Earnings season has been wonderfully hectic as always. In case you missed it, I’ve written 30+ detailed reviews so far. These reviews of a single quarter deliver the level of detail that others would call a “deep dive” and do so in a way that’s easy to digest:

I also liquidated a position today and added to a couple others. My updated holdings and performance can be seen here.

Snowflake’s overarching platform is called the Data Cloud – a “single foundation to eliminate data silos.” This infrastructure unlocks affordable data storage, organization, querying and learning at gigantic scale. It offers these services with elastic compute capabilities to allow for flexible scaling up and down of usage. The architecture naturally separates the functions of data storage and consumption, unlike legacy data warehouse solutions. That means data consumption capacity is untethered from public computing resources, which helps with controlling costs and waste, handling diverse workloads and resolving potential scale bottlenecks. Under this framework, customers can store as much data as they want without the requirement for immediate processing.

In Snowflake’s case, the storage is done in a centralized data repository in its Data Cloud and processed only as needed. Data is utilized virtually, which removes the need for dedicated hardware. Snowflake does all of this for clients in a managed fashion to minimize client effort and infrastructure needs. There are a few key products to know & track:

The Snowflake Data Warehouse is where structured data is stored and (on command) processed. Structured data is formatted data. It’s utilized for record keeping and report creation. Data can be easily fetched via structured query language (SQL).

Snowflake Data Lake does what the warehouse does for unstructured data. Unstructured data is unformatted and used to uncover new insights and patterns.

“Our North Star is to deliver the world's best end-to-end data platform powered by AI.”

CEO Sridhar Ramaswamy

“Snowpark” is its developer platform and data-equipped playground to build new things. It enables working with data in any source code language. With it, they can process and visualize data (through Snowpark functions) and build apps (through Snowpark Native Apps). GenAI models ingest absurd amounts of data. Snowpark Container Services allow GenAI models to run closer to the data that they require. This enhances performance, expedites model training and diminishes costs. 

Cortex AI is another important new product. It’s what Snowflake calls its “AI layer” and offers a slew of GenAI-powered tools to (as Snowflake always says) bring AI, application-building and analytics right “to a customer’s data.” That conjoining routinely lowers data transfer and storage costs. Cortex AI offers unstructured text summary, sentiment analysis, helps beginners write SQL etc.

Snowflake Data Sharing is its secure product for, as the name indicates, sharing data among the rest of Snowflake’s participating users. As more opt in, a compelling network effect of relevant data builds and Snowflake’s value proposition deepens.

Unistore is Snow’s hybrid table product, which can ingest, store and organize both transactional and analytical workloads. It has long been an analytical workload specialist within the data warehouse part of the business, and this unlocks transactional workload demand.

Snowflake’s revenue model is consumption-based in nature. This means visibility compared to SaaS business models is not as strong. It also means customers can more easily scale down (or up) usage when times are bad (or good).

“I feel guidance is appropriate… Given the scale we're at right now, we think a 3% beat should be considered big. And I feel good about the guidance that we've set for the quarter and the year.”

CFO Mike Scarpelli

It thinks product gross margin will improve in the coming years as it enjoys “easier access to GPUs and more AI revenue from the costs it’s incurring.

Progress:

One year into CEO Sridhar Ramaswamy’s tenure and Snowflake is in a much better spot. The company quickly turbo-charged innovation cadence to deliver 100% growth in product delivery for calendar 2024. It consolidated go-to-market teams, focused on early career hires, cut redundant middle management layers, added more performance reviews and sought to amplify “operational rigor” in every facet of its business. This is already “yielding meaningful margin gains,” which frees Snowflake to more aggressively reinvest in more product improvement. 

