Former Twitter CEO Building AI Web Infrastructure, Waymo’s Freeway Expansion | Nov 14, 2025

0:13Welcome, everyone, to the Informations TITB. My name is Akash Pasricha. It is Friday, November 14th. We have got a fantastic show lined up for you today. First up, I am talking to Parag Agarwal, former CEO of Twitter and now founder and CEO of AI infrastructure startup Parallel Web Systems. I'm excited for that conversation. We're then talking to the founder of an NVIDIA challenger, Dmatrix, about how they plan to compete with NVIDIA. And we're then bringing on our Deputy Bureau Chief of Finance, Corey Weinberg, to talk about his future story on Grindr and its escalating boardroom battle. And last but not least, we are debuting a new segment today called The Editor's Cut.

0:55Each week, a rotating group of senior editors from The Information will join me to take you inside their edit meetings, helping us understand what the big questions are facing specific pockets of the tech sector. Today, we are talking about autonomous vehicles, which promises to be an exciting discussion. I'm very excited. It's a very big show. And so let's get right on into things. One of the big questions surrounding AI agents is how they will be able to pull the most up-to-date information from all around the Internet. Parag Agarwal, the former CEO of Twitter, is tackling that exact question with a new company he founded two years ago called Parallel Web Systems.

1:32The company this week announced it raised$100 million at a$740 million valuation. And I want to bring on Parag to talk all about it. Parag, welcome to TITV. It's great to have you here. Thanks for having me. This is going to be a fun discussion. I'm really excited about this. I want to start with what your company does. And where I actually want to start the discussion is when I go to the parallel website, I see a little toggle at the bottom that says, are you a human or are you a machine? And I was playing with that toggle this morning. And I want you to tell us a little bit about how that toggle relates to what you're building at Parallel.

2:07Well, the web, as we all experience it, we use it. We do it in browsers. We do it in apps. It's been built for humans. When you think about publishing on the web, what do you think about? You think about, okay, how will people come to my website? How will people find it? How will people navigate through it? What actions will they take? Increasingly, it's actually going to be AIs that use the web. when we started the company, the motivation was that it was clear what the future would be. AIs would use the web a lot more than humans ever have. And so now the web has to evolve for its new customer, AIs.

2:45And our company's name is Parallel Web Systems. A parallel web built for AIs is how we view the world. We publish our website for humans and also for AIs. and it's a nod to what we see the future will be for all websites. Right. So, okay, so you see that being the future of the internet. Now talk to me about what the actual products are that you're developing at Parallel. At Parallel, the products we're building allow anyone building an AI agent to search and access the web. So if you think about any work that happens that you might do in your browser on the web. Now, let's say we build a AI agent to help you in whatever work you're doing.

3:31That agent must have a collection of tools to be able to do that work. Now, one approach is to give the agent the exact same tools that we humans have been using and let it go use them. And people do that too. We believe that the web is a interesting enough, hard enough, big enough problem that it's worth making a tool which is optimizing for AIs. And as much as all of us think that AIs are very similar to humans because they sound like us when we chat with them, they're not. They're limited by context windows. We need to put high signal, amazing information in front of them. So we're building tools, optimizing for AI to be able to more effectively use the web.

4:19And so in terms of your primary customer group, are you selling this tool to companies that are building AI agents to sort of help them access the web better? Or is this a tool that can actually be used by website publishers to help them optimize their own websites for the AI agents that will be crawling their websites? Our primary products today are for people building AI agents so that they can access the web. in order to do so we build if you might a search engine for ai agents which allows them to search the web or we allow them to pull content from the web our customers end up being all kinds of ai agents in all kinds of spaces so if you think about the space that everyone talks about all the time coding agents coding agents occasionally need to go look up documentation or they need to search the web to figure out when they get stuck and how to go beyond that or to pull data You think about someone building a finance agent.

5:16You think about someone building an agent for insurance or risk. You think about someone building an agent for a consultant or for a lawyer. And you look through all of these disciplines, even people doing like scientific research agents, they all need the web because all of these jobs need the web. So these agents now use our tools and our technologies in order to do work. to access the web better and more accurately. Yes. How do you deal with the issue of paywalls, for example? We are our website. We are known to have a paywall that it's pretty hard to get around. You have to subscribe. But how do you deal with all the information that sits behind paywalls?

