Inside Claude Opus 5, Open-Source AI Regulation, Google’s Chip Financing Strategy

0:13Welcome everyone to The Information's TI TV. My name is Akash Pasricha. It is Monday, July 27th. We have got a big week of quarterly results ahead. Wednesday, we've got Microsoft, Meta, Robinhood. Thursday is Apple, Amazon, Coinbase, and Red Hat. and tomorrow we have PayPal, which should be interesting to see in light of reports that there has been acquisition interest for that company. We'll be covering all of those results this week, so stay tuned to our shows. Today on the show, we are unpacking initial reviews on Anthropics Opus 5 model. We're bringing on the head of AI at Clay to talk about that.

0:52We're then going to speak with a former top White House official about her view on the debates that Kimi K3 has sparked around AI policy. We've also got some new reporting on how Google is taking a page out of Wall Street's ambitious approach to debt as it seeks to finance its AI buildout. And we're going to close out the show with a conversation with our partner, PwC, for a broader look on where Fortune 500 businesses stand on open source AI and how they're using it. It's going to be a busy show and a busy week, so let's get right on into it. Anthropic on Friday released its latest Opus 5 model, which means developers were very busy over the weekend seeing what the new model is made of.

1:32I want to bring on Jeff Barg, the head of AI at AI sales engineering company Clay, for his thoughts on all this. Jeff, welcome to the show. It's great to have you here. Thanks so much. Yeah, thanks for having me. So, busy weekend? Were you playing around with the new model? Yes, yes. Busy weekend. I think we spent a lot of time both evaling the model for our internal use cases, how our customers will use it, and as well just trying to use it as a personal coding agent. So it was a busy weekend. And so what's the review? Yeah, the headline is it's really good. I think it's always interesting to see what happens with an Anthropic launch because sometimes you see some benchmarks and then they don't replicate in either vibes or in actual eval results from our internal evals.

2:20I think here, the internal benchmarks that we had pretty much matched what Anthropic said publicly, which is really close to Fable's performance at a fraction of the cost. And we were seeing, I think Anthropic said about half the cost of Fable. We were seeing even more than that. So maybe a third or even less of the cost of Fable and really close to the performance across some of our hardest evals. What about 5.0 GPT 5.6? Yeah, I think it's a really hard, they're playing in a really hard space because actually 5.6 comes in just around the same as both Fable and Opus 5. And it's even cheaper than both of them.

3:01So I think, you know, Anthropic's really playing in this hard middle ground where they have the frontier. Fable seems to be really doing well at, you know, frontier performance and writing in quality of code, et cetera. But in this kind of tier two where, you know, you have, I don't know, just the daily workhorse model, it's not really a tier two, but your kind of daily model for doing things, I think Anthropic has a, faces some pretty steep competition from OpenAI. Interesting. So, I mean, that's the benchmarks and the evals. What about the vibes? I mean, this is kind of an interesting soft metric, I guess.

3:38And I think, I mean, when it first came out on Friday, it seemed like some people weren't too bullish on the vibes here. Unpack that for why do you think that was the case? Yeah, it's been interesting. I think every Opus launch has had this kind of reception on arrival. I'm thinking back to the Opus 4.7 launch, where I think it was maybe based off of a new pre-train. a lot of our internal users at Clay really didn't like Opus 4.7, and that's when actually a lot of folks started moving over to Codex and the GPT series models. I think with Opus 5, the vibes are good, but I think most users internal to Clay will continue to use Fable.

4:20I think the vibes on Fable are just so immaculate and just sort of effervescent. Like, there's just a... There's something to the quality of the writing and the sort of intangibles that are really hard to replicate, I think. And I guess, but you're saying that even acknowledging the tremendous cost of Fable, you think people are still going to be using that? I think so. I mean, it's really hard to say for daily users, especially on non-enterprise accounts. So really the Claude Max user base. I think a lot of people are going to move to Opus 5 and probably get pretty good performance there. but there's just something really special to Fable.

5:00I think inside of Clay, we'll see a lot of users stick with that. Yeah. What do you make of the timing of the release here? One of the things that we talk to developers a lot about is this cycle of when your company is in the headlines, when your company is getting good headlines versus not so great headlines, when you decide to release Opus 5. I mean, does it feel like strategic timing to you in terms of the release schedule? Hard to say. I think the recent thing that we think a lot about internally at Clay is just all of the open weight models and how close that those are getting to the frontier.

