AI Cyberattacks, Anthropic’s $500B Bet and the New Cost of AI Agents

0:00AI is exposing vulnerabilities in some of the most widely used software on the internet. Cyber criminals are using networks of agents to automate attacks on financial institutions. And a growing debate is emerging over who's responsible when AI generated software creates security risks. Meanwhile, the economics of AI are coming into sharper focus. Anthropix IPO disclosures reveal hundreds of billions of dollars in computing commitments. Investors are scrutinizing the companies building that capacity. and Elon Musk is exploring a merger that could bring nearly every layer of the AI infrastructure stack under one roof.

0:34As always, the new cycle never ends when it comes to AI, and we've got Worldwide Technology Chief Technology Officer Mike Taylor here with us to help us make sense of all of it. It's Saturday, October 24th, and this is the AI Proving Ground podcast from Worldwide Technology. Let's jump in.

0:57Okay, Mike, we're going to start in cybersecurity. This isn't a case of deja vu. It's just the nature of the news cycle these days, in which two developments this week illustrate how AI is changing both the speed and scale of cyber threats. First, researchers using frontier AI models found critical vulnerabilities in PHP, a programming language I think a lot kind of either forgot about or didn't know, behind as much of 70 % of the web that would allow attackers to target anyone using those websites. Now, that prompted a rush to deploy protections, and some organizations even went as far as temporarily taking their websites and services offline.

1:35Meanwhile, a hacking group called Cyber Cartel used a swarm of AI agents to automate attacks against more than 400 financial targets in Latin America. Now, taken together, we have AI accelerating vulnerability discovery on one side, and on the other, attackers are using autonomous agents to exploit weaknesses at scale. You've been saying for weeks, Mike, that we need to take security operations from human speed to machine speed and that the volume of vulnerabilities may eventually level off. But, you know, the time available to remediate them will keep shrinking. And we've also talked previously about how organizations can and should automate patching.

2:10So if you're out there listening, we encourage you to go check out those episodes. But, you know, Mike, I am curious from your perspective right now, what will ultimately determine which side gains the upper hand here? Is it going to be access to the most capable models or is it how effectively organizations can integrate AI into their security operations or something else? Yeah, so I think it's the perfect topic to start with this week, given the scope and really the anatomy of what was going on here, right? It had, as you outlined, all the components and elements that we've been talking about for a while, right?

2:44Some level of agency to go out and detect or find this vulnerability. we had, you know, in some cases, I think some of the key technology providers were sort of whispering to their customers that they were aware that this was out there. And I do think an element of that is likely public-private partnership related, you know, in terms of how some of those things come forward and agencies that monitor this sort of activity, but 22 million plus devices, you know, and less than a day really to try to react to that. And, you know, I think the major outcome from my standpoint in this is that while there were impacts, you know, to this based on this event, boy, unless I missed it.

3:31And I know you're looking at this stuff, too, but from a catastrophic sort of, you know, circumstance, I feel like we responded well. I said we just as a, you know, as a larger tech community here. Now, that's not to say there weren't potentially things that were harvested that could come back and or that services weren't interrupted, meaning, you know, people had to make difficult decisions around just turning things off. But, you know, when when you target 400 plus customers or I'm sorry, companies rather in Latin America and and you make out OK in this and not in a catastrophic sort of way, maybe it's a half a point for the good guys in this one.

4:12I don't know. But certainly events like this should lead and prompt more discussion, more alignment with technology leaders and their executive teams and their boards around what are you doing to be able to remediate and respond to these things as they come up? Because that one day circumstance, you know, in this in this particular issue with with cyber cartel, I think we're going to see more and more of this. So it's going to happen more and more frequently. And then, you know, shoot, PHP has been around forever. Why didn't we pay attention to it earlier? You know, I do think there's some effort here that we've got to look at across both privately controlled and open source components that make up large pieces of our digital assets, you know, within our enterprises.