5:58I'm thinking of news publishers, but there's also other websites that have this. Yeah, we currently do not go behind paywalls. We only access the open web. So content that can be accessed without paying, content that can be accessed without logging in. And so your paywall, your content isn't available to AI agents. One of the reasons we started the company is to incentivize the web to remain open. The open is a miracle. And I think it'll require some real business model innovation to incentivize you all to You're puzzling content in a way that you want AI agents to access it. And it's possible because AI agents can generate a lot of value.

6:46If we figure out a way for them to share that value with people creating amazing content, because that content is really valuable to the AI agent's customer, then you will have incentives to allow AI agents through our system, access all of your great content. because ultimately you are going to go through an AI agent to get to your customer. Right. So that's the challenge at hand. What is the solution? How will you incentivize publishers to do that? The solution will come soon. We don't have a solution to announce today, but let me tell you why. You're working on it. You're working on it. We are working on it.

7:28There are hints at a solution on our website. Whether you look at the human version or the machine version, there's an about page on there. But I'll give you the broad contours. The broad contours are our thesis AI agents will add a lot of value. When you add a lot of value, we also believe data published on the web is really fundamental. And there's some really high quality data that is really important. Now you combine intelligence with high quality data from multiple sources. And we believe the whole is much bigger than the sum of parts. Now, if you create a large amount of value through bringing things together.

8:03And then if you can figure out how to share this bigger pie you created, then you can create incentives for collaboration. Now at parallel, what we're working on is figuring out what are scalable ways of creating these incentives for people to keep their information, to keep their data open and welcome AI agents instead of block them. Now, I do want to go back to the incentives thing, because this is something we've talked about on this show is what is the right format for these incentives, for these deals that companies strike with publishers? We've seen a shift in a flat fee model towards more of a usage-based model, depending on how much you actually access the data.

8:49How are you thinking about it? Is usage based the direction that you are thinking of taking your own line of products? How do you think this is the right way to structure these licensing deals? I think it's much more akin to how much value this data provides, which has to look into not just was it looked at, it is the unique value of some data. Is it a very unique data point versus one of 100 ways you could have accessed this information? And so a lot of factors have to go into understanding how a certain piece of information or data or a publisher or a website must be valued. And that's some of the research we've been doing and it's looking very promising.

9:33I want to ask you sort of about the AI agents landscape at large. There are so many agents, every enterprise software company, the public, the private ones, everyone has their own agents. And we've talked on this show about how the agents are overlapping with each other, certainly with the public software companies too. I mean, And we actually had a great table that we published in a story this week about it was a matrix between all the enterprise software companies and all the different types of agents they could offer. And we colored in the boxes to show where each player is playing. And newsflash, everyone's playing everywhere.

10:05You know, there's a lot of overlap. How do you see this story playing out three, four years from now in a world where all these companies are offering products that very much step on each other's toes in a ways they haven't for the past five years? Well, I think competition is a wonderful thing. People building, it's very clear that there is a real opportunity here. There are a lot of problems to be solved. And we now have technology which can solve those problems. In our view, it will come down to AI agents delivering reliable, high-quality work. The reason at Parallel we focus so much in all of our technology, all of our products, we sell it on one thing and one thing alone, which is quality.

10:46If you get accurate answers and if you get work done at high quality and higher quality than anyone else, then you have the right to win. And we believe that everyone is going to go try to keep improving their systems and improving their quality. And I think ultimately, when things settle down, the highest quality solutions will prevail. So the highest quality solutions prevail. Do you see there being some kind of a shakeout three, four years from now? Do you see there being consolidation between some of these big enterprise software companies? We go from the Bessemer Cloud 100 to the Bessemer Cloud 30.

