← back to rankingUPST · Upstart Holdings, Inc.
Financial - Credit Services · mkt cap $3.1B · calls: Q1 FY2026 vs Q4 FY2025
81.0 conviction · conf-adj 74
conf 4/10 partial
enthusiasm:27.0 · trend:8 · quantifies:12 · impact:4 · under_radar:5 · credibility:5 · business_impact:10 · disruption:0 · commitment:6 · confirmation:4
Enthusiasm latest 9 / prev 8 (rising)
Upstart’s AI thesis is underwriting-centric: proprietary models and repayment data widen approval/pricing separation, lift originations at constant risk, and underpin capital access; enthusiasm is high and rising as Q1 adds operational AI (servicing, QA, dealer offers) atop Q4’s model/LLM launches. Management quantifies model and automation impacts more than AI revenue lines, which lends credibility on credit and funnel economics but leaves AI P&L attribution indirect. Analysts rarely ask about AI by name; questions on speed, charter, and margins draw technology answers rather than a separate AI product narrative.
GROUNDED NEXT-FY IMPACT vs CONSENSUS
Grounded on actual base — revenue $1.1B · net income $0.1B · net margin 5.0% · diluted EPS 0.45
These are next-fiscal-year annual uplift estimates, not next-quarter numbers.
Aggregate next-FY est. rev uplift: 3.5% · next-FY EPS uplift: 3.5% · vs analysts: inline · priced in: medium (model's call-read: high; verdict above is the hard-data one used for ranking) · confidence: 4/10
| Claim | Figure | Arithmetic | Next-FY Rev % | Next-FY EPS % |
|---|
Personal loans model accuracy lead +1.4ppt vs benchmark
other · soft | +1.4 ppt accuracy lead (Q1) | Proprietary model-quality metric; no disclosed $ or volume base to map to $1,075.5M revenue without assuming an approval-rate lift. Overlaps the originations claim; not summed. | | |
Models-to-manage 173.6%; 87.4% inaccuracy remains
other · soft | 173.6% models-to-manage; 87.4% remaining | Internal model-coverage/complexity index; R&D roadmap disclosure with no revenue or cost bridge. | | |
~3.5% more originations at equivalent risk (post-default recovery model)
revenue | +3.5% originations at equivalent risk | Originations are Upstart's primary revenue driver: 3.5% × $1,075,521,000 = $37.6M incr rev -> rev_uplift_pct=3.5. EPS sized on CONSENSUS ADJUSTED NI base $181.3M (not GAAP $53.6M, per thin-margin/distortion rule). At default current adjusted net margin ~16.9% ($181.3M/$1,075.5M), incr NI ~$6.4M -> eps_uplift_pct ~3.5%. Model deployed Q1 FY26 -> full-year run-rate, full 3.5% applied; assumes originations proportional to reported revenue (upper bound if subset-only). | 3.5 | 3.5 |
Doubled daily AI-assisted servicing conversation volume
productivity · soft | 2x daily conversation volume | Servicing-cost efficiency disclosed, but no servicing-opex $, cost-per-contact, or human-vs-AI differential -> cannot compute after-tax saving. | | |
HELOC >1/4 of loans fully automated
productivity · soft | >25% automated | Automation share for a small home-equity segment; no HELOC opex or revenue base disclosed to dollarize. | | |
HELOC time-to-close 6 days vs ~40 industry
other · soft | 6 days vs ~40 days industry | Competitive speed proof point; throughput ratio disclosed but no HELOC volume, capacity, or conversion lift % -> no $ bridge. | | |
Auto retail remote signature ~1/4 of transactions
engagement · soft | ~25% adoption | Adoption rate only in a small auto-retail line; no disclosed lift to originations, take rate, or segment revenue. | | |
Model separation accuracy +1ppt to 172.2%; >100x on loans not previously won
other · soft | +1 ppt to 172.2%; >100x on new wins | Underwriting separation metric; '>100x on previously-unwon loans' implies new-volume capture but no disclosed $ volume or win-rate delta -> unquantifiable. | | |
Partner channel originations +24% QoQ; -34% API latency
revenue · soft | +24% channel orig QoQ; -34% latency | Anchored %, but one channel's QoQ (not annualized) and channel share of $1,075.5M not disclosed -> no obtainable base to size company revenue; one-quarter burst != next-FY run-rate. | | |