5:02But maybe half a point for the good guys. That's what I'm chalking up to right now. Hey, we'll take half a point, you know, whenever we can get it. Regarding the PHP stuff, I read a quote that said, you know, a lot of the vulnerabilities were perhaps brought up decades ago by some researchers, but that those vulnerabilities were perhaps too complicated for humans to crack, and then they just became forgotten about. So I guess it's just another call to action that, you know, don't assume that, you know, any nook or cranny is going to go unnoticed by AI. Fair? Fair. And the other piece, I think, is just the extension of if the cyber cartels out there doing this, you can assume others are as well.

5:43I really want some focus and attention on where are they getting the compute to do this? and you know if it was a specific set of php vulnerabilities they threw this at maybe they didn't need thousands of gpus to go figure this out but at the end of the day let's kind of work together figure out where where are they we know we know they can get an obliterated model great where are they doing the actual compute functions to go out and do this and then where in the world is this stuff running from because i still think there's a big question out there of who are these people where are they operating from and what kind of assets are they able to use to get to get to their own version of machine speed, right?

6:21Yeah, well, meanwhile, I mean, there's another debate unfolding around who ultimately bears any type of responsibility for these risks that these models can introduce. Interestingly enough, Ben Horowitz on a recent episode of the All In podcast, he was sharply critical of Anthropik over security concerns, calling the Frontier model provider, quote, one of the most negligent companies that he's ever seen in terms of security, end quote. Now, I mean, given all that's taking place, these types of questions growing about the responsibility AI providers have when their technology creates vulnerable software or behaves in unintended ways was bound to come up.

6:56A few weeks ago, you made the point that front-tier model providers are increasingly in the business of trust, but you've also argued for independent standards to help enterprises evaluate these systems, what type of accountability should an enterprise reasonably expect or demand from an AI provider before trusting that their technology, you know, with those critical applications, or what do we have to verify for ourselves? Yeah, you know, and I think this is, I think we use the car analogy on this maybe a couple weeks ago, you know, right, is like, look, the difficult thing, I think, in terms of the frontier model providers, which, again, are moving more from just strictly frontier models that they're building to now more capabilities like code, you know, legal agency that they're creating, you know, within their tools.

7:44And now these, you know, claw-based sort of functions. We talked about, you know, MuseDot, et cetera, they're in there. So one, it's, there isn't one size to fit it all. And in the sense of a model, you know, I think of the model as something that the provider of and the builder of and the user of have a set of shared accountability for in terms of, you know, the car company doesn't, you know, get sued when you're when you're a negligent driver, you disobey a law or God forbid you're, you know, under the influence of something and you hurt yourself or somebody else. There's a component of this where these models can be used for bad purposes.

8:21And in the in the interest of using them for good purposes, we've got to assume some of that risk or we're just going to sit there forever and not make any progress. on the things that really matter around drug discovery, bettering humanity, all these other kind of, you know, righteous and noble sort of things that we can chase after, even in the context of growing a company or a business. But I digress for a second. I think Horowitz is hinting at something that I think is really important to the trust piece of this is this isn't about technology. I mean, it is. It's enabled by technology. And that same discussion that he was having, you know, with regard to Anthropic and some of the things that he doesn't agree with in terms of the context and, you know, even trying to bring like consciousness, is this thing conscious or not?

9:10You know, I look at that and say, the same burden of proof that these model providers are going to have in the tools that they give to you as a personal or enterprise user, you're going to have to establish the same level of trust with your executive team, your business owners. or, you know, if you and I are booking trips or manage, like, you know, there's some level of introspection we're putting on these things before I put a username and a password, you know, into a, into a memory, a prompt or a skill. And so there's, there's a lot there. I know I just, you know, kind of went through, but I would really hone in on the fact that the model companies are going to, and must be accountable for how their products behave or describing specifically the compensating controls that are required in order to make that more trustful and safe.

