cost of-revenue
Model and inference bill for AI features in the product
0.28% to 8.4% of the quarter’s revenue
Incremental total: counts in full.
Matched line moved : the whole line, not this channel.
our inferencedisclosure: bounded· motive: product-defensive· before LLMs: new
The CFO put AI expenses in cost of goods sold in the tens of millions of dollars and, with hosting, among the largest items, with no period. The CEO said the cost per Video Call fell from at launch to under , at least lower, mainly by moving to open source models, and that all AI use in China runs on local models. Gross margin of beat the expected and the year-end expectation rose to from . The size shown is the ledger's own estimate, : the phrase is in the words-to-numbers table, and with no period it converts to the quarter through the fixed undated band. The former estimate, a judgment share of cost of revenues, is no longer used. The counterparty is mixed: third-party model providers, and the hosting and compute providers that serve the open source and local models the company is moving to.
Evidence: 18 quotes, 11 figures, 2 confounds, 5 from before coverage
Third-party AI model costs carried in cost of revenues: what the company pays to serve Video Call, the Max features and the AI-generated content in the learning product. The filing names them beside payment processing and hosting fees and does not split them out.
Why this motive
Inference cost in cost of revenues for features moving into the standard paid tier: the cost per Video Call fell far enough to give it to Super subscribers (claim c2), and the CFO keeps the AI cost savings in margin while putting Video Call out to all Super users (claim c9). The product-defensive tell, as in the prior quarter.
Before LLMs: new
At the anchor the 10-K already listed generative AI costs inside cost of revenues beside payment processing and hosting fees, with no amount; cost of revenues was for FY2024. The CFO said on the Q1 2024 call that every use of the Max AI features carried an incremental cost. The bill exists only because of LLMs, so the tag is new whatever the anchor shows.
“Cost of revenues predominantly consists of third-party payment processing fees charged by various distribution channels in addition to hosting fees and generative AI costs.”
“Other business purchase commitments consist of hosting costs, web services and generative AI costs.”
“We talked about the fact that for MAX, there's incremental cost for every use case, for every usage of it. And that does lower the gross margin of the MAX product.”
“Further, our ability to continue to develop, maintain or use such technologies may be dependent on access to specific third-party software and infrastructure, such as processing infrastructure for the training of our own machine-learning models or the use of third-party AI models.”
“Now, that said, we expect that the cost of generative AI, the cost of large language models, will go down. So over time, we'll probably be able to roll out everywhere.”
Figures
- Gross margin (three months; the same 10-Q row also carries the six-month figures) · 2026-CQ2
- Gross margin, prior year (three months) · 2025-CQ2
- Gross margin expected for Q2 2026 when guided · 2026-CQ2
- Gross margin above the expectation, percentage points · 2026-CQ2
- Gross margin guided for Q3 2026 · 2026-CQ3
- Gross margin guided for fiscal 2026, raised on AI cost trends · FY2026
- Gross margin expected at the end of 2026, per the CFO · 2026-CQ4
- Gross margin expected at the end of 2026 on the Q1 call, per the CFO · 2026-CQ4
- Cost per Video Call when the feature was first put on the platform, per the CEO · FY2024
- Cost per Video Call now, upper bound, per the CEO · 2026-CQ2
- Fall in the cost per Video Call since launch, at least · 2026-CQ2
Reported line it is matched to
Cost of revenues was of revenue against a year earlier, against the prior-year share; gross margin rose from to . The 10-Q attributes the increase primarily to lower per-unit third-party AI costs.
2026-CQ2: 2025-CQ2: 2026-CQ2: 2025-CQ2:
What else could explain it
- line composition: Cost of revenues is mostly payment processing fees and hosting; the AI part is not split.
- mix shift: Subscription revenue grew faster than the rest, which changes the line mix without any AI cost moving.
Quotes
“When we first started adding Video Call to the platform, the first time that we put a Video Call on the platform, I remember the team that was working on it told me, "Okay, we can give this to users, but it's going to cost, like, $0.30 per call to give to users." That was expensive, and this is why we decided to put it behind our most expensive plan, which is Duolingo Max.”
“The good news is that, through a lot of really hard work, we've been able to bring down the cost of Video Call. It is now under $0.01 per Video Call. The reason for that is mainly a move towards open source models. It's just a lot cheaper to do that, and we don't see a loss in quality in there.”
“We're developing a lot of features, and there may be some that, because of costs of them or something, we end up putting them behind Duolingo Max. That is not the goal. The goal really is to try to give, particularly speaking features, which are the most expensive ones to provide.”
“I do expect that what'll happen is that the cost per usage of AI, whatever it is, the cost per token, if you want to call it, will come down for us because we will continue moving more and more to open source.”
“Internally, of course, we're still using many models from, say, OpenAI, Anthropic, and we'll continue doing that.”
“What we're basically saying is by going up about a point and a half on that adjusted EBITDA margin, we are seeing AI cost savings that will give us a bit structurally a better margin, even with our goal to put voice call out to all super users over the course of the year.”
“For gross margin, we now expect to end the year closer to 70% as compared to the 69% we initially expected, as we drive more AI content into our products, offset by AI cost savings.”
“Yeah, our expenses on AI in the cost of goods sold are in the tens of millions of dollars, so they're significant to the cost of goods sold. Hosting is another big cost for us as well, but those are the two biggies going through the cost of goods sold.”
“The reality is that in China, we simply cannot use the AI models, the kind of U.S. AI models. We have to use local models. That's by law.”
“At the highest level, our view inside the company is that it is in our best interest as a company to use open-weight models as much as possible. If I had a magic wand, I would try to move everything to an open-weight model. It's not always possible because sometimes the frontier models are more advanced. From a company standpoint, it is just significantly better because it's way cheaper to use open-weight models.”
“Gross margin was 72.6%, a small improvement over the prior year, and was ahead of our expectation of approximately 71.0%. This reflects our measured pace of AI-powered feature expansion (such as Video Call), as well as AI cost efficiencies.”
“We expect a gross margin of approximately 71.0% in Q3 and approximately 71.6% for the full year, better than the trajectory we outlined on our Q1 call based on AI cost trends.”
“The increase was primarily attributable to an increase in subscription gross margin, reflecting continued reductions in per-unit third-party AI costs.”
“Cost of revenues predominantly consists of third-party payment processing fees charged by various distribution channels in addition to hosting fees and third-party AI costs.”
“It just turns out that for a lot of applications, you don't need the absolute smartest model. The reality is the quality's indistinguishable for many applications that we use.”
“We do that in China. All of our usage of AI uses Chinese models.”
“Cost of Revenues and Gross Margin. Total gross margin increased to 72.6% from 72.4% during the three months ended June 30, 2026 and 2025, and total gross margin increased to 72.8% from 71.8% during the six months ended June 30, 2026 and 2025.”
“That's certainly what we have. What I would say is that our expectation is that over the next some amount of time, that portfolio will be weighted a little more towards open-weight models than it is today.”
By quarter
- Q1 2026direction only · our inference · product-defensive · $3.9mn to $28mn
- Q2 2026bounded · our inference · product-defensive · $833k to $25mn