Verification models lower default rates by 0.8%
other · soft | -0.8% default rates | Credit-quality improvement supporting funding/volume; a default-rate change, not a disclosed % of revenue or loss-provision $ -> no defensible $ without inventing loss intensity. | | |
Auto LLM document verification: -12% funding times
productivity · soft | -12% funding time | Operational speed/efficiency; no auto-segment opex or $/day cost base disclosed to dollarize. | | |
Training data >100M borrower repayment events
other · soft | >100M repayment events | Data-scale moat metric; no FY26 $ uplift bridge. | | |
Platform credit performance vs treasuries (651 bps / 608 bps)
other · soft | 651 bps (Q1) vs 608 bps (Q4) | Investor/lender return proof point on vintages, not management revenue or EPS guidance; no P&L conversion in claims. | | |
UMI ~1.4 vs ~0.8 (2021); identical loans 43% less likely to default vs 2021
other · soft | UMI 1.4 vs 0.8; 43% default differential | Macro-backdrop/model-quality context framing outperformance; not a forward FY26 increment in $ or % of revenue. | | |
Internal AI productivity lift (explicitly no $)
productivity · soft | qualitative, no $ | Management explicitly gives no $, headcount, or opex base; ties to forward operating-leverage/25% terminal-EBITDA path but unanchored -> null per methodology. | | |
Assumptions: All claims adopter-side; no supplier revenue. Only '+3.5% more originations at equivalent risk' is dollarizable: originations are Upstart's primary revenue driver, so ~3.5% volume -> ~3.5% of $1,075.5M revenue (~$37.6M). EPS sized on CONSENSUS ADJUSTED NI base $181.3M (2025), NOT GAAP NI $53.6M, to avoid the thin-margin (~5%) GAAP denominator distortion. Incremental net margin set to the default current adjusted margin ~16.9% (giving eps_uplift_pct ~3.5%); X used GAAP margin off GAAP NI and Y used a higher 20% margin (-> 4.15%) on the consensus base — consensus uses the methodology-default margin on the consensus base, the more conservative defensible figure. Phasing: model deployed Q1 FY26 -> full-year run-rate. Revenue base ~$1.08B -> %s flagged low-confidence.
Top line: Only one claim sizes cleanly: +3.5% more originations at equivalent risk ≈ +3.5% of $1,075.5M ≈ +$37.6M revenue, since revenue is overwhelmingly origination/platform-fee linked. Other top-line-positive signals — partner channel +24% QoQ, >100x accuracy on previously-unwon loans, -0.8% defaults, HELOC/auto automation — point the same direction but lack disclosed volume/share bases and remain unquantified (soft).
Bottom line: Flowing the $37.6M incremental revenue at a 20% incremental net margin = ~$7.5M, or +4.15% on the consensus adjusted NI base of $181.3M. Critically, the bottom-line productivity claims (doubled AI servicing volume, -12% funding times, qualitative internal AI productivity) are all real but carry NO disclosed opex/servicing base, so they cannot be dollarized — and management itself says the internal productivity lift has 'no $ amount.' On the thin ~5% GAAP margin I deliberately did NOT compute EPS% off GAAP NI ($53.6M), which would inflate the figure; the adjusted-base 4.15% is the honest read.
Consensus FY2026 revenue $1,410.4M is +31.1% over the $1,075.5M base; consensus FY2026 EPS $2.31 is +38.6% over 2025's $1.67. The single identifiable AI uplift (+3.5% revenue, ~+4% EPS) is small relative to — and almost certainly already subsumed within — that 31% modeled growth, which analysts build precisely on continued model improvement of this kind. The math does NOT point clearly above consensus; it sits inside it.
MODEL CONSENSUS (impact)
partial
Agree all claims soft except +3.5% originations. Reconciled EPS to consensus-adjusted basis with default margin (~3.5%); minor type relabels; low confidence on a ~$1B base.
