Oracle’s Shockingly Small Cloud Margins, OpenAI ‘Super App’ Goal, AI Copyright | Oct 7, 2025

0:13Welcome, everyone, to the Informations TI TV. My name is Akash Pasricha. It is Tuesday, October 7th. We have got a great show planned for you today. We are breaking down everything we learned yesterday from OpenAI's Developer Day with an expert panel. We are also talking about a big scoop that we have about the gross margins behind Oracle's AI cloud business. We're then pivoting to talk about AI copyright with the CEO of Runway. And we are also talking with Whoop about the fitness tracking company's newest product. And finally, we have got our very own Elon Musk reporter coming on to talk about the challenges facing Tesla's big robot ambitions for optimists.

0:53We've got a lot going on, so let's get right into our first guest today, folks. OpenAI put on a big show for its demo day yesterday, unveiling new ways it is vying to be a super app, among other new tools that it launched. It was an event that all of Silicon Valley was closely watching, given the power that OpenAI has amassed in the technology sector. And so to break it all down, I want to bring on an all-star panel. Stephanie Palazzolo is our AI reporter at The Information, and Matt Schumer is the CEO of OtherSideAI. Welcome to you both. It's great to have you. Thanks. Great to be here. All right.

1:28So, Steph, let's start with you. What was the headline coming out of yesterday's Developer Day for you? Yeah, I think for us, the biggest headline is OpenAI continuing to kind of try to make Chat2BT this super app where you're not just going this to help you write emails for work or to write funny little stories for, but you're going to Chat2BT to look at new houses you might want to buy using Zillow or ordering your groceries or going shopping. And so that was really seen with their new apps in ChatGPT launch, which basically lets you access a lot of external apps, like Zillow, Spotify, DoorDash.

2:11There's a number of them through ChatGPT. Matt, that was Steph's headline. What was your headline? Pretty similar. I think a lot of what was announced is technically for developers, and there is a lot of developer-specific stuff, but it really is going to enable new use cases, new ways of using this technology for consumers. Some of it probably will work, Some of it probably won't work as we've seen in the past. I think that pattern is going to happen again. It's exciting. I'm excited to see what actually sticks. And Steph, one of the things you talked about in your AI agenda newsletter this morning was that we have seen flavors of this in the past.

2:44I mean, this is not a net new idea for OpenAI. Yeah, totally. I mean, I think that was one of the most interesting takeaways from yesterday. I think sometimes you go into these events expecting, you know, huge like tech breakthroughs or crazy announcements. But a lot of this did feel similar to some other announcements we've heard in the past. For instance, apps in ChatGPT, there's some similarities with ChatGPT plugins, which a company actually announced in March 2023. And we also saw other similarities between some of their announcements around agents and making agents easier to build. This is a thing that they've been talking about for a very long time.

3:22And so Xander's India yesterday kind of felt like OpenAI basically saying, hey, we've been talking about these ideas like agents and ChatGPT as a super app for a very long time. Like this year, let's try to make that a reality. Right. So Matt, yesterday we were all watching the demos that they had and the agent tools, the ability for developers to now make their own agents much more easily. I mean, this sparked a whole parade of tweets and our posts on X saying, hey, you know, OpenAI just killed all of these startups that were trying to make agents a lot easier. What's your take on that? I think it's a very cool product.

4:03I think it shows what's possible. But I don't think it kills the agent startups. If anything, it's going to boost them. Right. The agent startups have sort of a niche where they sit between providers. OpenAI just serves OpenAI models. You can email other models in their platform. But when you're actually deploying with AgentKit, from what it seems like, you can just use OpenAI models. I've been building agents. I know many other people have been building agents for years now, I guess decades in AI time. And very few people that have built agents at a high level are using just OpenAI models or just Anthropik models or just Google models.

4:39It's usually a mix using the best model for what it needs to be used for, for each specific thing. Right. You don't want to use OpenAI for everything because OpenAI models aren't the strongest in everything. right there's sort of like difference in cost difference in capabilities difference in just their strong suits so i think some companies will use this because they don't really care about the difference you know they don't need the frontier performance for everything but when you're talking about this killing all these agent startups i i don't think so i think if anything it's going to boost them because you need to have startups that allow you to use you know open eye anthropic google but you don't think open eye will open things up at all later on to allow people to use other models at all?

