OpenAI Astra Sparks AGI Debate, Anthropic Pushes Into Payments Tech & How AI Threats Shift Corporate Budgets

0:13Stephanie Palazzolo:Welcome everyone to the Informations TI TV. My name is Akash Pasricha. It is Tuesday, September 8th. Today on the show, we are going to talk about the initial reactions to OpenAI's Astra model. We also have some exclusive reporting for you on Anthropics to push into payment-related tools. We'll then discuss our reporting on a joint venture that Code2 is talking about with AI chip startup Maddox. We're also looking at how escalating cyber threats from AI are forcing big companies to rethink where they spend on cybersecurity. And to close out the show, we're going to dive into how AWS is drawing spend away from Microsoft Azure following its overhaul of its own AI services and Amazon Bedrock.

1:00Stephanie Palazzolo:It's going to be a great show, so let's get right on into it. OpenAI's newest Astra model has been out for a couple days now. There has been a lot of chatter about whether or not this model is AGI. I want to bring on Stephanie Palazzolo, who authors our AI Agenda newsletter for a conversation about that. Stephanie, welcome back to the show. It's great to have you here. Great to be here. Okay, so we haven't asked you yet about Astra and how people are finding it. What are the initial reactions? Yeah, so, you know, initial reactions, people are definitely very impressed. I think Astra, you know, OpenAI spoke a lot about how the jump between the last model, GPT 5.6, Sol, and Astra was going to be quite a big one.

1:47I think that's definitely true. So, you know, Astra, I think, really shines in areas like, you know, 3D modeling and software engineering. So I spoke with some people that were early testers of it, and they said, you know, one difference that some of them noticed, I think, between this one and Fable 3.1, which is Anthropik's latest model, is that they said that with Fable, it was better whenever maybe you didn't have like a super detailed idea of what you wanted it to do. So maybe you have a more higher level task like, you know, make this website look prettier or like make this app more user friendly.

2:28Versus I think Astra was a lot better whenever you had very detailed instructions and you wanted to do something very specific. And so just some pros and cons. But I think overall people are very impressed with Astra and seem very excited about it.

2:40Stephanie Palazzolo:Seems like they were really pushing the voice capabilities in their commercial. I mean, is it better on that front as well? Yeah. So it does seem in the commercial there is, you know, in the kind of demo video, there was a lot of hype and emphasis around voice and computer use, which is the model's ability to like navigate your computer. Overall, like I have heard good things about that. But honestly, I think more people, at least I'm talking to, are more interested in trying it out for like software engineering versus voice. Got it. Okay, so now the big question that I guess I was seeing trending online over the weekend was, is this model AGI?

3:19Stephanie Palazzolo:I mean, what has OpenAI said in terms of whether or not it thinks it's AGI? I saw Jensen seems to be endorsing the idea that it is AGI. Is it that good? Yeah, well, you know, definitely OpenAI wants you to think so. So whenever I wrote my kind of brief on this, I did mention one interesting thing was during the press briefing, you know, Greg Brockman was there. He was very much pushing that this is AGI, edit the call by saying kind of in a joking way, like, welcome to the AGI era. But I think this is like obviously a very different shift than it feels like maybe with the last couple models. Like, OpenAI has been very careful about their wording, saying, like, you know, we're getting close, we're 80 % of the way there, versus this time they seem much more, like, willing to just straight up say, like, we believe this is AGI.

4:12Stephanie Palazzolo:Is the reason for that, is that at all related to the whole, you know, reframing of the agreement with Microsoft, that they can now say that we are at AGI freely? because I'm sort of trying to think about the communications angle here. They've removed that from the partnership agreement with Microsoft. So that no longer impacts what they have to pay Microsoft in terms of revenue share. That's one angle to this. But I'm also thinking about the social implications of all this. Why do you think they're so willing to say that this is AGI now? Yeah. So as you mentioned, there's definitely less risks to declaring AGI now that that, you know, term has been removed from the Microsoft agreement.

4:55I think in terms of, you know, socially, why now? I mean, it does, I guess, like, taking a step back, I think whenever you initially asked me the question, like, is this AGI? I think my honest answer is, like, does it even matter whether it is or not, right? Like, this is, I think it's very clear that this AGI thing is, like, a spectrum or, like, a continuum versus, like, a, okay, after this point, we're post-AGI versus pre-AGI. So I think in the last couple months, there's been kind of this building agreement that we are in the AGI continuum, right? Maybe we don't know exactly when the moment was that we hit AGI, but it's very clear that we are in that general space.

5:36So I think at this point, a lot of people have kind of accepted AI can do a lot of the things that humans can. And that is like a definition of AGI. So I think like people have come more to terms with the fact that like the tech we have today is like probably AGI or like basically almost AGI. And so I think there's like less risk to them admitting that. And it feels like now the kind of big milestone that everybody's scared of and cares about is like recursive self-improvement or like super intelligence. And like we kind of have moved the bar once again. So people are less worried about it. Like AGI feels less scary to people.

