NU › analyze
NU · Analyze
Differential re-analysis · 2026-09-28 · Price $12.49 (−8.6% on the day) · Market cap ~$60.3B
Verdict: WATCH — conviction 6.5 → 7.0. The business got materially better than the July file believed, the price fell 13%, and the two claims that capped the last verdict both resolved in NU's favour. What replaces them is not a credit problem — it is a capital-allocation problem that did not exist eight weeks ago. Fair value $10.00–12.50, base ~$10.75–11.50. Entry $9.50–11.00. Trim 14x fwd.
1. What this updates
Baseline: Output/Stocks/Financial Services/NU/analyze-2026-07-31.md — WATCH, conviction 6.5,
fair value $9.50–12.50, price then $14.33. That file's core call was: "excellent business, wrong
price, and a mechanical downgrade catalyst 13 days away."
Both halves of that call have now been tested by events, and they went opposite ways.
The event list since 2026-07-31
| Date | Event | Ledger rows touched |
|---|---|---|
| 2026-08-06 | Mexico full-bank conversion executed on schedule; formalised $4.2B Mexico investment commitment through 2030 incl. $2.5B capex over 4 years | S9, St5, NEW |
| 2026-08-13 | Q2 FY26 print — first $1B+ quarter. NI $1,061.1M (+49% YoY, +17% QoQ), diluted EPS $0.2162, ROE 33%, risk-adj NIM 12.4%, ETR 14.17% | T1, T2, T3, T7, St1, St2, St6, P2 |
| 2026-08-14 | Stock opened +13.1%, closed +9.3% on 156M shares — largest volume day in the series | J2 |
| 2026-08-14 | Q2 13Fs. Berkshire still absent. Hedge-fund holders 104 → 92 | positioning |
| 2026-08-14/25 | Three insider sales, ~362k shares at $14.71–15.81 (CRO Fragelli, CGO Junqueira ×2). Zero open-market purchases | T9 |
| 2026-08-31 | First-ever Investor Day announced — 2026-12-08, NYC | NEW |
| 2026-09-10 | US market entry launched — 3.50% APY deposits via FDIC-insured partner bank; "Nu Global" multi-currency account across 35+ countries with stablecoin yields and BTC/ETH | NEW, S7, T7 |
| 2026-09-16 | FOMC +25bp to 3.75–4.00%, first hike since 2023, 12-0. Median dot implies one more; 16 of 18 expect ≥1 more by YE | P-rows (discount rate) |
| 2026-09-16 | COPOM CUT Selic to 13.75% — 5th consecutive cut, unanimous | T8 macro |
| 2026-09-16 | Itaú BBA: Outperform → Market Perform, PT $20 → $18 (Pedro Luduc) | positioning |
| 2026-09-24 | BCB Relatório de Política Monetária: odds of breaching the 4.5% ceiling 79% → 90% | T8 |
| 2026-09-25 | Lula signs MP 1,393 — "Desenrola 3.0." State buys R$150B of aged household debt by reverse auction at 90–95% discounts | NEW, T3, T8 |
| 2026-09-28 | Sarah Wilson (ex-Circle GC) becomes Chief Legal Officer, reporting to Vélez | NEW |
| 2026-09-25→28 | Monzo. Sky/Kleinman: NU in talks to acquire the UK neobank at £8–10bn (~$10.6–13.3B). Stock −8.6% | NEW, S13, T4 |
Independent full-window sweep (per pitfall-differential-reanalysis-inherits-the-baseline-event-gap):
I re-swept NU's 6-K list back to 2026-04-01, not only since the baseline. The baseline has no event
gap — it caught the 2026-07-20 Porto Real 6-K, the 2026-06-04 buyback authorisation and the CFO
transition. The only pre-baseline item it missed is the 2026-07-06 AGM notice, which is immaterial.
It did, however, miss that ~$500M of the $1.0B buyback had already been executed by 2026-06-30,
which was disclosed only in the Q2 financial-statement 6-K and is material to the per-share row.
2. The delta ledger
❌ RETRACTED — wrong when written
| # | Baseline claim | The error, and its cause |
|---|---|---|
| P2 | "Consensus FY2026E of $0.886 is built on a tax rate management has already disowned… this forces a mechanical downgrade cycle that has nothing to do with credit." Ranked the #1 key risk, "near-certain, dated." | The downgrade cycle did not happen, and the arithmetic had two defects. (a) Reasoning error — it held pre-tax income fixed. Q2 ETR did rise, 8.69% → 14.17%, and NU still beat, because pre-tax income grew +29.6% QoQ and absorbed the whole step. The file itself computed that consensus required "a +30% sequential step-up held for three quarters" and judged it implausible; Q2 delivered +29.6%. (b) Vendor error — the $0.886 it attacked was Yahoo's epsCurrentYear field, not the live consensus. Yahoo's own eps_trend shows FY26 consensus was $0.808 60 days ago; the live figure today is $0.853, while epsCurrentYear still reads $0.88833, frozen and unchanged for two months. The baseline therefore overstated the consensus it was attacking by ~9%, making the gap to its own tax-normalised $0.72–0.77 look larger than it was. Outcome: FY26 estimates were revised UP, not down — 8 upward revisions and 0 downward in the last 30 days; mean PT $17.98 → $18.69; consensus rating unchanged at 1.64 across 22 analysts. Third-party check: no analyst asked a single tax question on the Q2 call — the word does not appear in the transcript. |
| J2 | "Do not initiate before the Aug 13 print… the asymmetry is negative. The SoFi precedent says the market is currently penalising fast-growing digital lenders on provision flow regardless of stock quality. A clean quarter may get partial credit; a soft one gets punished twice, on credit and on tax." | The print was rewarded, hard. NU opened +13.1% on 2026-08-14 and closed +9.3% on 156.4M shares, the heaviest volume in the series. The SoFi read-across was a single-analogue inference applied across a different credit cycle, a different tax situation and a different quarter, and it inverted. (Honest qualifier: the pop did not hold — the stock round-tripped over the following fortnight. But "do not initiate before the print" was a directional call about the print, and the print's own verdict was decisively positive.) |
