Transcript 0:00 Welcome everyone to a new episode of the Finops Weekly Podcast. And today, we are going to talk about, about the nitty and gritty about AI financials, all the details. 0:09 We are going to be very low level, very deep dive into what's bringing that even, you know, uh, talking about from the tech perspective to financial side, to even like, you know, more, more hardware things, more, more interesting things. 0:24 And to help me with that, we have Daniel Almeida, the CEO of Maia Labs. Daniel, how are you? Great. Thank you very much, Victor, and thank you for having me. Yeah. It's, it's a pleasure to, to have you here. 0:34 And you know, we, uh, been talking in, in, in before about the, the, the podcast and preparing all this because we are going to talk about all this AI that is, you know, the trend of, of everything. But we... 0:45 I think we have a very special, uh, line. So, you know, we both are-- have a technical background, right? And we... But we are focused a lot on, on financial. 0:55 So on your perspective, like, how is the AI financial pattern being repeated from FinOps, like from the cloud side, but, you know, in a different kind of timeline that we are seeing these days? 1:07 Yeah, that's a great question. So, uh, I was, uh, fortunate enough to, uh, cover the, uh, cloud adoption cycle, you know, in Portugal, but also a little bit in Spain in my, in my previous, uh, employees. 1:18 And, um, it's curious to see a little bit of the, uh, same fear of missing out or the hype leading the enterprise in the way that they're adopting, uh, AI. 1:28 So i-i-- we are seeing, for example, a big difference regarding cloud, which is in the cloud era, it was more about the technology stakeholders in the room when you were discussing, you know, cloud adoption. 1:41 Now, with AI, the business stakeholders are in from the day one, and often they are the ones leading the efforts to adopt AI, and the technical stakeholders come at a later stage. 1:51 So this is a little bit different, kind of the, the velocity and the pyramid has, has inverted itself. 1:57 But then we see some also some things very similar to the cloud, which is the big players, you know, subsidizing the adoption, trying to remove the friction. 2:06 If you think about, for example, OpenAI and Anthropic, you cannot see a, a billing granularity dashboard. You know, so you don't know what is the user is spending more on their tokens, what are they doing. 2:17 It's very much, you know, just burn tokens, consume tokens, and then we'll see, we'll see how it goes. 2:24 The other thing about the hype and the performance, I said to you, is that I'm seeing the same thing again, which is, uh, lack of strategy. 2:32 So it reminds me of my times at, at the cloud world, which was we have this application on-prem. It's not a very good application on-prem. 2:39 It's not even ready to the cloud, and we don't wanna do the re-architecture or the re-platform, but let's just move it to the cloud. 2:46 And then the customer just ended up having, you know, a, a, uh, the same application with suboptimal results in the cloud costing more. 2:53 And we are seeing the same thing now with AI, and we actually at Maia Labs, we usually say, you know, uh, "Don't do AI decoration, do AI transformation." 3:03 Meaning is that for you to implement AI, you need to be able as an organization to think about strategically the way that you re- you want to reorganize your company. 3:13 So it's definitely, uh, there are some, some, some coincidences and some things that are very similar in the technology adoption cycle, but others are very different. 3:21 And just to complete, just to complete my first answer would be that this is also, in our opinion, the first technology cycle where the technology is being matured as the same as the adoption model. 3:33 So there is not a specific adoption model everyone can say, "Follow this, and you're gonna have success." So that's also a big change. Yeah. No, definitely. 3:42 I, I see a lot of differences in, like, in how AI-- like the speed of AI evolution itself and the, the adoption and the growth that we are seeing, like exponential compared to what we s- let's say, saw in cloud, where the costs were a problem much later in the, in the journey compared to the adoption where, you know, all the compute. 4:03 Like most of the services that we use today have been there for... since the beginning. And th- these days, like we couldn't, let's say, leave room for, let's say, for adoption and for playing a bit more. 4:16 Like we only had, I, I think, a year only playing with, with AI to before, because we're an issue [chuckles] and, you know, foundations have been created around it. So 4:26 talking about the, the cloud and the AI difference, like 4:29 what is the disconnection that you observe between how companies are now managing the cloud cost, which is like way more, uh, advanced, a way more advanced practice versus the AI spend? 