One of the ways Snowflake re-positioned its go-to-market for better success was by focusing on selling partners to enable them to create joint solutions for the large batches of customers these partners represent. For example, Fiserv is using Snowflake to power the data analytics platform for their 2,500+ bank and credit union customers. If Fiserv is using Snowflake, there’s a far better chance of their clients doing so too. Blue Yonder, a supply chain company, is using Snowflake to facilitate its AI insights engine. With Snowflake, the company can more effectively mine value from the vast sum of data that its 3,000+ clients represent. This is a great way to make its selling motions more efficient and deliver customer wins in chunks, rather than one-by-one. Along similar lines, as talked about last quarter, it made it a point of emphasis to work more closely with AWS, Google Cloud and the massive marketplaces those public cloud vendors boast.

AI:

Aside from collaboration products like Notebooks, AI is where the risk of Snowflake falling behind had been seen as most pressing. Through Cortex AI and its data lake innovation, it’s quickly addressing previous holes in its offering. 

This quarter, the company debuted Cortex Agents to position it for the emerging Agentic AI wave. The toolkit provides processes to create powerful agents and “orchestrate seamless planning and execution of tasks” — across all company data. This is where access to both Anthropic and OpenAI, as well as leading open source models really shines; it gives developers more choice to build with the models they like the most. Snowflake is the only data platform to offer both Anthropic and OpenAI models. The product seamlessly integrates with Cortex Analysts and Search for world-class querying scale, precision and accuracy. It also includes Snow’s intricate analytics capabilities to rapidly uncover patterns unseen by the human eye.

ServiceNow, Amazon, Microsoft, Meta and many other enterprise software firms already offer similar products, and Snowflake needed to have their own option. They see the overarching company theme of “bringing data to your work” as a key differentiator here, as agents are only as good as the data they’re trained on and Snowflake offers an overarching destination to get whatever data a company needs. They’re not alone in this capability, but they’re not behind in it either. As part of the expanding Microsoft and OpenAI model partnership, Cortex Agents is available for Microsoft 365 Copilot and Teams customers.

This product is complementary to the Data Agent product Snowflake debuted in November called “Snowflake Intelligence.” Snowflake Intelligence is sort of like Cortex Search paired with Cortex Agents, but for data specifically. This is a dedicated platform within Snow’s ecosystem that unlocks conversational querying from a company’s databases. It also offers seamless ways to build data agents to automate task completion in areas like data analytics and streaming.

Today, Snowflake helps clients create chatbots as the “essential building blocks for a strong AI foundation.” It sees Cortex Agents as the next phase of this push, as it moves to goal-oriented, agentic service. As of now, 4,000 customers are using its AI/ML products on a weekly basis, with AstraZeneca leaning on it to realize its goal of 20 new medications by the year 2030. Furthermore, Snowflake’s own engineers are readily using Cortex in their internal work. This is automating code creation and giving salespeople better customer context to make their interactions more effective. That ties closely with its Customer 360 product, which uses Snowflake’s cohesive access to data to provide more complete customer data profiles.

“We are seeing broad adoption of AI… We expect that to turn into real revenue over the coming quarters.”

CEO Sridhar Ramaswamy

Flexing its Strengths:

Snowflake shines in its ability to seamlessly, quickly onboard customers and get them up and running, with little headache and very low upfront costs. From there, clients generally grow consumption as they get used to the platform and all of the utility it provides. The construction of its data cloud also eliminates the need to maintain “expensive engineering resources,” which allows Snowflake to routinely deliver 50%+ cost savings for migrating customers.

To augment its ease-of-use and frictionless strengths, the company debuted “SnowConvert.” This product is meant to automate code conversion and modernization, to enable companies to embrace cloud and position them to capitalize on GenAI’s potential. MongoDB, Amazon, Microsoft and a few others are working on similar offerings all with the goal of accelerating cloud migrations. More than 80% of global infrastructure is still on-premise and the potential here is gigantic. Anything companies can do to make the leap easier, less disruptive and more seamless is a positive.