11:30Like, you know, how do you see this shaking out? No, I think it actually goes in both directions. So there is going to be some amount of consolidation, but there's also going to be a large amount of new bespoke software created. Now, if you look at how easy it is to build personal applications or applications specific to a department within a company, that's also happening, right? The ability to customize software instead of all of us using exactly the same custom software is also increasing. So what we expect is actually much software that works for you. Now, at the top end, there is going to be some consolidation, but broadly, there are going to be more personal pieces of software that we will have access to.

12:16And we will not just choose from a desk, from a shelf. Instead, pick up ingredients and make our own. That's part of why our product is an API for people to be able to compose and build and fit into whatever their I want to ask you very quickly about the old world that you came from. You spent more than a decade at Twitter, and I read that you still call it Twitter. You don't call it X. Go both ways. Go both ways. Okay. What I want to ask you is, you seem to be using the platform still a lot. I wonder how your experience at that company, you spent over a decade there, certainly in the senior most position of that company.

13:00How did your experience at that company over the years inform the way that you're building parallel? I also spend my time as the most junior software engineer. But I have to tell you, that's not listed on your LinkedIn. Your LinkedIn has three titles, Distinguished Software Engineer, CTO, and CEO. So I don't know. I can also fix that. I think your experience do shape you, right? I think the big thing you have to learn from an experience like that in building a company like this is which part of that experience to bring to a new context versus not. Now, Twitter was, when I joined it, a rocket ship on the consumer side.

13:43We could not roll in servers fast enough into data centers to keep the site up. Now, that is a very different context to work in than a pre-product market fit company when I started parallel two years ago. And at Twitter, through the pandemic, Twitter was a remote friendly environment. Parallel started religiously in person five days a week with this one, because you need extreme velocity, agility. And it's really fundamental the first year or two in figuring out what the customer wants, what the technology is, and you need to switch directions every other week. And so instead of saying we'll take how we used to do things in a place and bringing it to a new place, you look from first principles.

14:31In fact, if you think about how we think at parallel, it's informed by at Twitter when I built products, I would imagine a human on a phone opening up the Twitter app and what their experience was and try to explain in every piece of technology we built to serve that moment and that customer, right? Now, we're like, okay, as a result of that, we learned these patterns of building distributed systems or how we build things. Turns out, if you think first principles and said, okay, the human is no longer the customer. It's actually an AI agent that's our customer. How would we build that technology differently?

15:12And where will the biggest differences lie? And let's go in a way there. So you don't apply what you did there. You in fact question that what might be inbuilt assumptions in your mind as a result of having done that for 30 years now. So in some ways you're building basically something that is somewhat opposite to what you built at Twitter, which is at Twitter you were building for people. In this world, you're building for everyone but people. Exactly. And I think that's the whole premise of the company that the customer for the internet, for the open web is changing. It's been built for humans as customers, and now it's going to be AIs.

15:53So what do you do differently? Can I ask you, which do you find more fulfilling? Because on a human level, building something for people and then building something for machines, look, my opinion, I would prefer to build for people, right? And I know you can make the argument that it's people at the end of the day who are using the agents and it's built to benefit them. But I mean, did you like building for people at all a little more than building for machines? I love building for people. And when you build for machines, you are building for people. I take for granted that AI serve humans. I don't question that.

16:32I believe that to be true. So you are ultimately building for people. I do find it more exciting in this moment to be building for machines. Because I believe the scale of how machines will use information, be able to aggregate information from different places, put connections together, find patterns, and drive insights, innovation, knowledge creation is going to accelerate. So by building for people, the amount of impact you get to have, the amount of value you get to create for people is higher. That's what's exciting about building for machines. Right. Last question for you before I let you go.

17:21With your experience from the social media world, I think you're the perfect person to ask this question to, which is that in the era of AI, now we have the advent of AI-generated content. This is video, this is text, this is images. And look, every dinner that I go to, the topic is that we are going to live in a world where we're going to see more of this content. It's going to flood our social media feeds in a way that we can't even imagine. And there are some concerns with that, obviously, right? It's the question of, well, how do you know what's real? How do you know what's right? How do you know what's verified?

