Anthropic's $40 Trillion AI Ambition and the Race for Control

0:00One of the most anticipated headlines of the year, Anthropik's IPO filing, offers a revealing look at the business of AI from a$40 trillion view of knowledge work to rapidly advancing cyber capabilities. Meanwhile, both Anthropik and OpenAI are looking to deepen their hold on how the enterprise works, and Google is having a bit of a TPU moment. What should enterprise leaders make of those moves and what decisions do they change? As always, we're joined by Mike Taylor to help make sense of all of that. It's Saturday, October 31st, and this is the AI Proving Ground podcast from Worldwide Technology.

0:33Let's jump in.

0:42Okay, Mike, Anthropic's public IPO filing was certainly the big story this week, and I want to start with what it tells us about the company's ambitions versus the ambitions of its customers. Anthropic estimates that knowledge workers generated more than$40 trillion, that's trillion with a T in economic value globally last year. And the frontier model provider thinks AI will expand that massive output in the coming years. And this is where it gets important. They want to participate in that growth. There's an interesting article from the Wall Street Journal that really articulates how they arrived at that figure and how the calculus behind total addressable market has really evolved over the years.

1:20But their math does lead to a much bigger conversation, I think. Most businesses probably start with a budget for AI tools and subscriptions, but increasingly the pitch is about changing the cost of an entire business process or doing work the company couldn't afford to do before. In other words, businesses bring its people, data, and its knowledge of the business to making AI useful. But according to Anthropic, it sounds like AI suppliers are now pursuing a share of that output as well. That was really long-winded, so I'm going to get to you now. How should leaders actually be thinking about how much of the resulting value of AI belongs to the technology provider and how much of it should be retained by the business?

1:57Yeah, one, it validates a huge market that we can go after here together. And those efficiencies, I think, are in large part still to be recognized within our enterprise customers. I think the consumer technology theme moving from, you know, things that we're playing around with into what we're calling kind of workforce to apply to AI solutions within our customers. I don't know, you know, 40 trillion is the right number, but it's massive. And I think very little of it is going to be derived from kind of borrowing or stealing from other markets, if that makes sense. I think we're talking about, you know, real productivity gains on top of things that we're already using.

2:44We've proven out, I think, the SaaSpocalypse or the collapse of those things is not necessarily, you know, completely replaced with or taken over by this$4 trillion market that they're talking about. And then the robotics layer of that or the physical AI layer of it and superintelligence related to that, I think, is another piece of it. So I think about this, the$40 trillion market as a function of market value that they're creating, not just for the model and frontier providers, but what that's going to mean to ultimately, I think that the cashflow and productivity of businesses as they get more efficient to me is the most interesting part.

3:26And we've seen announcements, whether it was going back and these are just things out there in the public domain, but companies that are leaning into this in terms of investment, you see Eli Lilly, what they're doing with AI and what they're doing in conjunction with NVIDIA and Jensen's support, you see what that's done to, you know, their market valuation just as one little example. And I think you could start to wrap your head around that this is tens of trillions of dollars of opportunity, how those pieces come together. I think this is a good framework probably for us to pressure test over the next, you know, three to five years in terms of exactly how it shakes out.

4:04But you have enough examples to point to here that it doesn't seem irresponsible to think that we're tens of trillions of dollars that are out there to be chased after. Yeah. And, you know, these very large numbers are not alone with just Anthropic. I mean, certainly the most recent string of earnings reports that we've seen come in from the likes of Microsoft, from the likes of Alphabet. I mean, these are solid double digit growth figures here. And yet the market is still a little bit jittery in terms of how it reacts. Is that mostly from the, you know, the infrastructure investment side of things?

4:37Or how do you make sense of that? Yeah, you know, we were talking this week, I think, you know, we're in the spend money to make money phase, right? And, you know, a bit of this in a supply constrained market, the commitments that companies are having to make to get access to power and shell are scary, I guess, you know, as you think about it from a market perspective. The part that I balance in that, though, is that, yes, while these companies were, you know, their stock prices were impacted this week, it has far more to do with how the investment side of their business is impacting their cash flow side.

