The Information's TITV
00:21:44 · Akash Pasricha, Paddy Srinivasan
[00:21:44] S0: And help me understand. So $18,000,000,000 per megawatt. So when you say yield per megawatt, that's... At the end of the day, that's what you're taking home is higher than what the other Neo Clouds might be offering? Or is it is it higher pricing? Like, just break down that metric for us. [00:22:03] S3: Yeah. It is not higher pricing. It is... We make $18,000,000 per megawatt on a blended basis for all of our hardware. And the reasons for that, the the reason why we command a premium is not because our prices are high, our prices are very competitive, but we have multiple ways of monetizing it, Akash. So we, of course, we we can sell GPUs, but we don't do that very often. We sell services, software services on top of that so that we can monetize our hardware in different ways. Like, we sell data services. We sell compute agent sandbox services on top of our infrastructure. We sell tokens on top of our GPU infrastructure. So all these are software enabled infrastructure plays where we can get a premium on top of others who are just renting their hardware. [00:22:53] S0: And so maybe this sort of relates to that, but then what gives you confidence once compute prices start to stabilize in this supply and demand environment? What gives you confidence that you will still be able to command the same demand once these dynamics equalize?
00:22:53 · Akash Pasricha, Paddy Srinivasan
[00:22:53] S0: And so maybe this sort of relates to that, but then what gives you confidence once compute prices start to stabilize in this supply and demand environment? What gives you confidence that you will still be able to command the same demand once these dynamics equalize? [00:23:14] S3: Yeah. So, I'm very confident because what we are monetizing is software and not just hardware. The need for more sophisticated software is increasing by the day because most companies are moving into productionizing agentic systems. And agents need memory, services, agents need complex orchestration, agents need secure operating environments called sandboxes. So there is a significant amount of software value addition that we provide which is not easily copyable. So as the supply demand imbalance kind of reaches an equilibrium over the next eighteen, twenty four months, we'll see if it reaches there in the next couple of years with many data center projects getting into scrutiny. I think we will probably be in this supply demand imbalance for longer than we want to. But regardless, we feel confident because we are a software company and we are monetizing a software platform, not just reselling hardware. [00:24:20] S0: Will that scrutiny comes maybe in an increased form because of the election results, will that scrutiny affect your data center ambitions at all? You talked about the facilities that are coming online for your own business on the earnings call.
00:24:20 · Akash Pasricha, Paddy Srinivasan
[00:24:20] S0: Will that scrutiny comes maybe in an increased form because of the election results, will that scrutiny affect your data center ambitions at all? You talked about the facilities that are coming online for your own business on the earnings call. [00:24:33] S3: Yeah, exactly. So we are very, very careful, Akash, in terms of the data center partners that we work with. We try to work with tier one data center providers that have been doing this for dozens of years, and we are very careful in selecting campuses for the most part that are already built. So most of the data center capacity that we have announced coming online in 2027 are already in campuses that are already energized and permitted in tariff. So we are adding a data hall or we are adding an adjacent building in an existing data center campus. We are not in this, Oh, let's bring on a gigawatt of new facility where we have to break ground and work with the local county and things like that. So we typically work with data center providers who are really top notch and who have been doing this development for a number of years. And we try to stay away from the hotspots of extreme data center scrutiny in the country today.
00:25:32 · Akash Pasricha
[00:25:32] S0: Right. I want to shift the conversation slightly to dig into the demand side of the equation here. And, look, we had chat, which was sort of a big moment for the world, I guess. Then you had these coding agents that sort of commanded a lot of the attention. Now there's been a conversation around, well, what is the next big use case for AI that will continue to drive demand and accelerate it? Look, personal assistance has come up certainly with Muse and Instinct. It's become a popular category. Elon has been very clear on X that the economics of some of their own personal assistants haven't been that great. And so I guess there's a question as to how long these personal assistants can stay at the prices that they are, also how good they are. What do you think is sort of the next big use case beyond coding agents that will command a lot of these tokens and compute that people are building for?
00:26:37 · Paddy Srinivasan
[00:26:37] S3: Yeah, great question. My fundamental belief is there are two different segments of the market. I think the knowledge worker workflows will increasingly become more and more agentic. It could be in horizontal use cases like marketing agents or sales agents and things like that. We're already starting to see that. There'll also be vertical agents for doing the legal work that has also started, or vertical agents for doing ad production, multimedia, generative media type of use cases. And then the second big bucket that I am starting to see emerge is the physical intelligence and agents that will enable physical devices, whether it is a surveillance device or a humanoid or a drone or any kind of physical device that needs intelligence is the next wave that is about to come on board. So I feel like the proliferation of these use cases is going to continue at a rapid pace, but there will be normalizing of do you really need the latest multi trillion parameter frontier model to drive all of the use cases. So for example, I have a bot now trying to hunt me down an iPhone Duo. Like, do I need Tables 5.1 for that? Or can I use open weight model to do that grunt work? Think there will be a lot of routing of tasks to intelligent routing of tasks to the best appropriate model for a given task. So I fully expect the cost to normalize, but there will be a proliferation of agentic workflows in both knowledge worker, but also in the physical world.
00:28:26 · Akash Pasricha, Paddy Srinivasan
[00:28:26] S0: I was trying to track down a tweet this morning that I feel like I came across a couple days ago about how maybe cybersecurity and finding vulnerabilities and addressing vulnerabilities, I think somebody had mentioned that they expect it to be something like 50% of all token use. Is that a future that you think is in the cards? [00:28:48] S3: Absolutely. Absolutely. We are living through that. I don't know if you saw, but the past month was dubbed as the Black October, given how many cyber incidents we all had to deal with as the industry. So I think this is only going to escalate. That's [00:29:03] S0: a you great think 50% is reasonable, you think? [00:29:06] S3: I don't know what the exact number is, but [00:29:09] S0: Clearly, I feel like I don't either. I just... It was... I I remember seeing it. [00:29:13] S3: No. But that's a great use case for a must have a frontier model both for playing offense and defense. That will, like you cannot compromise on that. So I think we will find an equilibrium where there are some frontier tasks for which you need a frontier model, and then you'll have a super long tail of tasks for which you can route it to the best available model.