Google Plans New ‘Frozen’ AI Chip, Preventing Scams in the AI Era, 160 Software Startups For Sale

0:13Welcome, everyone, to The Information's TI TV. My name is Akash Pasricha. It is Monday, July 20th. Today on the show, Google is developing a new server chip that would allow its Gemini AI models to run more efficiently. We'll talk to the Information's Google reporter who broke that story. We'll also look at how AI is changing the cybersecurity sector, for better or for worse. We'll talk to the CEO and co-founder of Savvy Security. Plus, last week, the Information published a list of more than 160 enterprise software startups that could be acquired in the near future. We'll speak with the reporter who put together that list.

0:49It's going to be a great show, so let's get right on into it. Google is working on a new server chip called the Frozen chip, complementing its TPU offering. I want to bring on Erin Wu, our OpenAI Google reporter who broke that story with our Asia correspondent, Channer Liu. Erin, welcome to the show. It's great to have you back.

1:08Erin Woo:Hey, thanks for having me. Tell me about this new chip that Google is working on. Yeah, so Google is working on a new chip. It's informally dubbed Frozen V2 inside the company. and essentially the idea is that it could be way more efficient even than the TPUs at the time of launch. They're projecting it could be six to ten times more efficient than the TPUs at the time of launch and that's because it works slightly differently by etching the model itself into the chip. So fewer decisions have to be made at the time that the model is running. Etching it into the chip. So I mean so why is it called frozen?

1:44Erin Woo:So the name frozen has to do with the idea that the model is kind of frozen into the chip to some degree. So essentially, the way that a regular chip works, it's generalizable, which means that it can do a lot of different things. But that means that there are a lot of different decision points that the chip has to run. And so this chip makes things more efficient by combining some, or sorry, by like, presetting some of that essentially. And so it doesn't have to make all those decision points. And so how would it be different than from the TPU? Yeah, so the TPU is more of this generalizable kind of chip versus these frozen chips, which would specifically run this model architecture.

2:23I see. So, I mean, if this frozen chip, I mean, if it's really only, I guess it's only meant to run Gemini models, sounds like. So is it only meant to be used internally by Google then? Is that the idea?

2:38Erin Woo:I mean, so a lot of Google Cloud customers also run Gemini models. And so this is something that could theoretically also be run by external customers. But this is still something that's pretty early. So it's not expected to launch until 2028 at the earliest. So I think a lot of these strategy questions might still be in the process of being worked out. Right. Right. Because, and I mean, the reason I'm asking this question is because we, of course, know that the TPU was mostly used internally, but now they're at the point where they're really trying to outfit customer data centers, if I recall correctly, with these TPUs.

3:15So, I mean, your point is well taken that it's not just people at Google running Gemini. It's all sorts of customers. So that could very much be the strategy. How does this compare with other inference folks' chips? You know, we've had the CEO of Sama Nova on the show. OpenAI is working on their own chip. How would this compare to that?

3:36Erin Woo:Yeah, so a lot of people are trying to do the same thing, which is essentially driving down the cost of inference because everyone's in this big compute crunch. So Google's chip works a little bit differently because of this first design. It's actually something somewhat similar to what this Canadian chip startup, Talos is doing. So they're also trying to like, like specifically like etch a model into the chip so that it is more efficient. Has Google been compute constrained? Is this a way to get at that problem? Yeah. So they talk about that constantly on earnings calls, the idea that like their cloud division is compute constrained.

4:15Erin Woo:Like Google reports earnings this week. I would expect them to also talk about that this week. And tell me, TSMC capacity, that I imagine is also top of mind here for this new chip? Yeah, so Google is going to have to find fabrication space for this chip. It's not totally clear to me that it's going to be like a one-to-one exchange in the sense that like in order to make these chips that like seals from TSMC space for the TPUs, like that's definitely possible. But it's also been suggested to me that this is enough of a different fabrication process that it might use different space. So it might not be a zero sum game for them.

4:54Great. Well, Aaron, I want to thank you for coming on. That is Aaron Wu, our OpenAI and Google reporter here at The Information. AI has made it much easier for hackers and scammers to deceive people. Our next guest wrote a book on that topic. Patrick Coughlin is the co-founder and CEO of Savvy Security. He is the author of Dark Side of the Boom. I want to bring on Patrick to talk about his work. Patrick, welcome to the show. It's great to have you here. Thanks for having me, Akash. So I want to talk a little bit about the broader cybersecurity landscape at large, and then we'll zoom into what you are doing at Savvy.

5:29You know, I was thinking about this question here. Is safety keeping up with AI innovation right now? What do you think? Yeah, look, I mean, for the last 20 years, cyber criminals, nation state actors, they saved kind their most sophisticated campaigns for exploiting government agencies and large enterprises where they thought they could extract the most value. I think what's changed in the last few years is these same organizations and new ones that are popping up are deploying AI across every stage of their operations, which is making it cheaper and easier and ultimately more financially feasible for them to move down market and really lever increasingly personalized, polished, and convincing scams at everyday people.