5:14But, you know, again, as an optimist, I look at this and say, look, you know, record growth within Microsoft and Alphabet. Amazon, it's hard to say record growth just, you know, in terms of the overall business. But AWS itself, including these offerings, was up, I think, 39 or 40 percent. So if you looked at it, you said, okay, cash flow and the impact of cash flow based on debt that they're creating and that they're going to have to service into the future is going to require that they continue to grow at these sorts of rates or more. And I do think it is a little bit of a, you buy these assets, you secure these assets.

5:55And Anthropic is an example of this. You can support, you know, their growth quarter over quarter is, you know, 400, 500%. And you start to think if these traditional cloud providers had access to the power and shell and some of the infrastructure that they had out there, they have an opportunity to do the same. Now, they don't have the model platform and some of the other pieces that are out there, but as open weights become more important as the ability to develop new models. And as we look at this space between GPU as a service and traditional SaaS, I think we're going to have a lot of room for very industry and function specific models that are going to get created.

6:38and the Microsoft's, Alphabets, Amazons of the world, along with some of those NeoCloud providers are going to profit from that. They're going to be able to host it based on being able to lean in first on the PowerShell commitments to run the infrastructure. So yeah, I get why the hit is there. We talked about this even as it related, you know, a few weeks ago here to CoreWeave. This is for me far more of a valuation pause perhaps than it is an indication that fundamentally the business model here is broken. and that we're in some massive, massive bubble. Yeah, well, certainly in any case, Anthropik's economic ambition, it's starting to take shape along with others.

7:18Anthropik's IPO set for mid to late November. Meanwhile, another interesting part of that filing shows how these providers intend to become more involved in the work that we're all doing itself. Anthropik described an increasingly integrated business between the models, the development platform, and applications like CodCode or Cowork. I mean, effectively, Anthropic is saying it's providing both the intelligence and more of the tools employees use to get work done. Now, that can make AI easier to deploy, certainly, but it also makes the business relationship with Anthropic deeper and potentially harder to unwind than simply switching models.

7:54Meanwhile, on a similar track, Oracle's giving us another example. It's newly announced AI workspace is designed to let employees describe a business goal inside of ChatGPT, which brings together connected Oracle applications, data, and tools to help accomplish all that. Like the Anthropic integration, this could save employees from navigating a lot of separate systems. But as I already mentioned, it also means our relationship with an AI provider extends well beyond which model answers our questions. Now, optionality and flexibility have been well-used keywords in our discussions that we've had over the last few weeks.

8:28Where does optionality matter most as these platforms take on more of the work? And how do you preserve it without giving up any of the benefits of that integration? Yeah, I think we're, you know, we've seen this with other SaaS providers, I think ServiceNow, Salesforce, and, you know, now Oracle in that front. They're going to have to place some bets as they are with frontier model providers that are going to create frontier-like experiences within their SaaS or ERP or regardless of where they're hosted applications. And so to me, this is an indication of progress. It's an indication that, you know, building a true frontier model versus, you know, hosting systems that provide core value like SaaS and Oracle and others do, they've got to place their bets there.

9:16Oracle's building out massive infrastructure to support these capabilities. There's a two-sided economy here, I think, in all cases with these providers where they're not only consuming those services, but they're building out infrastructure at scale that's going to participate in the development of the models as well that's happening there. And I think the bigger question for us and for our customers are, you have ERP systems, you have data sets that you have out there, and they live in all sorts of different platforms. architecturally you're going to have AI capabilities that reside natively within applications and you also though need to be thinking about an architecting for AI that's going to persist and utilize data across disparate data sets systems that maybe aren't naturally connected into a business process flow within your organization because you want to gain insights you want to create an omni-channel experience for your customer you want to take customers in a maybe a commercial banking sense that you also want to be consumer banking customers.

10:23And so figuring out how you connect those things, I think, is going to be the real opportunity beyond what's going to be embedded into, you know, Oracle's Fusion, NetSuite and other database products that they offer out there. not probably very few that run all of their business on oracle or salesforce or you know service now just you know as examples so lean into that i think you know it's good good for all of us that they're investing in that but also understand i think there's going to have to be a bigger picture and a broader architecture to to get to where we're starting to aspire to think about how we run and operate our businesses yeah keeping with oracle is a little bit of the pull through here.

