agent 时代，应用的账本归谁

So, human will say, I want to spin up Atlas, and then they will use Atlas in the customer adds that we continue to get, and then they scale over time. I mean, one of the labs told me that they actually came via self-service channel, and then we noticed it, and then we put team on it to make sure that they can scale nicely and so on. So, with this managed MCP functionality, Tyler, I'm optimistic that we will continue to see new customer adds. Now, you cannot game prompting. You just cannot with these agents. If you use certain words, of course, it will pick MongoDB in the prompt. If you don't pick certain words, it will default to a Postgres or things like that. So, yes, the teams are working hard because we also want right use cases to be built on MongoDB. If it is some employee-facing apps with 1,000 users, we really don't want that workload. We want the workload that can scale a quality workload versus a quantity workload.

One of them really cool and one of them (03:53:34). One of them that's cool is it's always modeled on JSON, right? That makes our migration easier later on. The mistake that the models make is that it picks Postgres on JSONB most of the time, which to Pablo's point, that's what we're trying to change. So when they scale and reach some sort of exit velocity just like Emergent did, right, easy migration and we'll get them up here. So I'm happy with the enterprise. I'm happy at the mid-tier.

We can prompt LLM JSON and it'll respond because JSONs are – I mean, LLMs are designed and built around JSON. So we think those attributes are just as important. And that's why we're building all this capability around like price performance, Infinite and Agent Engine to provide all the wrapping to enable agents. We recognize that we may have been a little slow like there's other companies like Supabase who have done some work around making – being (03:58:27) much more agent friendly. Those are easy gaps for us to fill, and you'll see that – us address that pretty quickly. And just to then answer that. So, there's memory of today which I'll talk about and then we'll talk about memory tomorrow or now today because we announced it today. Before today, when we talked about memory, you're 100% right.

Again, good usage, but it needs to monetize more as we look at that. And then of course, you have the AI natives. So underneath the covers, I mean, look at the AI natives are 3% to 5%. It means the core by definition as a percentage continues to call it slow down but the dollars continue to grow. And we look at that over the next three years and we feel very good about growth in both of those and we really haven't baked in a lot of upside for enterprise AI. It's the thing we talk to you, folks, the most about in the question earlier on when is that going to pick up. And at some point, keep in mind AMP is great. Dev did a great job of through that.

Sure. So again, thanks for having us, Matt. So let's talk about the workload growth as it relates to the durability. So in answer to your first question, we've not seen very much there's hasn't been a big difference between the workloads that we signed historically versus today. There is always some nuances by geo or by sector. What's really driving the growth is three things. One is as we continue to increase our go to market focus on the larger enterprises, as those workloads grow, we have two jobs.

Agentic workloads need memory that's transactional, high velocity, and able to retrieve the right context at the right time. Adobe's Journey Agent is a clear example, a composite, multimodal AI agent that unifies Adobe's marketing suite and orchestrates end-to-end customer journeys for their global B2C user base, with MongoDB as the agent's long-term memory and reasoning layer. Adobe leverages the MongoDB platform, Atlas Search, and Atlas Vector Search together to power the sub-100 millisecond hybrid search the agent needs to act in real time. To be clear, our results today are driven primarily by core workloads, but we are seeing real and growing momentum from AI and agentic workloads, and believe MongoDB is purpose built to be generational data platform for the agentic era. Built natively into the platform, MongoDB's innovations in the core database, embeddings, and vector capabilities are moving us beyond a system of record to becoming the real time system of intelligence. That shift comes down to five core strengths. Number one, MongoDB is architecturally built for AI in two key ways.

That is where we believe a lot of AI workloads will be and we'll talk more about that, not only the Atlas growth, but the EA growth being driven by customers running AI workloads on-prem. If listen to all that back, it somewhat maybe sounds like you actually expect some of the large enterprise AI workloads to maybe take place first on EA as opposed to Atlas. Is that a fair maybe read back or how should we think about that? So, that is a fair read back and that's what we expect. And that's what customers are telling us, which is I'm not going to move all my data to the public cloud. I'm going to run those on-prem and maybe they train it in the cloud and then they run inference or other – or build agents on-prem. We'll see. We want to meet the customers where they are and I think that's a big part of the adds to EA.

Economics may not work. So I'm going to now run this in on prem, and our ability to run anywhere was a clear advantage. But what changed in 2026, from my standpoint, we are nine months in is confluence of three things. One is public clouds definitely have capacity issues that they are talking to, even regulated industries, whether it's banks and other places. Number two is certain workloads because of the cost related issues besides the capacity related issues, they want to run on prem now and then specific to AI, getting this workload AI ready. They would rather run it in-house and from a database standpoint, in a modern database standpoint, we are the only one who can say that you can run that. And one last example I would share, a large financial services firm right here on the US West Coast told me they will move even more workload to Atlas if we can do a proper failover between Atlas and EA.

It's Mike, so let me take that. So, we will talk more, obviously, as we guide next year. We'll have a financial session at Investor Day. If you take a look at this year's full year guidance with Atlas at 27% and EA now at 11%, if Atlas is around 74% now, that certainly should continue to increase as a percent, but we do expect EA to be a more durable growth driver. So, to that extent, it should continue to increase probably not at the rate we thought before because as you talked about, the AI push is both in Atlas and in EA. And we feel very good that it is an and not an or. So, we expect Atlas to continue as a percent of total revenue, but certainly EA also contributing much more than we thought when we started the year. Yeah.