6:16So, I mean, tell me, when you look at the fleet of models, and I ask this question because, of course, Timmy K3 has been the topic of conversation for the last couple of days. There have been a lot of observations that it's a really good model, some questions around, hey, is it as good for cybersecurity and safety as Fable, for example, What's your view on that? Are there certain models that are better for safety than for others? Look, I mean, cybersecurity is a cat and mouse game. And, you know, we've seen that for the better part of the last 20 years is bad guys innovate and then defenders try to innovate as well.

6:52And it's a cat and mouse game of trying to keep up. I think what's different now is this relentless pace of innovation coming out of the foundation model and the key AI labs that are creating these new innovations. And it's making hard to really understand where are we in this cat and mouse game and who's winning. What we do know is that we haven't seen nearly enough investment in innovation deployed at the consumer and deployed at families because the data is really proving it out. I mean, now the average American adult is encountering multiple scams a day. and Americans in this country have lost close to$200 billion in the last 12 months to these AI powered scams.

7:35And so we've got to shift our innovation and our effort and our conversation about the power of all these models and this technology, not just to include what we're doing for enterprises, what we're doing for government agencies, but how are we using that technology to make sophisticated cybersecurity a little bit more accessible for everyday people? And so going back to it then, I mean, is there a difference then between the models? Are open AI's models better than Anthropics for security or vice versa? How do you see that? It's all about how you use the models. And they're so close to each other and they're changing all the time.

8:12And so, you know, scammers today, cyber criminals today can use the same AI image generators. They could use the same models that are powering the same chatbots. They can do voice cloning off of three seconds of audio. All of these tools are available to scammers today. So you don't have to be running off of the most sophisticated models to have the impact on consumers that we're seeing. And, you know, for me, this came home personally when my mom was targeted with one of these fake kidnapping scams where cyber criminals were able to impersonate my sister. They spoofed her phone number. They potentially cloned her voice and convinced my mom that she had been taken hostage.

8:53And the only way to set her free was if she sent$1 ,200 over PayPal immediately. And that's what sparked the interest in writing a book for you, right? Yeah, yeah. So that sent me, after I spent the last 20 years working in national security and ultimately enterprise cybersecurity, when that incident happened with my mom, it sent me down the rabbit hole of looking at really what had changed in the underlying economy of these cyber criminal organizations that allowed them to deploy that kind of a sophisticated and personalized attack that we used to see for, you know, saved for those three-letter government agencies or for a Fortune 500 company.

9:33But now that kind of sophistication was being pointed at little Beth Coughlin in Prairie Village, Kansas. And so it made me think, you know, how many more people were being targeted with these types of scams? And what would this technology roadmap that we're seeing laid out in front of us from the AI labs, what would that mean for the cyber criminals and the bad guys who are using this technology ultimately to exploit consumers. What was the most surprising thing you found in your research as you went to go write this book? I think it was as I was buying scam kits on the dark web and looking into these organizations, and they really are organizations, what surprised me is how they are operating like businesses.

10:16In fact, these scam syndicates from from Eastern Europe to Southeast Asia to West Africa and even closer to home, you know, in some ways they're the most profitable AI early adopters that we have. Research shows that when these organizations deploy AI, they're able to generate nine times as much volume and four times as much revenue than their sort of less AI inclined peers. So while - Why is that? Why is that though? What do you chalk that up to? It's all about personalization. I mean, look, the power of this generative AI trend that we have is that it can talk like us. It can behave like us.

10:56It can create pixel perfect images that are impossible for us to detect with the human eye. That same technology is really powerful for making us more productive, but it's also an industrial grade weapon for people who want to deceive us. But I mean, I guess what I'm getting at is, are they just really good at using the tools? I mean, they're using it for nefarious purposes and, you know, so you could charge whatever you want, obviously. And so, you know, it's not lost. I mean, that revenue might be easier to come by in some cases there. But, I mean, are they particularly talented at using this technology in ways that, I don't know, above board enterprises are not?

11:39Or what do you chalk it up to? I think it's some of that. these organizations have always sort of piggybacked on big technology transformations and use those to to find new ways to exploit everyday people. But but I think what's more than that is is the tools themselves are perfect for impersonation and imposter scams. Digital fraud now is made up almost a third of imposter scams. So 33 percent of cases that are reported to the FBI or to the FTC are some sort of impersonation where a scammer pretends to be somebody you are likely to trust. whether that's a brand you know and love like Microsoft or Apple or PayPal or the postal service, or maybe it's somebody that you love.