| T6 (extrapolation only) | "LDR jumped +9.2pp in a single quarter… the direction is one-way and accelerating. At even half this pace LDR reaches ~75% within a year." | It did not repeat. Deposits grew +6.7% QoQ to $45.3B while the total credit portfolio grew +5.9% to $39.4B — deposits outgrew the book and the ratio was flat-to-better. Separately, the baseline's own 58.3% figure does not reconcile with its own inputs: $22.80B net credit portfolio ÷ $42.45B deposits = 53.7%, not 58.3%. The level is an undefined basis; the extrapolation built on it is retracted. The underlying structural concern (S4, no funding moat) survives on separate evidence and is CARRIED below. |
🔄 SUPERSEDED — facts changed
| # | Baseline claim | What replaced it |
|---|---|---|
| T1 | "Risk-adjusted NIM has been flat for two years (9.28–9.63%) while headline NIM hit a record. This is the single most important number in the analysis and it is not improving." | Q2'26 risk-adj NIM = 12.4%, +292bp QoQ from 9.48% — a record, and ~190bp ABOVE the 10.5% guide the baseline said NU had "never earned once in two years." Headline NIM 22.9% (+180bp). Management re-guided to "around the current level for the foreseeable future… we see it as sustainable" — while explicitly refusing to call 12.4% a floor ("I don't want to overcommit there"). Corroboration — this is not an NU artifact: MELI's NIMAL went 18% → 21% (+3pp) in the same quarter, explicitly attributed to a Brazil consumer spread recovery as provisioning normalised, with its 15–90 NPL near historical lows. Two independent Brazilian consumer lenders posted the same shape in the same quarter. Caveat carried forward, and it is the single most important open question in this file: management disclosed that Desenrola contributed ~⅓ of the beat versus internal expectation, and that 4/5 of that benefit landed in Q2 with ~20% left in Q3, after which it stops. A Desenrola-stripped run-rate is plausibly 11.5–12.0%, not 12.4% — still a step-change from 9.5%, but the level is untested until 2026-11-12. |
| T2 | "Cost of risk at 35.4% is the highest in nine quarters." | Cost of credit fell −9% QoQ in absolute dollars ($1,690.8M managerial / $1,482.2M IFRS) on a book that grew +5.9%. Independently triangulated from the vendor income statement: total costs below the pre-tax line FELL $51M (−2.0%) QoQ while net revenue rose $231M (+6.5%) — the entire +29.6% pre-tax jump is a cost/provision effect, not a revenue effect. ⚠️ The ratio itself is UNVERIFIED — NU did not disclose a comparable cost-of-risk percentage and I will not infer one. |
| St1 | "Q1'26 EPS of $0.178 is not a run-rate… Q2'26 is the first full quarter under both the guided 15–20% rate and the higher CSLL." | ETR came in at 14.17% — BELOW the low end of the 15–20% guide, despite LC 224/2025's CSLL step-up being live from 2026-04-01. The guide was neither reaffirmed nor withdrawn on the call. A further basis finding the baseline did not have: on the managerial P&L the same quarter shows a 34.9% ETR on $1,630.3M of EBT — and identical net income of $1,061.1M. The IFRS/managerial tax gap is a presentation artifact that never touched reported EPS. Use IFRS 14.17% as the clean, arithmetically verified figure. |
| St2 | "Honest current-state ROE: 25.4%, not 29–30%." | Reported Q2 ROE 33.0% (+4.8pp YoY); tax-normalised at 20% ≈ 30.6%; at 25% ≈ 28.7%. ROA at 20% tax 4.03% → 4.94%. And the quality of the equity build is better than Q1's: Q1's equity increase was 34% FX translation, whereas Q2's ΔEquity (+$661M) is almost fully reconciled by NI ($1,060M) less buyback (~$444M) — this quarter's book growth is earnings, not FX. |
| P8 | "$14.33 requires a terminal ROE of 26.2% at COE_BRL 14.5% — at the very top of the moat range, and above it under every other calibration." | $12.49 requires terminal ROE 21.7% at the same calibration (20.9% at COE 14.0%, 23.3% at 15.5%). That sits just above the moat analysis's 20% base case and comfortably inside its 15–27% range — a materially different statement from "at the extreme upper bound." Driven by price −13% and book +18% and a higher starting ROE. |
| T8 (Brazil macro) | "Selic 14.25%, and the market repriced hawkishly all year: Focus YE26 went 12.5% → 14.00%." | Selic CUT to 13.75% on 2026-09-16 — the 5th consecutive cut, unanimous, from a 15.00% peak. Focus YE26 Selic 14.00% → 13.50%. Brazil 10yr 14.25%, down ~47bp month-on-month. The hawkish-repricing framing is dead. ⚠️ But the inflation risk went the other way: BCB's own odds of breaching the 4.5% ceiling 79% → 90% (RPM, 2026-09-24), IPCA-15 September re-accelerated to 4.47%, above consensus, and COPOM gave no forward guidance, stressing "serenity and caution." |
| J-row (election) | "Largely orthogonal to NU's drivers — the consensus election trade is state-linked names (BB, XP), not NU." | No longer orthogonal, on three independent vectors. (1) Lula signed MP 1,393 "Desenrola 3.0" on 2026-09-25 — the state buys R$150B of household debt by reverse auction at 90–95% discounts, eligible band 720–1,645 days overdue and <R$10,000, i.e. squarely NU's aged retail paper. (2) Flávio Bolsonaro's "SOS Brasil" proposes refinancing consumer debt at inflation-only — a 0% real rate, which is an interest-rate cap in substance. Both leading candidates are running on household-debt relief. (3) BCB Governor Galípolo has flagged household indebtedness and is preparing macroprudential measures explicitly targeting credit cards and non-payroll personal loans — NU's two highest-yield products. The trade remains state-linked; the policy is now pointed at NU. Polling also tightened: Datafolha R1 Lula 40 / Flávio 36 (from 40–45 vs 31–36), R2 46/44 — a technical tie. |