Data Sharing:

As Snowflake’s customer count builds and more customers opt into open data sharing with the ecosystem, a compelling network effect builds. SNOW becomes an increasingly powerful partner for plugging into a giant network of enterprise and consumer-facing data, rather than solely unleashing your own information. This quarter, Stripe and Braze were among two named customers that opted into data sharing. Just like Fiserv and Blue Yonder working more closely with Snowflake can lead to winning more of their customers, this can accomplish the same thing. Want to be in the place where Stripe and all of its participating customers share their context? Then you’ll go with Snowflake. That’s the idea. Scale creates more stickiness… to create more scale and stickiness.

More on Data Engineering – Iceberg Tables & Snowpark:

Snowpark reached 3% of total 2025 revenue as expected. It’s not just a healthier core business leading to better results, but quick traction within key product priorities such as this one. And speaking of quick product traction, its Iceberg offering continues to perform very well. As a reminder, Iceberg is an open-source product that provides data storage, with Snowflake offering support for this external product safely within its platform. Iceberg proliferation was seen as a threat to Snowflake’s storage revenue (11% of total). At the same time, creating more data interoperability should boost overall data processing demand and support the other 89% of its business. Many thought this would be a net negative, but it has been convincingly positive for Snowflake overall. First, it is supporting demand for that 89% as expected as they “see workloads that otherwise would have been inaccessible to Snowflake.” Secondly, and even more encouragingly, they’ve “yet to see massive amounts of data move out of Snowflake.” This has not been a material storage revenue headwind. The encouraging developments prompted leadership to include consumption revenue from this product in its guidance for the first time.

“The Iceberg story is one that Snowflake can write. We've made all the right investments in the format, how broadly it is getting adopted. And yes, while there will be some customers who might want to move data that is in Snowflake over to Iceberg, there is vastly more opportunity for us to take Snowflake to where the data is.”

CEO Sridhar Ramaswamy

DeepSeek Quote:

“I think we are definitely headed to a world in which there are now several players, it's not 2 any longer, that are leading the charge when it comes to the world's innovation with AI. And the fact that it is a healthy combination of open source models as well as proprietary models, we think is a good thing.”

CEO Sridhar Ramaswamy

Just like for all other enterprise software, model competition and cost deflation is uniformly good for Snowflake.

Strong showing for a company that had been left for dead in the GenAI race. Ramaswamy has quickly turned this thing around and morphed Snowflake back into the innovation machine it had once been. Its AI product suite is quickly catching up to peers while its rapid work in Iceberg tables and Unistore has already led to more than $200 million in annualized revenue. Margins are recovering and momentum is clearly accelerating. I would love to see stock comp falling more materially and also would love to see the company showing GAAP operating leverage. But all in all, it was a good quarter.

Duolingo is a leader in language learning. It’s now expanding into music and math.

My Duolingo Deep Dive can be found here.

Duolingo will continue to invest more in GenAI and its video chat product in 2025. This will have a 300 bps impact on GPM during the first half of the year and 170 bps for the full year. As I’ll explain below, the current focus is on product velocity over margin optimization in the near term, and this is temporary. It sees GenAI as a massive opportunity and wants to spend the funds needed to capture it. Duolingo will also shift more hiring to the beginning of the year, which will mean more payroll growth and another small headwind. Still, EBITDA margin is expected to expand by 2 points Y/Y.

Based on Duolingo’s consistent track record of conservative guidance, I think profit estimates will probably be stable following this report. The company trades for 62x 2025 EPS estimates and 48x 2025 FCF estimates. EPS is expected to grow by 31% in each of the next two years. FCF is expected to grow by 25% this year and by 39% next year. 62x is the lowest earnings multiple Duolingo has traded at since turning profitable.