10:03At the same time, know that we as users, as companies, as people, we have to think differently too around, you know, the parameters, the harnesses that we build and think about, you know, the guardrails that we want to establish around these capabilities. And we're both going to have a role in this. Well, let's turn to infrastructure here and the economics of AI, because we're getting a clearer picture of just how expensive this race is becoming. Certainly, the big news this week was Anthropics draft IPO disclosures, which revealed remarkable revenue growth, but also hundreds of billions of dollars in existing potential future computing commitments.

10:43You know, these companies are effectively making enormous bets, these companies being the model providers on how much compute they'll need years into the future, even as their models become more efficient or the economics of inference can continue to change. You know, we've talked about, and Mike, you've consistently identified power and access to compute as major constraints here on AI growth. You've also been optimistic about the demand coming from enterprise adoption agents and eventually, you know, physical AI. When enterprises build around a frontier model, such as Anthropic, how should they think about the capacity and specifically now given the IPO stuff, the financial commitments and infrastructure partners behind the service that they're buying?

11:25Yeah, their reliability to serve and perform is going to be dependent on power and shell, you know, to build these capabilities out and keep up. And I think, you know, we've had pendulums of adoption from, you know, cloud in the past, right, you know, where even prior to the high-performance architecture of the GPU world, you know, we have today instances where the hyperscalers are moving very, very quickly just to keep up with customer demand outside of AI. And those relationships into the PowerShell companies and access to that was, from my perspective, one of the more, you know, telling. I know there was the, you know, the operating loss that they have and other things that were highlighted in this, but It was like 500 billion worth of commitment to compute, you know, that they've included in this.

12:16And, you know, you think about what they have built out now. They're, I want to say it was two and a half or three gigawatts of capacity that they have. They famously signed, you know, 300 megawatts with Colossus and everything else. But, you know, they're looking at how do they get to, and this is every one of them, by the way, how do they get to 10 to 12 gigawatts by the end of next year? And what's hard to keep track of is all these compute commitments, PowerShell commitments is one, how many of them are in pursuit of the same asset as they're talking about, you know, how they're going to get to that?

12:50Because it's a significant increase in the overall power capacity that we have within the United States right now. To think we're going to double that in a 12-month period next year is probably optimistic or dare I say foolish. So where is it all getting spread? But the ability to serve those tokens is what we're going to care about. Yeah, well, I mean, one company with some pretty ambitious ideas about where some of that new AI computing capacity could come from is SpaceX. You mentioned Colossus and SpaceX, you know, certainly making plenty of headlines of its own this week, as Elon Musk confirmed that the company and Tesla are exploring our potential merger set up special committees.

13:34You know, what's interesting beyond the concept of it being the largest corporate merger in history, potentially, is, you know, what these companies could bring together. We're talking about AI models, chip manufacturing, data centers, energy storage, satellite connectivity, robotics. I mean, just the list goes on. SpaceX is even exploring the possibility of putting data centers into orbit, which is always an exciting thing to talk about. But taken together, it's an ambitious effort to control nearly every layer of the AI infrastructure stack. I mean, for an organization, how much of the stack is strategically important to control for simply consume as a service from the likes of whatever that, you know, combined company might be?

14:15Yeah, I think the idea of Elon having that, or any company having capacity to think about models that they may develop, whether it's coding that they're already moving into. The robotics thing to me is the future of where this is going. That's probably not a secret to any one of our listeners that are out there today. The training of frontier models for the purposes of robotics is already out and active. I think the humanoid capacity of that is a little further off in the future. And thinking about a company that's really good fundamentally at manufacturing electrical devices that just happen to be cars for the most part, and now moving into, you know, robotics, whether they're, you know, back to the manufacturing use cases or they move more towards humanoid, even if humanoid is around specific functions, you know, that's an interesting combination there.

15:09And I think that's one that, you know, we've, we, and I say we, the market, we've seen this, or you sure probably felt this coming at some point. So, yeah, they're going to start a committee to do it. I mean, I think that's just a pretty clear gesture that's probably just going to happen. But they're going to, you know, signal that to the market, go through the motions. And I do think it will accelerate robotics and in particular the contemplation of humanoid robotics use cases that I think will be very exciting. You know, as we think about the evolution of enterprises, manufacturing, quality of life and other capabilities that I'm probably not even imagining as I sit here and think about what can be possible.