📉 DRIFTED — moving against the thesis, not yet breaking it
| # | Claim | Movement | Where it breaks |
|---|---|---|---|
| T3 | "NPL 90+ has held a 6.1–7.0% band for nine quarters and sits below the Q3'24 peak. On the back-book measure there is no deterioration." | 6.5% → 6.9%, +35–40bp, now at the top of the band and 10bp off the Q3'24 peak. The CFO volunteered it: "if you look at 90-plus over the past two years, the general trend is upwards, and that's being driven by the mix." The baseline's kill-criterion on 15–90 did not fire — that reversed to 4.8% (−16bp) — so the leading indicator improved while the lagging one worsened, which is the expected migration shape, not a new deterioration. | A sustained 90+ print above 7.0%, or a 15–90 that stops improving seasonally |
| T3b | "Coverage 249% of 90+; allowance at 16.2% of book is the highest reserve ratio in the series. NU is over-reserved, not under-reserved." | Allowance rose to $6,642M (+~$540M) but the book grew faster: allowance/portfolio ~14.4% and coverage ~209% (both inferred, neither disclosed). ⚠️ Single-source, inferred — flagged. If it holds, part of the risk-adjusted NIM beat is thinner relative provisioning, not better credit. This is the most important unresolved tension in the file and it is adjudicated in §3. | Coverage below ~180%, or allowance/portfolio below 13% |
| S8 / S10 | "~41% of net income sits in the one product line consignado is purpose-built to refinance away" and "secured lending is ROE-dilutive by construction." | Secured mix went backwards: 8.1% → 7.9%. Secured grew +30% YoY against unsecured +45% and cards +35%. Mix is shifting further toward unsecured — which is management's own stated driver of both the NIM expansion and the 90+ drift. Vélez escalated intent on payroll ("one of the leading players over the next 18, 24 months") but also disclosed that early cohorts were "very high risk," not the low-risk refinancers expected. Origination share update UNVERIFIED. | Secured below 7%, or a payroll push at the 10–15% FPD Lago disclosed in Q1 |
| T8b | "System delinquency 4.7%, highest since 2011; household debt/income 49.9%; debt service ex-housing 27.1%." | Every level metric that updated got worse: individuals free-credit NPL 7.6% → 7.8% (July), debt service ex-housing 27.1% → 28.9%, Serasa negativados 83.3M → 83.98M (~51% of adults). But the second derivative turned — September's monthly increase was the smallest of 2026, driven by 19.7M renegotiations May–Aug, +36% YoY. ⚠️ That is renegotiation- and now policy-driven, not organic household repair, and Desenrola 3.0 will mechanically accelerate it. BCB system-wide August/September delinquency: UNVERIFIED. | Renegotiation-adjusted delinquency resuming its climb once Desenrola washes through |
✅ CARRIED — re-tested, still true
| # | Claim | The check |
|---|---|---|
| S1 | ROA 2.5–3.1x Brazilian incumbents, surviving tax and leverage normalisation, with far less balance-sheet risk | Q2 ROA at 20% tax 4.94% vs Itaú ~1.8% = 2.7x. Equity multiplier 6.25x vs Itaú ~11.7x. Both inside the baseline's stated ranges and the gap widened |
| S2 | Moat NARROW — cost advantage real vs incumbents; switching costs structurally near-zero; no network effects | No new evidence either way. Brazil's PIX/Open Finance portability architecture is unchanged. Standing prior holds |
| S3 | Cost moat level elite, rate of change peaked; AI is moat-narrowing for NU | Efficiency 17.6% → 19.5%, FY guide reaffirmed at ~20%. The ratio moved toward the guide exactly as the baseline said it would. Strengthened by a force the baseline did not have: the US launch and Mexico's $2.5B capex programme are new fixed cost against the same ratio |
| S4 | No funding moat — deposit cost flat vs interbank; incumbents structurally cheaper | Independently corroborated from the income statement: Q2 interest expense +22.7% QoQ against interest income +11.4% QoQ. Spread capture (NII ÷ interest income) fell 70.3% → 67.3%. The funding side is getting more expensive relative to the asset side, exactly as claimed |
| S5 | Converting from a payments/float business into a lending business — quality of earnings deteriorating even where quantity is not | Strengthened. Secured mix fell to 7.9%; unsecured +45% and cards +35% vs secured +30%. Management named the unsecured mix shift as the driver of both the NIM beat and the 90+ drift — which is precisely this claim, stated by the company |
| S6 | BACEN is an attacker NU cannot defend against, and its agenda is now pointed at NU | Two fresh instances. BCB is studying a ban on revolving credit inside PIX Parcelado, rules expected November 2026; and Galípolo is preparing macroprudential measures on cards and non-payroll personal loans specifically. Knowledge/Themes/brazil-fintech-regulation predicted exactly this shape |
| S7 | 91.3% Brazil; single operating segment, no country P&L | Still a single segment. Still no country P&L, no country ROE, no Brazil CET1. Mexico ARPAC ($12.3) and a Mexico-only LDR (35%) were disclosed on the call — a marginal improvement, and note several aggregators are misreporting that 35% as consolidated |
| S9 | Mexico is the only accretive defence; worth ~+2.5pp terminal ROE | Executed on schedule 2026-08-06. 15.8M customers → 16M by end-July. NEW: a formalised $4.2B commitment through 2030 incl. $2.5B capex — the option is real but it is no longer free, and the baseline did not have this cost |