Conflicts reconciled
- eps_uplift_pct(originations): X=3.5 (GAAP margin off GAAP NI) vs Y=4.15 (20% margin off consensus NI) -> used 3.5 because consensus adjusted NI base + methodology-default current margin (~16.9%) is the sounder, more conservative figure
- est_eps_uplift_pct: X=~3.5/1.04 vs Y=4.15 -> used 3.5
- type(HELOC automated): X=engagement vs Y=productivity -> productivity
- type(-12% funding): X=cost vs Y=productivity -> productivity
- type(-0.8% defaults): X=cost vs Y=other -> other
| Field | Opus 4.8 | GPT-5.5 |
|---|
| Rev uplift % | 3.5 | – |
| EPS uplift % | 4.15 | – |
| Priced in | high | – |
| vs analysts | inline | – |
| Confidence | 5 | – |
| Top line | Only one claim sizes cleanly: +3.5% more originations at equivalent risk ≈ +3.5% of $1,075.5M ≈ +$37.6M revenue, since revenue is overwhelmingly origination/platform-fee linked. Other top-line-positive claims — partner channel +24% QoQ, +100x accuracy on previously-unwon loans, -0.8% defaults — are directionally favorable but UNSIZEABLE: no channel-revenue share or loss base is disclosed, so they stay soft rather than padding the headline. Net identifiable AI top-line uplift ≈ +3.5%. | – |
| Bottom line | Flowing the $37.6M incremental revenue at a 20% incremental net margin = ~$7.5M, or +4.15% on the consensus adjusted NI base of $181.3M. Critically, the bottom-line productivity claims (doubled AI servicing volume, -12% funding times, qualitative internal AI productivity) are all real but carry NO disclosed opex/servicing base, so they cannot be dollarized — and management itself says the internal productivity lift has 'no $ amount.' On the thin ~5% GAAP margin I deliberately did NOT compute EPS% off GAAP NI ($53.6M), which would inflate the figure; the adjusted-base 4.15% is the honest read. | – |
| Reasoning | Consensus FY2026 revenue $1,410.4M is +31.1% over the $1,075.5M base; consensus FY2026 EPS $2.31 is +38.6% over 2025's $1.67. The single identifiable AI uplift (+3.5% revenue, ~+4% EPS) is small relative to — and almost certainly already subsumed within — that 31% modeled growth, which analysts build precisely on continued model improvement of this kind. The math does NOT point clearly above consensus; it sits inside it. | – |
Rows highlighted where the two models disagreed.
QUANTIFICATIONS
Personal loans model accuracy lead vs benchmark: +1.4 percentage points (Q1) (Q1 FY2026, both)
“In Q1, we increased the accuracy lead of our personal loans model over benchmark by 1.4 percentage points.”
Models-to-manage / remaining inaccuracy: 173.6% models-to-manage; 87.4% of total inaccuracy remains (Q1 FY2026, both)
“Our models-to-manage now stands at 173.6%, while 87.4% of the total inaccuracy remains to be solved.”
Incremental originations from post-default recovery modeling: ~3.5% more originations at equivalent risk (Q1 FY2026 (vs prior model), topline)
“which drove approximately 3.5% more originations at equivalent risk levels relative to our prior model.”
AI-assisted servicing conversation volume: Doubled daily volume (Q1 FY2026, bottomline)
“we doubled daily AI-assisted borrower conversation volume”
HELOC automation share: >1/4 of loans fully automated (Q1 FY2026, both)
“In Q1, more than 1/4 of these loans were fully automated”
HELOC time to close: 6 days vs ~40 days industry average (Q1 FY2026, both)
“we achieved an average time to close of just 6 days from application to signing, a new record for us and a fraction of the industry average of roughly 40 days.”
Auto retail remote signature adoption: ~1/4 of retail transactions (Q1 FY2026, both)
“About 1/4 of retail transactions in Q1 use the remote signature capability we launched late last year.”
Model separation accuracy vs benchmark (Model 24/25): ~+1 ppt to 172.2%; >100x boost on loans not previously won (Q4 FY2025, both)
“increased our measured separation accuracy advantage over our benchmark textbook model by about 1 percentage point to 172.2% on previously observed upstart loans, but had an accuracy boost more than 100x as large on those loans we did not previously win.”
Partner channel originations / API latency: +24% channel originations QoQ; -34% latency (Q4 FY2025, topline)
“driving 24% more channel originations quarter-on-quarter, while lowering latency by 34%.”
