Story

The Full Story

Between Q1 2023 and Q4 2025, Duolingo went from a gamified language app selling discipline to investors, to an AI-first learning platform willing to sacrifice near-term growth for a long-term platform bet. Management beat every full-year guide it issued from FY2023 through FY2025 — the biggest by an order of magnitude in FY2024 — then, at the same shareholder meeting that reported a record FY2025, cut FY2026 bookings growth guidance from "nearly 20%" (what they said they could do) to 10–12% and stepped margin back ~450bps. The story now is credible but fragile: a founder who tells investors he is spending ~$50M of bookings to rebuild the free experience around Video Call and AI tutoring, after a spring 2025 "AI-first" memo backlash that management never mentioned in writing to shareholders. Credibility is high on delivery, medium on communication.

1. The Narrative Arc

No Results

The arc has three acts. Act 1 (2023): a scrappy AI-forward operator with perfect execution — user growth accelerates past 60% YoY, bookings up 45%, margins tripling. Act 2 (2024): the execution engine keeps compounding; AI moves from slogan to shipped product (Max, Video Call); the first subtle flags — a DAU guide-down and AI cost headwind — appear but get absorbed by beats. Act 3 (2025–2026): the strategy gets re-underwritten. An April 2025 "AI-first" internal memo detonates on LinkedIn and never appears in an IR letter; the product teams dial back the edgy marketing voice; and management formally sacrifices short-term bookings to reinvest in the free experience.

The clearest inflection point is Q4 FY2025 (Feb 2026). Twelve consecutive beat-and-raise letters give way to a deliberate step-down. Luis von Ahn frames it as "taking the long view" — the same operating principle that opens the FY2021 10-K — but it is the first explicit investor-facing walk-back in the dataset.

2. What Management Emphasized — and Then Stopped Emphasizing

Because a flat Heatmap only renders a single period, the matrix below is the full series. The pattern is what matters, not each cell.

No Results

Themes that rose: AI went from feature-level in Q1 2023 to existential in Q4 2025. Subject expansion went from nonexistent pre-Q3 2023 to the primary new-user engine by Q4 2025 (Chess launched 2025 and reportedly grew faster than any prior subject). Teaching efficacy and learning science — the "Duolingo Method" — quietly moved from defensive talking point to strategic centerpiece in Q3–Q4 2025.

Themes that faded: The pre-IPO founder equity hurdles language (the "$178.50 60-day VWAP" bullet) appeared every letter in 2023 and disappeared after the first hurdle was hit. Cost discipline framing — the defining message of 2023 when Duolingo was proving it could be profitable — got progressively de-emphasized once margins reached 25%+. Family plan callouts, a Q1 2024 centerpiece, quietly dropped by 2025.

The quiet pivot: starting Q3 2025, the word "teaching" replaces "monetization" as the controlling verb. Management explicitly says it shifted A/B test weights toward "teaching better and user growth" rather than subscription conversion. This is the first philosophical re-underwrite in three years of letters.

3. Risk Evolution

No Results

Three risk-evolution stories actually matter.

First, AI went from absent to dominant in 24 months. The FY2021 10-K has zero AI-specific risk factor. FY2022 adds a one-paragraph regulatory caveat. FY2023 adds a full product-risk factor explicitly naming generative AI, Duolingo Max, frontier AI launch partnerships, and "hallucinatory behavior." FY2024–FY2025 expand it further with the EU AI Act effective date (Aug 1, 2024), Colorado, Utah, and Texas state AI acts, and the Dec 11, 2025 federal executive order. The competition risk factor was silently edited in FY2023 to insert a parenthetical — "a new product could gain rapid scale at the expense of existing brands through harnessing a new technology (such as generative AI)" — that persists through FY2025. That one parenthetical is the single cleanest signal of how management's threat model changed.

Second, Apple concentration got worse, not better. Apple App Store share of revenue moved from 50% (FY2021) to 62% (FY2025). The risk-factor language is almost verbatim between the two filings, so the heatmap shows a flat 5. But the underlying exposure rose materially — the risk is the same, the magnitude is higher, and management never called it out quantitatively in prose.

Third, the profitability frame flipped. FY2021: "we have incurred operating losses each year since our inception." FY2023 onward: "we have recently achieved profitability, which we may be unable to sustain." This is a genuine de-risking. Transitional IPO risks — lock-up, limited operating history, freely-tradable-shares language — evaporated across the five filings, while the dual-class/controlled-company paragraph remained completely unchanged. Governance has not relaxed as the company matured.

Smaller shifts worth flagging: the Beijing-office risk line ("over 30 employees in Beijing makes it easier for the Chinese authorities to bring enforcement actions") appeared only in FY2023; FY2025 replaces it with softer but broader language about "foreign-affiliated technology companies." The FY2021 AWS outage ("over 5 hours") grows to ("over 12 hours") in FY2025, reflecting an actual October 2025 incident. COVID and Brexit disappear entirely.

4. How They Handled Bad News

For a 12-letter dataset, there is almost no "bad news" to handle. Every FY2023 and FY2024 letter is a beat-and-raise. The notable stress tests are three.

The FY2023 Q4 DAU guide-down (Feb 2024). Management preemptively guided Q1 2024 DAU growth down to "mid-50s" versus the 60%+ trend. Rather than defending the prior number, CEO Luis von Ahn framed it as arithmetic: "we can't expect our user growth to accelerate forever." Actual Q1 2024 DAU growth came in at 54%, matching the guide. Clean handle, no drama.