| S11 | The competitive field is thinning | Nothing reversed it. No new entrant; Mercado Pago's Brazil licence still deferred |
| S12 | The company is evergreen; the current profit engine is not. Terminal ROE 20% base, 15–27% range | Unchanged as a structural judgment, and it is the frame that made the Q2 beat readable as a cyclical inflection rather than a permanent re-rating |
| T4 | Per-share discipline: dilution +0.77%/yr, SBC falling in absolute dollars, a first buyback | REFRESHED upward — the strongest positive in the file. Shares outstanding FELL 4,861.7M → 4,830.7M (−0.64%), the first sequential decline ever. ~$500.4M of the $1.0B authorisation was executed by 2026-06-30 — 50% in under four weeks, disclosed only in the FS 6-K and never mentioned on the call. Diluted WAS −5.5M QoQ |
| T5 | Revenue CAGR strong and decelerating cleanly; EPS CAGR tracking net income | Q2 net revenue +50.2% YoY, diluted EPS +66.3% YoY. Deceleration continues cleanly |
| T9 | Insiders sold into a 41% drawdown and bought nothing | CARRIED and strengthened. Three more sales, ~362k shares at $14.71–15.81 — including the Chief Risk Officer selling 221,707 shares the day after the Q2 print. Zero open-market purchases. Every $0.00 line in the feed (Livingston's 2,194,587-share inducement award; eight director grants) is a code-A grant, not a buy — pitfall-yahoo-insider-purchases-counts-rsu-grants applies exactly as the baseline said |
| P1 | Yahoo's forward P/E runs off FY2027 consensus, not FY2026 | Still live. forwardPE 11.03 on forwardEps $1.1299 ≈ FY27 consensus $1.101. FY26 consensus $0.853 gives 14.6x. The understatement is now ~32% |
| St3 | Yahoo SharesOut = 3.81B is Class A only; true total is ~4.86B |
Still wrong today: 3,808,087,961. The correct field is impliedSharesOutstanding = 4,830,688,659. pitfall-yahoo-share-count-dual-class-fpi confirmed live |
| Vendor revenue | Yahoo/roic.ai FY25 revenue $10.63B is a net construct, understating IFRS revenue 32.6% | Still wrong today ($10.63B). Note the consensus revenue line uses the correct basis — yearAgoRevenue reads $15,774,741,000, matching IFRS exactly. The corruption is confined to the vendor income statement |
| Graham IV | Structurally invalid for a high-ROE, low-book financial — weight 0% | fin.py prints $6.71 against a $12.49 price. Still excluded |
| P7 | Sum-of-the-parts: "you get Mexico for free" is inverted | Strengthened — the $4.2B Mexico commitment and the US launch are now explicit cash costs against that residual |
| St4 | Banco Porto Real / BCB approval vs the 2026-11-28 naming deadline — binary, brand-existential, apparently unpriced | CARRIED and now urgent. No BCB approval as of today; no 6-K since 2026-09-10. Two months to the deadline. Coverage is essentially zero while Monzo absorbs all attention — the "apparently unpriced" characterisation holds. NU's own position is that "a marca não muda," but that assertion depends on obtaining the licence via Porto Real, which is the unresolved item |
🆕 NEW — no baseline counterpart
| # | Finding |
|---|---|
| N1 | ★ Monzo. Sky News (Kleinman, 2026-09-25), with Advent International separately eyeing a minority stake. Reported, not confirmed, not denied — NU: "does not comment on rumours or speculation"; Monzo declined. No 6-K. Stage: heads of terms; Monzo advised by Morgan Stanley, Qatalyst, Slaughter & May; a sale to NU is one of several board options and the FT reports NU may end up taking a stake rather than acquiring. Detail and arithmetic in §4 |
| N2 | US market entry launched 2026-09-10 — 3.50% APY deposits via an FDIC-insured partner bank, plus "Nu Global": multi-currency across 35+ countries, stablecoin deposit yields, BTC/ETH. National bank charter applied 2025, preliminary conditional approval January 2026. The baseline only had Vélez's downside cap ("less than 100 basis points on our efficiency ratio" if no product-market fit). It is now live and spending |
| N3 | Vélez: NU is "not in a mode of optimizing for net income" but "investing a lot in growth." Read with Monzo, the US launch and the $4.2B Mexico commitment, this is an explicit capital-allocation posture change and the most consequential sentence on the call |
| N4 | Desenrola 3.0 / MP 1,393, signed 2026-09-25 — R$150B reverse auction at 90–95% discounts, nine days before round 1. Also a ~R$42.3B unbudgeted package landing mostly in 2027 — fiscally expansionary into the inflation risk BCB just raised ceiling-breach odds on |
| N5 | First-ever Investor Day, 2026-12-08, NYC. A likely venue for a guidance reset or a formal multi-year framework — and, if Monzo is real, for the strategic case |
| N6 | Sarah Wilson, CLO, effective 2026-09-28 — ex-General Counsel of Circle, where she co-led the June-2025 NYSE IPO and ran M&A. Reports to Vélez. A securities/M&A/stablecoin profile hired in the same fortnight as the Nu Global launch and the Monzo reports. Suggestive, not evidence |
| N7 | NU is not participating in the Brazil rally. EWZ +16.9% YTD, Ibovespa +5.1% over the past month on an explicit election trade — while NU is −30% YTD and near its 52-week low. But decomposed: through 2026-09-25, NU −5.2% vs ITUB −4.1% and BBD −5.5% — indistinguishable from Brazilian bank beta. The idiosyncratic damage is the Monzo day alone |
| N8 | FOMC turned. +25bp on 2026-09-16 to 3.75–4.00%, first hike since 2023; median dot implies one more; 10/28 odds 65–72%, above the 60% in the brief. US 10yr 5.17% (above 5.00% for the first time since 2007), also above the ~4.96% in the brief. DXY +1.3%. EM HY credit showed its "first meaningful crack" of the year, and EM local-currency debt −0.88% in the week to 9/19, almost entirely FX |
