ChatGPT Nears 1B Active Users, Nvidia-Backed Reflection AI Looks to Become Open-Source Champ

0:13Stephanie Palazzolo:Welcome, everyone, to The Information's TI TV. My name is Akash Pasricha. It is Wednesday, July 29th. Today on the show, we have a scoop on ChatGPT's latest user data. We're also going to unpack an in-depth profile that my colleague wrote on Reflexion. AI, an up-and-coming AI lab. And we're going to close out the show with Jitu Patel, the president of Cisco. The company has joined forces with dozens of others on the Open Secure AI Alliance. We'll talk about his views on AI safety and open source AI all together. It's going to be a great show, so let's get right on into it. The information has exclusive reporting on ChatGPT's latest user data.

0:55Stephanie Palazzolo:I want to bring on Stephanie Palazzolo, author of our AI Agenda newsletter, to walk us through what she found. Stephanie, welcome back to the show. It's great to have you here. Great to be on. Okay, how many active users does ChatGPT currently have? So as my colleague Erin and I reported last night, ChatGPT is right on the verge of hitting 1 billion weekly active users. So again, this is the number of users that are using the popular chatbot at least once a week. The reason why it's a little bit hazy is because, you know, typically whenever companies are keeping track of these numbers, they do have to do some kind of last minute counting for like bot accounts and things like that.

1:38But it does seem if they haven't hit it already, it's, you know, on track to hit that number in the next couple days.

1:45Stephanie Palazzolo:So when were they expecting to hit that initially? So, you know, obviously that's kind of the important point here. One billion is a super impressive number. That's a huge proportion of the world's population that is logging into ChatGPT and using it at least once a week. But the important thing is that the company originally was hoping to hit this milestone at the end of last year or by the end of last year. So, you know, it does come around seven months after the company was originally hoping to hit that usage number. And so what's been the issue? Do we know? So I think there's a couple things that are coming into play here.

2:23I think first, as Sam Altman told staff at the end of 2025, at this point, Google was really starting to do well. Gemini was really catching on with developers. Sam told staff, we're going to expect kind of rough vibes.

2:42Stephanie Palazzolo:Yeah, code red. I mean, you broke that story. Code red, things are not looking great for ChatGPT, and then what? Exactly. So, you know, lots of worries about Google. Then since then, Google maybe hasn't been as much of a threat, although definitely is obviously a very big competitor. People are using its AI tools more than ever through Google search AI mode. But since then, you know, Anthropic has really caught on. So its coding tools like Cloud Code are doing incredibly well. Lots of people are using, you know, Anthropic's Cloud Chatbot and its other AI tools. And so, you know, I think both of those, you know, just basically rising competition, as well as a kind of slowdown that we've previously written about that happened after OpenAI's GPT-5 launch last year.

3:29These are all contributing factors to why it's been maybe harder than expected for OpenAI to pass out last, you know, 100 million or so in users and why it took them, you know, much longer than expected to.

3:42Stephanie Palazzolo:What about paying users? I mean, nearly a billion in weekly active users. We know that you can log on for free and use it. What proportion of that are people that are paying for the platform right now? And how does that compare to the proportion of paying users for all these other platforms as well? So it's a little hard to say because obviously a lot of this data is kept very much under wraps by all these startups. But, you know, the last figure that we've written about is that ChatGPT, there's more than 50 million people that are paying to use the chat bot. So, you know, relatively like a small proportion of people are paying, which, you know, is maybe not great for OpenAI.

4:23I'm sure they'd love to increase that proportion. However, it's also important to point out that they have other ways to monetize non-paying users like ads, for instance, which is something that they have been trying to push more into in recent months. It does seem like, you know, from our reporting that this does seem to be kind of a higher proportion of paid users than some other AI tools, though.

4:47Stephanie Palazzolo:Yeah. So, I mean, look, ChatGBT is certainly one business vertical that the company has. We also know that Codex has been gaining steam. And so I wonder if you've heard of the company really shifting its priorities at all in terms of prioritizing growth in other areas where chat GPT growth might end up. And I don't want to say stalling. I mean, look, it's still growing fast. I mean, look, any platform to have, you know, one eighth of the world's population, I don't know how many people are they can expect. But have there been growth in other areas of the business to compensate? Yeah. Yeah, so this has been a really big theme for OpenAI in the last year or so.

5:31So, you know, as Chat2BT growth has kind of slowed and tapered off a bit, we have seen the company really double down more on selling AI to enterprises, as well as, you know, tools for coding and knowledge work, things like Codex and more recently Chat2BT work. So you can kind of see them trying their hand at the Anthropic playbook here and being much more focused these days on selling AI to enterprises because that's an area that maybe they haven't tapped into as much. And obviously enterprises are maybe much more willing to pay for AI tools than the average consumer is. So, so far, we have done some reporting on this.

