Upstart Q3 2025: Speed Bumps, Strong Signals
- Henry Dierkes
- 4 minutes ago
- 4 min read
Editor’s note for investors/builders: Upstart just delivered one of those quarters that looks messy at first pass and more interesting the closer you get.
What actually happened?
Upstart posted $277M in revenue (+71% YoY), an 11% GAAP net margin, and 26% adj. EBITDA margin. That’s a healthy P&L—just shy of their own revenue guide. The miss wasn’t demand; applications hit a three-year high (2M+). It was the model tapping the brake: approvals came in tighter, average loan sizes fell, and conversion slipped. Classic “AI system sees a risk blip, slows down.”
Translation: the business is running, but the underwriting brain chose caution over speed—temporarily.
The model did what it’s supposed to do (and learned)
Upstart’s macro index flashed hotter by ~0.2 points. The system responded by approving fewer borrowers, pricing slightly higher, and shrinking average ticket sizes. That dinged fee revenue, but separation (who pays vs who defaults) remains at all-time highs, and calibration (predicted vs actual losses) held up. If you believe Upstart’s pitch—foundation-model-style underwriting that updates in near-real time—this is what it looks like in the wild: occasional false negatives (too conservative), not false positives (bad credit).
They also shipped real plumbing: cut calibration-driven conversion swings ~50% and trimmed pricing latency up to ~30% in Q3. That matters. If you’re going to let an AI steer underwriting at scale, you’d better damp the oscillations.
Prime is weird, non-prime looks… normal?
Counterintuitive datapoint: non-prime (<660 FICO) looks closer to pre-COVID norms; prime (720–750) is where defaults remain elevated vs pre-COVID. Why would that be? Non-prime got punched first and already cut spending. Prime consumers may only now be adjusting. Upstart’s models felt that and tightened T-prime even as others leaned into volume. That’s the edge they’re selling: react before the lagging charge-off stats show it.
New products are real—and increasingly automated
Auto, HELOC, and Small Dollar Loans aren’t side quests anymore.
Auto: rooftops nearly doubled QoQ; originations +357% YoY to $128M. Upstart also introduced an “auto-secured personal loan” (lower APRs, likely better conversion).
HELOC: +323% YoY to $72M; instant property valuations and automated approvals ramped from <1% in June to 10% in September and 20% in October.
SDL: +294% YoY to $138M; instant funding now lands cash in ~90 seconds for most approved borrowers.
Automation overall hit 91% of loans end-to-end in Q3. That’s the compounding loop you want: more products → more data → more automation → better unit economics.
The balance sheet hang-up (and why it shouldn’t define the story)
Let’s address the bear case cleanly: loans on balance sheet rose to $1.23B, with R&D products (auto/HELOC/SDL) now ~71% of that. Critics say: capital-heavy, can’t move the paper, multiple deserves a haircut.
Here’s the counter: year-to-date originations are $7.8B; only ~$424M net added to balance sheet. That implies ~94.6% of platform volume funded by third parties. Q3 platform funding was essentially all external; the balance sheet growth likely reflects R&D purchases while new forward-flow/securitization lanes get papered. Management says multiple agreements are targeted across these products by year-end into 2026. We’ll hold them to it—but the funding pipes for core personal loans already look oversubscribed.
Also worth flagging: net interest income printed $19M vs $5M guided, and co-investment fair values stepped up—both hints that realized credit performance is fine under the hood.
New economics lever: dynamic take-rate optimization
Upstart rolled out a ML-driven take-rate model that captures more value when their APR is clearly better than legacy, and compresses when the spread is thin. Concretely: if legacy prices a borrower at 29% and Upstart can clear at 24%, offering 25–26% both wins the borrower and improves unit economics. If legacy is 18% and Upstart is 17%, price it close to 17% and win on conversion. Subtle, but over millions of loans, a big deal.
What still bothers us (and how it resolves)
Three friction points:
Model opacity – inevitable with ML, but communication matters. The fix is more instrumentation, not fewer neurons: show the pathway from early macro signals → pricing → realized losses.
R&D balance sheet – investors need to see the transition: signed forward flows, active securitization, visible runoff. One or two marquee deals across Auto/HELOC/SDL would reset the debate quickly.
Narrative control – Wall Street isn’t giving benefit of the doubt. Better prep, cleaner Q&A, and concrete KPIs (automation rate, latency, conversion volatility bands, time-to-fund) would help.
Direction of travel
Ignore the noise, track the slope: applications at a three-year high; automation still compounding; new products maturing; funding capacity for core personal loans intact; underwriting chose safety during a macro wobble and engineering reduced the wobble. That’s not a broken story. That’s an AI-lender taking its lumps while getting faster.
If management lands the forward flows and offloads R&D paper, Q4–Q1 could look very different from the headline narrative around this quarter.
What we’ll watch next
Signed forward-flow capacity specific to Auto, HELOC, SDL
Conversion volatility vs Q3 baseline (post-calibration upgrade)
Automation share for HELOC (can 20% → 30–40% quickly?)
Net interest income trend as co-investment marks roll through
Mix of T-prime vs non-prime approvals as prime normalizes
Resources
Rebellionaire Upstart Q3 2025 Earnings Review (PDF) — full analysis, charts, and notes. (Download at the top of this post.)
Upstart Q3’25 earnings call transcript and investor materials.
Upstart Macro Index (UMI) methodology.