| N9 | The BRL tailwind is gone. USDBRL 5.078 (7/31) → 5.220 — BRL −2.7% since the baseline, and the 12-month appreciation has compressed from ~9.5% to ~2%. Explicitly attributed to the narrowing BR–US rate differential. The FX flattery in reported USD results is spent |
⏳ UNTESTED — could not verify or could not corroborate
| # | Row | Why |
|---|---|---|
| U1 | Cost of risk as a %, credit yield % (baseline 62.7%), coverage % and allowance/portfolio % as disclosed, and Q2 ROA | NU disclosed dollars, not ratios, and I will not substitute the "credit = 41% of gross profit" figure, which is a different metric. My 209%/14.4% are inferred and labelled as such |
| U2 | Consolidated loan-to-deposit ratio on a primary-source basis | The baseline's 58.3% does not reconcile with its own inputs (53.7%), and the Q2 figure is single-source secondary. ⚠️ The widely circulated 35% is Mexico-only and is being misreported as consolidated by several aggregators |
| U3 | Consignado / payroll origination share (baseline: NU 0.4% vs Itaú 18%) | Qualitative escalation only. No updated share table exists in reachable sources — the same gap the baseline flagged, now UNTESTED for a second consecutive pass |
| U4 | Brazil standalone ROE / country P&L / Brazil CET1 | Still never disclosed. Last Brazil ROE was ">40%", Q1'24. Second consecutive pass untested |
| U5 | NTN-B 10yr real yield and the DI curve | The baseline's stated blocker, still unresolved. ANBIMA and B3 pages not reachable. Partially mitigated: I now have an actual Brazil nominal 10yr of 14.25%, which anchors the COE band from below rather than leaving it a free parameter |
| U6 | Brazil 5yr sovereign CDS, EMBI/CEMBI spread levels, EM equity fund flows in dollars | Qualitative EM-weakness evidence only |
| U7 | BCB system-wide delinquency and household debt/income for August–September | The August 28 Nota de Crédito exists; the table could not be extracted. The largest remaining gap on the credit question |
| U8 | Itaú individuals 15–90 (baseline 3.0%) and Bradesco PF 90+ (baseline 5.4%) | Figures surfaced conflict with the baseline's and were not asserted. Direction is corroborated — Itaú flat, Bradesco retail worsening |
| U9 | Monzo consideration mix (cash / stock / mixed) and financing | The single most decision-relevant unknown in this file. "Likely cash-and-stock" appears only in unattributed aggregator commentary. §4 presents an all-stock sensitivity as a bound, not a forecast |
| U10 | Tencent/Sequoia/SoftBank positions and lock-up overhang; Berkshire exit date; UK redomiciliation status | All carried forward untested from the baseline. On redomicile: no reporting links the Monzo talks to it. That connection is inference, not reported fact |
3. How the close calls were decided
3.1 Is the 12.4% risk-adjusted NIM real, or is it thinner reserving?
This is the central adjudication of the pass, because two rows point opposite ways: T1 says the margin inflected; T3b says coverage fell from 249% to ~209% and allowance/portfolio from 16.2% to ~14.4%. Release reserves and you manufacture exactly this print.
Weighed for the inflection: - The pre-tax line, computed independently of any disclosure. Costs below the pre-tax line fell $51M (−2.0%) QoQ while net revenue rose $231M. Provisions genuinely fell in dollars on a book that grew 5.9%. - MELI, an independent issuer with an independent auditor, posted the same shape in the same quarter — NIMAL 18% → 21%, explicitly a Brazil consumer spread recovery, with 15–90 NPL near historical lows. This is the corroboration the command demands before a claim moves. - The leading indicator improved: 15–90 NPL 5.0% → 4.8%. Reserve release does not move a leading delinquency bucket. - Itaú's cost of credit was stable at 2.7% of average loans with 90d NPL flat at 1.9%.
Weighed against: - The coverage and allowance ratios are inferred, not disclosed, and both moved the wrong way. - Desenrola contributed ~⅓ of the beat versus internal expectation and is ~80% spent. - Management refused to call 12.4% a floor — the most informative thing said on the call. - Every Brazilian household level metric still got worse; the Serasa improvement is renegotiation-driven, and Desenrola 3.0 will make it more so.
Decision. T1 is SUPERSEDED — the "flat for two years" claim is factually dead and cannot be carried. But the level is explicitly not underwritten at 12.4%. The working assumption is a Desenrola-stripped 11.5–12.0%, and T3b is carried as DRIFTED with its inference flagged, so the next pass tests reserving rather than inheriting a clean bill. Neither row is allowed to resolve the other — they are logged as a live tension, and Q3 on 2026-11-12 is the test.
3.2 One event, or four independent confirmations?
The command warns against counting one force several times. Q2 is one event and it moved T1, T2, St1, St2, St6 and P2 together. The tax normalisation, the pre-tax surge, the ROE step-up and the consensus upgrade are one fact — provisions fell — refracted through six rows. I have weighted it as one, which is why conviction moves half a point and not two.
Symmetrically, N1/N3/N2/S9's $4.2B are also one event: a strategic decision to spend the balance sheet on multi-front expansion. Monzo, Nu Global, the US launch and the Mexico capex commitment are four expressions of the same posture change, not four separate risks.