5:16I don't see the incentive, which makes me think it's not going to happen. And there hasn't been much that we've seen from them that would indicate that that would be the case. Maybe I'm wrong, but that's why companies like OpenRouter exist. Okay, Steph, do the people in your orbit agree? I mean, do the agent startups still have a chance here? No, I think overall, I definitely agree with Matt. I mean, I think a lot of these agent announcements are in some ways targeted at non-developers or people that are maybe not super technical. We saw this kind of very easy to use like drag and drop interface where without typing a line of code, you're able to make these agents work.

5:56And so I think that is a little bit more targeted at business users or just typical everyday consumers that want to make agents, not necessarily the really technical developers that might be using some of these other agent tools that Matt was just talking about. Right. Matt, I want to ask you, let's go back to the super app part of this, because that is, I think, one of the most interesting parts of yesterday's announcement. I'm trying to think about it from the perspective of the companies that are allowing their apps to now be accessed through ChatGPT. And our co-executive editor, Martin Pierce, made a good point in his newsletter last night about what is the incentive for something like a Zillow to participate in this.

6:39And, you know, I'm sort of thinking about the same thing, which is that, hey, you know, it allows people, a wider audience to access my tool. I get that. But you're sort of losing control over the entry point as to how people actually start navigating your software. What is the incentive for them? I mean, do we know anything about the economics here about is Is OpenAI getting a cut of the subscriptions? What do you think of all that? I can't personally speak to the economics. I'm sure there is some information out there and I'm sure we'll learn more. The way that I think about it is there are going to be two phases.

7:11The first phase is proving out if this is even a thing, right? We saw GPTs. We saw chat GPT plugins, right? Neither of those really took off. This may take off. This may not. In the initial phase, it's going to be about distribution. Can this get more users to my app? Developers are going to explore. Can we find new interfaces that work better in this way? Then the question is, does this actually stick? Are people using this? Is this getting distribution to the news? If it keeps going and keeps going and keeps going, eventually this is going to be the common way in what people access apps. I don't know if that's going to be the case.

7:42I'm actually somewhat skeptical. Let's say we actually do get there and this is the way that people access apps. But skeptical why though? Skeptical why? I don't know. The interface doesn't feel as perfect as it needs to feel for something like this to truly work. This is threading a very, very tricky needle. and I think it needs to be just utterly incredible. Like so much better than anything else. So you said the UI for you wasn't, I mean, you saw the demos yesterday. It's not there for you. It didn't blow me away. It didn't feel like it was perfect. Okay. And maybe it's good enough. But my feeling is once something like this works, if we get to that point, at that point it's going to be like, you know, not having an iPhone app.

8:19You need to do it. You need to do that to keep your business going because that's where the people are. But we're not there yet. The first phase is just distribution and then does this thing even work. And Steph, what's your take on the incentives piece? I mean, these companies, what is their incentive to participate in this other than, hey, I am attaching myself to ChatGPT, why not? Yeah, I mean, I think Matt just kind of started to touch on it at the very end. I think there is a sense for some of these companies of, you know, what else am I supposed to do, right? who are going to TACPVP to make playlists on Spotify or to look up information about local restaurants and to book dinner reservations.

8:54Like you either meet them where they are or one of your competitors does. And I think companies would rather like be in the conversation than, you know, let one of their rivals take that opportunity away from them. Right. I do agree with Matt where, you know, it's a bit early to really say for sure. So for them, I think this is a way for them to get in the conversation without, you know, immediately having to deal with potential bad consequences of, you know, losing advertising. Right, right. And I mean, look, the economics of this is certainly what's most interesting to us and what we know least about.

9:28Matt, I want to finish with you. There were so much other things that were announced yesterday, not the least of which were new APIs. You had the new GPT-5 Pro model that they launched. Just give us your sort of highlight reel here. What were your big takes on some of those other announcements? I know we could spend three hours on this. Yeah, I'll run down the list. But give us your hot takes about the rest of it. Voice One's pretty simple. It's a nice upgrade. It's not a game changer. It's slightly cheaper. It's going to be a slightly better model. That's great, but it's not going to enable any new use cases.