Verification model default reduction: -0.8% default rates (Q4 FY2025 launch, both)
“launched a brand-new architecture of our verification models that lowers default rates by 0.8%”
Auto LLM document verification — funding time: -12% funding times (Q4 FY2025, bottomline)
“automating document verification processes with LLM, thereby reducing funding times by 12%.”
Training data scale: >100 million borrower repayment events (Q4 FY2025, both)
“the scale of our training data crossed 100 million borrower repayment events for the first time”
Credit performance vs treasuries (platform proof point for AI underwriting): 651 bps average spread last 12 vintages (Q1); 608 bps (Q4) (Trailing 12 quarterly vintages, both)
“The average return of our last 12 quarterly vintages of loans exceeds U.S. treasuries by 651 basis points”
UMI macro backdrop vs 2021 peak revenue year: UMI ~1.4 vs ~0.8 in 2021; identical loans 43% less likely to default in 2021 (2025 vs 2021 comparison, both)
“statistically identical loans were 43% less likely to default back then as they are today, that our 2025 results surpassed 2021, in spite of that UMI backdrop”
Internal AI productivity (forward/ongoing, not dollarized): Qualitative productivity lift; no $ amount (Q4 FY2025 — ongoing into margin path, bottomline)
“productivity lift. We're already getting and will continue to get from adoption of AI internally”
PAST (realized)
- Q4: Radically improved AI over the last 3 years; Model 24/25 raised separation accuracy ~1 ppt to 172.2% with >100x boost on loans not previously won.
- Q4: Verification model architecture cut defaults 0.8%; first voice LLM for manual verification; LLM doc verification cut auto funding times 12%.
- Q4: Training data crossed 100M borrower repayment events; 2025 credit returns beat 2021 despite worse UMI — cited as proof of AI-first strategy.
- Q1: Extended models to predict post-default recoveries using AI instead of historical assumptions.
- Q1: Personal loans model accuracy lead vs benchmark up 1.4 ppt; models-to-manage 173.6%.
CURRENT (now)
- Q1: Doubled daily AI-assisted borrower conversations in servicing/collections; expanded to mobile and AI-powered payments; deployed AI QA on service calls.
- Q1: Auto dealers can generate firm AI-powered offers across multiple vehicles from one application; ~25% of retail transactions use remote signature.
- Q1: >25% of HELOCs fully automated; ~6-day application-to-signing close vs ~40-day industry average.
- Q4: Applying AI to auto secured personal loans; tripling automated HELOC underwriting share; partnership API/model work driving channel originations.
FORWARD (guidance)
- Q1: AI will transform credit for lenders — winner of the technology/modeling race wins the market; Upstart intends to widen technology lead on core personal loans.
- Q1: Bank charter partly to engage regulators directly on applying AI in lending.
- Q4: Expect AI-powered lending led by Upstart to rapidly gain share in home, auto, and new categories.
- Q4: Expect AI-enabled lending to dominate categories over ~a decade; Upstart aims to be among the leaders with the most sophisticated models.
- Q4: Expect continued internal AI adoption to drive productivity and operating leverage beyond 25% terminal EBITDA margin.
TRACK RECORD — PROMISE vs DELIVERY
58/100 track record mixed 6 calls reviewed
Upstart makes few hard, dated AI scorecards; it reliably ships concrete model milestones (e.g., unified unsecured underwriting) but tends not to close the loop on bold multipliers (10x AI) or percent efficiency targets (50% calibration/marketing gains). Net AI promise credibility is mixed: real delivery on engineering milestones, soft follow-through on headline metrics.
Move to a single underwriting model for both unsecured products in the near future — promised Q4 FY2024
delivered In Q1 FY2025 management reported the unified model was live, with SDL immediately benefiting from embeddings and other core ML innovations.
Win rate higher than any other digital competitor in each product by end of 2025 (on-target credit) — promised Q1 FY2025
partial Management said personal loans were running ahead early in 2025 and later cited best-rate leadership (e.g., HELOC ~300 bps vs fintechs in Q3 FY2025), but later calls never reported segment-by-segment win rates vs all digital peers.
10x leadership in AI in 2025 — promised Q4 FY2024
quietly-dropped Repeated as priority #1 in Q1–Q2 FY2025, then largely absent as a scored objective; later calls cited model wins and separation metrics without revisiting the 10x goal.
