Halyk Bank multiple. MercadoLibre returns. One of those is wrong.

• Priced like a country bank. KSPI trades at 7.8× trailing P/E on 58% Kazakhstan-only ROE. Nu Holdings trades 21.5× on 30% ROE; MercadoLibre 43.5× on 36%. Kaspi is best-in-peer on margin, ROE and leverage at the same time.
• One customer, three engines. Payments, Marketplace and Fintech each generated ₸355–433B of FY2025 net income — three roughly-equal profit pools sharing one 15.7M-MAU base. CAC paid once, monetised three times: take-rate on ₸44.2T of payment volume, 10.5% commission plus ads on ₸6.7T of GMV, and NIM on an ₸11.7T loan book funded ~93% by retail tenge deposits.
• Why the gap exists. Kazakhstan country tag, a January-2025 Türkiye acquisition that flat-lined consolidated net income, and an unrebutted 2024 short-seller file. Each is a real reason to discount — none is a structural reason to discount this much.

The model whose controls already failed once is the one being tested.

• Coverage collapse. Allowance against gross non-performing loans fell from 99% (FY2023) → 90% (FY2024) → 80% (FY2025). Gross NPLs grew +91% over the window; the allowance grew only 53%. Rebuilding coverage to 90% takes ~₸89B of additional provisions — roughly 8% of FY2025 net income.
• The Cost-of-Risk number says the opposite. Management reports Cost of Risk at 2.2% (FY2025) and 0.7% (Q1 2026). Coverage falling and CoR flat cannot both be the truth. The 1H 2026 IFRS interim, due Aug–Sep, prints the answer.
• And the auditor hasn't blessed it yet. The FY2023 internal-control material weakness sat in this exact expected-credit-loss model. Management certified ICFR effective for FY2025 — but the first independent Section 404(b) auditor attestation only lands with the March-2027 20-F. Until then, the data-flywheel claim rests on a model whose controls have failed once in production.

A Kazakh compounder became a two-country bet in fifteen months.

Before (2018–2024). From FY2019 to FY2024, revenue compounded at ~38% and operating margin held a 75–79% band — a profile only a handful of listed businesses match. The story was clean: one country, one super-app, one customer monetised three times, with excess cash returned as dividend.

Pivot (Oct-2024 → Apr-2026). Kaspi signed in October 2024 to buy 65.4% of Hepsiburada in Türkiye for ~$1.13B, closed January 2025, and absorbed ₸429.7B of goodwill plus a ₸89.6B Türkiye segment loss. FY2025 dividend zeroed, FY2026 guidance switched from net income to Adjusted EBITDA, Rabobank A.Ş. licence purchase pending. Tencent took a strategic minority in April 2026; two IG bonds raised at Baa3/BBB−, the second oversubscribed 3.5×.

Today. FY2025 consolidated revenue +60%, net income +3%. Kazakhstan-only underlying net income still grew ~18% — the headline flatness is funding-cost squeeze plus Türkiye drag, not a broken franchise. The unanswered question: does Hepsiburada earn back ₸430B of goodwill, or is Türkiye a five-year ad-hoc impairment risk?

The thesis is not binary on any single print — but three dated events do most of the work.

• Aug–Sep 2026 — 1H 2026 IFRS interim. First disclosure that prints NPL coverage trajectory inside the 2026 cycle. Coverage rebuilding above 85% with CoR ≤ 2.5% confirms the data-flywheel; staying at/below 80% with CoR drifting above 2.5% confirms the bear.
• 2H 2026 — first NBK base-rate cut window. Governor Suleimenov said cuts only possible "in the second half of this year." Each 100 bps cut reprices retail deposits faster than the 9.3-month loan book — mechanical Fintech NIM expansion the FY2026 +5% Adjusted EBITDA guide does not assume.
• March 2027 — first Section 404(b) auditor attestation. Independent verdict on whether the FY2023 ECL-model material weakness is truly remediated. A clean opinion closes the load-bearing forensic concern and likely starts the Halyk-to-Nu re-rate; a qualified opinion re-opens the entire under-reserving file.
• And the Hepsiburada test runs alongside. Management has committed to FY2026 Türkiye Adjusted EBITDA breakeven. Order growth went −11% (Q1-25) → +19% (Q4-25); engagement per consumer is still ~one-quarter of Kazakhstan. Any 2026 quarter back into Adj-EBITDA losses puts ₸430B of goodwill on the impairment clock.