5:37So I think there, it's really the timing of Opus 5 seems to relate to that in a pretty interesting way. I think we're spending a lot of time really migrating a lot of our workflows to open weight models. And there's been obviously a lot of movement in that space. Well, and say more about that. Which models are you using in the openweight category? Yeah, I think the openweight models are much harder for us to think about. I think if we look at our internal evals, openweight models tend to be very spiky. So some of the evals they do really, really well at, and some of them they're much worse than many of the Frontier models.

6:13We use GLM 5.2 and Kimmy Cage 2.6 pretty extensively at Claynow. We just launched it in Claygent, which is our main research agent, but we also use it in Sculptor, which is our go-to-market engineering agent. So more of a typical kind of coding adjacent use case. I think there, you know, we're seeing for GLM 5.2, sometimes it's at the top of our evals, but it's truly a fraction of the cost of even, you know, the cheaper open AI models. So it's really kind of a pretty interesting, it occupies a pretty interesting space, at least in our, you know, in our model usage. So what percent of model use at Clay is now open-weight models versus frontier models?

6:58We're still majority frontier models. I would say it's increasing as a percentage fairly rapidly. And I think we're investing a lot of time on the engineering team here as well. And so I'm going to go back to you're using Kimi-K. Is it the most recent Moonshot model that you're using or is it models before that? 2.6. 2.6, okay. Yeah. And so, I mean, walk me through sort of your considerations here on using models from overseas versus models from the U.S. I mean, Clay, just to recap, I mean, you guys are building out the go-to-market engineer, the sales engineering role, I guess, that AI has now enabled.

7:41So, I mean, you have a lot of data from customers about their leads, you know, their notes from customer conversations, stuff like that. And this is all fairly sensitive stuff. So how do you think about the decision to use models from overseas from a security perspective? Yeah, and it's important to note we host all of our models with Base 10 and Fireworks. So all of the models are, yeah, yeah. So we host with inference providers that are located stateside. But yeah, it's an interesting consideration. And I think, especially with sort of the push or the potential conflict between the Trump administration and some of the open weight models, I do think it's very important for the US to have strong open weight models, like the ones that we're seeing from NVIDIA and Thinking Machines and others.

8:29Great. Last question for you, Jeff. I mean, when you look at Opus 5, I wonder what it tells you about Anthropic's strategy overall with their suite of models? I mean, you know, I sort of think about which model, which company is positioning themselves at like in the premium model category versus the cost leader category. Fable is an expensive model. You know, when you think about the Sonnet, the Haiku, the Opus, and then the Fable family, where could you see Anthropic taking this suite of models given what we're seeing with Opus? yeah i think anthropic strategy is very interesting here because um we're seeing basically they're moving away from the lower end of the market here um it sounds like they're not going to have a haiku five and i would say for us we really think about opus five is almost the spiritual successor to sonnet 4.6 actually like a lot of our workloads that we were running on sonnet 4.6 we're going to run those on opus 5 now because the increase in cost actually does come with a material increase some performance in a way that we weren't really seeing with earlier iterations of these models.

9:39And so I kind of think about it as all of these models are effectively moving up a weight class, right? Anything that we were running on Sonnet, we'll now consider Opus 5 for. And any of the frontier workloads that we have, where you're really looking for the best in class performance, we'll really consider Fable or kind of Fable class models there. Or you could just move everything open source, which it sounds like everyone's interested in doing anyway. Yeah, it's absolutely, I think that's, it's really in the conversation now, and I think really something we think about actively in a way that a year ago we were, you know, we were using open source models a year ago, but it really wasn't in consideration for these frontier workloads.

10:20And it's a material cost advantage for you guys to use open source? Absolutely. Yeah, I think both in terms of the percentage of savings, but Clay just has a lot of inference spend across the board. And so any kind of marginal improvement is material to our business. Great. Well, Jeff, I want to thank you for coming on. That is Jeff Barg, head of AI at Clay here on TITV. Thank you. Moonshot AI's Kimmy K3 has sparked much debate around open and closed source AI. And meanwhile, OpenAI's hugging face hack debacle has raised a lot of new concerns about AI security. I want to bring on Teresa Payton for her view on all this.

