AI Infrastructure Management and ROI Measurement

I love that you brought up costs. I'm going to do an audience question right now. Do you know how much you are spending on your AI infrastructure? Do you know what the relative costs of the output of what you are, the models you're running, the programs you're running, and the AI you're building. So what's fascinating about the world we're in right now is how do you measure cost? What is cost? What is the value of what you're doing? Is the cost your GPUs? Is the cost your tokens? Is the cost the number of queries that you're able to do? So I would love to talk about that. First big macro question, how should I even think about cost? I actually have a talk later, talk about the cost. Again, really agnostic side, right? We don't host any models. It's really purely from, you know, pricing point of view because we have token index is one thing we do keep track. Like, do you always need the latest or greatest model? Maybe it's, maybe not. Maybe you want to use, you know, those open source models where it give you more deterministic, faster. There's no thinking involved at all times, right? Not much thinking involved. So it really depends on, right, the total cost aside from a token, other, there's human costs, right? There's other, but everything else is much, much lower compared to actual token costs and the input of a token, I'm going talk about all those things, but in general. So I think another thing people don't consider is the human piece. If you keep running something, the human being, you're expensive, engineers are very expensive. So that's one of the things I want to highlight. She's very expensive. Well, maybe I'll be replaced by AI in a few years. Yeah, I think that makes a lot I agree with that. There's a lot of like hidden costs that are not necessarily discussed. Just maybe just to echo on that, there's also a lot a lot of, like, fixed costs, let's say, in power and cooling, like running data centers. Those are really expensive as well that are not necessarily accounted for when you're but this might also be applicable only to bigger companies as well. Yeah. So for wearing my pre Lambda hat, cost is opportunity cost. Okay? So in the the mode that I am seeing conversations happen right now, and this is not me speaking from direct engagement with Lambda's customers, at least not yet. When you have a clear value proposition, it becomes much easier to justify what you need and where. But I think our company has to be very conscious of the costs that, Elisa mentioned, right? Building a working with our colo providers, work to get our capacity, potentially building, in the future. All of those things are, a financial game as much as they are a physical constraints game. And so, being intelligent in terms of picking the right infrastructure at the right time in the right place, that's our best bet in terms of managing cost, because our customers are then going to influence what we need to do next. Maybe just to add on to that, I really like that idea. I think something else that we don't have a good sense of is like what is the ROI of our capacity? Like how do we measure the sort of like upside of using our capacity that is very expensive? That's not always easy to measure either. I would say margin is hard to measure to begin with. The ROI is even more challenging, we don't have costs and yeah, it is very, the whole chain is very murky at this point. Yeah. And from an ROI perspective, should I be thinking about is it the number of tokens that I'm that we are that we're generating if we're doing inferencing? Should I be thinking about how many users I can support simultaneously? Should I be thinking about GPU utilization? And this idea of what is the currency of AI? And if I'm an enterprise, how do I know how to understand if this is good or bad? I think so it does vary, I think, depending on the machine itself as well. I I think token is, like, a very common Costa, I would say, unit here. But then you also do have, let's say, like, your, you know, network cost, IO cost, or your storage cost, which could be measured in terms of, like, know, your petabytes. So I think the unit varies a lot depending on what type of machines we are talking of. Yeah, I'm going go back to the opportunity cost discussion. At least I learned I did my basic training in that at Google. So, but I think the issue there for an enterprise is understanding, I don't know, it sound like a broken record, understanding what you need to do. What is AI going to accomplish for you? That dictates effectively your input understanding of what it's worth and then what you're prepared to spend. And I would just caution, I think, doesn't matter what metrics you pick. Just know that what doesn't any metrics have blind spots will lead you to a path that could not be the right may not be the right path. If you do token outputs, not all token the same quality, and the precision can be different, right? Like the longer the output doesn't mean it's better, right? Another thing I will caution is utilization. How do you precisely measure that as well, right? So, if I were anyone, I would try to measure, I will be a hoarder. I'll measure everything, and then we'll try to triangulate and see what kind of thesis they come up with. As a business owner, right, what I do when I have an AI product is I don't I'm not concerned, maybe it's the wrong thing to do, I'm not concerned about cost. I'm concerned about output. I'm concerned about will my user pay for this, right? Once they prove its validation, I was like, okay, let me try to lower the cost. My GAC, my default insights use the best, the latest and greatest. Once I achieve what I want, we'll we'll break the workflow down to different models, we'll try to do different things, but right now I'm not concerned about cost. So you brought up a couple of interesting things, this idea of, like, what are what are our blind spots? What are the things we don't know yet? And there's also a piece of this that is there's only so much new here. There's only so much new. We are building infrastructure for AI in the world that did not exist years ago. So there's some blind spots in that, but there's only so much that's new. There's a lot that's pattern recognition.