4:41 Because we are seeing a lot of differences and specifically the practice even though can leverage a lot of, of it. Uh, we have a lot of disconnection between one and the other. Yeah, for sure. 4:51 I think also, like, the, uh, AI labs and the, the enterprise are chasing different metrics. Mm-hmm. 4:56 So, for example, the other day, I just saw, you know, that OpenAI was saying that the enterprise consumption of tokening-- uh, reasoning tokens was up three hundred and twenty percent in a year. 5:07 But for, for companies, I think what they want to see is what is the cost per resolved case. Mm-hmm. You know, so it's not really how much-- how many tokens am I spending. 5:16 Am I spending less and less tokens to do the same thing? I think that's also what the enterprise are looking for. Um, I have, I have a small anecdote to, to, to talk about this, about, about the, uh... 5:27 I would say the kind of-- There was, uh, this moment, uh, where during a lot of time, people didn't really talk about tokenomics or concern, you know, about models. 5:36 It was like, you remember, like all the token leaderboards, right? People saying, "Oh, you need to spend tokens." Doing what? I don't know. Just spend tokens. 5:44 And I was speaking with a, with a customer, very, very big customer, um, and, um, they said, "Well, we built this inter-internal application to do, you know, some questions about travel requests and things like that." 5:57 And I was just asking, you know, this, this, this big enterprise. 5:59 I was saying, "Okay, so how do you manage, you know, the, uh, the, the model allocation, you know, regarding the, the questions that your thousand users do on a daily basis?" 6:09 And the guy said, "Well, everyone uses GPT 5.5." And I said, "But, but why?" And they didn't really thought about that, you know. So everyone was just saying, "Look, what is the travel policy to Brazil using GPT 5.5?" 6:22 And I think this, this connects, Victor, with, with our previous answer, which is because we are still early in the cycle, there's a lot of subsidizing and a lot- not a lot of granularity, the bill is not hurting so much in the pocket, you can use GPT 5.5 for everything, you know. 6:36 So but this is, this is an, an example that the, the discipline, it's not coming in, in, in the beginning of, of the adoption, which for me, uh, doesn't make sense because I don't remember any like serious company would run cloud like that, you know. 6:51 So, uh, there was a lot of, uh, attention, you know, to right sizing, you know, all of that, making sure that you didn't leave machines running during the weekends, things, things like that. 7:03 Um, and last but not least, I think the providers are not also helping on that because they don't provide financial metrics, so they want to sell you more tokens, right? 7:13 But you can-- as a company, it's very difficult for you to understand what's the return on investment on this. 7:19 And that's why I was reading this number the other day, fifty pers- fifty-six percent of the CEOs report no financial benefit from AI. 7:27 And I think this should be a concerning number, you know, because the return on investment side is still not being answered. Yeah. I, I agree. 7:36 And also, like, what, what we saw, I think it was earlier this year or maybe this year, um, like, the, the MIT saying, like, a report about, like, I think ninety percent of the AI projects didn't provide any benefit or anything else. 7:53 And we are seeing that as a constant, and I think that's the main 7:57 challenge that everyone is trying to focus on, especially because of, um, measuring, always measuring business output, even though it's not recommended, and we say, like, forever and, uh, we repeat that a lot, it's difficult to 8:10 measure it. And other thing that it has is that it's not a m-magic formula. Every business have to measure a specific set of KPIs, specific set of results, and it's not like, "Hey, cost per token." It's like 8:24 nobody cares about the, the amount of tokens as long as the, the result is, uh, you know, we are still providing like a, an output for, for the company that it beneficial, uh, in terms of the cost itself, you know, like margins and stuff. 8:37 Like, basic business still applies to, to, to that. And I, I think that one of the interesting things is about the usage. So now with AI, like, everything seems to be needed to apply AI to everything. 8:54 Uh, and I think you, you have a very interesting insight that I, I think also share. 8:59 It's about, you know, whether we decide to eliminate something, to delegate it to AI, or to automate it, like, the old-fashioned way, which provides a, a, a result that we expect. 