Duolingo: Netflix/Spotify 10 Years Ago.
Q4 2024 ER Digest
Many of you have been asking me to update my Robinhood deep dive. While I circle back to it, I suggest you read this Robinhood deep dive written by Satishan Muthusamy - an alumni of my Tech Stock Goldmine course. His work is of great quality and I believe you will enjoy it.
Edited by Brian Birnbaum and an update of my original Duolingo deep dive.
There's a breed of companies that are going to see their earning power increase exponentially at a marginal cost in tandem with the progress of AI models.
Duolingo is one of them, together with Palantir, Okta and Mongo DB.
Duolingo’s math and music verticals are gaining traction after their launch in November 2023, paving the way for a substantial increase of Duolingo’s operating leverage. As has been the case with Spotify, deploying new verticals beyond language learning allows Duolingo to non-linearly increase ARPU (average revenue per user) at marginal cost. Math and music now have 3M DAUs, representing 7.5% of total DAUs in Q4 2024.
As you can see in the graph below,DAU growth has been spectacular over the past four years.
Further, Generative AI is accelerating the speed of Duolingo’s value creation. Duolingo’s business comes down to two things: creating content that optimizes learning outcomes for students; and monetizing the engagement. Behind each new vertical, it’s generative AI that now enables Duolingo to improve the learning experience at marginal cost in an increasingly automated fashion. Together with A/B testing, GenAI allows Duolingo to “10X” the speed of content creation and new features.
See what Duolingo CEO Luis von Ahn said about this during the Q4 2024 earnings call:
And it [Generative AI] also allows us to go way faster, particularly in content creation. So the amount of content we’re able to generate compared to 2 years ago is way more.
I mean, it’s like 10x or more than that. So that’s one thing.
The other big part is in just feature experimentation or kind of live features like, for example, the Video Call feature. That adds expenses because we have to query the large language model kind of in real time. The way we’re seeing it is, for now, we’re not even trying to optimize those costs. We’re just trying to go as fast as possible in generating the best possible features.
So the way to think about it is there’s going to be kind of an upfront cost for AI, which will happen probably throughout this whole year, but it’s because we’re in a unique opportunity, a unique time in history where we really want to develop the best possible AI features over time.
AI will be a tailwind for Duolingo so long as it continues to strengthen its proprietary data moat over time. As of Q4 2024, Duolingo’s 40M DAUs and over 80M MAUs were generating data that explains what educational content works for them and at what stage of their learning journey. As the user base continues to grow and Duolingo deploys more verticals, the company creates an increasingly unique and inimitable dataset that will likely lead to an increasingly unique and inimitable AI teacher.
Until now, the consensus has been that LLMs (large language models) aren’t good at maths. However, during the latest earnings call Duolingo CEO Luis von Ahn explained that these models are now good at maths, because they can reason. According to Luis, this will be transformative for teaching maths in general because the subject is so multi-dimensional. Having a model that can reason in this manner is not only a step forward for Duolingo, but for the world in general.
See Luis’s remarks during the earnings call:
Now in terms of uses for AI, for Math in particular, AI is really going to be amazing.
[…]
Somebody here said, math is like a combination of 1,000 subjects. It’s just a bunch of different things. And historically, the first version of the large language models, were not very good at math.
But of late, we really – the large language models have added reasoning. So they’re actually pretty good at math now, and I think that’s going to really accelerate how much content we put in there. So you’ll see over the next year or so that the pace at which we are adding content to the Math course is going to improve quite a bit. And that, of course, will help us have a lot more users.
Duolingo is a prime example of a company whose odds of increasing power are rising in correlation with the non-linear improvements in AI models. As was the case with Palantir, Duolingo’s business has been strengthened by the rise of LLMs. As AI evolves from LLMs to Agentic and Physical, the power of Duolingo’s features stands to increase exponentially at marginal additional investment.
To leverage new models, Duolingo must simply continue enhancing its proprietary data moat. So long as Duolingo feeds its models unique data, features will be increasingly competitive in the marketplace. In the short term, the increasing use of AI in Duolingo’s app will drive higher inference costs. But as the technology improves and inference costs decline, unit economics will radically improve.