10:59Teresa was as Chief Information Officer at the White House under President George W. Bush. She is now the CEO of Fortalus Solutions. Teresa, thank you for being here. Thanks for having me. Okay, so tell me a little bit about your view on the open source, closed source AI debate that has been playing out. What's been your reaction to it? Yeah, I mean, I think it's a really interesting debate that we have going on right now. What I am advising business executives is, obviously, given where we are with AI right now and not knowing what puts a firm in the penalty box and what allows the firm to come out of the penalty box, it's a good idea to have more than one vendor and to think about that from the standpoint of resiliency, recoverability, security, but also just making sure from a supply chain you don't get cut off from something that's considered mission critical.

11:49The argument that's happening right now is really interesting to watch unfold because as you see, some of the heavyweights, many of which are in sort of the United States, Silicon Valley, in AI, you're starting to see two different camps emerge as to whether or not we should allow sort of this open source and open weight models, many of which are lower cost, maybe not as good, but maybe good enough, versus sort of the closed model concept. And, you know, I think there's a lot more to unpack there as to which where do you stand on that debate? Should there be regulation on open source? There should be regulation of governance and guardrails on any of the models, whether it's open source.

12:31We use open source code bases today, not AI, but open source code, which has really helped propel development teams forward over the past several years. But we need governance and guardrails for both, not just one versus the other. You know, I've been trying to figure out, so all this conversation came up because Kimi K3 was so powerful, people were very impressed by it. You know, the conversation initially was around regulation of open source models coming from China and from overseas broadly. That then drifted into a conversation around open source altogether. And I've been trying to figure out why it is that companies have taken this approach that we need to remain open on regulating open source AI.

13:16when the Kimi K3 issue was a little more isolated than that, right? I mean, this was an issue of the model coming specifically from China that was the concern. So why did these two issues get conflated? Yeah, it could be that there's just so much going on right now as it relates to AI that the issues got conflated. You know, what I would say is this. You know, businesses have to look at dollars and cents. They have to look at their bottom line. And as we see, the closed models ratchet up in price. And in some cases, many firms are saying, hey, what used to only cost me a couple tokens in the past and sort of the newer closed models that are being released, it's costing me more tokens and costing me more money.

13:56And they're looking at more alternatives that are more cost effective and maybe just as good or just good enough. So it's very interesting to sort of see almost like an anti-competition thing sort of evolving here in the discussion. What I would just say is like free market competition is good. What I would always caution is be careful, make sure you get what you pay for here. And you want to make sure that you've got your own internal governance and guardrails to make sure that your business is safe and secure if you're using open source models, regardless of their origin story, which country they were built in.

14:34Why do you think China has developed such an advantage in open source models specifically? Well, there's a couple of things at play here. First of all, they were banned from using some of our most powerful chips. And, you know, sort of the necessity is the mother of invention here. So they had to get by with chips that weren't as good as some of the chips that many of the American companies had access to. You also see over the years that China, although they deny this, they are incredibly good at doing sort of copycat technology. So sort of studying, looking at what's being used, and then being able to leverage intellectual property and doing, you know, just good enough and copycat technology.

15:18I have heard from people who've gotten their hands on the model that the models are not as advanced and sophisticated and optimized for token use as some of the American closed models are. But for many businesses, that may be just good enough. Now, I think the debate is still out whether or not there is a true security risk with these models just because they come from China. I would love to see tests run in the lab, and I would like to see the results of those tests. So you haven't yet made an opinion on whether or not you think there is a security risk from these models? I think it's okay to assume there might be, but let's, before we start accusing just because a software product had an origin story of China, let's put the proof behind the details.

16:05Let's do the work. Let's get in the lab, see what the vulnerabilities are, and then let businesses decide for themselves, based on the facts, can they actually mitigate any security vulnerabilities that have been found? Right now, we're just saying, oh, it's from China. It must be dangerous. Right. You really need to see those results. Have you, has your team or the orbit of companies that you work with, I mean, you work in the cybersecurity space very closely. That's what you've built your practice on. Are there any early indications, results? Have you ran any early tests that could give us a view into that at all?