3.3 The baseline as a witness
Put to the file: "You concluded the tax normalisation forced a downgrade cycle because consensus was $0.886 on an 8.69% rate. Estimates went up instead. Which premise failed?"
Answer: both, and in different ways. The consensus figure was a stale vendor field — the real number was $0.808. And the inference held pre-tax income constant across a step-change in provisions. The file was right that the tax rate would rise and right that Q1 EPS was not a run-rate; it was wrong that either implied a downgrade. Facts moved and the reading was wrong — which is why P2 is RETRACTED rather than SUPERSEDED. Note what this does not impeach: every structural row the baseline asserted survived re-testing, and its three flagged vendor traps are all still live today. The file's fact-checking was excellent; its forecasting was not. That distinction is why conviction rises rather than falls.
3.4 Where forces balance, the baseline stands
- S2 (no network effects, near-zero switching costs) drew no new evidence either way. CARRIED.
- S12 (evergreen company, non-evergreen profit engine) is the frame that let the Q2 beat be read as a cyclical inflection rather than a permanent re-rate. Re-weighting it on one good quarter would be exactly the flip-flop the command warns against. CARRIED.
4. Monzo — the force that changes the question
Status: reported, not confirmed, not denied. One journalist's anonymous sourcing (Sky, 2026-09-25). Bloomberg, Axios, Telegraph, CityAM, Forbes and Retail Banker International are all downstream of it. Breadth of coverage is not independent corroboration. No 6-K eight days into the leak.
What Monzo is (FY to 2026-03-31)
| Monzo | NU | |
|---|---|---|
| Revenue | £1,712.3M (+39%) | ~$22.8B FY26E |
| Adj. pre-tax profit | £172.6M · net income £86.3M | $1,061M per quarter |
| Customers | 15.2M (→16M) | 139M |
| Deposits | £25.7bn | $45.3B |
| Total equity | £1,283.5M | $13.25B |
| ROE | ~6.7% | 33% reported / ~30.6% at 20% tax |
| Last mark | $5.9bn secondary, Oct 2024 (last primary $5.0bn, Mar 2024) | — |
The arithmetic that explains the −8.6%
At £10bn (~$13.3B), Monzo is ~2.2x its last mark, ~115x pre-tax profit, and ~7.8x book.
NU would be issuing equity at 4.55x book to buy a bank at ~7.8x book earning one-fifth of its return on equity. That is value-destructive on a book-for-book swap before any integration question is asked.
All-stock bound at $13B (a bound, not a forecast — U9):
| Now | Pro-forma | Δ | |
|---|---|---|---|
| Shares | 4,831M | 5,872M | +21.5% |
| FY26E EPS | $0.849 | $0.718 | −15.4% |
| Tangible BVPS | $2.74 | $2.55 | −7.0% |
| Goodwill created | $409M existing | ~$11.7B | ~29x |
| Blended ROE | 33% | ~16% on stated book | roughly halved |
Why this matters more than the dilution number. The cleanest positive in the entire baseline — and the row I just marked as most strengthened (T4) — is per-share discipline: 0.77% annual dilution, SBC falling in absolute dollars, a first buyback running at $500M in four weeks. A 21.5% stock issuance is the direct negation of the single best thing about this company, executed by management that has never integrated a bank, into a third simultaneous regulatory perimeter (BCB via Porto Real, CNBV in Mexico, OCC in the US — now add the PRA and FCA), while ~$500M of an authorised buyback remains unspent.
The counterweights, stated honestly. NU has zero UK/EU presence and Monzo supplies a UK banking licence and an EU platform. Monzo is genuinely profitable and growing 39%. The FT's read — that NU may take a stake rather than acquire — would be a far smaller commitment. And Advent's parallel interest gives NU a path to walk away without losing the option.
How this is treated in the verdict. It is not in the fair-value model, because the structure is unknown and modelling an unsourced consideration mix would be false precision. It is treated as a live, one-sided risk to the per-share thesis that caps conviction and changes what a dip means. See the standing rule in §6.
5. Thesis persistence and conviction delta
Structural rows: 12 of 13 CARRIED or REFRESHED = 92%. The business is what the baseline said it was. Not one description of the model, the moat, the funding structure, the regulatory exposure or the geographic concentration failed re-testing — and several (S4, S5, S6) were corroborated by new independent evidence.
Trend rows: 4 of 9 CARRIED or REFRESHED = 44%, with 2 SUPERSEDED in NU's favour (T1, T2), 2 DRIFTED (T3, T8b) and 1 partially RETRACTED (T6).
Combined Structural + Trend persistence: 16 of 22 = 73% (baseline pass had no comparator).
The shape is unambiguous: the structural analysis held almost perfectly, and the cyclical read was wrong — in the optimistic direction. That is the classic mis-timing profile, not a broken thesis. It is also the reason conviction moves up rather than down despite three retractions: what was retracted was the forecast, not the description.
Conviction 6.5 → 7.0
Raised by, in order of weight: 1. T1 SUPERSEDED — the file's own designated "single most important number in the analysis" reversed, corroborated independently by MELI's NIMAL. 2. P2 + J2 RETRACTED — the #1 ranked risk did not exist and the near-term asymmetry was backwards. Removing a phantom risk is a genuine improvement in the setup. 3. P8 SUPERSEDED — the price now requires a terminal ROE of 21.7%, just above the moat base case, versus 26.2% at the very top of the range. The stock did the work. 4. T4 REFRESHED upward — the first-ever sequential share-count decline, with 50% of the buyback spent in four weeks. 5. St2 REFRESHED — tax-normalised ROE 25.4% → ~30.6%, BVPS +18%, on earnings not FX. 6. T8 SUPERSEDED — Selic cutting, not repricing hawkishly.
