AI Lenders Are Growing. Better Loan Returns Need a Different Test.

The signal

Platform expansion is established. A cross-platform ranking of realized underwriting performance is not supported by the accessible evidence.

Growth is real, but its denominator is the platform

Upstart’s third-quarter release reported $4.1 billion of originations, up 44% year over year, and $371 million of revenue, up 34%. Pagaya reported $3.6 billion of network volume, up 29%, and $390 million of total revenue and other income, up 12%. Both releases describe meaningful expansion in business activity. ^[1][2]

Those figures answer how much business passed through the platforms and how much income they recorded during the quarter. They do not answer how the newest loans will perform over their full lives. A platform can grow because it adds partners, broadens products, approves a different borrower mix or secures more funding. Several of those changes can occur before a new loan cohort has experienced a full credit cycle.

Nor is total revenue divided by new lending volume a clean borrower interest rate or platform take rate. Income can include fees, servicing and other items associated with loans from different periods. A ratio can be easy to calculate and still answer the wrong economic question.

The investor’s first task is therefore to maintain two ledgers: the platform’s income statement and the borrowers’ repayment record. Strong evidence in one does not erase the need to inspect the other.

A projected return can be useful without being fully realized

Upstart’s earnings presentation described an approximately 660-basis-point premium over the two-year Treasury from an equal investment in twelve recent quarterly vintages. Its footnote states that the return includes assumed future cash flows, net of servicing fees, and excludes the newest quarter because of limited seasoning. The scope is personal loans, with specified smaller-loan products excluded. ^[3]

This is a valuable expectation measure, but it should remain labelled as a mixture of observed experience and projected future payments. It cannot be presented as twelve fully completed vintages delivering a realized excess return. Treasury yields also describe a different credit and liquidity exposure; a premium is compensation to assess, not automatically underwriting skill.

The strongest constructive reading is that the company’s model, incorporating information available at the reporting date, points to attractive prospective investor economics. The stronger claim—that the outcome has already been earned and can be generalized to future borrowers—requires more evidence.

A fair test revisits the same vintage as it ages. Did the expected loss move? Were realized collections consistent with the earlier forecast? Did the projected return improve because borrowers performed better, the required discount rate changed, or assumptions changed? A single current curve cannot answer all three.

Third-party funding does not necessarily remove platform risk

The Upstart presentation says third parties funded more than 95% of the platform’s outstanding loan balance, with 68% through commitments of up to two years. This is a stock-of-loans funding measure, not the share of one quarter’s originations. ^[3]

Its filing also describes arrangements in which Upstart makes or receives payments when underlying loan performance differs from initial expectations, subject to contractual limits. Certain beneficial interests are valued using projected cash flows. Those structures mean that selling or funding a loan externally does not always sever the platform’s economic exposure. ^[4]

The point is not that committed capital is undesirable. Longer funding commitments can make a platform more resilient and help avoid abrupt origination stops. The question is what the commitment costs, what performance protection accompanies it and how losses are allocated when the loans disappoint.

For equity investors, the platform’s fee growth should therefore be read alongside retained assets, performance-linked arrangements and funding renewal terms. For loan investors, the question is whether a prospective return reflects borrower performance alone or also a contractual protection supplied by the platform. These are different routes to an attractive investment outcome.

Pagaya’s funding progress is evidence of capacity, not final credit outcomes

Pagaya’s shareholder letter states that it raised $3 billion through ABS securitizations in the quarter, with an increasing share committed for a year or longer. Management connects that funding infrastructure to the expansion of its lender network. ^[5]

That is evidence that capital-market access supported activity. It is not the same as saying every securitized loan has performed well, or that all future funding will be available on the same terms. A securitization can allocate risk among investors with different priorities, allowing the platform to finance a pool whose underlying credit risk remains.

The constructive case is that more durable funding and partner distribution allow better credit selection to be deployed at scale. The countercase is that growth depends increasingly on the terms required to attract capital, while public platform profitability does not isolate the eventual economics of each loan pool.

Neither interpretation should be settled by the amount issued alone. The next useful evidence is the relationship between collateral performance, required credit enhancement, financing cost and the platform’s retained exposure across successive transactions.

A tranche upgrade is not an underwriting score

KBRA’s September review covered twelve Upstart transactions: eighteen note classes were upgraded, six affirmed and one downgraded. The agency linked upgraded securities to increased credit enhancement and excluded three young transactions. A separate review covered nine Pagaya motor-loan transactions, affirming forty classes and upgrading ten; it also identified increased enhancement and excluded recently closed deals. Both reviews used July collections reflected in August distributions. ^[6][7]

These are encouraging observations for many rated securities, but the unit is the note class within a transaction. Multiple classes backed by one pool are not independent observations of underwriting quality. An improvement in a senior tranche’s protection can occur as its balance pays down and its protection relative to the remaining collateral rises.

The two reviews also concern different collateral and structures. Comparing their raw upgrade counts would reward the number and design of note classes rather than establish which platform makes better loans. The single downgrade should not be generalized to every Upstart borrower either.