16:37Yeah, the early tests that I've seen have been looking at if you've downloaded the model and you're using it in your own sandbox, how can you actually protect and secure it? that has its own downsides, right? Because you have to make sure you can update the model and that you're not going to kind of break the threads of things that you've built by updating the model. But most of the focus right now has been on how effective these models are in solving problems. How long does it take? How expensive is it? As far as coming up with sort of the major national security vulnerability, my team is still in conversations with teams doing those tests.

17:14We have not heard anything definitive that says, aha, here it is, here's the smoking gun. But this takes time. It takes time to find those hidden backdoors. Is this like weeks or months? I've never done any of these tests before, so I mean, how long should we expect it to take? Yeah, the analysis oftentimes only takes a few weeks to do, but I will tell you that even the best trained software looking for these vulnerabilities powered by the best practitioners, you still might miss something. Right. Well, and to that point, so then let's talk about the OpenAI hugging face debacle from last week.

17:57Were you surprised when you saw that headline? I wasn't surprised. You know, a couple of things about what this is and what this is not. Yes, obviously what it is is AI jumped the guardrails, went and hacked another company. The other company, in order to be able to protect and defend itself while it was trying to figure out what was going on, they actually were almost like firefighters locked in a closet, couldn't get out. They actually use open source, open-wide models. I mean, that's what I read in the report this morning. It ended up being that open source models from China were actually better at protecting them.

18:31Yeah, exactly. Now, what we do want to talk about here is this is not a commercial product running a commercial enterprise that jumped the guardrails. This was tests going on in a lab that jumped the guardrails and hacked another company. So I think we have to ask the lab a lot more questions, you know, just like we ask labs that work with live viruses, you know, what protocols are you taking to make sure it doesn't just jump through the doors and start infecting the public. We have to do the same thing here. I'm starting to feel like there's not enough of those kind of like hazmat-like situation governance and guardrails on the tests that are being done in these labs.

19:10So sadly, I'm not surprised that it happened, but there's a lot of lessons learned that we can take away from this, starting with, if you're going to be doing these tests in these labs, make sure you've got the right governance and guardrails so you're not attacking another company. Let's not do this again. Let's not repeat the same mistake. Right. You know, I want to ask you about regulation, broadly speaking, given that you were in a government position for a little while. We've been thinking on the show here about the levers that the government could pull as it relates to governing AI, regulating AI.

19:45On one hand, you have the chips, which has been a point of regulation in the past. We also have the data centers and where the data centers are actually located. You have the software. Which lever do you think is most likely to get pulled here as the government seeks to regulate AI if they decide on delineating between open source and closed source? How do you actually do it tactically? Well, I would say the challenge that we have right now is incentives are not aligned. So you have the incentives of businesses, incentives of national security, the incentives of, so you have the businesses that are the AI companies, then you have business executives who are trying to figure out how to implement AI.

20:31Then you have ordinary citizens like you and me. So we have all these different stakeholders, and our incentives are not aligned. And so if we can find those common grounds where we can have alignment, then we will have alignment around some key issues around where do data centers get built and how aesthetically pleasing and environmentally friendly and neighbors. As it relates to regulation around safety, security, ethics, and resiliency, we have to align incentives. So, for example, many of these companies today are not inherently profitable. What if they were given R &D tax credits that didn't expire or had a shelf life of 100 years so that one day when they are incredibly profitable, they made the right decisions now because they had the right aligned incentives?

21:18That's where I see the disconnect. So there's a lot of things to focus on right now. But if we actually strip all of that away and actually sit down and say, where are incentives aligned? Where is it that we all want the best for each other? And then get the things in place around that. But I'm asking, so when you ship it all away, I'm sort of asking, where do you think the government should focus its efforts? Should it be on regulating the chips? Should it be on the data center location? Should it be on the reviewing the model release cadence schedule? Like, which lever do you think they should pull on first?

21:54I think this is a tough one. I mean, obviously, AI is another technology, just like we have data centers today for, you know, processing of banks information or healthcare information. We need data centers tomorrow for AI and other technologies coming. So what I would say is that we need to focus on the governance and guardrails of the actual technology itself, but that pretty quickly it needs compute power to run. So we really have to have this debate now, sooner rather than later, around data centers, compute, chips. And candidly, many of these are age-old problems. We've been kicking the can down the road for a while.