Product Iteration Machine & AI:

Outperformance was credited to more of the same. Duolingo continues to rapidly iterate every single piece of its product to ensure it’s always getting better. Constantly split-testing every single variable is in its DNA. It leads with product, rather than advertising and relies on word-of-mouth growth to power the vast majority of its success. It then supplements that with efficient marketing (from social media to the Super Bowl) to create viral moments and demand accelerants. The company is delivering an unmatched scale, growth and margin combination because of this mindset. Not one-off events… no fad products… just rapid, structural, margin-accretive growth. 

The culmination of these experiments leads to truly impressive trends. DAUs as a percentage of monthly active users (MAUs) rose by more than 4 points Y/Y to 34.7%. Furthermore, 25% of its DAUs have a usage streak of over 365 days. This was 20% just two quarters ago. Sticky-sticky.

How far can this hyper-efficient playbook take them? The ceiling is fortunately unknown. User growth is not correlated with market maturity. Its most mature market (Latin America) just delivered an 80% Y/Y DAU explosion. This just goes to show how long the runway is for language learning if it stays on its current path. Just 2% of global language learners are currently Duolingo DAUs. And this potential doesn’t even include encouraging expansions into math and music (more later).

“People often ask me, “What’s the secret behind Duolingo’s success?” The truth is, there’s no secret —just our mission and a process.”

Co-Founder/CEO Luis von Ahn

More on AI & New Subjects:

Now, AI is allowing Duolingo to run product experiences ever faster. Luis sees their velocity here as in a great spot. As I’ve argued for years now, AI is not an existential risk to this business… It's a massive opportunity to create better products and more financial success.

AI also enables Duolingo to greatly boost the amount of content that it’s offering across its apps. Over the last few quarters, leadership spoke a lot about how this is improving their advanced English content and helping them attract more experienced learners to that product. This quarter, it focused on how promising this technology is for its budding math product. Today, Duolingo doesn’t have nearly enough content for this subject. It can only teach 3rd, 4th and 5th grade math and knows it needs to offer K-12 for this to realize its potential. The issue is that creating this type of content is especially hard.

As Luis explained, building a beautiful lesson for the coordinate system doesn’t entail transferable material for other topics. They more frequently have to start from scratch, instead of letting the concepts more neatly build on one another like with music and languages. Enter GenAI. While early foundational models (FMs) were bad at math, new reasoning models are excellent at it, and create so much opportunity for Duolingo to exponentially accelerate its content creation pipeline for this subject. That’s the plan for 2025 and beyond as it “lays the foundation” to turn this and music into two more promising growth stories. And that might not take long.

Notably, Duolingo Music and Math combined have 3 million DAUs — 14% of language learning. Both are growing more quickly than language learning and that is expected to continue. Rounding out the currently bare-bones offerings for those nascent products should directly support steady demand growth. Down the road, it’s easy to see how more subscription packages can arise from this, just like Spotify has done with audiobooks and podcasts.

Within its original language learning niche, GenAI allows DUOL to reach fringe use cases that hadn’t been large enough to pursue. For example, now it has a Spanish class for Korean learners, instead of requiring them to learn it through English. There are several examples like this that were all too small to go after and are now becoming attractive.

Finally, Duolingo will keep investing in GenAI for internal process optimization and to enable more future scaling and margin expansion. Whether it’s automating more pieces of content creation, split-testing, back-office operations or anything else, it sees this technology as raising its margin ceiling over the long term.

Duolingo Max:

Duolingo Max, its newer, more premium subscription, outperformed internal expectations and was the main source of the beats. This is already up to 5% of total paid subscribers, with a healthy mix of brand new users and Super Duolingo (other tier) upgrades. It’s now available to most of its customers and traction everywhere is excitingly strong. It isn’t yet offered in China, as OpenAI models are not allowed there, but it will soon introduce the product there through other partnerships.

A lot of this great success is based on the fantastic launch of its GenAI-powered video chat product. This, along with accelerating overall content creation, is where a lot of the focus will be during the new year. The tool is fostering far more effective learning for speaking use cases, as it allows users to practice talking without feeling judged or embarrassed.