6:13It seems to be going well. I have heard, I have talked to a decent number of developers who have switched over, for instance, from Cloud Code to Codex in recent months as OpenAI's models have gotten better, and especially as they have proven to be, you know, in many cases, a lot cheaper than Anthropix models. but at the same time it still is very much a super tight race and you know right now it is very much neck and neck and it's not clear if open ai is going to be able to you know match or surpass anthropic when it comes to kind of ai for enterprises and remind us i mean sarah

6:50Stephanie Palazzolo:fryer had some comments about this earlier this year where she sees the future of the business in terms of revenue breakdowns. She, of course, is the CFO of OpenAI. What did she say with her expectations there? So earlier this year in January, Sarah Fryer did say that around 40 % of the company's revenue was coming from businesses, and she expected that proportion to grow to 50 % by the end of the year. So that is obviously a very significant chunk of the business that is coming from enterprises, which I think even in January, that was somewhat surprising for people who had mostly seen OpenAI as this, you know, very consumer-focused business.

7:32But I think at the same time...

7:34Stephanie Palazzolo:And it kind of goes to the point that you were saying earlier, which is that, you know, although they've fallen short of this metric, which is very much, it's a consumer-focused metric, maybe inside OpenAI, I mean, you know, where they started with Code Red, maybe it doesn't actually matter to them as much because at the end of the day, I mean, it's businesses that that's going to help get them to profitability. And you shouldn't ignore it, but I'm just saying in terms of converting it to revenue, you know, if it's only 5 % of people who are paying for it, probably not the highest ROI to focus on that necessarily.

8:09Totally. I think that's super fair. And we do see, you know, OpenAI even now moving away from this kind of just like kind of simpler chatbot product where you're just going back and forth with a chatbot and more into things like chat GPT work and codecs, where you see their products, you know, basically completing actions on the user's behalf, which is very, you know, attractive to enterprises who want to like automate things like payroll or sending emails or scheduling, right? So all these things are definitely much more targeted at the enterprise, you know, side of the business.

8:46Stephanie Palazzolo:Right. Steph, before we let you go, I want to get into very quickly another story that you published with amira fradi our co-executive editor uh you did some analysis today on the uh the stripe open router acquisition that everyone is expecting to happen uh we of course have reported that 10 billion dollars is the price tag that they are considering you took a look at the valuation and what that means in terms of multiples uh what did you find there so yeah i think this is one of the most potentially expensive acquisitions that we've seen in quite a while. So as my colleague Amir and I reported earlier this morning, OpenRouter, which is a startup that basically helps developers choose between a bunch of different open and closed source models, was recently generating around$140 million in annualized sales.

9:43So, you know, obviously that's not nothing, but it's a far cry from, you know, the billions of dollars that we see other, you know, application and infrastructure companies making. I think what makes that really surprising though is comparing that$140 million number to the, you know, around, you know,$10 billion price tag that we've reported on. So comparing those two numbers, that means that if the deal goes through, Stripe is going to be paying more than 70 times open routers revenue, which is quite expensive, especially whenever you compare that to other recent deals like SpaceX and Cursor, where Cursor was not only making significantly more revenue, but when we compare to the price tag, that ended up being more around a 22 times revenue multiple, you know, a lot lower than what's like.

10:36Stephanie Palazzolo:Nearly three times as much as Cursor on a multiples basis, which is kind of crazy, especially when you consider that the Cursor deal was, that's, you know,$60 billion was a big check to begin with there. What do you think is accounting for the premium here? What does Stripe see in Open Router? So I think there's one element of this, which is just that, you know, Are you going to say FOMO? There's probably a bit of FOMO. I think, you know, as we reported, it does seem like there were a lot of buyers around the table. You know, Databricks is a name that we have written about. It does seem like there were at least a couple others.

11:13So definitely there can be a situation where, you know, with M &A, I don't think anything, I don't think everything is very rational. It's like, you know, we can attribute this part of the price to like this value that we're getting. I think companies and buyers are bidding each other up. So I think that could definitely be an element that was happening here. I think on the other hand, Stripe obviously is involved in payments. It has a good kind of insight into what its customers are paying for AI models. And I think OpenRouter could be a nice complement to that business where not only are they handling like payments and seeing, you know, that data on what businesses are spending on AI models.

11:55But now they can also get into the business of like helping those customers choose which AI models they want to use when. And I think it could be, you know, a nice way for Stripe to add this kind of AI flair to its narrative, which I think every company wants to have these days. And as some of my other colleagues have written about, it doesn't like they have a pretty hefty, you know, war chest of cash. And, you know, they need somewhere to kind of put that cash. And, you know, M &A is always a good place to do that.