15:51Yeah, two other announcements in the infrastructure space that I wanted to make sure that I brought up with you that caught my attention. NVIDIA introduced DSX Pod, a factory-built approach to deploying AI infrastructure with the goal of dramatically reducing deployment timelines, which has been a struggle for many. And separately, a startup called Double AI, of which NVIDIA is actually an investor, is working on improving code execution via code execution models, which is an emerging class of models designed to operate inside running software and get more useful compute out of existing hardware.

16:23So while everyone is talking about building more capacity, there's also a growing focus on deploying it faster and using it more efficiently. Where do you see the bigger opportunity for enterprises today? Do you think it's more adding capacity or is it improving how effectively we're all using the compute that we already have? Yeah, I think it's going to be adding capacity, but probably in different ways. You know, the idea that we will all be consumers of what I'll call kind of NEO or GPU as a service. I think everybody's going to have some on-premise and that may look more like an on-premise data center.

16:58And then I think this POT announcement to me signals not just for the big Verirubin factories that they're going to go out and create, but, you know, we already have, I'll call them, you know, small compute modular devices that are purpose built for robotics. I think there's a lot of room between the big data center and robotics or edge use cases. And we're going to need a number of different form factors to fill that void in. So this to me is a strong signal of that. I think it's, for me, validating in terms of where we're spending our time on this concept of an AI factory. How do you build an AI factory out?

17:36How do you, within the rack, control as much as you can in terms of the cabling and plant that goes into that? At the end of the day, in a data center context, those things are going to be daisy-chained together with other compute storage and high-performance compute architectures. And so, to me, it's, you know, the Lego block was maybe the GPU, then it was, you know, the DGX device. Now it's the DG, you know, the DSX, rather, pod that they're building out. And it's optionality, you know, for all of us. and should simplify at the end of the day, the deployment into an enterprise use case, whether it's edge and something, you know, small purpose built or, you know, a smaller backend that needs to be in your data center to support some of the things that you're doing in particular in training.

18:20Yeah, optionality, certainly a key word that I've heard you revisit time and time again here. Let's bring this back to enterprise adoption. Gartner's IT symposium was this week, made lots of headlines, and it had an interesting statistic that I saw, which was just 13 % of CIOs said that their organizations have realized significant or transformational value from AI. And Gartner warned that AI risk perhaps is increasing faster than AI value. Maybe a pretty obvious statement based on just a lot of the headlines that we're seeing these days. But Mike, you know, we've talked a lot about this. And you've made the distinction between individual productivity gains and applied AI solutions that actually change how a business process operates.

18:59Specifically, you've mentioned things like, you know, things we're doing here at WWT, like Proposal Studio is an example of moving from those isolated use cases into measurable business functions. If we know that redesigning workflows and measuring business outcomes are key to getting value from AI, why do you think that so few organizations are doing it successfully, you know, as you think about that stat from Gartner? Yeah, I think it's new. And, And, you know, I, you know, this isn't a Gardner thing for me. This is just I think there it's easier sometimes to fear monger into a headline than to build into one that's practical.

19:34And so I'm not arguing with any of the numbers. You know, they always do a thorough and accurate job of sort of what they collect. But to me, it's a balance of what do you talk about, you know, in that. And so, you know, I think it was just over a year ago. It was like April or May of last year. You know, it was the infamous report, you know, 95 % of organizations are failing at AI deployments. And, you know, we look at that and we're like, yeah, it's new. Okay, it's new. Your data is not ready. And this isn't an IT project. This is a business transformation project. And so building these bridges across the technology leadership to the line of businesses and really integrating those teams with forward deployed engineering models and other things that are going to speed that up, the objective that where I see it successful is I'm in a room with a customer or with our teams and I don't know who's in the business or who's in IT.