Engagement for this feature continues to briskly rise. Meshing world-class OpenAI models (Duolingo is a preferred partner), the largest language learning dataset on the planet and an innovation-driven company is clearly working.

Notably, the video launch is Duolingo’s first product that has generated 2x the demand for non-English learners vs. English learners. This is raising propensity to buy in markets across the globe beyond rates enjoyed with Super Duolingo.

In 2025, it wants Lily (the video chat character) to get more personal. It wants her to get better at knowing what to say and how to react, so that you want to use the product. Duolingo is all about making learning as fun and engaging as possible. If you’re having a good time, you will advance more quickly, stick around longer and deliver more company revenue.

As briefly mentioned, Duolingo Max is a margin headwind because of the cost associated with using Large Language Models (LLMs) in this video product. That’s a concession leadership is eager to make, as overall profit dollars are higher than Super Duolingo and lifetime value (LTV) is too. It remains fixated on shipping products and upgrading the experience as quickly as humanly possible to fortify its product lead amid the GenAI revolution. This means forgoing efficiency opportunities here and accepting the temporary margin pressure. While the margin headwind will persist through the beginning of the year, GPM is expected to rebound to levels similar to 2024 by year’s end. It has a concrete idea of everywhere it can enhance margins within this business. It’s just not focused on that yet. Based on the team’s track record of generating highly profitable growth and still expecting 2 points of 2025 EBITDA margin expansion, I say go for it. I get excited when a team like this sees more productive places to pursue more growth. I get even more excited when that happens while margins move higher.

“We know that these costs can be optimized in a lot of ways. And even if we do nothing, the cost will be optimized because the cost to query LLMs is coming down every month… We're prioritizing using the latest AI models to deliver a high quality experience.”

Co-Founder/CEO Luis von Ahn

Interestingly, as costs to serve the video tool inevitably fall, it plans to optimize Max pricing in countries like India. The cost of querying right now means their subscription is too cost prohibitive there. That will change as innovation from firms like Meta, Nvidia and others continues to drive deflation. Some of that cost savings will be used to lower prices in several countries where it knows that will improve the overall business (like India).

More Notes:

The family plan is now 23% of total subscriptions. This comes with higher retention and lifetime value.

It sees an opportunity to raise retention levels for resurrected active users. Churn is higher for that specific cohort; that’s because Duolingo needs to improve on understanding that the learner forgot a lot of what they were taught. Despite delivering masterful results, leadership is always talking about things they can do better. They’re always trying to improve.

Advertising revenue is just not a priority for the company. It is squarely focused on subscription growth and is increasingly using ad real estate to promote its own subscription offerings instead of other companies. That’s an ad revenue headwind and a good business decision.

I thought this was a very good quarter. While I would have obviously preferred outperforming EBITDA guidance, management has a perfect track record of sandbagging guidance and I expect more beats this year. GenAI investments or not, the superb momentum throughout this business should yield thriving demand and fixed cost leverage to offset part of this 3-point margin headwind.

Duolingo is a special company. Very few firms can maintain 40% growth at this scale and even fewer can do so profitably. This is a rare bird. Duolingo is doing so with a 40% FCF margin and a 2024 Rule of 40 score of 78 using FCF. I can count on one hand the number of publicly traded companies consistently delivering results this upbeat. I’m highly encouraged by math and music already having 3 million DAUs and see those two subjects as two compelling growth levers to join its language learning centerpiece. Everything is going well for this company, and stellar leadership should mean that continues.

I admire this firm, this team and everything they’ve accomplished so far. They focus on all of the right things, meet all of their promises and boldly pursue new ideas while fixating on margin expansion. They balance approachable quirkiness on the outside and surgical internal focus on the inside. They add a small, healthy dose of paranoia to this formula to always ensure they’re staying ahead of the competition. And? They make me comfortable investing my money with them.

As of right now, I have no plans to add and zero interest in trimming.