So I think token growth and Mongo DV growth are not linearly correlated. That's a good observation. And I would not expect Mongo DB consumption to grow at the same rate as token consumption. Um, I think in fact, token consumption could grow three to four times whereas Atlas consumption grows two times. The relationship is positive but it's not one to one. The reason is that the tokens are consumed every time an LLM performs the inference, whereas Mongo DB is primarily involved in storing and retrieving the data, storing the embeddings, managing memory and supporting operational workflows. And many additional tokens can be generated without any proportional increase in database activity. For example, if a company upgrades from a smaller model to a large reasoning model, token consumption may increase dramatically, whereas MongoDB consumption will remain unchanged. Similarly, longer context windows can increase the token usage without needing additional database operations. Whereas MongoDB benefits is when AI application becomes more agentic. The agents typically need memory retrieval, conversation history, tool output, and vector search. And as these agents become more autonomous, the number of database interactions per user request increases. And that's why the agent adoption is more important to Mongo DB than raw token growth. And if let's say token consumption grows by three to four times by year end, I would expect the AI related MongoDB consumption to grow two times, not three to four times. The database layer captures only part of the economic activity generated by these AI workload. I think another way to think about it is that token growth um measures inference intensity whereas Mongo DB growth measures data persistence and operational complexity. So this is why these two metrics are related, but they scale differently. And long term, I think uh actually the relationship uh only increases.

回 答 者：消耗量确实在增加。值得注意的是，消耗增长很大程度上源于架构效率不足。例如，如果智能体与向量数据库之间没有缓存层，或者编排设计不合理，就会产生额外消耗。供应商通常会将消耗曲线作为AI价值的证明，Snowflake也会这样展示，但对此需要谨慎看待。部分消耗与业务影响成正比，但如果不重视架构设计，很多消耗实际上来自不必要的架构复杂性，通过优化管线本可避免。

回 答 者：在相关能力等方面，PostgreSQL是免费的，但AWS、Azure、Google等超大规模云服务商，以及Databricks和Snowflake，也在推出与PostgreSQL相当的产品，这进一步加剧了竞争。MongoDB通过收购Voyage，在嵌入、重排序和运营数据存储等方面形成了差异化。本季度尚未出现重大影响，但客户的决策周期正在延长。部分客户即使已经使用MongoDB，也不会直接将新应用或新工作负载部署在MongoDB上，而是会进一步分析预计成本，评估是否应考虑这些新产品及其成本差异。我们预计，这种影响将在未来几个季度逐步传导至市场，届时才能看到实际采用情况。

我认为，至少还需要几个季度，才能判断是否会产生负面影响，最早也要到今年年底，即Q4。我们看到相关讨论在Q2开始增加，Q3已经形成项目储备，客户将在Q4作出一些重大决策。因此，我认为相关变化将体现在数据中。Snowflake和Databricks的Q4业绩至少会反映出这一趋势，MongoDB受到的负面影响也会在Q4显现，这可能就是我们看到影响的时间点。

I mean, we're in an environment where things are changing so rapidly. And so, I think it's just par for the course now to deal with change. And I would say to the third point about someone new coming, I would say clearly we would want someone who really understands the business and not just make change for the sake of change. Some of the change that CJ made, people were just tired. Some of my team who I was working with me, they had also been on the journey for a long time. So it wasn't like they suddenly need a new skill set. They were just tired. It was a long time, so they wanted to take a break.

It's just is, is Atlas important? Absolutely. And we focus on it a lot. I think there's a perception, which is we will only be an AI winner if Atlas... Yeah. ...accelerates. And we struggle with that greatly because that's not what our customers are saying. They're saying, we want to deploy, run anywhere. We might want to self-manage that. And I think going forward it's going to be really interesting.

We have increased our long term target to mid-20%. Okay? This is an important metric, so let's talk about this a little bit. Our old target was 20%-plus. Our new target is mid-20s. Keep in mind we're using the same guidance framework that we've talked about the last three quarters. If we've been clear about one thing, it's the guidance framework that we have for Atlas and for EA and we've not changed that. We continue to feel very good about Atlas.

So, just to level set with everyone, well, what I started talking about maybe like three years ago was the fact that when we had lots of customers who felt trapped on legacy platforms and it was easy to move the data, it was very hard to rewrite the application code. And that was the big tax (03:54:55) and because these were lot of old brittle systems, a lot of stored procedures, it was a very, very heavy complex job. With AI, we thought that that problem would get solved. What we underestimated candidly was that the maturity of the AI models took longer than we had expected. Obviously, we've seen rapid innovation of model capability and performance. And so, Mike being the great CFO he is, he's good at managing expectations, says investors are little tired about hearing AMP, so next time we talk about AMP, let's have something really hard to deliver. I would tell you that we're really excited about what we're seeing. We made an announcement of Cognition this morning, who's now our partner, who's helping us essentially work with us to help migrate and they've become very successful as a coding tool to help migrate customers and deal with that application logic problem. And obviously, we provide the tooling to migrate the data as well as put in all the guardrails to make that production system.

**Expert:** We would always typically start with the first application coming up through a development team. It's very much a land-and-expand motion. You start with a single application, a single team, they are successful and that becomes a reference for other parts of the business. Typically, the sales motion was, we would call it, finding smoke as where there's smoke, there's fire. It's like find where there's engagement and people using MongoDB, make that project successful and then use that as a launching point to go higher within the organisation. It always felt like a much more bottom-up approach. The one thing that could possibly change that is the work that they've been doing on application modernisation. Another challenge for MongoDB vis-a-vis Postgres is that there are plenty of companies that want to get out of their Oracle contract because Oracle is expensive and nobody likes working with them. Many of them can just lift and shift to a relational database like Postgres. If they want to move to MongoDB, they need to rewrite their application code. That is a huge lift until AI came along. I know that they have several software and services pilots going on to help customers modernise their applications, get off of Oracle and get on to MongoDB. That is very much a top-down approach where you're going to a CTO or CIO and saying, "Hey, you have this initiative to get off of Oracle, we can help you do that."