6:14Stephanie Palazzolo:Yeah, yeah, yeah, yeah. I do want to get to another story that you published as well about Anthropic. It was kind of interesting. And this is a story about Anthropic expanding into tools that are related to payments. It's not directly payments. Help us unpack your reporting there. Yeah. So this is, I guess, a very different vibe than deciding if we're at AGI or not. But we published a story last Friday about, as you mentioned, Anthropic getting more into kind of payments-related tools. So I guess essentially how the story came about is we noticed a bunch of job postings that have been put up fairly recently, basically where Anthropic is looking to hire more people that would work in an area like, you know, billing and fraud detection and treasury services.

7:04And These are all kind of in the realm of payments and billing. And it's generally like all related to, you know, my customers like paying me for my service. Like how do I keep track of that payment and make sure I'm charging them the right amount and all these different questions. And so this is very interesting for a company like Anthropic because, you know, this is typically in something that, you know, a company like Stripe would offer. And so the fact that Anthropic wants to build more of it in-house has some pretty interesting kind of implications there.

7:36Stephanie Palazzolo:So, I mean, this is – how much is this related to the story here of Stripe buying Open Router and Stripe expanding on the AI Labs turf versus the AI Labs expanding into payments? I mean, is this the two of them kind of going at it here? Is this actually a rivalry? what do you think? Yeah, so again, this is like, you know, a caveat here. This is like kind of our analysis versus, you know, any kind of direct reporting. But it does seem like the fact that Stripe bought OpenRouter, at least our read of it, is that it was maybe more of a defensive move, right? So essentially what's kind of going on here is, you know, as Anthropic is growing, as companies like OpenAI are also growing, they're getting super complicated.

8:28And you can imagine that kind of charging and keeping track of payments for AI is pretty complicated and maybe even more so than traditional software, where it's just this monthly fee, basically, that you pay. Now it's like usage-based, it's based off of tokens, you can like cash tokens. It's very, you know, it's a very, you know, complicated thing. And so now we see Anthropic wanting to build more of that in-house and maybe potentially, you know, at some point move some of that off of external vendors like Stripe and use more of their in-house tools. So if we see companies like OpenAI and Anthropik doing that more, what position does that leave Stripe in, right?

9:06Like, you know, in the AI boom, Stripe doesn't want to kind of get left out in the cold. So what might make a lot of sense is for them to buy a company like OpenRouter, where, you know, they still get kind of a piece of this like AI token economy and still get, you know, this company OpenRouter that's very quickly growing and is really at the center of this like open source AI business.

9:28Stephanie Palazzolo:Right, right. And I mean, just going back to Anthropic's push here, you sleuthed through some of the job postings that Anthropic had related to this. I mean, how far are they in this push? What specifically are they hiring for? What did that tell you about that? So it does seem like they're still pretty early. So some job postings that we found are, you know, there were a couple for billing. There was another one focused on financial fraud. There was one that basically asked for a person who would overhaul Anthropix order management, provision, billing, invoicing, and other related systems. So this is kind of the whole process of like what happens behind the scenes when you're paying for your Claude subscription or paying for tokens, for instance.

10:17There's other roles related to Anthropics treasury systems, which is an area that Stripe does not get into. So it does seem like they're still in very early stages of finding people that are going to be able to work on these sorts of systems.

10:31Stephanie Palazzolo:Right. Great. Well, Stephanie, it's certainly an area to watch. I want to thank you for coming on. That is Stephanie Palazzolo, author of our AI Agenda newsletter here at The Information. The information has exclusive reporting that investing giant Cotu is looking into establishing a new joint venture to help chip companies address component shortages. Maddox is a chip startup at the center of that story. I want to bring on Valita Pau, who reported that story along with Stephanie Palazzolo and Phoebe Liu. Valita, welcome back to the show. It's great to have you here. Hi. So what is this joint venture that Cotu is looking into right now?

11:10So this is a first-of-its-kind joint venture that Co2 is looking to work with, this chip startup called Maddox. It's essentially helping the startup to finance their purchases of all these important chip components that the startup needs to make their own chips. So think about memory chips or logic dice or even manufacturing capacity the startups need to get at, say, TSMC, which is super constrained. and you'd have to get TSMC to make your chips. So Code2 is trying to be the center of that and trying to help the startups secure these manufacturing capacities, supply these chips.