11:06I mean, these platform decisions, they sit on top of another set of choices, one of them being where does the computing actually happen and what does it cost? Oracle's announcements this week extended into that layer. They announced plans with Google to bring Google's TPUs, those are their purpose-built AI chips, to Oracle cloud infrastructure beginning next year in 2027. Now, TPUs have been used internally by Google for a long time, and some Google cloud customers have had been using them for over the last several years. But this deal would make Oracle one of the first hyperscalers outside of Google itself to offer them.

11:41The information, meanwhile, reports that OpenAI plans to buy TPUs, while Anthropic has disclosed substantial commitments to TPUs as well. For enterprise leaders, Mike, I think there are two ways we could think about this directly, if you're operating your own models or through the price, performance, and availability of the AI services that we might be buying. Are these choices something that business leaders need to be evaluating directly or something that we should expect an AI provider to manage for us? I think we're going to have circumstances where that we'll be using these. At the end of the day, we'll be using these models in some of these clouds.

12:18We're not even going to know what they're running on, and frankly, we're not going to care. Our cost will be a function of the cost of the model and the cost of the infrastructure that's there. And, you know, TPUs are, I think, you know, Google's been very open about this. Their entire Gemini suite runs on it. They have the opportunity between DeepMind and the model development and then the hardware orchestration, figure out how these things operate and function, you know, in sort of a harmonized fashion between the software and the hardware that they're running on. And the piece with the TPUs, I think Cerebrus is in this boat.

12:49You know, NVIDIA is leaning into this as well as we're going to have high performance architecture consumed within these data centers, whether they're ours or they're hosted or they're, you know, NEOs, whatever. But I think more so we're seeing these use cases evolve to where we want or need that compute to be closer to the consumer or the point of data generation, for lack of a better term. And all of these, you know, TPU solutions, the Cerebrus stuff that's out there and the investments that NVIDIA is making are going to give us optionality in conjunction with, but also outside of, you know, the physical AI and robotic circumstances to where I think, you know, it's reasonable to believe you're going to have some form of high performance architecture running inside of a branch, an office, a clinic or surgical hospital.

13:42You know, like that to me makes a ton of sense. And these are going to be options to be able to do it. I personally feel like, you know, Google's delay in bringing the TPUs to market was more just the excessive demand that they had internally to build out Gemini and all the capabilities that have gone into, you know, their consumer and office productivity products that maybe they're at a point now where they're like, okay, we can share some of this with other people. because it's being adopted at a very fast rate now that it's available in the, you know, out to portions of the consumer and enterprise markets.

14:16Yeah, quick follow-up in that regard. I mean, over the last few weeks, you've alluded to a good foundational mindset, which is to choose infrastructure around the work, evaluate it in your own environment, and expect the right deployment mix to change as adoption grows. I mean, more options give customers more ways to match the infrastructure to whatever workloads they're running, but they can also introduce another decision into an AI journey that's already complicated. I mean, some companies are still figuring out where AI delivers value, if at all. Others are running enough of it that performance and costs are becoming much bigger considerations.

14:49If an organization came up to you today, Mike, and asked whether they should be looking at TPUs, what would you need to understand about their business or their infrastructure or their workloads before you could give them some type of useful answer? I think it's a couple things. I think the first is right now, you know, what are the use cases that you're trying to take on would be a big one. The second would be, you know, the TPUs, the products, you know, AMD and NVIDIA. And, you know, they all have products now that are more, call them inference focused. And also, I would say, you know, facilities and infrastructure friendlier, if that makes sense.

15:30And so, you know, if it's taking the big data center providers two years to get a generator, it's probably going to take, you know, me as an enterprise or even some of the Fortune 500 as an enterprise, it's going to take you two years to do that. And then you're also fighting the supply-demand curve with companies where their business is actually providing that power and shell. And so how do you compete? So what we see in most of our use cases in customers, about 80 % of their demand today is more in the inference side versus training. That doesn't mean that won't change. But, you know, as you think about providing inference, you have a lot more options there in terms of the, you know, intensity by which you had.