Held back from higher by: 1. N1 Monzo — a one-sided, unquantifiable risk to the best row in the file (−15% EPS, −7% tangible book at the all-stock bound). 2. N3 + N2 + S9's $4.2B — the posture shift from self-funding compounder to multi-front global expansion, in management's own words. 3. T3 + T3b DRIFTED — 90+ at the top of its band, with coverage inferred to have thinned. 4. Election no longer orthogonal — both candidates on debt relief, plus a named BCB macroprudential package aimed at cards and non-payroll personal loans. 5. N8 — the US rate cycle turned against every high-multiple EM financial. 6. St4 unresolved with two months to a brand-existential deadline. 7. My own analysis proved volatile. Three retractions in eight weeks is evidence about this file's forecasting, not only about the company, and per the command that belongs in the conviction number. It is worth roughly half of what the favourable rows would otherwise justify.
6. Updated verdict
WATCH · Conviction 7.0 · Fair value $10.00 – $12.50
Valuation method. Residual income run entirely in BRL-nominal terms per
principle-currency-consistency-roe-vs-coe, then applied to USD book. The improvement over the
baseline is methodological, not just arithmetic: rather than treating cost of equity as a free
parameter across 14.5–17.5%, each (COE, g) pair is calibrated off Itaú's live P/B and ROE and
then applied unchanged to NU — so the discount rate cannot flatter one and not the other.
| g (BRL) | COE_BRL (calibrated to ITUB) | BEAR 26→18 | BASE 28→21 | BULL 29→24 |
|---|---|---|---|---|
| 7.0% | 14.00% | $7.68 | $11.76 | $16.47 |
| 8.0% | 14.51% | $7.18 | $11.24 | $15.98 |
| 9.0% | 15.03% | $6.67 | $10.74 | $15.52 |
| 10.0% | 15.55% | $6.16 | $10.25 | $15.10 |
| 11.0% | 16.06% | $5.62 | $9.77 | $14.73 |
The base case is remarkably stable at $9.77–11.76 across the entire surface — far more robust than the baseline's single-point outputs, precisely because the calibration is self-consistent.
| Case | Value | Prob. | Driver |
|---|---|---|---|
| BEAR | $5.60 – $7.70 | 25% | Risk-adj NIM proves to be a Desenrola artifact and reverts toward 10%; coverage erosion confirmed; consignado crowd-out at base rate; terminal ROE 18% |
| BASE | $9.80 – $11.80 | 50% | Desenrola-stripped risk-adj NIM holds ~11.5–12%; ROE fades 28% → 21%; COE_BRL 14.5–15.5% |
| BULL | $14.70 – $16.50 | 25% | 12.4% is a real new level; Mexico and the US scale without ROE dilution; no large stock-funded acquisition; terminal ROE 24% |
$12.49 sits ~9% above the top of the base band and comfortably below the bull. That is a very different statement from July's "30% above base, inside the bull case."
⚠️ The unresolved COE tension, named. Calibrating to Itaú produces a COE_BRL of 14.5–15.5% against a Brazil nominal 10yr of 14.25% — an equity risk premium of only +26bp to +130bp over the government bond. That is thin. A build-up COE would be several hundred basis points higher and would cut fair value by roughly a third. I keep the market calibration for the same reason the baseline did — a model asserting the whole liquid Brazilian banking sector is mispriced more likely has a wrong discount rate — but this is a known, unresolved weakness in the valuation, not a solved problem, and U5 (the NTN-B real yield) is the fetch that would settle it.
Multiples at $12.49
| Value | |
|---|---|
| P/B | 4.55x (was 5.53x) |
| P/TBVPS | ~4.7x |
| TTM P/E | 17.0x |
| FY26E P/E (consensus $0.853) | 14.6x |
| FY26E on tax-normalised ~$0.845 | 14.8x |
| FY27E P/E (consensus $1.101) | 11.3x |
| Implied terminal ROE required | 21.7% at COE 14.5%, g 8% |
| P/B per ROE point vs ITUB | 0.144 vs 0.096 — a 50% premium (was 75%) |
Action
| Zone | Price | Note |
|---|---|---|
| Genuinely attractive | ≤ $9.50 | Below the base band under every calibration |
| Accumulate | $9.50 – $11.00 | Straddles the base case |
| Hold / no action | $11.00 – $13.50 | Fair to modestly rich. Current $12.49 sits here |
| Trim | ≥ 14x fwd (~$15.85 today) | Requires terminal ROE ~23.5%+, upper half of the moat range |
| Hard trim | ≥ 16x fwd (~$18.10 today) | Above the moat ceiling |
Note on the trim, per CLAUDE.md's valuation rule. The baseline's fixed $15.00 trim fired — NU closed $15.14–15.68 between 2026-08-25 and 2026-09-10, peaking at $15.68 on 2026-09-03 — on a stock whose fair value was rising through that window as book and ROE grew. That is exactly the failure a fixed-dollar trim produces. It is replaced with 14x forward, which the site recomputes from live EPS each build.
★ Standing rule — the Monzo gate
If NU trades into the $9.50–11.00 accumulate zone while the Monzo structure is still unknown, the verdict is DEFER, not ACCUMULATE. A 21.5% stock issuance would move fair value by more than the dip would, so buying the dip because of the deal risk is buying the risk at a discount to nothing. Wait for a 6-K, a denial, or confirmation that it is a minority stake. If it is confirmed as an all-stock acquisition at ~$13B, conviction goes to ~6.0 and the entry band drops to $8.00–9.50 on the diluted per-share math.
Break triggers — any one moves the bear case to base
- Risk-adjusted NIM below 11.0% in Q3 (2026-11-12) — the Desenrola-stripped floor. Management refused to call 12.4% a floor; this is the number that tests whether it was a level or a subsidy.
- 90+ NPL sustained above 7.0% — through the nine-quarter band top and past the Q3'24 peak.
