk than unsecured personal loans, which meant Upstart’s AI thesis could be applied to a much larger addressable market. The company built partnerships with auto dealers, captive finance arms of auto manufacturers, and banks to provide real-time underwriting at point of sale.
The years 2020–2021 were expansive for Upstart. The Federal Reserve slashed interest rates to near zero, credit became cheap and abundant, and loan origination volumes surged across the industry. Upstart’s platform benefited from the boom—volume grew, take rates improved, and the stock soared. Investors in growth-stage fintech were hot; venture capital flowed into the sector, and several competitors emerged (Affirm for point-of-sale lending, Klarna for buy-now-pay-later, and others).
The 2022–2023 contraction and cyclicality
The thesis collided with the credit cycle in 2022–2023. As the Federal Reserve raised interest rates sharply to combat inflation, credit suddenly became expensive, and loan origination volumes collapsed. Consumers took on less debt, credit availability tightened, and default rates began to rise. For a company like Upstart that earns revenue per loan originated, a drop in origination volume is immediately damaging.
More problematically, Upstart’s credit models began to underperform. The machine-learning models had been trained on years of loose-credit-environment data. When the credit cycle swung hard and quickly, borrower behavior changed—defaults rose faster than the models predicted—and lenders, stung by unexpected losses, scaled back their purchases of Upstart-originated loans. This is the core risk of AI-driven credit underwriting: the models are only as reliable as the data they are trained on, and a regime change in the credit cycle or the macro environment can invalidate them.
Upstart cut costs aggressively in 2023, reducing headcount by roughly 40% to improve unit economics. The company shifted its language from ambitious growth targets to emphasis on profitability and conservative underwriting. This is a common pattern when AI or machine-learning companies encounter a downcycle: the hype recedes, the model’s limitations become apparent, and the company hunkers down.
The business model and revenue drivers
Upstart’s revenue streams are: (1) per-loan funding fees paid by lenders or loan originators, typically $10–50 per loan depending on type and size; (2) origination revenue (a percentage of the loan amount), which can be 0.5–1.5% for auto loans; (3) servicing fees if Upst holds or services loans; and (4) data and analytics services sold to banks. The bulk of revenue historically came from per-loan fees.
The cyclicality is severe: in boom times, low rates and abundant credit mean loan volumes are high, and Upstart scales rapidly. In busts, volumes collapse, and the company faces fixed costs that become unsustainable. The leverage to credit cycle is higher than for most fintech because Upstart has no deposits, no bank charter, and no net-interest-margin business to stabilize on; it is a pure-volume play on origination activity.
Geographic expansion is also part of the strategy. Upstart initially focused on the U.S. but has begun pilot programs in other countries to diversify origination sources and reduce U.S. cycle dependence.
AI as competitive advantage and liability
Upstart’s value proposition rests on the claim that its machine-learning models are better at credit prediction than traditional methods. If that is true, it is a durable advantage: better models mean better risk-adjusted returns for lenders, which means lenders choose Upstart. If the models are not better, or if they are only better in certain market environments, the advantage vanishes.
The 2022–2023 experience revealed that Upstart’s models, like any machine-learning credit models, can degrade when the environment changes. This is not unique to Upstart—all machine-learning models face this risk—but it is particularly acute for a company whose entire business model depends on proving superior predictive power. Rebuilding trust with lenders after a downturn where the models underperformed is difficult and slow.
Competitive landscape
Upstart competes with traditional banks’ in-house underwriting, other fintech underwriting platforms, and larger players like the captive finance arms of auto manufacturers and bank holding companies. Some of its early partners (like Uptake for equipment lending) have merged or failed. New competitors continue to emerge, and the market for AI-driven credit underwriting attracts venture capital and large technology companies’ attention.
The barrier to entry is lower than Upstart would hope: lending and credit assessment are not new problems, and dozens of startups are building machine-learning credit models. Upstart’s current advantage, if any, is in the breadth of lending products it covers and the relationships with lenders that it has built.
Path forward: profitability or scale?
Upstart faces a strategic choice: compete on being the best, lowest-cost origination platform (which requires scale and network effects), or build a more stable, diversified business that can survive downturns. Aggressive pursuit of loan volume in a soft market can be profitable short-term but risks model degradation and lender losses; conservative underwriting preserves reputation but means slower growth and lower market share.
The company’s long-term value depends on whether AI-driven credit underwriting is genuinely superior to traditional methods across a full credit cycle, and whether Upstart can maintain lender relationships and trust through booms and busts. If AI credit models prove durable and reliable, the market for underwriting services is vast and Upstart could become a critical infrastructure layer. If the models are brittle and environment-dependent, Upstart risks being one of many failed fintech experiments that worked in one market regime but failed to compound over a full cycle.
The broader thesis—that AI can improve financial decision-making—is sound; the execution and cyclicality risks are substantial.