16:15And some of the questions will come down, would you rather have no inference or would you rather have some inference and then figure out how you sort of retro fit that into data center facilities and spaces that you have, limiting your capital exposure to do so, and now offering, call it on-premise, sovereign to you, your country, your organization capabilities. And I think that that's where a lot of this stuff is going to be headed over the next six to 12 months. The gap between the frontier models and the open weight models, we talked about that continuing to close. We just released this week, I was on with our teams this week on harnesses that we built to try to test these models out against different use cases from development to sentiment analysis to language processing and other capabilities where our teams are pretty impressed with how all models are performing, but particularly just the distance closure between the open and the close.

17:18So all those things to me lead to on-premise in some way, shape, or form. Not everything. Won't be everything. And I think all of these provide those options. Yeah. Well, I mean, in any case, these infrastructure decisions, they're going to depend on where a customer is today and what they're ready to take on next. But, I mean, important to note that time is not always a luxury that we have in hand. And nowhere is that tension more palpable than in cyber. Back to the Anthropic S1. Anthropic said Claude's success rate on ExploitBench, that's a test of its ability to exploit software vulnerabilities, rose from 40 % to 91 % in roughly four months.

17:58I mean, fair to note, we already know that newer models are better at exploiting weaknesses than earlier models. We've talked about automating, patching, and compressing maintenance windows probably on every single episode that we've had here together. I suppose we'll keep hammering that message home until it's less of an urgency. But I want to bring up an interesting wrinkle in this that I heard from Nikesh Aurora, the Palo Alto Network CEO, who was on the All In podcast earlier this week. Nikesh, who, by the way, will be with us next week when we're at the Worldwide Technology Championship. He'll be a speaker at our Business Innovation Summit.

18:32So shameless plug, go and watch that on the Golf Channel. The tournament won't be able to have the Innovation Summit. We'll release those videos at a future date. But he was discussing the recent PHP security incident. He described a customer that could patch some websites while others were running software so old that fixing them required taking them down and rebuilding them altogether. And he warned that incidents like this could become increasingly frequent, putting more pressure on organizations that already struggled to remediate them. Now, Nikesh is describing a different problem than simply the patch faster mentality.

19:04He's saying some systems need to be rebuilt from the ground up. But how do you decide, Mike, or how should organizations decide which of those systems can be protected temporarily even and which can we can no longer wait? We may have to just rebuild or get a replacement. Yeah, he did a great job, first of all. And I think it was a good balance of people processing technology and what they're talking about. because we can instrument these harnesses and even helping some of our customers respond to this PHP issue that they've been dealing with. The first barrier we run into is we can run all these scans and figure out, okay, where does this vulnerability exist?

19:49What we then run into is who actually owns that system? You know, is it something, you know, we had one example we were working on with a customer And we found a whole host of things that we think weren't actually their devices. They were just someone who had sort of attached to their architecture and done some registration in other countries around things. So we can run these things, do these patches, but things like who owns the system? Who has the credentials? Where are they stored? Are they in our secrets password management? Are they off in some other area? And so the hygiene that goes into this, I cannot underscore how good, strong process hygiene just feeds into machine speed like capabilities because you can push the button, proliferate that out.

20:41But you got to have your secrets management taken care of. You got to know who owns these applications. And you're pressed a bit here where if you can't figure those things out quickly, then you're faced with a decision of do you turn it off? You know, do you just have to quarantine it, turn it off and see if somebody says something, which is the worst spot to be in. So I think that's where we are. And, you know, the PHP is an example of that. We had a couple other zero days that came out this week as well on perimeter based technology providers that that's the coaching we're given right now. You know, if you had to do this tomorrow, could you actually get in touch with or find the people who understand what this system is doing and be able to verify that it's functioning properly after we patch it and reboot it?

21:31Excellent. And for those that are interested in what Nikesh has to say at the Business Innovation Summit, we'll be sure to release at least a portion of that session. That should be fantastic. We'll release it here on the AI Proving Ground podcast. Mike, I know we're up on time here, so I did want to be able to let you go. Wish you a happy Halloween. Are you a full-size candy bar giver-outer, or what's your strategy? Yeah, you know, we do all right in the neighborhood. Yeah, yeah, we do all right. We've got – I'm a big Almond Joy fan. Nobody seems to like those, so this is the best time of year.

22:02Yeah, no, I'll bring all mine to the office and hand them over to you because those things are disgusting. But, Mike, appreciate the time, and we'll see you next week.

22:12Thank you.