9:11 So can you, uh, dive a little bit more about this, the thought that you have? Sure. So we, we usually say at Mayo that, uh, the cheapest token is the one you never spent. So that's kind of the, the rule, you know. 9:22 So before thinking about AI, let's look at first principle. Let's look what is the business objective. And business in a way they operate on two, uh, two routes. One is to, uh, be more efficient, save money. 9:34 One is to sell more, have more revenue, you know, have more money. So when businesses come to us with a business challenge, and we often got, you know, that, "Oh, we need an agent for this. We need an... 9:45 " We often say, "Well, actually, let's, let's look at this process, you know, from first principles." 9:50 So i-if we, if we are facing something which is, you know, rules-based, repeatable, there's no judgment calls, then it's a deterministic case. We can just use good old software. 10:00 It will do the job, it will be cheap, it will be reliable. So you don't need any AI to put on that, you know. 10:06 If, if actually you're going to deal with ambiguity, unstructured or multi-model, you know, context management and reconciliation of information, okay, maybe we can use GenAI and a, and a, and a good old agent as well to, to do that. 10:19 Then the last, the last part, which is also very important, we saw some customers getting really, um, difficult, uh, 10:26 issues about this was when there is, you know, uh, external contact, meaning contact with customers or external entities. 10:34 And when there are no exceptions or accountability that you're, you're-- you, you cannot tolerate hallucination, you need to put human judgment, so have human in the loop. That's very important. So just another example. 10:48 We, we worked with a customer, you know, a couple months ago. They said to us, "Look, we need to optimize, you know, our technical orders part. We have five people now doing this. 10:57 Ev-every technical note takes two hours to do." And they said, "We need an agent." 11:02 And actually, we built, you know, a composite workflow, which has deterministic automation, it has GenAI, and it has human in the loop, and we were able to reduce this for ev-every technical note taking ten minutes only. 11:16 So if we would just follow, you know, the hype train and just go directly to AI, we would over-engineer something. It would probably hallucinate more. It would cost more to the customer, you know. 11:24 So I think this is kind of the, the way that we like to work. 11:27 But to do that, as a customer, you need to be able to, you know, take a really a deep dive into your business and say, "I'm not, I'm not gonna chase, you know, the bells and whistles. 11:38 I'm gonna chase what is really the pragmatic approach to solve this issue in the most efficient way of resources and, and time." 11:44 So most of, most of the issues is not about cost or a routing problem, it's that you built something that shouldn't exist. You built-- You, you picked up a, a broken process, and you put a chatbot on top. 11:57 So because you needed to, you know, comply with the board directives, or you were fear of missing out because of the market, everyone is saying that they have five hundred agents, a hundred agents, and you feel that you're missing out on, on this. 12:09 So, uh, yeah, w-we start by pushing back, you know. So if the client asks for a dashboard, we say, "What is the decision that is you're trying to make?" You know, so we try to... 12:18 It's a little bit counterintuitive, but we try to delete before adding something, you know. So question, delete, simplify, and accelerate, and then automate in, in, in, in the end. Yeah, no. 12:30 It, it's, it's a really good approach, and I think it's super, super interesting because after all, it relates to... 12:36 I think with the business, people will relate a lot to it because if you read, I think this was, like, the four-hour workweek book when the, the Tim Ferriss, uh, mentions the, you know, eliminate, automate, delegate. 12:50 I think that's the framework. Um, so the first thing is to eliminate. Like, do we need to do this in the first place? Uh, because there, as you mentioned, like, there is no cheaper token than the one that is not spent. 13:01 So it's the same thing with if you, you have to inherit... That inherits from the process. Like, if this process, this thing is not needed, it's not bringing any value to the company, that's it. 13:12 And also what you said, um, if this is going to... We are going to implement AI, which doesn't bring, uh, deterministic result because of the nature of the LLMs, we don't need to... And we need a deterministic result. 13:25 Like, we need something that is predictable, such as code. Um, then use code. 13:31 And maybe use AI to produce or make that process faster because that's something that I've always pushed through, is, like, the capability that AI provides us now to create code in a way better, uh, in general way, 13:47 uh, has been, you know, proven. 