10:01Sora is quite interesting. Having that in API is fantastic. There are going to be a new breed of video gen wrapper startups that pop up that will find opportunities. Whether they last, that's an open question. but it's definitely going to be interesting to watch. And the pricing is actually really good for what it is. It's in many ways cheaper than VO for a better model. Not the case for the Pro model, but for the regular model at 720p it is. That's going to open up some new use cases. GPT-5 Pro in the API is very interesting. Most people, this won't matter for. This won't be in the consumer app.

10:32Because it's expensive. It's expensive. It's super expensive. Yeah. But for the blank check, you know, enterprise buyers that will pay anything for something better, it's fantastic. And I'm going to be using it sometimes to get better outputs. I love the Pro model. It's fantastic. The sleeper of the day actually is GPT-5 fine tuning in private beta. I haven't seen much on this, but that is going to be super exciting. When you get a GPT-5 class model and you can actually fine tune on top of it using reinforcement fine tuning or similar techniques, that is powerful. That being said, as I always say, your GPT-6 model will be a fine tuned GPT-5.

11:06but we have some time before that. So that's going to enable some very interesting applications at lower cost. Great. Well, Matt and Stephanie, thank you so much for coming on. It was a jam-packed day and we look forward to having you on again as we see some of these tools start to get used and what more and more people are thinking about it. Thank you again for coming on. That is Matt Schumer from OtherSide AI and Stephanie Palazzolo, our AI reporter at The Information. Oracle has turned into one of the hottest stocks of 2025 after executives projected$381 billion in revenue that will come from renting out cloud servers to AI giants like OpenAI over the next five years.

11:42But today, the information published a story that has some pretty staggering findings about the profitability margins that cloud business has. I should say that Oracle stock is sliding this morning on the news. And so I'm very excited to have two of the reporters on who wrote that story. Anissa Gardizi is our cloud and compute reporter here at The Information. And Anita Ramaswamy is our financial analysis columnist. Big news today, folks. Oh my goodness. Where do we start? Anissa, tell us what you learned. Hey, Akash. So yeah, we had this story this morning that hopefully will help readers really understand the profitability around Oracle's new GPU cloud business, which is really where executives have said the future growth is going.

12:27And what we found is that in the last couple of quarters, the gross profit margin on this business fluctuated between 8%, 20%, and 14%. And we're sort of shedding some new light on what it's really like for the companies pivoting to this business. Software companies are used to much higher margins, but as Oracle makes the pivot into GPU cloud, investors should expect lower and potentially lumpy margins, at least as they get going. Right. And we should make clear, I mean, for people who don't know the margins landscape that well, this is low. Okay. Software margins are typically 70, 80%. I mean, these gross margins are, are, I mean, you know, they are razor thin as we outlined in the story.

13:13And he said, do we know anything about why these margins are so thin right now? Was this expected? Yeah. I think one thing I'd point out is that, you know, we have oracles, but, you know, So the other cloud providers don't actually publicly release these numbers. And so it's very hard to find apples to apples comparisons. Companies like Amazon, which traditionally has a net grip margin in the 30s, that's across its whole entire cloud business, inclusive of GPUs, but also more traditional workloads and other services. So we have the benefit of seeing Oracle's numbers, but not exactly the full picture on what the entire GPU landscape looks like.

13:55I think investors thought that, you know, GPUs were going to be lower margin for Oracle, but based on our conversation, these were lower than expected. Okay, so Anita, talk to us about why these margins are so low. What are they spending money right now on? I mean, I know the chips are expensive. Yeah, absolutely. So the chips, like you said, are the biggest cost, and the data center is really expensive in general. There's not only the chips from NVIDIA that Oracle has to buy, but there's also all of the networking equipment, there's servers, There's a lot of stuff that goes into the data center in terms of hardware.