20:28They're leveraging the tools to increase their technical expertise. the technical folks have they're not having to sit in front of a screen and code nonstop they can engage with the business better to understand adapt and overcome as they're moving through these things so you know the the stat that they threw out in there i was like this is the stat i would have you know maybe i'm overly positive on it but it was they said uh 58 percent of leaders were reporting some positive outcomes from ai right and so i look at 95 percent failure a year in a couple of months ago to 58 % value. And I'm like, isn't that the headline?

21:05Or it feels like it could be. Now, are those an individual who's independently gaining what we would call that workforce AI use case versus something that's in the applied category like Proposal Studio, RFP Assistant, the different capabilities that we built out or some of the use cases we're doing? No, it's a bridge to get there. But that 58 % number to me was the one that stood out. is we're gaining enterprise traction. It ties back to what we were just talking about. As that grows, it's going to grow exponentially at more intensity than the consumer market has. And we need the PowerShell and all the magic from NVIDIA to keep feeding the beast as we build this stuff out.

21:48Yeah, Mike, I'll try not to have hurt feelings as a former journalist about your crack about fear-mongering headlines or anything like this. You would have never done them. But I do think you're right. I do think you're right. Right. And I like the optimism there with the 58 percent mark instead of the 13 percent. We are getting short on time here. I want to get to one last story that I think is relevant to a lot of the stuff that you've mentioned about flexibility and optionality. The information reported that Salesforce is preparing to charge customers when AI agents built outside its platform, including those powered by the likes of, you know, Claude or OpenAI, when they access and perform work in Salesforce applications.

22:25starting as early as November. According to this report, customers could be required to register those outside agents with plans to charge based on their interactions with Salesforce systems. And the company hasn't disclosed the actual pricing yet. But for decades, enterprise software has been more about, you know, seat based and licenses. Salesforce, along with other software providers, they appear to be looking for ways to monetize the work agents perform, as well as, if not in place of humans. I mean, Mike, we've talked a lot about tokenomics and the complicated relationships between model providers, cloud platforms, and enterprise software companies.

23:00What happens when enterprises also have to now start to think about paying for every action an AI agent takes inside their existing software platforms? Does that fundamentally change how we calculate the ROI or how we think about value? Yeah, I mean, they're going to have to price that to make sense or they're going to, you know, it's a fine line. And it's not just Salesforce, you know, that, you know, I mean, it's the larger software as a service and general, you know, traditional software markets have to contemplate when you go from, you know, per seat licensing and organizations that are primarily comprised of human-based usage.

23:38How do you intend to price or account for that. The other side of that coin is, and I don't believe this necessarily, but there are people out there who'd say, oh, well, you can just one shot Salesforce and go and completely displace them. That's not been our experience. And while it's easier to integrate with or build agency that interacts with it, all those things are true. But I think these companies are smart enough, one, to signal that and then start maybe getting some feedback on how should that be constructed and priced. But to me, it's if you are a cost center or P &L owner, you're going to have a line item or a piece within your P &L statement that's going to include agents and the cost of agents in that.

24:28And so I'm more balanced on, okay, yep, they're going to work. They're going to find a pricing mechanism for us to be able to integrate with this, number one, we have to be diligent on designing for efficiency. You know, it shouldn't be in some recursive loop unless there's really value in doing that. And, you know, you already have data that you're potentially extracting from those. Is it cheaper to pay for the data extraction, throw agency at it? For some use cases, probably. For other use cases, you know, you're just going to be able to use, you know, use whatever licensing device that they come up with.

25:06But I think this is a signal. I think they want to collect some feedback from their customers on it. And then I would just say as a P &L owner out there, you need to be thinking about the mix of human and technology agency that's going to exist within your organization and finding the best of both of those. Well, Mike, as always, we appreciate the time you share each and every weekend with us. We'll be talking next week. I believe it'll be close to, if not, on Halloween. So get that Halloween costume ready. You got it. Thanks, Brian. Take care.