13:49 And y- we can now create a-all these deterministic workflows much faster, thank to, thank to AI, and then we need to put AI where it's needed, like creating images, uh, you know, creating summaries, all, all these things, and, uh, that instead of we can do with deterministic. 14:07 So I think I would definitely support your approach. First, to eliminate things. Like, if you don't need it, remove it because you only get debt and more things to account for and, like, it's, it's a mess. 14:21 And one, one other thing that we've seen, and this I think got recently update, um, from, from... I think it was, uh, JR mentioned it in, in one of the, uh, Tokenomics brief. It's about the, the gateway. So, 14:36 uh, we've seen that people that are using gateways are getting more visibility, like significant and better results. We are seeing enterprises implement it. We have seen OpenRouter recently bought by, by Stripe. 14:47 So we are seeing, like, these are really, really important things. So what's your tho- uh, like, your thoughts on, on AI gateway and how it helps, you know, having better visibility? Yeah. 14:57 W-we've been, we've been using our own internal AI gateway called Brain since January. We didn't rebuild it from, from the ground up. We, we built it on top of Little LLM, you know, the open source- Yeah... technology. 15:08 And we had this... I think it's sometimes, you know, because we are a very an AI, AI-native company, when we speak with those large enterprises, we expect to see some, some things. 15:17 And when we're asking them, you know, "How do you route your models? How do you do that?" And the, the majority of them, they were u- actually using commercial tools. They were not able to use it. 15:24 So I think the, the, the, the first principle that to justify, uh, uh, um, an, uh, AI gateway, which is a guardrail, in our opinion, is that you cannot manage what you cannot see. Mm-hmm. 15:36 So if you don't see it, you're not gonna manage, right? And we made a bet internally that LLMs will be a commodity and a technology that's gonna be swappable. So it's a, it's, it's a technology you swap easily, right? 15:47 So and for that, you need to be able to have this gateway to create, you know, this kind of, you know, highway with different lanes according also to the, to, to the task. 15:57 There, there is a, there is a, a, a thing that we usually say, which is there... 16:00 And there's a principle in law, which is the principle of equality, and people often think the principle equality means treating everything the same way. 16:07 Actually, what the principle of equality says that things that are equal are treated, treated in the same way. Things that are differ-different are treated in a different way. 16:14 Meaning, for some requests, you know, you need to start from the request. Is it too complex? Is it gonna, is it gonna ask for several tools? 16:22 Is it your data or IP very important you don't wanna go into, you know, to the internet or to the external app? And for example, in our case, we're gonna talk this a little bit more later. 16:32 In our case, for example, we have certain requests that we serve internally in our own infrastructure. For example, we have two, uh, very small hardwares. 16:39 We have an NVIDIA DGX that we bought a couple months ago, and we bought a Max 2D Ultra, and both of them are running open weights model. 16:46 One of them runs one of the Gemma models from Google, the other one runs a Qwen model. 16:51 And some of the requests that we do, they go directly there because, one, we don't want to expose that information, you know, uh, to the cloud. 16:59 Second is it's grunt work, so it's, it's repetitive work, you know, that we can serve easily and cheaply through the, the, the, the gateway. 17:08 Of course, for certain tasks, the frontier technology is still very important, and you're still gonna, gonna, gonna use it. But this, th-this decision also to use, you know, and to create Brain came from our own pains. 17:20 In the beginning, we were using Cursor. You know, when Cursor appeared, you know, they were very strong, and we had a month that we just lost complete control of the costs. You know, the cost just twice... 17:31 They, they went up twice. And for a small startup like ours, every, every penny counts, right? 17:36 So and then we also understood that when we, when, for example, we left humans to decide what is the best model to work with... 17:44 For example, we did this with OpenRouter in the beginning when they, when they, when they w- they went up. 17:49 We had a user that we said actually to them, "Look, you're gonna use this model or this model," and this developer was just using Opus, you know, all the time. 17:57 So if we default to the user decision, the user al- uh, uh, most of the times defaults to the best model, not to the model that is, that is important for, for the task. Yeah. 18:07 So every request needs to be priced, attributed, and budgeted as well. So this is, you know, this is the part where it's actually important, the, the AI gateway. 