Replit CEO's Contrarian AI Bet, Elastic's Search Moat & Musk's OpenAI Suit | Aug 29, 2025

0:13Welcome everyone to the information's TITV, my name is Akash Pasricha. It is Friday, August 29th. It is the end of a busy, busy week, but we have got a great show planned for you today. We are talking about a big story that the information is publishing today about the Elon Musk versus Sam Altman legal battle. We're also talking to the CEO of Elastic. It's a company you may not have heard of, but the company reported earnings yesterday, and we are going to talk about where the CEO is taking his business. But I want to start with a conversation that we have for you about the financial margins of AI and where we go from here.

0:50Revenue growth certainly hasn't been a challenge for many of AI's highest-flying companies, but one of the biggest topics right now is how much startups have to spend to service their customers. I want to bring on two people who have followed this topic very closely. Amjad Massad is the CEO of Replit, and Sri Mupiti covers venture capital, OpenAI, and everything AI here at The Information. I want to bring the two of them on. Welcome to the both of you. It's great to have you. Amjet, let's start with you. So look, this topic of financial margins and AI, you've been pretty outspoken about it, and it's been kind of an interesting topic.

1:25You know, one of the things that we've sort of come to understand is that the prices of AI as it relates to the models and accessing these models, those started to come down, but they haven't continued to come down. And so I wondered if you could talk to us, but why do you think that's the case? Yeah, well, first of all, it's great to be with you, Akash and Sri. So a few reasons. One, with regards to, actually, they've been going down, but I think Anthropic particularly has a great product. They seem to be ahead of the competition on coding, and that makes it so that there's less competition, and therefore, there's no pressure to reduce prices.

2:15For example, GPT-5, if you remember, OpenAI reduced all three prices about, I think, 90%, which is wild for any kind of reprice. GPT-5 came out of the gate quite affordable. But I think for a lot of AI coding companies that depend on Anthropik, I think that that's been the topic of discussion. And yeah, frankly, you know, that's how the market works. If you have a great product, you have pricing power. If you're selling commodity where there's a lot of other competitors, you don't have pricing power. And so from your end, so just to be clear, have you seen prices continue to come down or have they plateaued a bit?

2:58uh they they fairly plateaued um okay and you you chalk that up to the product basically and the product being a good product as it is that's my sense uh i don't speak for anthropic they're a great partner uh of ours but my sense is that uh inference continue to uh get optimized and and get cheaper and you would expect that to uh reflect the price of the product but again with any market, if you have what people call a moat or pricing power or what have you, this is where you can command higher margins. Right, right. Sri, what's your take on this issue? What are you seeing with founders? Yeah, I think that the expectation is that we are not at scale in terms of where AI could be in the future.

3:52And so while there are high costs now, things will eventually come down and i think that's the bet that both investors as well as founders are making and so right now you might be able to subsidize users and be able to just get them on board on ai products and services but over time be able to adjust pricing to be able to better reflect the uh the business model and actually like make revenue as a result i think a lot of investors liken it to like past waves of technology where for example like when the hyper scalers were getting off the ground uh they similarly had similar issues to what we're seeing now and so i think that's the bet that a lot of investors and founders are hoping to uh see play out as well right and go ahead if i may jump in that that's not the bet we're making and i've been sort of vocal about this i i uh the the prediction that i've made is either that uh prices will plateau because some model will get ahead.

4:48Or there's an oligopoly that emerges where there's a few players and then there's this implicit price collusion. Not that they're talking to each other, but you see that in the hyperscalers, right? Compute prices continue to go down, storage continues to go down, but you don't see a commensurate price go down for using AWS, Azure, and Google Cloud because those are the choices that you have. The bet we're making is that we want to build enough convenience and enough technical depth around the product in order to have a moat, have pricing power. So Replit is the only product on the market today that ships with a development environment that we've been building for many years that has all the packages loaded in, batteries included, all the library has a full IDE behind the product as well.