12:25Stephanie Palazzolo:So can I just ask you one question, though? I mean, you wrote about the risks to OpenRouter's growth and the sustainability of the revenue in the story. And, you know, I want you to sort of share with us what those risks are. But more importantly, if Stripe buys OpenRouter, then it really becomes a game of pointing that technology to all of Stripe's customers. And so after explaining just the risks, I want you to then opine on whether or not you think the Stripe acquisitions could sort of nullify those risks or maybe mitigate them in sort of a joint union of sorts. Yeah. So I think in terms of the risks, I mean, there's a couple of things here.

13:02I think first, there's been a lot of skepticism around businesses like OpenRouter, which are essentially, you know, helping developers choose between existing open source models and, you know, existing open and closed source models and providing this kind of centralized hub that makes it easier for them to pick and choose and do that all in kind of one place. I think some people would argue, though, that, you know, how much value really is a company like Open Router adding to that process? And at the end of the day, some people might even describe them as just kind of reselling other companies' APIs.

13:39So I think initially it's kind of maybe not entirely clear what the value add is there. I think another issue too is that, you know, from our reporting, a lot of customers are using OpenRouter to access, for instance, closed source Anthropic models that are out there. And, you know, Anthropic has definitely not been shy about, you know, cutting off access to its models to companies that it feels like maybe are somehow using those models in an unfair way.

14:06Stephanie Palazzolo:um so or pointing them elsewhere basically saying hey like we're giving you access to this model but but open router is saying it's actually not the model you should use why would anthropic even allow access then in that case yeah and you know anthropic can also say hey why why don't these customers just come directly to us you know why are they going through this like weird third party that is reselling our models right like right like it's kind of just leaving money on the table here so in that case i do think like stripe is not going to be much help there because no not at all i mean these these are still like you know there's the business of routing is doesn't really matter how big and in fact i mean you could you can even argue the bigger the bigger the owner of the company and i'm just brainstorming here i mean stripe is in the payments category anthropic doesn't really have much in the payments layer but but to the extent that they ever start to compete on anything, the risk seems even more likely in that case.

15:05Yeah, I think that's definitely true. So I do agree with you that I think Stripe will provide a lot of distribution benefits. And obviously, Stripe has a huge base of customers, which I'm sure OpenRouter would love to have access to, but definitely has yet to be seen whether they're going to be able to help with any of those risks. It seems like maybe not with at least some of those issues.

15:26Stephanie Palazzolo:But the bankers love it. The bankers do love a good$10 billion deal. Yeah, I know. It was pretty crazy to see the kind of discourse after we first broke that news. And just because, again, the company was last valued at around a billion. And so that's the outcome for the company's founders, for its employees, and definitely for the VCs that backed it at a much lower price. So firms like Andreessen and Menlo Ventures that got in way earlier. Right. Great. Well, Stephanie, I want to thank you for coming on. That is Stephanie Palazzolo, author of our AI Agenda newsletter here on TITV. reflection ai is an up-and-coming ai lab with a lot of momentum it has raised billions of dollars from investors 800 million dollars from nvidia alone it is just one of eight ai companies assigned a deal with the defense department joining google microsoft and aws and yet it hasn't yet released a model to the public our ai and robotics reporter rocket drew profiled the company in a story out today i want to bring on rocket to share more with us about what he found Rocket, welcome back to the show.

16:32Stephanie Palazzolo:It's great to have you here. Thanks, Akash. Great to be here. Why did you decide to laser in on Reflection specifically? Yeah, I mean, it's exactly what you just said. There were all of these buzzy announcements that mentioned the company, including ones from the Department of Defense, where Reflection is standing shoulder to shoulder with these giants, these heavyweights in AI, but Reflection is comparatively not well-known at all. So that got us thinking, you know, what is it about this startup that's getting them so much traction? They're signing these big deals and they have these big ambitions for themselves.

17:03We wanted to know a little bit more about the company.

17:04Stephanie Palazzolo:And I want to get into the backstory of the company, but one of the most surprising things in your reporting that shocked me was you found that, I mean, the model that they're developing, it's expected to be quite good, right? Yeah, that's right. But good relative to the open weight models that are available in the US. So it is still the case that Chinese models are pretty far ahead of the open weight model that are available from developers in the US. At the time Reflection comes out with its first model, its model will likely be on par for Western models, but not for Chinese models. Of course, they hope to exceed Chinese models in the future.

