Duolingo
DUOL
Open
Growth
- +5.2%unrealized, USD
- 139.72average cost, as supplied
- 147.03last · 2026-08-20
The record
Thesis
Written 2026-08-21. Figures are Duolingo's Q2 2026 results, reported 6 August 2026, and market data from that week. I bought this position in November 2025 and laddered into it through February 2026 — this is the argument as I hold it today, not a reconstruction of what I thought then, which I did not write down.
I hold this because I do not see the business slowing down, and the market has priced it as though it already has. The re-rating I expect is a consequence of that, not the reason for it.
Daily active users since 2021
millions of daily active users, at each year end
Scroll the chart sideways →
Show the numbers
| Period | Value | Note |
|---|---|---|
| Q4 2021 | 10.1% | |
| Q4 2022 | 16.3% | |
| Q4 2023 | 26.9% | derived |
| Q4 2024 | 40.5% | |
| Q4 2025 | 52.7% | |
| Q2 2026 | 58.7% | latest |
| 2028 | 100% | target |
Users are growing, and growing faster than last quarter. Daily active users rose 23% to 58.7m in Q2 and accelerated from Q1, against monthly actives above 160m. Whatever the market is worried about is not showing up in whether people open the app.
Where the growth is, and where it is not
growth, per cent — period stated on every row, because they are not the same
Scroll the chart sideways →
Show the numbers
| Measure | Value | Note |
|---|---|---|
| Daily active users | 23% | Q2 2026 vs Q2 2025 |
| Revenue | 18% | Q2 2026 vs Q2 2025 |
| Paid subscribers | 17% | Q2 2026 vs Q2 2025 |
| Bookings | 11% | full-year 2026 guide |
That gap is a dial, and management holds it. This is the part that convinces me they know what they are doing. Luis von Ahn has described the trade directly: cutting daily energy from 25 units to 24 raises revenue, because more people run out and subscribe, and lowers daily active users, because it frustrates them. One unit. They know the elasticity of their own product to that resolution, and they have been tuning it for years.
Right now they are turning that dial toward users on purpose. They have pushed advanced AI features down into the lower and free tiers, made the app less aggressive about prompting conversion, and set a target of 100 million daily active users by 2028 against roughly 59 million today. There is also a mechanical effect: a three-month free trial pushes booking recognition out by a full quarter. So a good part of the bookings gap in the chart above is a choice and a timing artefact rather than demand disappearing.
I want to be careful with that argument, because "we are slowing deliberately" is also what a company would say if it were slowing involuntarily. What makes me believe this version is that the DAU line accelerated in the same quarter — you do not usually get that by accident while the money line slows.
The international story is the part I think is most underrated. China is now the second-largest market by daily active users and the fastest-growing region, monetising at levels comparable to France and contributing roughly 5–6% of the business — helped by a brand partnership with Luckin Coffee that reads as a local company's move rather than a foreign app's. Asia is the fastest-growing region overall. A business adding users fastest in the largest markets on earth is not a business at the end of its growth.
The brand is a moat, and it shows up in the accounts rather than only in the marketing. Sales and marketing ran at about 12.1% of revenue in 2024 — low for a consumer subscription business, because the product and the owl do the acquisition that competitors have to buy. A rival launching a language app has to pay for every user Duolingo gets talked into.
The habit system is the other half of that moat, and it is measurable. The streak is not a gimmick; it is the retention mechanism. A single streak-revival feature re-engaged 15 million users in the quarter. That is the asset an AI chat interface does not have and cannot quickly build.
The expansion beyond languages is real where it is working. Chess launched in 2025 and reached roughly 7m daily active users inside a year. Music absorbed a $34.5m acqui-hire of NextBeat for game design and retention expertise. The stated ambition is to be the default place people learn, not the default place people learn languages.
On AI I take the other side of the argument that has cut this stock. The fear is concrete — a GPT-5 demo built a language-learning app in minutes, and the shares are down about 66% over the year. My view is that a model is not a curriculum. What Duolingo owns is a structured path, a difficulty curve and a habit loop, backed by more data on how people actually learn than anyone else has. The evidence so far reads my way: the Max tier is about 8–9% of paying subscribers and already produces 12–16% of subscription revenue, and Q2 raised the full-year adjusted EBITDA margin outlook to 26.5% as AI serving costs fell. Users are paying more because of the AI, not leaving because of it.