5:49We ship with a deployment environment, including virtual machines, autoscale deployments. We ship with database, with authorization out of the box, with security primitives. And so I think it's a mistake for investors and founders to bet on something that is not in their control. our bet is that enterprises will continue to want to pay a premium to the token price for a product like ours because of the convenience that we bring to the organization and the amount of productivity that we unlock so we hear from our customers constantly whether they're using replet for their operations which is one of the primary use cases the other use cases are for for their product developments and product development velocity we hear crazy numbers Like we saved$300 ,000 by not having to buy the SaaS product, by building it internally and building it really quickly.

6:47Or we're saving so much money in the order of also hundreds of thousands of dollars of engineering hours because our product team is able to iterate really quickly and put a prototype in front of customers to get feedback. So relative to human labor, Replit and all of AI is very cheap because that's really the comparison. It is if you spend two or three dollars per hour on a product like Replit, maybe five dollars, you're saving, you know, 100, 200, 300 dollars per hour and probably a lot more because those products can, in many cases, move a lot faster than a human developer. So that's really the comparison.

7:31It is not between different products. It's the human labor that you're automating. So Amjad, if I catch your drift here, are you saying that costs have basically come down as low as they're going to go right now and customers are going to be willing to pay more? That's the bet that you're making. Yes, and even if token prices go down a little more, we'll consume more tokens because the more tokens we consume, intelligence we create. And that's the current scaling because pre-training, which is the method for training the base models, have sort of plateaued in how much intelligence it could provide.

8:12And now new scaling factor is what's called inference time scaling or test time scaling, where you can loop the LLM more and more for it to think in order to produce more intelligence. Right. I mean, Sri, I find this argument interesting because as you said, I think the the the consensus or i guess the hope among maybe not umjad but other founders seems to be no no no we're betting on these prices coming down and we're not seeing coming that come down so this seems like a bit of a contrarian take right shari yeah i agree i think maybe the time horizon is something that investors and founders might uh be hoping um that eventually these will come down and so maybe there's differences there um but i think that you're seeing it really play out in terms of the gross margins of these companies itself where uh for example like lovable uh even though they were able to scale from 1 million to 100 million in ar within eight months which they announced last month um they still have uh roughly 35 gross margins as does um stack blitz which is another competitor in the space they have about 40 gross margins as of last july or sorry last month um And so what's interesting about that is these gross margins just represent the cost of serving free users.

9:29So inference costs for free users, but don't actually include the cost for, sorry, inference costs for paid users, but doesn't actually include the cost for free users. And that's interesting because when you really account for all of that, that would really imply that their gross margins would dip much lower. And so I really do hope for the sake of these businesses, long term investors, as well as companies, see these costs going down over time, because I think that then implies what profits that they'll eventually be able to make as well. I think our strategy is different in that we're doing a PLG play.

10:05It's an age-old play in tech where you're willing to lose money and free users, perhaps the users in the lower price packaging, but you're making money on enterprise. because what you could be spending on market or acquisition or distribution, that money you're spending on the users you're acquiring via those subsidized products. And it is better than also, in many ways, PLGs has many advantages, such as having these power users and champions that bring us into the organization. For example, we have customers at Zillow, at Coinbase, and many other places. And invariably, it is one user that brings it to the organization.

11:01They create tremendous value. They get attention for that. Other people pay attention. In the case of Zillow, it was like a product manager and the operations team that brought it tremendous value for the organization. The VP of product heard about it. They wanted it for their employees. And so you start spreading into these organizations. Right. And so our margins on the enterprise is 80 % plus. And our bet, and we're leading in the enterprise among our peers, our bet is that's going to dominate our business over the long term. Amjad, I wanted to slightly pivot the discussion a little bit to talk about data.

11:40You know, one of the conversations that we had a couple weeks ago was with Bobby Samuels, who founded a company called Protege. And they have this sort of data marketplace where they've got companies selling data on one side and then basically the model providers buying data on the other side to help them train all their models. You know, we haven't talked too much on this show about the data that application layer companies sit on and how valuable that might be for a model provider. You are an application layer company. I mean, how do you think about the data that all these applications are sitting on?