12:22These tools make it very easy to sound like us and to confuse us and ultimately to make it easy for scammers to help us depart from our hard-earned savings. Right. Tell me about the company Savvy Security that you've started. What is the ambition of this company in the context of what you're focused on? Yeah, so like I said, you know, there's been a ton of innovation for enterprises and the government around, you know, safe AI and creating defensive tools to protect against AI-powered hackers and nation-state actors. I think what we haven't seen in the last 20 years is the same kind of investment and innovation on behalf of the consumer.

13:04If you look at the consumer cybersecurity market, you know, it used to dwarf the enterprise cybersecurity market. If you go back to the times of the McAfee's and the Norton's, but for the last 20 years, that market has stayed relatively flat. And meanwhile, the threat actors are shifting their crosshairs down market to focus on the consumers more because they can profitably extract value there. So we need a whole new wave of innovation on behalf of the consumer. And we're starting with the places where the threat actors are having the most success. So text just surpassed email as the most common vector for fraud for consumers.

13:39It's been email forever, now SMS and text, and we all feel it. We feel it every day in the scammy and spammy texts that we get. And voice calls are a huge issue when it comes to the big ticket losses. If you look at the five and six figure losses, the scams don't always start with a call. But at some point, you know, a victim gets on the phone with a scammer and ultimately they lose significantly more money. And so we're starting with how do we filter out scammy texts? How do we provide more tools to consumers to understand, you know, when they're on a scam call or a spam call? So do I just do I sign up and give you my phone number, my email, and then you basically protect, you deflect these scam calls?

14:22How does it work? Yeah. So we started with a free tool. It's available at scamwise.com. You can submit a screenshot of a text message. You could take a picture of physical mail that you don't know whether that's real or not. You could forward an email to it. You could just talk to it. And what we heard from users when we launched that a few months ago is that they wanted to see that in an app. So we're now processing, you know, over 100 ,000 suspicious messages every month through Scamwise. And we just launched our Savvy app where we've built in more proactive protection. So you don't have to actually come and submit the message.

14:56we can actually deploy things on your devices to filter out those messages before they even reach you. Do you think, you know, I'm thinking about the, again, going back to the models, the position that the government has taken, that the government has said, hey, we have these voluntary frameworks in place. We would like to work with the AI labs on how they release these models. We, of course, know that these staggered releases have become a little more common now. Do you think that the government should have a role in AI safety as it relates specifically to the models that are released? Look, I think we need a whole host of things.

15:36We need more innovation. That's what we're doing at Savvy, but we also need smarter policies. We need more rituals and our families to talk about these things in different ways. And of course, regulation is going to be a part of that. Now, whether it comes from the industry or whether it comes from the government. I think there's different ways that you can help solve the problem. But, you know, right now we're way behind. We are so far behind in the kinds of policies that we need to keep up with this technology that is getting into the hands of bad guys. This isn't the first technology we've created that has dual use, that can do wonderful things for people, for companies, for enterprises, for governments, but can also do scary things when they're in the hands of bad actors.

16:16Right now, it's harder to buy Sudafed in this country than it is to get an API key for the most advanced model and to turn it into the perfect weapon for scammers. And so I think we can do more to increase the friction so that it's harder for these organizations to get their hands on the most cutting edge technology and then turn that against consumers. Right. Great. Well, Patrick, I want to thank you for coming on. That is Patrick Coughlin, the co-founder and CEO of Savvy Security here on TITV. Earlier this month, the Information put out a special report covering the 160 enterprise software companies that we think are likely to be acquisition targets.

16:56This is an annual project that we've been working on for a few years now. It is a telling snapshot of the software sector. I want to bring on Alix Couture, the reporter behind that story. Alix, welcome to the show. It's great to have you here. Hi, Akash. Tell me, Alix, how did you put this list together? So I worked as a starting point with a list provided by PitchBook. It was a 300 startup list. And then I looked for any updates. I really screened all the startups. I looked for updates in fundraising. I looked for updates in valuation. I looked only at the startups that haven't raised money in the past two years that are privately owned and that are valued more than$1 billion.

17:37dollars. So then after that, I looked at a lot of companies that sat at the intersection of the consumer market and the enterprise market, and I took those one off. And then I reached out to all the companies to look for any updates and make sure that they really fitted the criterias. And then I ended up with a list of 160 startups, which is a lot because it doubled since last year. And And that's because a lot of would-be buyers have backed away from software companies because they redirected their attention to AI. And also because a lot of those companies are profitable and they actually don't need to raise money.

18:19So they do fit the criterias. So I want to talk about the companies that are on this list. And we'll go through a couple names in particular in a minute here. But just broadly speaking, I mean, across these 160 names that you identified, what were the themes that stood out to you about who these companies are and what they focus on? Well, one thing that surfaced a lot, both in conversations with bankers, investors and founders and on the list are data startups. A lot of startups are working with data. They put together data softwares, but also have proprietary data sets. And they think that data as AI is taking the center stage, they think that AI is dramatically more valuable, that data is dramatically more valuable.