DN did not obtain the detailed dashboards behind these releases: both displayed an ABS Premium Subscription requirement. We therefore report the public review outcomes and their limits, rather than invent a twenty-one-pool loss table from the number of transactions mentioned.

SoFi shows how the reported loss rate can change

As a consumer-credit comparator, SoFi reported a 2.26% annualized personal-loan charge-off rate. It also estimated an approximately 3.4% all-in annualized net charge-off rate, including recoveries, when excluding the impact of late-stage delinquent-loan sales. Its discussion explicitly noted the role of portfolio growth as well as underlying performance. ^[8]

The difference is 1.14 percentage points, or 114 basis points. These are different presentations of credit experience, not evidence that one number is secretly the correct result and the other fabricated. The company provides the adjustment so readers can see how transaction activity affects the headline ratio.

A growing loan denominator also changes the age mix. Newly originated loans have had less time to default than older loans. That can improve a current annualized ratio even while a particular older vintage develops less favorably. Conversely, a shrinking portfolio can make losses look larger relative to the remaining balance.

SoFi is included to demonstrate the measurement issue, not as a matched non-AI control. Its borrowers, products, funding and accounting differ from the platform pools discussed above. A credible comparison would align those features before claiming an underwriting winner.

The test that would justify a stronger conclusion

Better underwriting should show up in comparable loan cohorts over time: similar borrower risk, loan terms, origination environment and months on book, with a consistent definition of realized principal loss. Pricing also matters. A higher return earned by charging a much higher rate is different evidence from a lower loss rate at comparable terms.

The strongest reason to remain constructive is that the platforms are expanding, obtaining funding and reporting useful performance information. The strongest reason to withhold a definitive superiority claim is that the available measures combine projections, changing portfolios and security structures. Both can be true at once.

The illustrative loss-budget tool keeps the denominator fixed. For a hypothetical original pool of 100, realized net principal losses of 4 and remaining principal of 50, a lifetime loss threshold of 8% leaves room for 4 of additional loss—8% of the remaining principal. This is an arithmetic threshold, not a forecast that the remaining loans will lose 8%.

Investors can use that distinction to ask sharper questions. What has already happened? What is still expected? Which party bears the remaining exposure? And what future data would make the current expectation wrong? Those questions turn AI underwriting from a growth narrative into a testable credit proposition.

The evidence, side by side

Static-Pool Loss Budget

Hypothetical static pool; no prediction of actual borrower losses.

Numerical guards check positive denominators, finite non-negative inputs and static-pool principal consistency. User must verify loss definitions, recoveries, no new advances, pool scope and matched seasoning; the calculator cannot verify these.

Scenarios to watch

Underwriting improvement survives seasoning

Matched vintages deliver lower realized losses at comparable pricing and borrower risk, while issuer support does not rise. Confidence in selection quality strengthens.

Conditional DN interpretation

Structure supplies the protection

Rated securities improve because enhancement increases, while underlying pool losses are unchanged or worse. Security performance and underwriting performance diverge.

Conditional DN interpretation

Growth masks an older cohort

A large flow of young loans lowers a current portfolio loss ratio, while older cohorts deteriorate. Vintage tracking reveals what the aggregate cannot.

Hypothetical mechanism

Method and limitations

Research cutoff: 13 November 2026 UTC. DN reviewed eight primary documents: two platform earnings releases, Upstart’s presentation and filing, Pagaya’s shareholder letter, two KBRA public surveillance summaries and SoFi’s earnings release. The evidence panel consists of company disclosure observations and two review cohorts, not a reconstructed loan-level dataset. The linked premium dashboards were inaccessible without a subscription, so the proposed 20–40-pool matched comparison was not completed. We did not combine auto and personal-loan collateral, count note classes as independent pools or convert projected cash flows into realized returns. The SoFi difference is 3.40%−2.26%=1.14 percentage points.

Public summaries do not provide the complete borrower mix, collection histories, pricing and losses required for causal comparisons. Rating-agency opinions incorporate projections and structural protection. Issuer financial performance may depend on retained interests, fees and funding arrangements beyond direct loan ownership. The SoFi estimate is management’s counterfactual presentation, not a DN reconstruction. No claim is made that AI lending is inferior or superior to conventional underwriting on a matched basis.

Primary sources and evidence register

Source dates and reporting periods differ. Several documents from one issuer are one source family, not independent confirmations. The evidence download contains the observations used in this article.

Frequently asked questions

Do higher loan originations prove better AI underwriting?

No. They establish more lending activity. Better underwriting needs comparable realized credit outcomes after allowing for borrower mix, pricing, loan age and financing structure.

Are Upstart’s illustrated investor returns fully realized?

The cited presentation says the measure includes assumed future cash flows. It should be read as a blended observed-and-projected measure, not a completed-vintage return.

Why do rating upgrades not settle the question?

A tranche can gain protection through increasing credit enhancement. Multiple tranches can share the same pool, and collateral types differ across transactions.

Why was the larger vintage panel not published?

The public summaries did not provide enough comparable performance data, and their detailed dashboards required a premium subscription. Publishing a ranking would imply evidence DN did not obtain.