18:17 And then alf-also you need to take care if you're gonna talk a little bit more about opt-optimization, you need to take care of subscriptions. 18:23 So we don't know until when, uh, inference subscriptions are gonna be, are gonna, gonna exist. 18:28 But so far, if you compare consuming directly from the API of the third-party providers or, or leveraging the subscriptions, you should go af-actually for the, for the subscriptions. 18:39 So this is, you know, a little bit of the discipline, you know, from, from, uh, FinOps, you know, which is attribution, seeing what is being spent on what, you know, doing the kind of return on investment now, uh, applied to, to tokens. 18:51 It's, again, we, we-- in the beginning, we felt a little bit, you know, going backwards, but we see that clearly this is, this is now where the market is going, you know. 19:00 Every serious, you know, AI earner currently has a, a routing capa-capacity. You're seeing more and more companies offering those routing capacities. Ramp has one, Wonderful has one. Everyone is going to, to this market. 19:13 And I anticipate that the labs at, uh, at some time, maybe post-IPO, will also have their routing capabilities, you know, just to s- to change models. Well, I, I, I... 19:24 That's a very good analysis of the, of the situation, and I also say that, you know, uh, the routing, uh, as you said, is super important because, uh, I think it reminds me of when we, you know, it's more broad and more, 19:41 I think, more affecting more resources and more people, so less technical. 19:45 But talking to, like, the tiers of the instances that we have, for example, on Compute, like we selected the instances based on whatever we consider because we think that we have enough space and we have enough resources, and then we right-size based on, on the, on the workloads that we have. 20:02 But this is, like, the order 10 of complexity because now every query could be different from one to the other. 20:09 And then you w- you can have queries that requires, like, a super advanced model, while others is like, "What's the weather tomorrow?" Um, 20:18 and, and then you need, like, you know, maybe Luna to be there or a local, uh, instance that is the cheapest of the cheapest that, that you have. So I, I think it's super important, and we'll see how that evolves. 20:30 And your, your thoughts on the, 20:33 on the providers being able to have their own AI gateway, like I, I definitely see that as a problem, uh, that they need to solve because, you know, now we've seen especially that, especially 20:46 enterprises are not, don't want to get logged in, and m-not everyone has a, probably none of the providers have a login feature that they need to, to do that. 20:57 And regarding subscription, I also see, like, I'm not sure how it's going to be in the future, but it's going to be interesting because we are seeing, like, a lot of changes in that, in that side. And while, 21:09 while it lasts, uh, I think it'll be good, but it's going to re- After all, I've talked with, with another FinOp. 21:17 This guy was more on the FinOp side, um, but he mentioned, like, he's been doing, uh, SAM for, for a long time, and the subscription management and everything is just SAM but for AI. So 21:30 it's going to be an interesting thing to see in the future. And I want to deep dive on one thing that you mentioned is the, the optimization part. 21:38 So we've talked, you know, first to, to eliminate, and then we optimize, and we briefly talked about that. So we've seen techniques, you know, the caching, the routing, the model, uh, local models, et cetera. 21:50 So what's the, the yield that you are seeing based on your experience with the company about these, these techniques? Yeah, I think so. When-- In the beginning, we actually, 22:02 as soon as we bought the hardware and started deploying, you know, these open weights models as well, we start, you know, making a little bit of a list. Okay, what do we need to do a-at scale? 22:11 It needs to be, you know, mechanical. It needs to be, it needs to be, uh, something that happens- A process. Yeah. They... Yeah. So the first thing is caching. 22:18 So we, on a, on a monthly basis, we process, you know, around three hundred and fifty million tokens. From those ones, two hundred and fifty million were cache reads. 22:28 So three out of four tokens, you know, never went to fresh compute. So this is, you know, the, the, the plumbing part, you know, done, done right. 22:37 So th-this is actually sometimes, you know, about, as I was saying to you, making sure that we treat things in a, in a different way. 22:43 Then as I, as I said, you know, comes routing after that, which is the most important thing, you know. And, uh, if the labs are not giving you that, then you need to create your own mechanisms to do that. 22:52 It can be with LittleLM, it can be with other, you know, harnesses that you can, you can use or purchase. But the, the, the simple example is from that customer I was telling you. 23:02 You cannot have GPT 5.5 to say hello and thank you, you know, or to do mechanical tasks or the same routine tasks. 