Cut unwanted month-to-month volatility in calibration-driven conversion changes by ~50% (from Q3 calibration work) — promised Q3 FY2025
quietly-dropped No later call confirmed the 50% volatility reduction; Q3 itself saw conversion fall from 23.9% to 20.6% on macro-driven model tightening, with no follow-up scorecard.
~50% uplift in incremental originations from the same marketing spend (causal AI targeting, early results) — promised Q3 FY2025
quietly-dropped The 50% figure was not revisited in Q4 FY2025 or Q1 FY2026; later quarters discussed marketing investment and mix without repeating the efficiency gain.
35% compound annual revenue growth rate over the next 3 years (AI-enabled lending platform) — promised Q4 FY2025
too-early After one quarter, Q1 FY2026 management said results put them on track for full-year 2026 guidance ($1.4B revenue, $294M adj. EBITDA); the 3-year CAGR endpoint is not yet testable.
PRICED-IN (REFINED)
MEDIUMEst. revisions falling · Fwd P/E -68.7 · EV/Sales 3.6x
AI claim maps to Servicing Fees, Net, Servicing Fees, Borrower Fees
Analyst price targets have been cut (last month/quarter avg ~40.6 vs last year ~49.9) and rating counts are roughly flat (strongBuy slipped from 3 to 2), so revision momentum is not rising despite aggressive forward revenue/EPS growth already in the consensus. Valuation is still rich on trailing metrics (P/E ~64, EV/EBITDA ~56) and EV/Sales ~3.6x, which embeds a recovery/AI-lending growth story even with a negative near-term fwd EPS. AI-driven efficiency and volume would most plausibly show up in Servicing Fees, Net and related fee lines, not as an unmapped total-revenue bump. Falling revisions argue against a fully priced-in rerating, but rich multiples and baked-in 2025–26 growth keep the AI upside from looking cheap—hence medium.
COVERAGE — ENTHUSIASM TRAJECTORY + CATALYSTS
8Q4 FY20249Q1 FY202510Q2 FY20259Q3 FY20257Q4 FY20259Q1 FY2026
AI enthusiasm across 6 calls — trend → flat
Core ML underwriting depth peaked with Paul and AI Day, then broadened into platform automation without losing conviction.
RECENT AI CATALYSTS & NEWS
BUSINESS IMPACT - QUALITATIVE MATERIALITY
9/10 qualitative impact transformational near-term · mixed evidence
Where AI matters: AI/ML credit underwriting and lending automation
Proprietary ML is Upstart's core product—not a cost tool—driving quantified ~3.5% more originations at equivalent risk, widening model separation vs benchmarks, and automating HELOC/auto flows (6-day close, >25% fully automated); without continued model gains the platform fee and originations thesis collapses.
Caveats: Only one clean P&L bridge (~3.5% originations uplift); most AI claims are model-quality or operational metrics without dollar attribution; Management has quietly dropped bold multipliers (10x AI, 50% marketing efficiency) while credibility on engineering milestones is mixed; If incumbent banks or horizontal credit-AI vendors close the modeling gap, platform take rates and partner dependence compress faster than originations grow
AI DISRUPTION / CANNIBALIZATION RISK tailwind · 3/10
AI commoditizes manual underwriting and legacy scorecards, which expands Upstart's addressable market rather than replacing it; the main threat is well-capitalized banks or vendors replicating alternative-credit models, but 100M+ repayment events, capital partnerships, and regulatory embedding create durable switching costs that generic LLMs do not erase near-term.
OPTIONS / MARKET STRUCTURE
option liquidity: fair
proxy inputs — dollar-ADV $157M · beta 2.262 · px $32.39
source: proxy (no options chain on FMP)
FMP /stable/ exposes no options-chain endpoint on this key, so ATM IV, bid-ask spread and open interest are unavailable. Liquidity below is a PROXY from dollar-ADV, beta and price level (a stand-in for option depth), not measured option-market data.
CONFIRMATION — INSIDERS · 13F · LANGUAGE
Confirming — insiders buying, institutions flat, management language 8/10 committed.