12:16and whether that could be useful. I think invariably there's use to them. We've actually proven that back in 2023. We trained a state-of-the-art open source coding model, and we trained it on open source code for the initial phase, and then we did the continued training using Reflit code, and we were able to achieve state-of-the-art. So since then, it just stopped being economical for us to train models because they're spending billions of dollars on them. And we're an application company and we're trying to build a product. And so it's a lot more convenient for us to just build a product. That being said, we still train models and do research on the margins.

13:00And oftentimes, you know, for example, we have a new product coming out that a big piece of it is an internal technology that is actually 10 to 15 times cheaper than, say, the similar technology from the big labs. So I think there is definitely, on the margins, obvious usage. That being said, I think there's this similar delusion between investors and application layer companies that it's a story that they tell themselves. Token prices are coming down. Our data is going to be so differentiated such that we're going to train models in the future that's going to be competitive with the big labs.

13:44Now, it might be true, but I think these stories are more like fundraising stories instead of actual strategies because these people just haven't even validated this hypothesis. Sorry, I didn't mean to cut you off, but if I catch you correctly, as it relates to, for example, the opportunity for Replit to sell that data in the future, Have you thought about that at all?

14:13Look, I don't think you can build a big business selling data, unless you're a data marketplace. I'm not sure what's the status of these businesses. But if you want to build a really big company, you have to create tremendous value. And selling data is not one of those. Right, right. Last question for you. Are you guys using your own models? or? At the moment, no. We use a constellation of a lot of different models. So we have a partnership with Google. Right. The backbone of our agent is Claude. Right, right. And I was just going to, in the future, could you see that being a possibility for you?

14:57Yeah, yeah. We did it in the past. Like I said, I don't think we're going to be able to compete with the core model right now, at least at the scale we're at. Right. But there are so many, for example, code search could be a separate model than the main agentic model. And we actually are using some of our own proprietary technology in order to do code search. So there's a lot of opportunities to train models. I just think it's delusional to think you can compete with these companies that are putting billions of dollars behind training code models specifically. Right, right. And that was kind of what I was getting at was perhaps could that be a way to further expand your margins?

15:40But I think going back to what you said. Yes, because you can take out these pieces that you can continue to charge for them at premium. But like I said, we can save like 10x by using some other specialized model. And that continues to contribute to your margin. But again, you're still using the core model, which is most of the cost in the middle. Again, the strategy of Replit is to build the technology around the model that allows us to create tremendous value on top of these models for which we have actual pricing power. So I think that's the mainstream of the strategy. Everything else is optimization.

16:22Right. Great. Well, a fascinating discussion. I'm actually really excited to have this exact discussion maybe two, two and a half, three months from now and see how things progress. because I think, first of all, as the new models come out, things change so quickly. And, you know, you also have this idea that models could get a lot cheaper, faster. You know, we have the deep seek side of all this that could change the game. So maybe we'll get you both back on two months from now. And Amjad, we can ask you if you're betting on the same strategy because it is a little bit different than what we've heard at the information, but I think it's very interesting.

16:56And so I thank you for coming on to share it with us. And thank you, Sri, as well. Okay. Well, Elastic is an enterprise search company that you may not have heard of, but the company sells its product to a number of big corporations like PepsiCo, Cisco, and T-Mobile. The company reported second quarter earnings yesterday. I want to bring on CEO Ash Kulkarni to talk about where he plans to take his business from here in the era of AI. Ash, welcome to TI TV. It's great to have you. Hey, Akash. Thanks for having me. So look, you guys, you grew revenue 20 % in the last quarter. You put out the guide that growth will slow to 14 % for the fiscal year.

17:37And so look, the earnings are what they are. They happen every quarter. The numbers might be different next quarter. I want to take a step back a bit and talk about where you want to take this business. Because from the outside, what it looks like is growth is slowing for the company. It has been slowing. It will continue to slow. And so where do you plan to take this business in the future? Yeah. So Akash, first of all, I just want to make sure that I clear the air on how we guide, how we look at the business. The business is doing incredibly well. There's a lot of strength in the business. As you rightly said, we posted 20 % growth on the top line, 18 % in constant currency.