19:12So there's a lot of data companies that are really interesting and that are very valuable to software companies, private equity firms and big tech companies. There's also a lot of companies both in productivity and sales and marketing. That's because there's low barriers to entry and it's easy to build and easy to distribute, but then it's more difficult to grow. So for those companies, only those that are really accelerating in AI will really succeed and find and exit. And, you know, the data category is interesting because we've been talking on the show about how front-end companies, application layer companies, CRM companies, for example.

19:58I mean, those are the companies that can probably conceivably be replaced a little bit easier with Vibe coding. data companies, I mean, you know, you could say that they could be ripe for acquisition because we could use that infrastructure. It also could mean that, you know, they could raise again. And so they certainly are a category to watch. Tell me, in this data category of companies, was there any one particular company that stood out to you as a takeover target? Yeah, in the data category, there's one that's called Project 44. So it's not a data company, but it's a logistics company that works with data.

20:39So it pulls shipment data from thousands of carriers, and it has both data graphs, so a data network powered by an algorithm, and gigantic data sets of shipment data. And that company is extremely valuable to big companies or to software companies or private equity firms because it has proprietary data. So it's really interesting for AI companies, for example, that want to train and run their AI models to have proprietary data sets. So that company, the CEO told me that it's been approached several times and that the number of inquiries has increased, has been multiplied by 10 since last year.

21:31So they are getting inquiries. We should point out this is not just based on our filtering of pitch book data. I mean, you know, this is based on real reporting that we are doing. Tell me about the biggest company on this list, Miro. What do we need to know about that company? Well, Miro is a productivity company, and it's valued$17 billion. So that's a lot, and it's the one with the highest valuation on our list. It's a productivity startup that really draws interest from both big tech companies and private equity firms. So basically, it's a visual workspace that helps hybrid teams work together, plan strategies, brainstorm.

22:22And it's really an example of a startup that has also been a buyer. So it acquired the AI productivity startup Reforge in March. So they're really growing very quickly. They're very profitable. They haven't said whether or not they were interested in an exit, so they declined to comment. But they are profitable and growing quickly, and we know that they are drawing interest. However, at this high valuation, an exit would not be easy because who would want to buy a company for$17 billion? Well, I mean, look, there is always the option that you get bought, but it's not for$17 million. And maybe it's a discount to your last funding round, which, look, we've seen it happen before.

23:11you said that Miro's founders, I believe you said that they did not respond to a comment to you, if I'm not mistaken? They didn't respond to the question, are you interested in an exit? The question I was going to ask is, of all the founders that you did talk to, was there any founder that was open and saying, yeah, I mean, we'll sell. You know, we'd be open to an acquisition. Well, a lot of them said they weren't interested. Some of them said, we're not interested, we're just focused on growing, but we would never say no. It depends on the price. So a lot of them say, well, they leave the door open.

23:56One in particular has said and confirmed that they weren't interested in Nexa. It's the HR company Papaya Global. It's valued$3.7 billion, and it really transitioned from being an HR firm to being a fintech firm that helps with payments. And that company is valuable and really stands out because it did a real transition from HR to fintech. And that one is drawing interest from PE firms and software companies. and it is open to an exit and it is in talks right now to sell the company and the sale could happen this year. So some companies haven't admitted it like this, such as this one. You mentioned software companies.

24:47So are big public software companies, big tech companies, I mean, is that the most logical buyer here? Who could you see as being an acquirer for these businesses? Well, there would be the big tech companies that run AI models. So they would be mostly interested in data companies because they need data to train and run their AI models and customers their AI models. So for data software companies focusing on data, it would be big tech AI companies. Then for software companies, it would be large software companies. But industrials like Schneider Electric also want to accelerate in AI. And they would also have an interest in buying software companies that have accelerated in AI.

25:40And then there's private equity firms. So the number of deals have slowed down this year in 26 at the beginning of this year. But they're also hunting and they're also looking to acquire software companies that have accelerated in AI. However, for those companies, so for private equity firms, it would mostly be startups that haven't accelerated fast enough in AI and they would sell at a discount because it's not the most attractive exits for most founders. And for all those buyers, they would be interested in software companies that really have accelerated in AI. Because as it is, just software companies doing software is not necessarily attractive to those companies.

26:28And they're really looking into those that have accelerated and doubled down on AI. Tell me, Alex, as you talk to so many people in the orbit of these companies, investors, founders, analysts, customers even, what did you learn about the state of the war on enterprise software that AI, in some cases, has positioned itself as? People say we could just make these applications on our own. Software stocks have yet to come up. We were talking to Carl Kerstedt from UBS about this last week. what is the state of the SaaSpocalypse for these smaller non-public companies right now? Well, it really depends.