INSIDERS buying 3 open-market buy(s) vs 20 sell(s) — net accumulation
INSTITUTIONS (13F) flat as of 2026-03-31: 51 new / 107 closed positions; 176 increased / 104 reduced; institutional ownership -8.90pp; -61 net 13F holders
MGMT LANGUAGE 8/10 committed Past-tense rollouts and hard metrics dominate; competitive AI framing stays belief- and future-tense.
commit “In Q1, we increased the accuracy lead of our personal loans model over benchmark by 1.4 percentage points.”
commit “which drove approximately 3.5% more originations at equivalent risk levels relative to our prior model.”
commit “In servicing and collections, we doubled daily AI-assisted borrower conversation volume”
VERBATIM AI QUOTES
“Unlike in some areas, the application of AI to credit is an unambiguous good for the consumer, saving them time and money to use on the parts of life that really matter.”
— Paul Gu, Q1 FY2026
“For lenders, AI will transform credit from a structurally commodity-like business to one where the player who wins the technology and modeling race wins the market.”
— Paul Gu, Q1 FY2026
“With a decade-long head start, we believe that race is ours to lose.”
— Paul Gu, Q1 FY2026
“As always, our most important growth lever is improving our underwriting model.”
— Paul Gu, Q1 FY2026
“In Q1, we increased the accuracy lead of our personal loans model over benchmark by 1.4 percentage points.”
— Paul Gu, Q1 FY2026
“Our models-to-manage now stands at 173.6%, while 87.4% of the total inaccuracy remains to be solved.”
— Paul Gu, Q1 FY2026
“This quarter, we extended the scope of our models to predict post-default recoveries, replacing the assumptions we've used historically with the full strength of our AI models.”
— Paul Gu, Q1 FY2026
“This fuller view of loan economics lets us serve more creditworthy borrowers, which drove approximately 3.5% more originations at equivalent risk levels relative to our prior model.”
— Paul Gu, Q1 FY2026
“We're also moving quickly to maximize use of AI across every part of the business.”
— Paul Gu, Q1 FY2026
“In servicing and collections, we doubled daily AI-assisted borrower conversation volume, brought that capability to our mobile app and expanded our AI-powered payment features.”
— Paul Gu, Q1 FY2026
“We also deployed AI-driven quality assurance tools to review customer service calls, giving us a scalable, consistent way to continuously improve the borrower experience.”
— Paul Gu, Q1 FY2026
“We also rolled out a new feature that lets dealers generate firm AI-powered offers across multiple vehicles from a single customer application.”
— Paul Gu, Q1 FY2026
“In Q1, more than 1/4 of these loans were fully automated, and we achieved an average time to close of just 6 days from application to signing, a new record for us and a fraction of the industry average of roughly 40 days.”
— Paul Gu, Q1 FY2026
“The Upstart team has built a highly differentiated AI-powered credit platform and the runway in front of us is enormous.”
— Andrea Blankmeyer, Q1 FY2026
“it's obviously, we think, a decade where there's going to be substantial advances in AI that AI is going to do a lot to transform consumer credit.”
— Paul Gu, Q1 FY2026
“we should have a direct relationship with the regulators in helping them understand what it means to apply AI in the context of lending, help them get that right and do that directly as opposed to trying to do it through a large number of intermediary financial institutions.”
— Paul Gu, Q1 FY2026
“we are very committed to being a model-first, model-led company.”
— Paul Gu, Q1 FY2026
“Paul said to me that most of the trillions of dollars in interest paid by borrowers in this world is unnecessary in the age of AI.”
— David Girouard, Q4 FY2025
“In addition to radically improving our AI in the last 3 years, we've also entirely rebuilt our capital supply for resiliency and competitiveness.”
— David Girouard, Q4 FY2025
“In the coming quarters, I expect AI-powered lending led by upstart to rapidly gain market share across these enormous unsecured home and auto segments as well as additional lending categories in the near future.”
— David Girouard, Q4 FY2025
“We've been building the teams, the skills and the technologies to deploy AI in lending for more than a dozen years, and I believe we are peerless in this regard.”
— David Girouard, Q4 FY2025
“I look forward to partnering with them to show the world the potential of AI-powered credit.”
— Paul Gu, Q4 FY2025
“This combination of scale, consistency and returns is only made possible by our use of AI lending, and we believe will be a growing draw for capital partners, banks and credit unions.”