23:08 So in that way, you really need to route requests considering the complexity, considering, you know, the cost, the IP, data sovereignty, all of that. 23:18 Then I think nowadays it's more and more relevant for organizations to adopt an hybrid strategy. And, um, it wa- it was the same thing at, at the cloud as well during the cloud times. 23:28 You know, the best strategies often were hybrid, you know, because, uh, some of the, the majority of the applications were not ready to be migrated to the cloud, and also the customers often didn't want to do the hard work of re-architecturing or re-platform these applications. 23:41 So it's-- We're not saying that local Victor is the state-of-the-art and you, and you gonna just now build a company and run, you know, on the VDS or Max Studios. 23:49 But in our case, for example, in our case, for example, we can-- we operate with a, 23:56 with a, a local, you know, that we use to serve, you know, some, the grunt work that I was telling you, and then we operate with Frontier as well, uh, for the most, uh, I would say demanding, demanding requests. 24:09 Then the other thing as well, it's, uh, even though it's not technical, it's something that happens a lot with AI, which is the, uh, overhead tax. 24:17 We call it the overhead tax, which is you need to be, you know- Doing a lot of tracking, reporting, triage, you need to end that as well to, to, to the agents, you know. 24:29 And so the work for a human needs to stay with a human, the work that is repetitive, you just need to ou-outsource that to agents as fast as possible. And then it's, it's, it's compression, you know. 24:39 So for example, uh, the tooling that we used, you know, eighty percent-- eighty-two percent of the tokens, they would, you know, they, they've been compressed. So it's about doing, uh, uh, uh, all of that. 24:51 So I think what we are doing now is again, not, not different to what the majority of the company are now seeing and, um, you can get really good results in terms of, you know, this hybrid architecture mixing local and, and frontier, because I think that's the right way. 25:07 For you to run, you know, open weights model, um, that are, you know, Kimi size, you're gonna need to spend a lot of money, and you're probably not gonna get the hardware that you need. 25:16 But for the grunt work, maybe with the small machines, you can run the Quens of this life will do the grunt work for you in a bea-beautiful way. Yeah. 25:24 No, I, I think that's, uh, like, a very good comparison with the hybrid, uh, cloud that we, that we had, and I think it really needs to be adaptable based on, you know, the, the type of work that you, that you have, and you can route. 25:38 Uh, that's why all the, the importance of the routing itself, like how to, um, send each of the requests compared to, and each of the workloads that use AI to, to go to the right place and go to the, to the right model and, 25:54 uh, also the, the caching and, and all those techniques. 25:57 Like it's, it's really important that to, to be advanced in, in the techniques because as you mentioned, like having, uh, that, uh, almost seventy, seventy-five percent about the, the, the tokens being cached and things like that, like it reduces a lot for the same amount of work, and it's super, super important. 26:16 And diving into one of the things that you mentioned about, you know, this hybrid architecture where you have the frontier models and the local models. 26:26 So one of the things that we saw, especially in the beginning of the AI, was that enterprise wanted to adopt all these frontier models because everyone was using it. 26:35 But there was, uh, data cost, uh, data risk, risks, especially about sending information. We have seen, you know, data leakages and AP-- uh, IP issues. So what's, 26:47 you know, your thoughts about that risk or strategy risks for companies and how are you seeing, you know, companies dealing with, with that? I think there, there are two different aspects here. 26:57 So one is the enterprise IP, the other one is the cost sustainability. Let, let's focus first on the enterprise IP. 27:04 You're already seeing a lot of companies, you know, um, going through, you know, building or fine-tuning or building on top of open weights models, building their own LLMs. 27:13 You saw, for example, Thomson Reuters, they just did this. You saw big law firms in the, in the US like Kirkland & Ellis, Latham & Watkins just doing this as well. 27:21 Some of, some of them with, you know, players like Palantir, others by themselves. 