22:34So put the governance and guardrails on the software, and then the rest should follow. Right. Great. Well, Teresa, I want to thank you for coming on. That is Teresa Payton, CEO of Fortalist Solutions, here on TI-TV. Google reported quarterly results last week. Our AI and finance reporter, Dakin Campbell, has new analysis on why Google's balance sheet might underrepresent the amount of risk that the company is taking on. I want to bring on Dakin to explain what exactly is going on here. Dakin, welcome back to the show. It's great to have you here. Yeah, thanks so much. Okay, so Dakin, you wrote about how Google is taking a page out of Wall Street's playbook, especially with respect to how it's thinking about debt and loading up its balance sheet.

23:17Explain to us a little bit about the analysis that you did. Yeah, so we were looking at their balance sheet and we noticed that they had, starting last year, they had a new kind of derivative that they were reporting on their balance sheet, which we had not seen before in Google's results. And it relates to these backstops that they are providing for data center leases. So Google has committed to about$44 billion worth of backstops on leases for data centers. So these are companies that are building data centers and leasing with a third party. In many cases, it's FluidStack. And in order to sell the debt that will allow them to build the data centers, they're getting Google to come in and basically guarantee the leases on the back end.

24:08The debt investors and people who buy the bonds are looking at the leases. That's the money that's coming in. And so they're worried about whether that money is going to keep coming in to help repay the bonds. And so Google is basically committing to backstop those. So they're the guarantor. I mean the – Yes. Neoclouds. It's almost like insurance. It's almost like insurance that they're providing. Or like New York City rental guarantors, which should be pretty – they're pretty steep requirements a lot of the time. Yes. Okay, so Google the guarantor. What is the magnitude of guaranteeing, I guess, that Google is doing here?

24:53Yeah, so they have committed to$43.8 billion of guarantees so far. They've said further in the footnotes that there may be another$24 billion in guarantees. So added up, that's$68 billion. currently on the books or coming. And there's an expectation among some people, among credit rating analysts that we talked to, that this will continue. So it's getting a few more quarters of this and you might be getting closer to 80 or$100 billion. Okay. So now let's get to the heart of the matter here. How much$100 billion may be coming in possible obligations if any of this goes south? How much of that is represented on its balance sheet?

25:53So because they have represented it on the balance sheet as a derivative, they basically look at this exposure. So right now it's 44. Just bring it back from 100. Right now it's 44. they look at how probable they think it is that they're going to have to pay out on these. And right now they think they've recorded it as a derivative on their balance sheet and they've got it marked at 815 million. So we're not at a billion yet. That's like what? That's like 1 by 44. That's like 2 % basically. Yes, it's 1.83 % or something. Yes. Okay. Okay. Is that realistic? I mean, you talk to experts about this.

26:38Is that fair? I think it is. It goes into the assumptions that Google is making about its ability to pay out or the likelihood that it's going to be paying out. It's a little hard to know exactly what those assumptions are. They are assuming that if they have to take over some of these leases, they'll get capacity for themselves, or they might get other things that will reduce their actual money out the door. But the thing to keep in mind is if they're in a place where they have to take over these leases, so they have to act in their guarantor, the entire AI ecosystem might be in a very different place.

27:25So they might be taking possession, for example, of a data center whose lease they take over at a time when they don't need the data center because we've decided that AI compute needs are much smaller than we expect they will be right now or companies aren't using AI as much as we expect that they will be. So a lot of it has to do with what the future looks like when some of these guarantees that they've provided take over. And it's really hard to know what that is going to be. Now, Google says that they say in their filings a little bit that in the case that they have to act as a guarantor, so they have to take over these leases, they have various measures that allow them to recoup some of their costs.

28:25So they can take over the lease themselves and run it as a Google data center. They can sublease it out. So maybe Microsoft at that point needs a data center. Google can sublease it to them. Or those are sort of two of the big ones that they have at their disposal. So let's talk about the other side of the story, which is that, I mean, Google just reported that it's burning cash in its latest quarterly results. It now you're telling us, you know, balance sheet. And by the way, we should point out, I mean, this is all above board. You know, they're not there. They don't they don't need to capture the full amount of these obligations on the balance sheet.