— Paul Gu, Q4 FY2025
“that our 2025 results surpassed 2021, in spite of that UMI backdrop, is a strong testament to the power of our AI first strategy.”
— Paul Gu, Q4 FY2025
“These models increased our measured separation accuracy advantage over our benchmark textbook model by about 1 percentage point to 172.2% on previously observed upstart loans, but had an accuracy boost more than 100x as large on those loans we did not previously win.”
— Paul Gu, Q4 FY2025
“we redesigned our partnership models and integrated new data sources into our partner APIs together driving 24% more channel originations quarter-on-quarter, while lowering latency by 34%.”
— Paul Gu, Q4 FY2025
“we launched a brand-new architecture of our verification models that lowers default rates by 0.8% and launched our first voice LLM for when manual verification calls are required.”
— Paul Gu, Q4 FY2025
“we began applying AI for the first time to our auto secured personal loan product.”
— Paul Gu, Q4 FY2025
“begin automating document verification processes with LLM, thereby reducing funding times by 12%.”
— Paul Gu, Q4 FY2025
“In Q4, the scale of our training data crossed 100 million borrower repayment events for the first time, 14 years after we started the company.”
— Paul Gu, Q4 FY2025
“We believe that proprietary training data, when combined with the ever better learning algorithms our team develops drives a long-run growth trend that persists independent of any ups and downs of the macroeconomy or competitive environment.”
— Paul Gu, Q4 FY2025
“our view is that AI-enabled lending as a specialty is going to be the dominant market share in all of these categories, probably a decade from now.”
— David Girouard, Q4 FY2025
“We don't think there's anybody else building the sophistication of models that we are.”
— David Girouard, Q4 FY2025
“we're pretty bullish about just how much productivity lift. We're already getting and will continue to get from adoption of AI internally and that we expect to continue to happen and give even more give us even more operating leverage.”
— Paul Gu, Q4 FY2025
ANALYST QUESTIONS ON AI
Q (Q4 FY2025, Simon Alistair Clinch): With auto and home, when we think about the long-term market share opportunity in businesses like this... has anything changed in the environment that means that those kinds of targets [e.g., double-digit auto share] are perhaps not achievable or are they very much in your sight?
A: David Girouard: "I think they're very much in sight." ... "our view is that AI-enabled lending as a specialty is going to be the dominant market share in all of these categories, probably a decade from now." ... "We don't think there's anybody else building the sophistication of models that we are." ... "we think the economics make it inevitable that an AI-centric platform will be used to originate the majority of loans in these categories over time."
Q (Q4 FY2025, David Scharf): As you think about the margin guidance for 2028, it's a notable roughly 400 basis point expansion from the 2026 level... Do you view 25% is an informal kind of steady state level of profitability at scale?
A: Paul Gu: operating leverage from growing the top of the funnel; also "we're pretty bullish about just how much productivity lift. We're already getting and will continue to get from adoption of AI internally" that should add operating leverage; "I definitely don't expect 25% to be some kind of terminal state."
Q (Q1 FY2026, James Faucette): HELOC product... 6-day process versus up to 40 days industry norm — where is that speed advantage showing up (conversion, CAC, loss selection, partner appetite, take rates)? What's driving mix — cross-sell from personal loans or DTC?
A: Paul Gu: "being able to run a 6-day process is huge" — better conversion and lower CAC; big operational savings as loans move from manual to fully automated; "technology and getting that 6-day process are just a huge part of making that happen"; heavier cross-sell from existing base given 20M+ rate-check accounts.
Q (Q1 FY2026, Vincent Caintic): Bank charter economic implications — blog cited ~$200M annual frictional costs and geographic limits; how achievable vs ~$294M EBITDA guide?
A: Paul Gu: much is missed revenue/TAM from state limits solved by national charter; direct origination costs from multi-bank network; separately "it's obviously, we think, a decade where there's going to be substantial advances in AI" in credit and Upstart should work with regulators directly "to apply AI in the context of lending" rather than through many intermediary FIs.
Q (Q1 FY2026, Dan Dolev): Comments on overall health of the consumer?
A: Paul Gu: consumer "largely stable"; "we are very committed to being a model-first, model-led company" and "we let our models detect all that's going on in terms of consumer repayment patterns" — no major news factors showing up yet in the data.