17:45Western open source models, to be clear. That's right. That's right. Western open source models, to be clear. So not the models you're seeing from OpenAI and Anthropic, but instead from companies like Mistral, from Thinking Machines Labs, which got into the game recently with their pretty large model inkling and then also smaller models that have come out recently like one from poolside they have a model called laguna that's specialized for coding so the point is that while anthropic has these ambitions and they are tracking to be on the western frontier every day that goes by it feels like there's more competition there are more models out there that are available even from western open source companies so reflection has better come out with a model soon if that's, you know, they're still tracking to be on the frontier.

18:29Stephanie Palazzolo:Right. I want you to walk us through the history of this company. So the co-founders, Misha Laskin and Yanis Antonoglu. I think I pronounced that right, hopefully. Yanis, please correct me. Come on the show and correct me, if you will. Tell me about these founders and why they are so popular in the AI ecosystem right now. Yeah, absolutely. They definitely bring with them a certain AI pedigree, especially because they both worked at Google DeepMind and they got a lot of experience there. So Giannis was one of the people who worked on AlphaGo, which was the pioneering reinforcement learning system.

19:08They taught an AI to play Go, which was once thought to be a game too complex for any machine to ever play, let alone master. And then in 2016, it dethroned the reigning world champion. It was a coup really for AI progress. It blew people away. So that's what Giannis brings to the table. And that experience he has with reinforcement learning is really attractive for reflection, bringing in new talent. You know, understandably, people are excited to work with him. Misha also worked at DeepMind and cut his teeth working on the earliest versions of the Gemini models. So he brings in some more of that experience with large language models.

19:46So together, they have some star power on the founding team for sure. And what was the original ambition for the company when they started it? Well, I think at the very beginning, maybe they didn't really know. Talking to some people who were working with them in the early days, they were trying a lot of things. Smart people will do smart things. Let's give them money. Right. The classic Neolab proposition. And in fact, I think that's one of the things that's interesting about them is there's been kind of this moment for Neolabs lately, but in some ways, Reflection was early to this game, realizing that there's probably room to get in and compete against these larger players like OpenAI and Anthropic.

20:24But to begin with, Reflection saw its angle as coding. I mean, if you'll remember a couple years ago, Cursor was on the rise. Coding was such a hot topic already. Their first product that they ever released was an agent for understanding existing code. So not for writing code, but just for understanding code. But that was not a model that they trained themselves in-house soup to nuts, right? That's sort of an entirely different proposition. So what happened was the company then made a dramatic pivot and decided that they were going to train their own models from scratch and they were going to open source those models.

21:00And in fact, they were going to be the champion of open source in the US. And that pivot happened after Misha met with NVIDIA CEO Jensen Huang a year ago out here in California. So Jensen floated the idea and then they made this big pivot essentially overnight and changed the strategy of the company. A year later, the model still hasn't come out. So that's what we're waiting on now.

21:24Stephanie Palazzolo:Do we have any sense for what Jensen said in this meeting to convince Misha on what frankly is now, I mean, as of today at least, it's, you know, one of the most competitive battlegrounds for models is the open source model category. Do we know what Jensen said to convince him to change the whole strategy? It is very competitive. Now, we don't know for sure, but what you should remember is that it's absolutely in NVIDIA's interest to have more players in this category, right? The more companies that are out there using AI, the wider variety of companies, the more revenue they can bring in for their AI chips.

21:59And also, it helps diversify their revenue away from these large, concentrated players like OpenAI and Anthropic, AI companies that are increasingly developing their own chips as alternatives to NVIDIA's. So you can see why it would be in NVIDIA's interest to have a player-like reflection out there. You also have to remember that a year ago, the territory did seem more open. You know, like DeepSeek had happened, but there weren't a ton of players yet, even on the China side, that were competing at that level in open source AI. So it still seemed, I think, more like it was anyone's for the taking.

22:35Now, to be sure, there's still a lot of room for reflection to get in and compete in this market. Right. It's been working on the model for a year. It could release a frontier model. And, you know, models are, they're not all the same. It could find a niche that works for it very well. But I think the proposition was probably something along those lines, that there's a lot of room for an open source US champion. And if that company has our backing, has the compute that it needs from NVIDIA, they're basically going to be set.

23:02Stephanie Palazzolo:But, I mean, let's get to the heart of it. I mean, they've raised so much money. They haven't yet been able to release. And I should say, you know, I'm saying they haven't been able to. It might be likely as a decision, right? I mean, we've talked on this show about timing, too. It matters. If Kimi K3 just came out, people are raving about it. You better make sure that it's maybe not competitive with Kimi K3, but at least competitive with the Western open source model. So maybe it's a strategic decision to not release it yet. But, I mean, what do we know about what's happening inside the company?