18:20And what we did was we raised our guidance for the full year by the entire beat, plus a little bit more. The goal was not to imply any deceleration in the second half. But that's typically what you do, right? At the beginning of the year, you set an appropriate guide with the right amount of prudence. And then as you execute quarter by quarter, you sort of set the expectations on how you are doing, how you're exceeding and so on. So I see our ability to continue to grow being incredibly strong because the demand for our products is very strong. We are seeing our competitive differentiation really hold out in the market.

18:59And where we are seeing our focus and where we are seeing our wins are in the area of primarily AI. So if you think about the work that's going on with companies building agents, with companies building conversational apps and so on, fundamentally, as you know, you need retrieval. You need some sort of data retrieval to ground these LLMs in the right context. And Elasticsearch, our vector database, and everything that we've built around it, our own embedding models, our re-ranking models, our chunking strategies, we see ourselves as a data retrieval and context engineering platform because that context, and especially context accuracy, as you know, is incredibly important in the world of, you know, agentic AI.

19:43So that's the area where we are seeing a tremendous amount of growth and it's helping us across all aspects of our business. Right. And so what do you say to people who say that, you know, you have these larger tech, larger database companies, for example, the Snowflakes and the Databricks of the world, they've sort of, they started to increasingly look at AI search or AI enterprise search, for example, as a category. And then we wrote about today a company, Pinecone, which was a vector database company, a technology that Elastic very much rooted itself in at the outset, right? Pinecone is a company that is, it's open to a sale, it's saying.

20:22And so you kind of have Elastic, which is sort of in the middle here. You're not as big, but you're certainly not as small. And my question for you is, you know, people who say that, well, Elastic might be better off as part of a larger company. What do you say to that? Yeah, I think fundamentally, those folks don't know what the hell they're talking about, just to be very blunt about it. If you think about our core strength, it's always been about search relevance. That's where we've spent the last 15 years of our existence. If you look at Elasticsearch, the reason why people use Elasticsearch is because it's all about unstructured, messy data.

20:56If you have data that you have a good schema for it, you understand what that schema looks like, There are tons of great technologies to put it in. If you look at Snowflake as an example, or Databricks, as a BI technology, they are exceptional. We use BigQuery within the organization from Google for all our BI needs. We don't use Elasticsearch because it's just not the right fit. If you are a data person, you know that you have to optimize and you have to use the right technology for the right kind of workloads. When you're dealing with unstructured information, you need the kind of search index-based platform that Elasticsearch has always been about.

21:32Because when you think about search relevance, the techniques that you use to determine relevance are very different from the very deterministic queries that you use in a database. When it comes to vector search and your reference to Pinecone, we've always said that vector databases are a feature. they're never going to be a business in and of themselves. When Pinecone first came out with their product and they are not the only ones, our core thesis was that if all you do is shove all your data, you vectorize them and think of vector search as a singular technique, you're going to fail. You're not going to get the most accurate results.

22:12You're not going to get the best context. You need to do a lot more. And we talked a lot about what hybrid search means, what re-ranking means. And you're seeing the world come around to our point of view, which is why the news about Pinecone, I mean, I don't know what's going on with Edo. And I saw the news break on the information. So congrats on that scope, that scoop. But from my perspective, you know, what I've always felt is that this is a market where context and relevance is going to win. That's been our secret sauce for the last 15 years. we haven't been spending our time focusing on OLTP transactions like some of these other larger database companies have been.

22:51So my belief is that every company going for every organization going forward is going to have their systems for structured data, but they're also going to have a system like Elastic for all their unstructured messy data, because that's what provides the right context to LLMs. Let me ask you a quick question, because I know you've got to run. Not Pineco, Elastic, I'm sorry, getting my words mixed up. It's a Friday though, so we're chugging along. Elastic has built its technology on open source software throughout its history. Where do you stand on the open source versus closed source debate as it relates to AI?