- Coverage confirmed below ~180%, or allowance/portfolio below 13% — would convert the NIM beat into a reserving story retrospectively.
- Monzo confirmed as an all-stock acquisition at ~$13B — see the gate above.
- BCB does not approve Porto Real before ~2026-11-28 — forces a rebrand across 139M customers of the most valuable intangible NU owns.
- BCB macroprudential package lands on cards and non-payroll personal loans with a quantified NIM impact.
- A post-election rate-cap or debt-refinancing mandate of the "SOS Brasil" (inflation-only) shape.
Upgrade conditions — what would take this to ACCUMULATE
- Q3 risk-adj NIM ≥ 11.5% ex-Desenrola with disclosed coverage back above 220% — that converts T1 from a one-quarter print into a level, and it is the single highest-value datapoint available.
- Monzo denied, or confirmed as a minority stake, with the buyback resumed.
- Porto Real approved, retiring the brand binary.
- Price into $9.50–11.00 with the above resolved.
- Investor Day (2026-12-08) delivering a country-level P&L — the disclosure gap that has now gone untested for two consecutive passes.
A portfolio-specific passage was removed from the public build.
7. What this pass did NOT test
Named UNTESTED rows: U1–U10 in §2. The ones that matter most:
- U1 / U3 / U4 — the ratios NU does not disclose. Cost of risk %, credit yield, coverage, allowance/portfolio, consignado origination share, and any country-level P&L. U3 and U4 are now UNTESTED for two consecutive passes — the system keeps deferring the same two questions, and §5's coverage judgment rests on an inference as a direct result.
- U9 — the Monzo consideration mix, which is the single most decision-relevant unknown in the file and the reason for the standing gate in §6.
- U5 — the NTN-B real yield, the baseline's stated blocker, still unresolved. It is the fetch that would convert §6's calibrated COE from a market-implied assumption into a built number, and it is the known weakness in the valuation.
- U7 — BCB system-wide delinquency for August–September. The August 28 Nota de Crédito exists. Without it, the "flow decelerating" read rests on Serasa plus one renegotiation statistic.
Single-source conclusions flagged, so the next pass knows which of today's calls are thin:
| Conclusion | Why it is thin |
|---|---|
| Coverage ~209% / allowance ~14.4% | Inferred by me, not disclosed. Carries the DRIFTED status on T3b and half the tension in §3.1 |
| The entire Monzo story | One journalist, anonymous sourcing, no 6-K. Seven outlets re-reporting one source is one source |
| "Likely cash-and-stock" | Unattributed aggregator commentary. The all-stock math in §4 is a bound, not a forecast |
| Consolidated LDR | Single-source secondary, and the baseline's own figure does not reconcile. Guard against the 35% Mexico-only number, which aggregators are misreporting as consolidated |
| Desenrola ≈ ⅓ of the risk-adj NIM beat | Management's own characterisation on the call, never quantified in basis points. It carries the entire §3.1 caveat and the 11.0% break trigger |
| Focus YE26 Selic 13.50% | Two Rio Times reports conflict (13.75% on 9/16 vs 13.50% on 9/28). The official Focus PDF would settle it. ±25bp, immaterial to the verdict |
| Post-print retracement magnitude | Low-tier secondary sources give 62% vs 78.5%. Direction reliable, magnitude not |
Conflicts logged and deliberately not blended: Brazilian delinquency exists in at least four incompatible series (total system 4.7%, free credit 6.4%, individuals-free 7.8%, and a 5.8% figure from a fourth definition) — they must never be compared across sources. Datafolha's R1 (Lula 40 / Flávio 36) conflicts with an unattributed Gramercy figure (Bolsonaro 42 / Lula 40); Datafolha is used. The brief's own US 10yr (~4.96%) and 10/28 odds (~60%) were both stale — 5.17% and 65–72%.
8. Sources
Primary: Q2'26 press release 6-K · Q2'26 financial statements 6-K · Buyback 6-K 2026-06-04 · Porto Real 6-K 2026-07-20 · Mexico authorization 6-K · US launch 6-K 2026-09-10 · Investor Day 6-K 2026-08-31 · Sarah Wilson · Q2'26 earnings-call transcript · Itaú Q2'26 MD&A 6-K · MELI Q2'26 8-K · BCB RPM Sept 2026 · BCB Copom statements
Monzo: Axios 09-28 · Bloomberg 09-26 · Forbes 09-27 — 2x last valuation, 115x pre-tax profit · CityAM · Investing.com — why NU is sliding
Analysts / macro / election: Itaú BBA downgrade · CNBC — FOMC 09-16 · Rio Times — Selic to 13.75% · Rio Times — Desenrola 3.0 / MP 1,393 · Gramercy EM Weekly 09-19 · Datafolha September · Flávio "SOS Brasil" · Seu Dinheiro — Galípolo macroprudential alert · Serasa — 83.98M negativados
Knowledge base applied: principle-currency-consistency-roe-vs-coe (valuation specification) · Themes/brazil-fintech-regulation (predicted the PIX Parcelado and macroprudential vectors) · pitfall-yahoo-share-count-dual-class-fpi (confirmed live) · pitfall-vendor-revenue-is-net-construct-for-banks (confirmed live) · pitfall-vendor-forward-eps-is-the-wrong-fiscal-year (confirmed live) · pitfall-yahoo-insider-purchases-counts-rsu-grants (confirmed live, 2.19M-share false positive) · pitfall-stale-entry-zone-suppresses-a-name (the $15.00 fixed trim fired on a rising fair value) · pitfall-differential-reanalysis-inherits-the-baseline-event-gap (independent full-window 6-K sweep run; no gap found)
Vendor (used and cross-checked): Yahoo Finance MCP, roic.ai MCP, .mcp/fin.py