17:54Does that concern you at all, the way that the content creation ecosystem could go in this future that people are painting? There are going to be challenges, certainly, as content creation is easier. I do believe these are all surmountable challenges. I think we went from everyone being able to publish on the web through blogs and by creating their own websites, there was some friction to there. Social media enabled everyone to publish with no friction. It created challenges that we had to grapple with and we continue to grapple with. And there is a huge continuum from only publishers who like had access to publishing companies being able to publish and distribute content way back when, to now everyone can publish on social media and now the friction is even lower.

18:53Content that looks and appears high quality and requires a little bit more to figure out what is actually high quality. right so uh that but we also have the tools to now solve for those challenges and there are always going to be it's it's it's always a battle to adapt right i think you're going to see a bunch of ai slop around and it'll be really worth the smartest people's mind to figure out how to have the highest quality content, the best content that people actually want and care about shine. And double click on those tools that you talked about. What in your mind is working? What is the solution that we should be developing to combat this challenge?

19:42I think some of these solutions end up being technology and some of these solutions end up being societal and the best once end up combining the two. So establish cultural norms. We will establish tools and technologies we use. People will establish ranking systems and how they incentivize content. People will establish reputation and brands. And you might go out there and you might say that we only create content this way. And if people trust that, you will win. On the other hand, someone else will go out there and take a different strategy. And if they're able to produce extremely high quality content using AIs and gain the trust of people on the other side and get distribution, they will win.

20:22And so I do believe that, like I'm a big believer in free markets. I'm a big believer in us being able to solve these problems together. It will not be a linear path. But I think even our technology, we build search rankers systems to try to figure out what's the highest quality content. We run into this challenge every day. We don't solve it perfectly. But there are many, many things we can do in order to help our customers get the highest quality, original, reliable, accurate information from the web, which is relevant to the question at hand. Right. So algorithms to combat algorithms, in short.

21:02Well, thank you, Parag, for coming on. It's a great, great conversation and I really appreciate you joining us. And I know you're early in your product development. And so as you release more and more products, please do come back on the show and tell us more about what you're building. We appreciate you coming on. Thanks for having me. Of course. Parag Agarwal. Okay. Another NVIDIA challenger is picking up steam and raising new funds for its rival chip. Dmatrix announced this week it has raised$275 million at a$2 billion valuation. The company says its technology is 10 times faster, three times cheaper, and three to five times more energy efficient than GPU-based systems.

21:43Joining me now is Sid Sheff, the CEO of the company. Sid, welcome to the show. It's great to have you here. Hey, Akash, how are you? Pleasure to hear you. I'm doing well. Let's talk chips. How about that? Love it. I can talk it 24-7. Well, I trust you're pretty good at it because you just raised a lot of money. Clearly, some folks believe in the way that you talk about it. Tell me, why is Dematrix's chip better than NVIDIA? As I talked about the stats up front, okay? You talk about faster, more cost efficient, it's cheaper. How are you doing this? Yeah, yeah. So I think let's peel the onion on that a little bit, right?

22:21And let's take a step back. So first of all, Dmatrix is focused on inference, right? We don't focus on training, which is the other piece of AI computing, which, you know, NVIDIA GPUs are extremely good at, right? When we started the company in 2019, we had a singular mission. Focus on the other piece, which is inference, where AI actually gets deployed, where monetization of AI happens, decisions get made, productivity is unleashed. And back in 2019, I'll be honest, when I would visit investors and talk to them about inference, it was a word that most people didn't even understand. People were like, what is inference?

22:57And now, post-ChatGPT, post-DeepSeek, the whole world is talking about inference. How do we make inference available to all of humanity? Every human on the planet wants to use compute for inference, specifically. Very few people want to train models, but everybody wants to use AI and use inference, right? So it's not really a one-size-fits-all problem. It is not like you can build one chip that serves all of humanity's needs for inference, right? So you really need something that is different strokes for different folks. And when we talk about how we are better at GPUs at certain types of workloads, certain types of applications, you know, the 10x, you know, faster, 3x, you know, energy efficient you know 3x cost efficient uh again the whole premise of the platform that we built at dmatrix was around optimizing for efficiency and when we talk about efficiency it's about doing more with less right and what are the three most critical resources that humankind cares about well it's time it's energy and it's money right so you can find a way to do more compute more influencing compute with less time in less time with less dollars with less energy uh that would be a clear win now we also realized that we really can't do this for every single application out there right because uh you know the gpu today is a more general purpose acceleration engine uh what we were building is a more dedicated accelerator for inference and we started our journey with a focus on what we call specific applications that run on smaller language models that's what we're focused on today and it's really narrowing in on a specific application what we are hearing from most customers is like look you know most applications enterprise applications today run on these smaller sub 100 billion parameter models call it right the frontier models uh that are you know we call it trillion parameter models uh you know more uh you know suited for agi right uh more general purpose ai so there is there is going to be a segment of company that chase AGI frontier models, that is not what the matrix is focused on.