14:27And the interesting thing about this is that it suggests to me that Oracle is actually pricing their offering really low. And that's why they're incurring a low margin on this business. They have signed up a bunch of customers with discounts. They've been offering their GPU rentals at a lower cost typically than other providers. And that's a big factor as to why the margins are as low as they are. But like Anissa said, we don't really know exactly the apples to apples comparison here because we don't have Google Cloud financial results broken out. AWS doesn't report their gross profit margin separately.

14:58So we can only really look at the operating margins. Another example here is CoreWeave, which is kind of an interesting one. And so I think as we see more and more of these GPU cloud firms that are doing server rentals come out and go public and there's more reporting, hopefully, from our newsroom about their financials. will have more insight into what is driving the pricing discrepancies. But, Anita, I kind of want to get into some of the line items here. You know, you're our financial analysis columnist. You study these statements all day, right? I mean, you have these expensive chips. One aspect is the depreciation expense on how quickly these chips do depreciate, how much that expense actually is.

15:37Is that in their control? I mean, is this something we can expect to get better? How do we think about how this number might improve over time? Yeah, it's a great question, Akash. So the depreciation for the chips is probably the largest category here. There's depreciation involved with all of the equipment in the data center. And the depreciation schedule in terms of how long you take a time period over, I suppose, is one form of control that companies have. But it's really only some extent of control. Because when you buy an asset, you take it, and then you kind of split the useful life over a certain number of years.

16:08and you're incurring that cost in different years. And so I think, you know, we might be able to see these costs come down over time, but it's really in the hands of NVIDIA. And that's what it's going to come down to. So some of the older chips, actually, Oracle is able to incur a higher margin on those because the older chips aren't as costly. But, you know, with some of the newer generations of chips, I just think if Oracle and if other cloud providers want to stay at the cutting edge, they probably will have to take this margin hit. Right. Anissa, I want to come to you to close this out. But does Oracle really have a choice here?

16:39I mean, it has to compete. We know the direction this technology is going. Is there an alternative or do they kind of just have to accept it? I think executives are pretty certain that they want to pivot into this business. And they're just sort of getting people used to what profitability might look like. And I think when Oracle is negotiating a massive multi-year deal with a company like OpenAI, you know, they're going to, you know, maybe lower the price to get that big deal because overall, that's going to generate them lots of revenue over those years, you know, more revenue than they would generate without the AI data center business.

17:19So I think partly, you know, they're just, you know, pivoting into this space and this is what Marge is going to look like. But I think they're going to be working pretty hard to improve them. And one thing that we noticed was that as NVIDIA's latest chips come out and Oracle installs them in their data centers, there's sort of a gap and a ramp up period where, you know, it takes customers a little bit of time to actually start using those chips. So my guess is that that's going to be a key area that they try to shorten so that they can, you know, improve these numbers. Right. And Anita, I should say one thing that you mentioned in the story, you mentioned some of the older chips that Oracle can use to sort of boost this profitability.

18:00This was something that you pointed out in the story. It kind of goes against what Jensen Huang has sort of thought about these chips and the idea that the newer chips will sort of make the older chips obsolete. Clearly, Oracle is saying, well, we can still get some usefulness out of the old ones. And in fact, we have to because otherwise our gross margins might be negative in some cases. I mean, they're still selling. Yeah, yeah. Great. Well, Anissa and Anita, thank you so much for coming on the show. It was a great story. And like we said, shares are sliding on the report. And so we look forward to having you on more to discuss how this story develops.

18:34That is Anissa and Anita from our newsroom here at The Information. Copyright has been a hot topic in AI. And now that AI-generated multimedia content has become the focus of even the biggest companies like OpenAI and Google, copyright is becoming all the more important to think about carefully. Today, Cristobal Valenzuela, the CEO of AI Video company Runway published an exclusive op-ed in the information about this topic, arguing that policymakers should treat AI-generated art just like they do any other art. And so I want to bring on Chris to talk more about what he meant in that piece. Chris, welcome back to the show.