23:28How do you think about that? I'm a big fan of open source. I'm a huge fan of open source. Naturally, why? It sort of creates a democratization effect that is incredibly good for the industry. It's a hard model, to be honest, because fundamentally it helps you get reach into a very large ecosystem of developers. That's the great part about it. But then, you know, you have to be honest with the community and you have to provide them with a free product that is really good. And then you have to work incredibly hard to build additional differentiated features that you can monetize and sell. And delivering it through a cloud service is another good way to monetize all the hard work that you're doing.

24:08So it's not an easy model, But what it does is it opens up the aperture, it brings more developers into the fore. And once you do that, those developers come with great new ideas. Like we wouldn't have gotten into the security business if it hadn't been for our community building these really interesting security applications on Elasticsearch. That's how we got into that space. So I'm a big fan. I think that same model applied to the world of AI, specifically in the area of models, this notion of open weight models and so on. I think that's going to be a huge boon for the entire industry. We are betting on it.

24:48Much of what we do, almost everything that we do, we make sure that it's in the open. You can go and look at our source code. Everything that we do is in our public GitHub repo. and we feel that that model just creates such a great tailwind for the entire industry that if you truly want AI to accelerate even faster, open source is going to be one of the most important pathways to it. Right. Well, Ash, thank you so much for coming on the show. I really appreciate it. And I'm excited to have you on again and digging into some of the very specific elements that you guys are exploring with Elastic.

25:19Because I think your expansion strategy is very interesting. That is Ash Klikarni, the CEO of Elastic. okay for our final segment of the week elon musk versus sam altman has been one of the most closely watched tech beefs in the ai era one of the most interesting developments as of late is the lawsuit that musk has filed against altman and against open ai the information is publishing a story today looking at how this lawsuit could play out and i want to bring on theo weight and rocket drew to talk all about it thank you to the both of you for being here i appreciate it uh I'll tell you what, you guys can decide who takes this first.

25:58I want to know why Elon Musk hates Sam Altman. I'm thinking Rocket might be a good one to answer. What do you think? Sure, I can take a stab at that. Thanks for having us on, Akash. I think it's going to be fun. We'll do a little mock trial here, I guess. I think it makes sense for me to explain the Musk side first, because Musk is sort of the first mover here. He's the one suing OpenAI to begin with. So to understand where Musk is coming from and his version of events, it helps to go back to 2015 when OpenAI was first set up. At this point, Musk and Altman, these two crazy kids, they're looking at AI and they're thinking, this technology could be a really big deal.

26:37But we're not sure we're headed the right direction. We think the company that is most likely to end up developing this technology is probably a big tech company. And they had some experience about that. In particular, Google had recently acquired DeepMind. So they were kind of a favorite in this race to develop the technology. And the two of them were looking at this thinking, we're not sure we trust Google to responsibly steward this technology, which could be, you know, according to them, you know, maybe the most transformative technology our species ever developed. So they thought we've got to do something about this.

27:08We'll make our own organization, but we'll be different because we'll be a charity. At the end of the day, we'll be a nonprofit charity. so we'll have a mission that's about developing this technology in the way that's best for the public, for all of humanity. And we can be held accountable to that mission. And this is where Musk comes in. Musk says, I'll give you some money. I'll give you a lot of money. I'll give you$44 million. So that's what he donates over the next few years. Unfortunately, it turns out AI is pretty expensive, right? They didn't know that at the time. $44 million is not going to cut it.

27:37So they have to start getting creative. And this is where things start to come apart for the two of them because they're starting to consider strategies like, well, should we issue a cryptocurrency or should we bring in an outside tech company like Microsoft, for example, should we bring on outside investors to raise more money? And according to Musk, he doesn't really like the looks of this. This looks like becoming the kind of company that their whole purpose of existence was to avoid becoming. So eventually it does end up being the case that OpenAI partners pretty closely with Microsoft. They set up a for-profit subsidiary to handle the commercial side of the business, develop the product, and bring in outside investors.