11:54Stephanie Palazzolo:And this would basically involve Code2 basically negotiating for capacity at these manufacturing companies because they're able to sort of take orders from all these chip startups together? Is that how it would work? So like the specific of these like negotiations is not like really worked out right now. Like it's still sort of like ongoing, but essentially I feel like, you know, this is sort of, you know, first exclusive to Maddox in some way to have a first right of refusals. Potentially like, you know, if it doesn't work out for like Maddox, like KOTU can resell those capacity to like, you know, other startups.

12:32But, you know, it's still like, you know, up in the air, like, you know, how specifically it works out. But like, you know, it's truly innovative in its own sense because we have never seen this before, that, like, you know, an investing giant is, like, you know, partnering with a startup to, like, you know, get ahead of these, like, supply constraints that, like, they may face down the road.

12:51Stephanie Palazzolo:Tell us about Maddox here. I mean, this is a chip startup that we have not talked about too much on the show. How far are they in their chip development process? How big is the startup? What do we know about them? So for a starter, this is a startup that Anthropic actually discussed buying earlier this year. And, you know, it's a startup, like many of the expiring NVIDIA challenges, you know, like they make chips. And then for Maddox's case, you know, they are saying like, you know, their chips can be optimized for both like, you know, inference and training, which, you know, like these startups, these are like, you know, focusing on like on inference.

13:29They have not really have their first chips yet. It's like scheduled to like, you know, tape out, which is like, you know, a first prototypes of their like chips. That will not be until the first half of like 2027. So, but it had raised a lot of money from, you know, like Gene Street, Leopold's situation of awareness with lots of kind of big investors behind it. And, you know, it was last value at$4 billion in February. And it's now trying to raise another round and are talking to investors to at least double that valuation or even more.

14:00Stephanie Palazzolo:Is KOTU an investor in Maddox? I'm not yet, but like KOTU is discussing investing is like a new round they're raising. But we're not sure if KOTU is going to invest yet and how much and what is even a valuation. It's just like new startup. But lots of their rivals like Edge has raised like a$21 billion valuation. So the market's very hot right now. And so if Maddox is looking their tape out for their chip is coming down the road here, sounds like this is a pretty proactive initiative for Maddox. It's sort of anticipating, hey, once we do get our chip to market, we don't want to be in a position where we are struggling to get capacity and stuff like that.

14:46I feel like it's one of the many challenges startups may have because, you know, when the whole world is very constrained on these like supplies and big chips, like NVIDIA or Apple or Google, they can just, you know, use their kind of leverage because they're like one of the world's biggest companies to lock down all this capacity from like the chip suppliers or like FTSMC or elsewhere. And so like this is effectively, you know, trying to help them to overcome this like challenge down the road because they're like a big financial backers behind this.

15:15Stephanie Palazzolo:So Valida, do you anticipate that we're going to see more joint ventures like this pop up between big investment companies, between chip startups? I mean, we've certainly seen the agreements where investment companies will help the hyperscalers with their own chip efforts. I'm talking about the TPU, the Google Broadcom deal, I believe it was, to help get the TPUs out into the market. Are we going to see more tie-ups like this with the startups? I felt like, you know, JVs are like now increasingly popular. And, you know, we don't expect this one to be like the last one in the specific chip financing venture.

15:59We know like, you know, some other investors may have also proposed this, like, you know, other chip startups, too. But like, you know, in this way, this one is the first one that's like, you know, reporter or like, you know, we know that it's like, you know, coming along. So we felt like, you know, as these big financial institutions are like trying to help expand the market and, you know, they will try to, you know, create some more creative or like innovative ways to help these startups to get what they need the best way they can. Great.

16:29Stephanie Palazzolo:Well, Valida, I want to thank you for coming on. That is Valida Pau, our Deals reporter here at The Information. with ai cyber threats on the rise following the open ai attack on hugging face companies are ramping up their security budgets for more on that i want to bring on my colleague aaron holmes for a conversation about a story that he published this week on that aaron welcome back to the show it is great to have you here hello so you have chronicled the cyber security sector for many years now. You were once a cybersecurity reporter, turned enterprise software reporter, turned Microsoft reporter, but cybersecurity is still near and dear to your heart.

17:10Stephanie Palazzolo:And every time you've come on the show, you have helped us understand how this sector very much fluctuates. I mean, there's booms, there's fatigue. Now AI is here. Where are we at in that story right now? Yeah, you know, I think we're about six months out from Anthropic sort of announcing mythos and warning the world that we were going to enter this new era of AI threats. And in the six months since then, there's been sort of a broad recognition across the industry that cybersecurity budgets need to go up. And I talked to almost a dozen cybersecurity executives for this story, and pretty much all of them said that they're increasing their budgets for the year ahead, anywhere from 20 to 30 percent in a lot of cases.

17:52And, you know, that money is going to both AI models to help protect against potential AI threats and also to new tools that they say they need to keep track of how their employees are using AI. Essentially, a whole new raft of security threats across the board that companies are in many ways scrambling to protect against.