As an example of this hypothetical progression, Duolingo users can now practice their language of choice by video chatting with an AI. Within Duolingo’s top tier subscription plan, called Max, and which tier boasts the highest LTV,Video Chat is the main driver of the growth. As more powerful models come to the market, this feature will eventually become indistinguishable from talking to a friend. Luis mentioned this during the call:
Yes. I mean in general, what we meant by that is, at the moment Lily is not necessarily your best friend yet.
We want that to happen.
With the onset of Physical AI and humanoid robots, Video Chat will likely evolve into a “physical” personal teacher. Generally, Duolingo’s intelligence and ability to drive student outcomes per dollar spent are likely to increase beyond imagination from here. Combined with the deployment of subsequent educational verticals, I believe Duolingo is likely to grow into one of the world’s primary educational institutions.
CFO Matthew Skaruppa shared during the Q4 2024 earnings call that they expect gross margins to decline 300bp in H1 2025 as Duolingo leans even more into AI and thus incurs higher inference costs. However, he also said that they expect margins to improve in H2 2025:
In 2025, we expect a temporary 170 basis point year-over-year impact on gross margin, primarily due to Max. In the first half of the year, there will be roughly a 300 basis point year-over-year impact as we prioritize rapid product innovation to drive Max adoption.
We expect margins to improve in the second half of the year as we work to improve AI costs.
Thus, my view is that so long as Duolingo’s process power remains world-class, its earning power is going to outpace costs tremendously over the coming decades. The depiction below is merely illustrative:
As further illustration of the positive impact of AI advancements on Duolingo’s business, consider their new course, which allows Korean speakers to learn Spanish. Mere months ago this was not possible. Students were forced to take a circumvoluted route from Korean to English and then English to Spanish. The required labor was excessive, the course not cost effective. It serves as an illustration of how AI enables Duolingo to deliver value where previously it didn’t make economic sense.
These dynamics will drive faster user growth across the three major verticals. And as the models get smarter, there’ll be an infinite number of levers Duolingo can pull to accelerate growth that previously didn’t even make economic sense.
Eventually, Duolingo’s plan is to allow AI features to seep down into lower tiers as inference costs drop, thus sharing those cost savings back with customers–a form of economies of scale shared–which will be interesting to track going forward. This approach will likely increase customer goodwill and accelerate KPIs across the three subscription tiers.
However, I should be clear that we’ve yet to see this process begin in earnest. See Luis’s remarks about this during the call:
It may be the case that, at some point, it gets cheap enough for us to put in a lower tier. We may even be able to give some of it for free. And in general, we will be testing whatever is best for the long-term, not just LTV, but the long-term health of our app.
And we know that, in many cases, giving things for free is actually the best thing for our app. So, we are just going to be testing a lot.
At the moment, our hands are a bit tied because the costs are just high – too high, but we will definitely be testing where to have these features.
Lastly, Duolingo remains a pristine example of the sort of company I teach folks to spot in my Tech Stock Goldmine course. As previously mentioned, the odds of exorbitant success for this company are meaningful so long as the process power remains world-class. And process power is the result of many different attributes, which you can learn to spot in my course.
Spotting process power early has been the key enabler of my 6X+ Spotify, Palantir 11X+ and 2.6X+ Hims investments. It’s taken me a decade to unlock the ability to spot this highly intangible quality in companies, but in my course–and in under two hours–I teach you everything I know in an elegant and powerful mental model that over 400 students have happily integrated to date.
Until next time!
⚡ If you enjoyed the post, please feel free to share with friends, drop a like and leave me a comment.
You can also reach me at:
Twitter: @alc2022
LinkedIn: antoniolinaresc
Thanks, Duolingo has great potential. I believe next 3-5 quarters will show how their investments to AI pay off, what kind of progress they show in math learning and how they can improve their offering to advanced learners. I have started a small position and believe there will be good chances to add after Q1 and Q2 if the story develops.
Excellent work, Duolingo has the potential to become mandatory learning material in all schools across the globe. We really are in the early innings of a gigantic company