The April 2025 "AI-first" email. On April 28, 2025, Luis von Ahn sent an all-hands email — then posted it to LinkedIn — announcing Duolingo would become an "AI-first" company, including phasing out contractors and permitting teams to hire new humans only if the work "could not be automated through AI." The backlash was immediate and external: "AI first means people last"; "I can't support a company that replaces humans with AI." A week later, von Ahn walked it back: AI is "a tool to accelerate what we do, at the same or better level of quality" and "I do not see AI replacing what our employees do (we are, in fact, continuing to hire at the same speed as before)." In the Financial Times he later conceded: "I did not expect the amount of blowback."

The Q1 and Q2 FY2025 shareholder letters — the two that followed this event — do not mention the memo, the backlash, the contractor wind-down, or the walk-back. Management chose silence in writing. The closest acknowledgement is Q3 FY2025: *"we posted less 'unhinged' content on our English-speaking social media accounts as we listened to community feedback."* That is the only trace.

The Q4 FY2025 strategic step-down (Feb 2026). This is the biggest tonal shift. Management guided FY2026 bookings growth to 10–12% after FY2025 delivered 33%, and cut the Adjusted EBITDA margin target by roughly 450bps. Von Ahn's framing: "we expect bookings growth of around 11% in 2026, compared to the nearly 20% we believe we could achieve if we operated like we have in past years"; "I don't take this decision lightly, and I know it may come as a surprise to some investors." That parenthetical — the first time in 12 letters he explicitly preempts investor pushback — is the key tell. The rationale: invest in the free experience (Video Call moves from Max-only to Super; $50M of bookings friction removed) to reach 100M DAUs by 2028.

5. Guidance Track Record

No Results

Three years, three big beats. The FY2023 Adjusted EBITDA margin beat (~17.6% vs. 4.5% initial midpoint) is almost hard to believe — but it reflects how drastically operating leverage kicked in once paid-sub penetration rose, and it established the credibility that carried management through the next two years. The FY2024 bookings beat was ~$70M above the initial guide midpoint; FY2025 was ~$70M again.

Management also raised full-year guidance within year in every quarter across FY2023–FY2025. The cadence was: set conservative at Q4, raise at Q1, raise at Q2, raise at Q3, beat again at Q4. Every cycle.

Delivery credibility (1-10)

8.50

Credibility score: 8.5/10. Three years of consecutive full-year beats on both the top line and margin, with transparent in-year raises, is unusually clean. The deduction is entirely on communication: the FY2026 downshift was proactive, but shareholders learned about the "AI-first" memo from LinkedIn rather than the IR letters, and the contractor wind-down — a reputationally significant workforce decision — was never acknowledged to investors in writing. Delivery is top-decile; narrative discipline has cracks.

6. What the Story Is Now

The old story was simple: a freemium language-learning app with a beloved mascot, a gamification engine, organic word-of-mouth marketing, and enough engineering culture to be a profitable business at scale. Every piece of that is still true and has been de-risked. Revenue over $1B, nearly 50M DAUs, 12M+ paid subscribers, 29.5% adjusted EBITDA margin, ~$305M of adjusted EBITDA — these numbers were a fantasy in 2021.

The new story is bigger and stretchier. Management is saying Duolingo is no longer just a language app but a multi-subject learning platform (languages, Math, Music, Chess) that plans to use frontier generative AI to become meaningfully more effective than a human tutor, and is willing to take a 450bp margin hit and 20 points of growth off the FY2026 guide to fund that transition. The target is 100M DAUs by 2028 — roughly 2x today. Video Call, the AI conversation tutor, goes from a premium-tier feature to free-tier bait. Content generation ("over 150 new courses via AI in Q1 2025") is supposed to be the wedge that turns language dominance into subject-wide dominance.

What to believe. Execution is real. The beats are real. AI variable costs did come in lower than feared (GM pressure of 100bps vs. the 300bps management warned about at the start of FY2025). The pivot to teaching efficacy in A/B testing is directionally right if you believe personal-tutor-quality AI is coming — and it gives Duolingo a bigger TAM if it works.

What to discount. The 100M DAU target is a management bet, not a plan with intermediate milestones. Apple concentration at 62% of revenue is a quiet structural tax that never gets addressed in prose. The "AI-first" memo incident shows Luis von Ahn can misjudge public communication — and the decision to never address it in an IR letter means the next communication misstep is likely to compound. Subject expansion beyond languages is still unmonetized ("we do not expect these new subjects to meaningfully contribute to monetization in the near-term" — Q3 2023; still true in Q4 2025). Founder dual-class voting control has not softened.

**Bull case:** a founder-led, profitable, high-retention consumer AI company that shipped before it needed to defend. If the Video-Call-as-free-tutor bet works, DAUs re-accelerate in 2027 and the margin comes back. Management is buying the option with a known price tag ($50M of bookings, ~450bps of margin).

**Bear case:** growth was always going to decelerate; the pivot narrative dresses up that deceleration as strategy, while Apple-fee exposure, dual-class governance, and the pattern of silencing uncomfortable topics in IR letters stay constant. If FY2026 misses even the 10-12% bookings guide, the story flips from "investing through the cycle" to "the bull case was wrong."

The honest read: this is a management team that has earned the benefit of the doubt on delivery but not on communication. Hold them to the 100M DAU-by-2028 number, watch the Q1 and Q2 FY2026 DAU prints for signs that free-tier Video Call actually moves top-of-funnel, and assume the next surprise will come from outside the shareholder letter — because that is where the last one came from.