AFYA — Afya Limited

Afya Limited (NASDAQ: AFYA) — deep-value analysis, REVISIT

Date: September 5, 2026 · Price: … (close Sept. 4, 2026) · Market cap: …mn ≈ R…mn · FX: 5.1237 R$/USD Regime: REFRESH over the August 14, 2026 thesis (verdict then: MODERATE BUY, GBL score … INTERESTING) Trigger: a single new filing — 6-K from August 24, 2026, accession 0001292814-26-004318

What I inherit and what I re-derive (refresh regime)

The regime is refresh, so I explicitly state what I take as-is from the August 14 report and what I redo.

I inherit, unchanged, since nothing in the filings moved it: the business description and moat structure (the MEC licensing barrier, the market price of a license at R…mn/seat from the FUNIC transaction), the 2018–2025 acquisition history, the analysis of the SoftBank preferred-shares buyback from November 3, 2025 (R…mn to retire financing at 8.6% and replace it with commercial notes at ~15.5%), the management profile and Thorndike grid with a score of 2.2/5, the October 2025 refinancing and resulting maturity ladder, the Moody’s AAA.br reconfirmation from May 5, 2026, the ENAMED timeline (SERES Ordinances 72–76 from March 16, 2026, 25%-seat cuts at four units).

I re-derive from scratch, since facts changed or the old figures were wrong:

1. The thesis chapter. The new filing isn’t a financial result — it’s confirmation that Bertelsmann is negotiating an Afya merger with Yduqs. This doesn’t adjust the old thesis, it rewrites it: the asset you’re buying at … might not be the asset you hold twelve months from now.
2. The entire valuation. The Monte Carlo simulation in the delta pack (mc-AFYA-20260904.json) is unusable: it ran with oe = …, i.e. FCFE in thousands of R$ given to a script expecting millions — hence the median of … margin of safety and a …% undervaluation probability. The error comes from teza-AFYA-20260814.json, where the oe assumption is written with the unit “thousand BRL.” I re-ran the simulation with correct units.
3. The discount rate. August’s report used r = …%. Selic is at 14.00% (Copom, August 5, 2026) and CDI at …%. A …% cost of equity is 50bp above the local risk-free rate — not a conservative assumption, an order-of-magnitude error in the risk premium. I raised it to …% and explain below.
4. The owner-earnings base. I rebuilt the FCF bridge on TTM through Q2 2026 and normalized it for cash taxes — because, between August 14 and today, a new, dated fact occurred: on July 31, 2026 the company fully paid the R…mn of Pillar Two owed for fiscal 2025.
5. The points left unverified. All three are now closed or reduced (see the red-flags chapter): the Schedule 13G filer is Doma Perpetual, not Bertelsmann; depreciation sensitivity is quantified; challenges against Ordinances 72–76 remain publicly unresolved, but the 20-F says the precautionary measures took effect after the H1 2026 enrollment period.

I don’t build the thesis on the price move. AFYA rose from … to … (…) since the prior report. That’s noise. What isn’t noise is the asymmetry: on August 24, the confirmation day, AFYA fell 0.8% (14.33 → 14.22) while YDUQ3 rose …% (7.87 → 8.88). Since then YDUQ3 has added another 8.6%, to R… — … from the last pre-announcement close, versus … for AFYA. The market read the transaction as a value transfer from Afya to Yduqs. This is information about the thesis, not about price.

Executive summary

The thesis, in one sentence: Afya remains a cheap regulated rent on trailing figures (8.6x profit, reported FCFE yield 10.5%), but as of August 24, 2026 the asset is no longer defined by its own flows, but by a negotiation led by the controlling shareholder, and correcting two model errors — the Monte Carlo unit and a discount rate below the local risk-free rate — moves the reference value from apparent undervaluation to moderate overvaluation.

What happened. On August 24, 2026, Afya filed a 6-K confirming that discussions with Yduqs Participações S.A. (B3: YDUQ3) regarding a possible business combination “are at an early stage,” following a material fact published by Yduqs the same day under art. 157 §4 of Brazilian Law 6.404/1976 and CVM Resolution 44/2021. There’s no binding agreement, contract, or commitment. Brazilian press (Valor, picked up by Money Times, BP Money, Seu Dinheiro) states the negotiator is Bertelsmann, not the Afya board, and that among the points discussed is the ownership structure of the combined entity itself.

Why it matters so much. Bertelsmann SE & Co. KGaA holds, per the Schedule 13D/A filed June 16, 2026, 61,076,080 shares — 34,074,134 Class B (ten votes each) and 27,001,946 Class A — i.e. 72.27% economic under the filing’s own denominator, ~69% relative to the 89.0mn common shares on the balance sheet, and ~93% of votes. Afya is a Cayman company with dual-class voting. The Nasdaq minority has no veto right, no Delaware-style judicial appraisal right, and no independent board with real power. When the controller alone negotiates a transaction that could turn a 44%-EBITDA-margin rent into an education conglomerate with a 33% margin and 50% more leverage, the relevant risk is no longer ENAMED — it’s the exchange ratio.

What the figures show. Q2 2026 results (filed August 13, already in the prior thesis) confirm the deterioration: revenue … YoY at R…mn, below consensus of ~R…; adjusted EBITDA … at 406.5mn, with the margin falling 180bp to 41.8%. Net profit still grew 14.0% — but the entire difference is tax-driven: the effective tax rate was 1.0% in Q2 2026 versus 9.0% in Q2 2025. At last year’s rate, net profit would have grown 4.8%, not 14.0%. Concurrently, net receivables climbed …% in six months, against half-year revenue up 7.0%, and the entire growth in the receivables portfolio is in the past-due buckets: the non-current balance fell 1.5%, while the “over 180 days” bucket grew 24.1% and the 91–180-day one 31.7%. DSO climbed from 69.8 days (dec. 2024) to 74.3 (dec. 2025) and 81.3 (jun. 2026).

Estimated value. The FCF bridge on TTM through Q2 2026 gives a reported FCFE of R…mn (externally verified: the company publishes an LTM FCFE yield of 11%, my calculation gives 10.5% — convergent within 5%). Normalized for cash taxes (the gap between R…mn TTM tax expense and R…mn actually paid closes — 89.9mn was paid July 31, 2026) and for capex at R…mn versus 298.9mn TTM and a guidance of 340–380mn, the base becomes R…mn. With g1 = 6%, r = 16%, gt = 4.5%, without subtracting net debt a second time (FCFE is already a post-interest flow), the base-case intrinsic value is …/share, i.e. … versus price. The five triangulated models give a range of … to …, with the median at …. Monte Carlo over 20,000 scenarios gives a median of … and an undervaluation probability of …%.

Verdict: HOLD IF YOU OWN IT, DON’T ADD. No new buying at … The divergence from August’s verdict (MODERATE BUY, 1… position) doesn’t come from a business deterioration — it comes from three model corrections, two of which are error fixes, not opinion changes: the Monte Carlo unit (error), double-subtracting net debt in an FCFE model (error — August’s report wrote nd = 0 in the rationale and ran nd = 1394), and the discount rate below Selic (opinion, but hard to defend in the other direction). Add on top: a merger negotiation led by a shareholder with 93% of votes, real CADE risk, and no published exchange ratio. The correct positioning is to not pay now for an asset whose definition is being rewritten.