18:12Stephanie Palazzolo:So when you say they're spending on the AI models, does this mean that the labs are the big winners here? Or are there specific pockets of the enterprise software cybersecurity tools that are also benefiting a lot from this? Labs are certainly big winners. You know, there's models that are specialized for cyber tasks like Mythos and, you know, OpenAI Cyber Models. But even outside of that, you know, there's a ton of companies that are using just the publicly available models from those labs to scan their software to look for vulnerabilities. And we're also seeing, you know, big incumbents like Palo Alto Networks, Microsoft, CrowdStrike all build new AI agent tools that in most cases rely on models from the lab.

18:59So, you know, a lot of the new spending going towards those types of products are also benefiting the labs at the end of the day. Right.

19:06Stephanie Palazzolo:And I mean, let's let's dive into the labs cybersecurity strategy here, because one way that you just mentioned is they have the models that are specifically built to help uncover some of these cybersecurity threats and stuff like that. It sounds like, though, they're also sort of diving deeper into cybersecurity altogether with product offerings that are a little broader than that, too. Right. Yeah, that's right. We're seeing them also, you know, put out these applications that, for example, OpenAI calls them red agents and blue agents, which are, you know, specific agents meant to try to hack into company systems and see what vulnerabilities they can find and then suggest ways to patch those vulnerabilities.

19:46And, you know, what's interesting is that's potentially going to disrupt this sort of older category of vulnerability scanners where we've had players like Tenable, Qualys, Rapid7, you know, using kind of pre-LLM technologies to do similar vulnerability scanning for companies. Of course, all of those firms are now also building, you know, LLM powered tools that in some cases use the bottles from these labs. But some of the cybersecurity executives that I spoke to said that they're rethinking, you know, how they choose between those more legacy vulnerability scanners and the fact that LLMs can now do a lot of that work themselves.

20:25And so which way are they favoring then in that tradeoff? You know, I think so far a lot of people are still in the testing phase, but I have certainly heard from some who say that they are getting more bang for their buck out of the LLM powered tools. At the same time, you know, some of these companies like Tenable, for example, has still seen a surge in revenue in the last six months, which, you know, Tenable CEO told me is because customers are increasingly, you know, wanting their vulnerability scanners, including the newer kind of AI infused products that Tenable is building. But, you know, for some other companies, they seem more on the rocks.

21:02For example, you know, Rapid7 actually had falling revenue in the most recent quarter. And they also had to cut 12 % of their staff, which they said was about trying to refocus their business on building AI native tools. So it's definitely roiling this sector of the industry and the broader cybersecurity world.

21:20Stephanie Palazzolo:When you talk to executives in charge of cybersecurity at these businesses, I mean, there's a lot of demand right now in this moment for protection. As we talked about at the start of the segment, I mean, this fluctuates, right? I mean, cybersecurity is well known to cause fatigue. Sometimes people just get tired of spending so much on protection, especially if they don't see the results. Do you expect this demand to last? How are the CISOs that you talked to thinking about what results they're trying to see and how they're measuring success? What do you think? So the perception right now is very much that the landscape is favoring hackers right now over defenders.

22:04That has a lot of these companies very nervous. And, you know, that's because in theory, any hacker with rudimentary coding skills can use a tool like, you know, OpenAI Codex to supercharge their ability to find vulnerabilities. You know, even though models from OpenAI in theory will refuse to carry out cyber attacks, we've heard that there are ways to kind of jailbreak that. There's also open source models that hackers are using. And so across the board, you know, I'm hearing from companies who are saying in the short term, we need to spend heavily on, you know, models to defend ourselves and other products to defend ourselves.

22:38And at the same time, you know, I think some companies are hoping that maybe in the next six months to a year, there will be more of an equilibrium where they'll, you know, have gotten their defenses up and hackers don't exactly have, you know, one foot forward against them. But in the near term, there's definitely a perception that companies need to quickly spend more to catch up with these threats.

22:59Stephanie Palazzolo:And how would they inevitably measure the success? It's kind of like you're almost measuring a negative, right? In a sense, it's like, should I be hacked, it will cost me this many millions of dollars to address it. And I guess I'm just trying to prevent it. So it's a little bit like you're trying to monitor something from not happening, right? It's true. Yeah. I mean, like the successful state in cybersecurity is nothing happening. But obviously, you know, which is hard to measure ROI on then, right? Right. Although there is, you know, a lot of companies are using these models to just scan their code for vulnerabilities.

23:40And they say the amount that they're finding is, you know, skyrocketing. You know, Palo Alto Network said that the amount of vulnerabilities they found their first month using Mythos was something like five times the amount that they would typically find in a regular month before that. So there is definitely like a rush in new findings from companies that are using these models to scan for undiscovered vulnerabilities.