27:26 But what they are seeing is something very, very similar, Victor, which is my I-- my moat is not better-- uh, better reasoning or better models. My moat is decades of proprietary data, 27:39 uh, the suppliers information, relationships, regulatory licenses, you know, all of the accumulated judgment. 27:46 Because, you know, you can be speaking with the two biggest retailers in Spain, and they work in the same area, but they operate in different way, and they have different IPs. 27:55 So if you, if you, for example, uh, just use the frontier model all the time and you keep training these models, like for example, you see the case of Figma, for example, you know, in the Figma design situation, you get into very tricky situations. 28:09 So I think that for m-- the-- for the more regulated areas, you will definitely see the companies that can afford to training or fine-tuning their own models to, to, uh, to leverage the scale. 28:19 And again, I'm not saying that they go- go-- they, they're gonna use everything, their own models. They're probably gonna have an hybrid strategy. 28:26 If you look at the Thomson Reuters case, for example, the model that they trained is for document processing, which makes total sense in their use case. 28:33 You know, they process a lot of documents, so they don't want to be sending that over to an external provider. 28:39 Then, um, I was, I was in a meeting a couple months ago with a, with a law firm, and they were saying to me, "Oh, yeah, we use Arvi, and we love Arvi." 28:47 And I was saying to them, "Okay, so if the same-- If there are five law firms in your street and five-- all, all of those five use Arvi, what's your differentiation?" They couldn't answer me. 28:58 So i-if, if everyone is running the same models, then nobody has an IP or n-nobody has an edge. So you cannot outsource your differentiation. 29:07 Your context, your, your, your, your layer of knowledge on top of the model is a valuable layer, you know. 29:14 So then at-- When you talk about the, uh, the costs, which is you cannot really continue to think that this status quo currently of high subs-subscription, you know, because the sub-subscription that you pay a hundred euros, for example, and it costs five K to Entropic, it's not gonna last this long. 29:31 Probably after IPO, they're gonna need to, you know, rationalize the cost. And also the hyperscalers will not continue to subsidize. We see, we see this during the cloud era. 29:40 A lot of startups got, you know, one hundred K of credits, two hundred K of credits. Nowadays, they're not giving any credits to anyone. You're only getting credits if you are buying the-- if you're burning GPUs. 29:50 If not, you're not getting credits. So I think for, for an enterprise as well, there's the, there's the risk factor there. 29:56 So you should probably have a strategy, you know, where you gonna do a business case, again, to understand what is the grunt work, you know, the repetitive manual work, and I can, you know, co-invest in some hardware and use some small models, open weights, that I can control all of that and just, you know, outsource, route the work to the, to the, to those, uh, small models. 30:16 I think in, in a way, it's... The, the-- What we usually say to customers, "Rent a model, own the intelligence, because that's your IP. It's, it's not the model." And the intelligence of the customer is already theirs. 30:28 It's already there, you know, so the job of the customer is To connect that intelligence and to keep that intelligence and to build on top of that intelligence. 30:36 So that's, that's what we, that's what we see i, in, in the market that there's already this trend happening, and it will happen, you know, more and more. 30:45 Of course, in Europe, sometimes it takes a little bit longer to disseminate, but I'm, I'm-- I don't have any questions that we're gonna see the big enterprises in Europe also doing this shift. 30:55 I know that's, that's a very good analysis, and I see, um, that, uh, only a few, 31:03 uh, are able to especially fine-tune properly and, you know, customize the models and have IP within AI because I had a, a, a, a previous experience with AI, like in, you know, in a different domain and in a different, uh, topic, but, uh, the, the principles were the same. 31:21 And the complexity of a fine-tuning process, of a retraining, 31:27 uh, having, like, good outputs and that, it's super complex compared to what we see as users for LLMs and what we learn and doing a prompt engineering and some whatnot. And that's or, you know, even compressing. 31:41 Like, those things are nothing compared to fine-tune a model properly to something, uh, given the data set and all the, the training. 31:50 It has a more scientifical part that we are far away as users to be able to do so, and that, as you mentioned, requires a lot of resources. 