19:11It's great to have you. Hey, thank you for having me. So what a great piece you published today in the information. I do want to get to what the crux of your argument is, but I thought maybe a good place to start is give us sort of copyright 101. I mean, how do copyright standards work right now for traditional artists, talking about musicians, filmmakers? Let's put AI aside for a second. We'll get there. How does it work generally? Sure, yeah. Look, it's based on this idea of human authorship, right? So if you as a human and an author can create and can express and can prove that what you've created was inherently human, and then you can copyright it.

19:52And I think that the medium on which it happens, it's kind of like a second aspect of that creation process. And so for photography, for jazz, for any form of expression, that's how we've evaluated and thought about copyright forever. And so what we're kind of advocating here, and we'll go deeper into it, is basically AI doesn't change that formula, and it should be holding the same standard as any other art form, basically. Got it. And so you're saying that, hey, copyright standards, the way they've traditionally worked is that you have to be able to prove that you made it. And the argument that you're making is that it doesn't really matter what technology you use to make it.

20:32This is just the latest tool, essentially. Correct, correct. And I think part of it is really understanding what AI is at its core. I think a lot of the assumptions and decisions as to how we interpret AI when it comes to copyright have been based on what we've seen AI do over the last, I would say, three years, two years. Now, it's really important to understand that when you think about media and AI, think about video generated with AI models and images and music and multimedia formats, you're seeing highly creative forms of expression, right? These are not machines that just create on their own.

21:03They're people behind the systems. The systems are there for advocating for artists. They're there to augment a specific creative process, but the ones initializing, starting that creative process are artists, first and foremost. And so if artists are the ones communicating via this medium, then artists should hold their rights to anything they create. Right. So I do want to get into some of the nuances here, but very quickly, I mean, you have tons of users using your platform right now, using AI to generate multimedia content. Has this been a big issue for them? It has. I think it's important to recognize as well that it's going to expand as use cases and usage continues to grow.

21:41We have tens of millions of users. We expect to get to hundreds of millions of users. I think broadly, AI video will become probably the most important modality of AI over the next couple of months. I mean, you're seeing this. Everyone is really betting a lot in video as both a reasoning kind of engine, but also as an art and a creative kind of system. And so we should expect even more pressure on how we understand and think about copywriting the age of AI. And so really, we want to make sure we can help guide and understand and have a bit of nuance in terms of thinking of this technology as a very powerful medium of expression and make sure that discussions don't get lost in simplifications or generalizations.

22:23So the interesting part here is you're talking about the content that's the output of people using AI to create this art. The question I have for you is what about all the data and all the content that goes into training these models? I mean, how do you sort of square your argument that, hey, you should be able to copyright the output with the fact that there's a lot of, you know, potentially copyrighted work that is being used? You know, I'm not talking about your company specifically, but generally speaking, right, there is a lot of this content that's used to train these models in the first place, right?

22:54Yeah, so those are two different, totally different conflict questions. One has to do with the argument of free use that I think the courts are going to be basically deciding very soon. I think there's arguments on both sides. I'm very optimistic on the for use because I think this is transformative at its core. What people are using and creating with AI is totally new and different from anything we've seen in the past. And these models are learning patterns around the world. Now, on the copyright, on the output side, that's the question I think that's most important. And I'll give you basically a good example of how this conversation has moved now and the focus has moved to output.

23:26The studios that we work with, the production companies that we work with, the Hollywood companies that we collaborate with, the argument where they care the most about the outputs. It's about how do you make sure that you can make something and then have the rights to either copyright it or use it in perpetuity. And that's really important. And that has no change whatsoever with the technology itself. It's just you should be the one owning that because you as a human created it. And one question I did want to ask is you talked a lot about video content that is the business that you're in in your piece.

23:56What about stuff like, you know, I don't know, spoken word poetry that you can just use an LLM for, you know, with ChatGPT? I mean, does your argument extend to that and saying, hey, I use this tool to write a poem, but I want to copyright that? Like, is that as far as we're going with it? Yeah, you know, that's a very interesting question. In a way, we have artists using chance. Surrealism has been experimenting with non-conscious ways of creating for years. And so if you take away the kind of like method of creation and just focus on what was the intention, what was the control of the person or the human created it, you can still create it.