25:01We are very focused on the enterprise class model models. Let me ask you this. So where is the product at now in its development? Have you finished the development of a chip? Have you sold the chip to customers? Yes, yes, yes. All of the above. The product is in customers hands. They are testing the product. It's being piloted. Okay, this is the Corsair chip is the flagship chip you have. So You've sold the Corsair chip to customers. Who's using it? Well, it's a combination. So I think our target customers are hyperscalers, NeoClouds, AI Clouds, Sovereign Clouds. So we have customers across the board.

25:38Right now, we have more customers than we can handle. So we're just really trying to pace this. Okay. Chip deployments take time, right? This is not like you throw something over the fence and the customer can take it and run with it. Right. From the time we give our chip to a customer, it takes them another 12 months to really deploy it in production, right? And there's a lot of support and hand-holding that we have to do. So we really have to make sure that we can only chew as much as we can digest, right? You guys are using TSMC to make the chip? That's right. We use TSMC to make all our chips.

26:08How hard is it to get TSMC to actually make the chip? We've talked to chip companies in the past. Look, capacity is an issue. I mean, everyone wants to use TSMC. There's only a couple of players in the world can actually make it. Any challenges actually getting or finding capacity from their end to make the chip for you? So we, you know, first of all, capacity is a constraint right now for the leading edge process nodes, right? So if you're building your chip on the latest and greatest process technology, the stuff that NVIDIA uses, Apple uses, you know, and other larger semiconductor companies use, absolutely, there is a shortage, right?

26:42But at Dmatrix, the way we built our platform, we built it in a way where the architecture did not really need to take advantage of the latest and greatest process node. So we are typically N-1, maybe N-2 generations in terms of process technology behind what the leading-edge semiconductor companies use. Because our architectural advantages will more than make up for the lag in process technology. That's number one. Number two, we don't use things like HBM memory, which is very, very popular with GPUs. We use other forms of memory technology. We don't use this thing called Silicon Interposer and Covast, which is, again, an extreme shortage and short supply.

27:20We have built the platform. The computing platform has been built with technology that does not really use any of the technology that are in short supply today. Got it. So you've sort of optimized around it. I wonder, you've just finished this funding round. How hard is it to find venture capitalists in this era that actually know chips really well? I can't imagine that every flagship venture firm has the expertise to really know what it takes to make a chip, right? It's a handful of luck. It's really, I mean, I can count them on my fingers. I've been doing this for seven years. There's not been a week in the last seven years I haven't pitched to an investor.

27:52And I can tell you it's a handful of investors who really can diligence deal. So there are some firms that say, like, it's just not our space. We don't know. Wait, it's important. We don't know. Many of them don't go there. Some of them want to go there because of all the, you know, there's just so much noise around chips and, you know, with what's happening with NVIDIA. Many of them want to go there, they want to participate, but they just can't pick the winners from the losers, right? It's very, very hard for them to do that. Right. And last question for you, talking about the winners versus the losers.

28:22Look, you are taking on Jensen Huang. There's no two ways about it. I know that it's a very specific application. How do you wake up, when you wake up in the morning, on a personal level, I mean, what does that feel like to be taking on perhaps the most powerful person in AI right now. Does it not scare you? Does it inspire you? Just walk me through your personal thinking. Now, I've been in the industry for 30 plus years, right? And, you know, I started my career at Intel, you know, and Intel was at the top of the world, right? And everybody was taking on Intel. Then I went, I've been at three startups.