31:59 I think your analysis is pretty, pretty good about, you know, you are renting the model. The, the model is not yours. It's something that, you know, someone else is developing. 32:08 Someone else is giving it to you in exchange for a price because the cost of doing that yourself is a lot. And then also being able to have, like, the local, the local version of it, which is, uh, again, the same. 32:22 You're renting the model, but in this case, you have the, the proprietary, uh, infrastructure to run it, and you are able to ingest the data. 32:30 Um, so I think that's, that's pretty, pretty good and, you know, uh, one thing that I wanted to maybe, uh, you know, dive in, it's like, 32:39 uh, for, for people that don't know you and for people that, you know, are a little bit new to, to, to the ecosystem, like, how do you know all that? 32:48 What's been your, your trajectory to, to know all of these tips and tricks about the, the AI cost? Yeah. 32:54 I think we, we-- So we, we, as you said, we are, we are quite new to the scene, you know, even though I don't look new. [laughs] But we're quite new to the scene. 33:02 We, we've been, you know, we have one year old as a company, but we say that actually, you know, every year at mile lab is like a dog year. So we're-- I think we are seven years old already. 33:11 We, we try since the beginning to be, you know, practitioners. 33:13 We believe that this technology really, uh, forces you to experiment a lot, try different things, you know, experiment, see what it works, see what it doesn't work. 33:22 We started, you know, as a, as a product company called Orbit Platform at the time. Um, but we had a customer that said something really curious about to-- really interesting to us, which is, "You know what? 33:34 Your product is great, but it's from two thousand and thirty. The year is two thousand and twenty-six. I'm in two thousand and fifteen. Who's gonna take me to two thousand and twenty-six?" 33:42 So we actually, you know, came from principles, and then in, in, in, in the, in the beginning of April, we co-founded, you know, mile lab, so we did a pivot, and now we are an applied AI lab. 33:53 So we have products and services. So for example, we have customers that bought, you know, brand, the, the LLM routing, uh, product from us. 34:00 But we have other customers, you know, where we do, you know, for example, reskilling, upskilling. We do executive trainings. Then we do use cases. 34:08 One of the things that we do and that the, the market has been receiving, been, been receiving the message very well is the customers, you know, especially the enterprise customers, are tired of getting the company saying to you-- saying to them, "Look, I'm gonna help you implement AI. 34:22 We're gonna do great things together." But then they come, you know, selling the old ways in a new packaging. 34:28 So customers are saying, "Look, I don't, I don't care if you put here a team of ten people, or they're gonna work five hours or six, six hours. I want an outcome. I want a milestone." You know? So we work in that way. 34:38 We usually work in very small cycles, twelve weeks from idea to production, and it's milestone. So we assume, we assume, you know, an objective, and we compromise wi-with, with, with, with that. 34:48 We s- we, we get that commitment, and we take the risk as well. So we believe that in this world it needs to, to be that. We are, uh, we are around twelve people, you know, uh, in, in Portugal. 34:59 And, um, it's been, it's been a great journey. 35:00 You know, I have two brilliant co-founders with me, Luis and Rui, and the rest of the team doing a great job, and we are working with customers, you know, also across Europe and, uh, trying to spread the word. 35:10 Very much, you know, pragmatic AI, what I was telling you before. So let's focus on the business issue, and then let's decide, you know, how the technology goes. Not forgetting a very important criteria, which is people. 35:21 Companies are made of people. Otherwise, they're just, you know, a billboard or a slogan. So we need also to take people in this journey, and the best projects that we have are co-creation. 35:33 So all of these success stories that we have is also because our customers trusted in us but also took the time to co-create with us. 35:40 That's, that's cool, and I think it's also, you know, helpful to be in a journey with, with the, with the customers and being able to do great things together. I think it's a, it's a really nice approach. 35:51 So, well, uh, that's, that's an impressive work and, you know, uh, super helpful that you share all the knowledge that you're getting from, from your adventures to today with us. 36:01 So, Daniel, thanks a lot for, for being here today, and thanks for sharing everything with us. And, you know, thanks for, for taking the time. Thank you, Victor. Always a pleasure. Hope to see you soon. Take care. Yeah. 36:13 Take care, and see you in the next video. Bye-bye. [upbeat music]