What the AI tier already carries
share of the total, per cent — a share, not a growth rate
Scroll the chart sideways →
Show the numbers
| Measure | Value | Note |
|---|---|---|
| Share of paying subscribers | 8.5% | about 8-9% |
| Share of subscription revenue | 14% | about 12-16% |
Worth noting that Max is still unavailable in China pending approval of large language models there. The fastest-growing market is growing without the AI tier at all.
The financial position means I do not have to be right quickly. About $1.25bn of cash against roughly $55–92m of debt, so net cash near $1.16bn. Free cash flow of $226.4m over six months, up 20%. Against a $6.3bn market capitalisation, the operating business is priced at an enterprise value near $4.9bn.
And then, last, the price. 12.5x trailing earnings and 13.4x free cash flow are the ratios of a mature or distressed business. If the growth above continues to show up, I expect the multiple to expand as that becomes undeniable. But I am buying the growth. The multiple is what makes being early cheap rather than what makes the case.
Where this thesis is weakest
Stated with the same date, because a thesis that lists only its strengths is worth nothing.
Bookings are still the number that hurts. Guidance is 10–12% bookings growth for 2026 against more than 30% sustained through 2025. Bookings lead revenue, so the 18% revenue growth converges downward toward that unless they reaccelerate. I have argued above that much of this is a deliberate turn toward users, and I believe that — but the deceleration is real either way, and if I am wrong about the intent then I am wrong about the whole position. Calling it "still high growth" would be false.
Net income fell 26% year on year to $33.2m on higher operating expenses and tax. Free cash flow is strong and I weight it more heavily, but the earnings line is going the wrong way.
Gross margin is guided down to about 69% as AI serving costs land, even as the EBITDA margin outlook rose. The AI tailwind has a cost side.
The brand moat is not free. Sales and marketing rose about 39% in 2025, and the idea that Duolingo does no paid acquisition is a myth — it runs consistent paid social. The moat is that it spends less than peers, not that it spends nothing.
The expansion is not uniformly working. Chess is working. Management has said Math and Music will not improve in 2026, so one of three legs is carrying the diversification story.
What would make me wrong
- A general model that ships structure, not just conversation. My case rests on a curriculum and a habit loop being the hard part. If a general-purpose model ships a graded path with retention mechanics and holds users on it, the moat I am describing is not there.
- Bookings that do not reaccelerate. If 2027 guidance opens near 10% again, this is a mature business and 12.5x trailing earnings is roughly the right price rather than a mistake.
- Daily active users rolling over, especially in Asia. Engagement is my strongest evidence; if it decelerates while bookings stay slow, both halves of the argument fail at once.
- China stalling on regulation. The fastest-growing market is growing without Max. If approval never comes, or local competition is favoured, I lose the growth I am relying on most.
- Margin compression that does not stop at 69%, which would mean the AI features cost more to serve than users will pay for.
Latest movements
- 2025-11-14 — Opened at $189.82. The first trade in the ledger.
- 2026-01-21 — Added in two large lots at $148.00 and $147.91. Position size +550%.
- 2026-02-03 — Added at $120.50. Position size +15%.
- 2026-02-04 — Added across four tranches, $118.00 to $119.50. Position size +17%.
- 2026-02-05 — Added across five tranches, $116.50 to $118.50. Position size +14%.
- 2026-02-27 — Added at $95.00, the final and deepest average-down. Position size +25%.
- 2026-04-30 — Trimmed 20% at $107.00. The first sale ever made. Realized −19.3%.
- 2026-07-30 — Trimmed 50% at $135.90, the largest sale in the ledger. Realized +2.5%.
- 2026-08-13 — Trimmed 25% at $140.00. Realized +5.6%. Recorded rationale: "Reduction for diversification."
- 2026-08-18 — Trimmed 33% at $140.00. Realized +5.6%.
- 2026-08-19 — Trimmed 50% at $150.00. Realized +13.2%.
- 2026-08-19 — Added back at $146.88 the same day. Position size +100%.