Subscribe
Share
Share
Embed
Demystifying the conversations we're already having here at RRE and with our portfolio companies. In each episode, your hosts, Will Porteous and Raju Rishi, will dive deeply into topics that are shaping the future, from satellite technology to digital health to venture investing and much more.
Raju: Hello, listeners and viewers. Welcome to another episode of RRE POV. I’m Raju Rishi, and I’m joined by my partner, Will Porteous, and today we’re going to discuss the growing pains of AI. There is no question that AI has already had an incredible impact across the world and promises to have an even greater impact, but there’s definitely been some growing pains, and we’ll unpack some of those today and give you our take on how they’ll evolve. So, there’s sort of four or five major areas where AI has been constrained.
And you’ve heard about some of these, and probably all of these in the news, and we’ll kind of unpack each one in order. So, the first one is hallucinations; the second one is cost, the cost of running AI; the third is talent bottlenecks, which we know it’s like everything in life is a people business; the fourth is sort of copywriting and legal bottlenecks; and the last is enterprise adoption just being harder than expected. So, if that’s okay, Will, I’ll just, you know, move forward.
Will: Sounds great. I know we’re both excited about this discussion because as we look bottoms-up through the RRE portfolio, we see our companies confronting these realities as they’re scaling. And they’re clearly poised to unlock a tremendous amount of value both, for their customers and for their investors, but these bottlenecks are what is pacing the penetration of AI and the success of the category right now. So, it’s a great list, Raju, and let’s get into it.
Raju: Yeah, let’s get into it. I think it’s going to be fun. I mean, I just got to say, like, you and I have had the pleasure and the hardship [laugh] of having lived through, you know, four inflection points that have just been so powerful, I guess, but also so chaotic. And every one of them was the same, where people were just like, well, this PC just came out, and the transistor, and then, you know, the internet, [unintelligible 00:02:13] AI and mobile, they’ve all kind of created, you know, consternation. But they get there in the end. And some of them take longer than others, and some of them have a bit more impact upfront and a bit less downstream, but they all have had, you know, massive tectonic shifts in human behavior. With AI, let’s start with hallucinations first. You and I both know, models can sound so convincing.
Will: [laugh]. Well, I, you know, all of my responses seem to begin with, “Well, you’re a 53-year-old man, so…” [laugh]. That’s what ChatGPT and Claude preface everything with. So, yeah, I’m right there with you. You know, I think we hear a tone of certainty in what we get out of a lot of AI that masks a degree of well, in some cases uncertainty, in other cases confusion.
Raju: Yeah. It is a lot like my wife tells me, “You speak as if you are 1,000% confident, [laugh] even though it’s a wild guess.” [laugh]. So, I think AI has been emulated behind me. So, they have a habit of periodically providing the wrong answer, and this is especially problematic in law, medicine, finance, and enterprise workflows.
And you know, consumers have a little bit more latitude, you know because we’ve lived through generalized searches on the internet and seen some links that just don’t provide the right answer, so there’s a little bit of fallacy there, but you know, we do want accurate answers, right? We want to be able to trust it. You know, I don’t—do you experience this? Like, when you do a search using AI or a request, you make a query using AI, do you validate it, or do you just sort of assume it’s accurate?
Will: Increasingly, I validate it. I look for other independent sources. And this touches on kind of one of my favorite topic areas historically, which is sort of accuracy and certainty and quality of information. When is something that’s presented as a fact actually a fact? And we live in a world where the large language models that we deal with—at a consumer level, at least—are incapable of saying, “I’m not sure, I don’t understand. It’s uncertain based on what I know.” [laugh].
So, to your point earlier, like, they come at it with a degree of 1,000% certainty when they’re probably well less than that. And so, I think it is important for people to find, to seek external confirmation on a lot of things that they’re starting to rely on the AI models for.
Raju: Yeah, have you seen this Instagram, he produces a bunch of Reels, it’s called Husk.irl. You got to check it out. He’s kind of hilarious. He bullies the AI into making mistakes and then, you know, he’s recording it.
And he has some where he asks, you know, like, the AI, you know, “How many R’s are in strawberry?” And you know it says, “Well, you know, there’s two R in strawberry.” [laugh]. And you know, like, he’s like, “Two Rs? You sure there’s not more than two Rs?” He goes, “Yeah, you’re right, there’s, like, five Rs.” And then it just keeps going back and forth. And you know, he’s like, “Seven Rs,” and he’s like, “So, you’re sure? A hundred percent sure? Seven Rs in ‘strawberry?’” [laugh]. It is hilarious. You got to, he does one where he, like, pretends like he’s dying in quicksand, and he asks, you know, the advice of the AI [laugh] on what to do. He goes, “I think I’m sinking. Only my head’s over.” And the AI goes. “I got you.” You know? It’s like—[laugh].
Will: Your AI buddy there to help, but certainly not sure [laugh].
Raju: Yeah. Yeah, you’re not going to call it in a 911 situation, that’s for sure. So, anyway, there are things that the industry is working on, and I just want to kind of point that out to our audience. The rollout of technologies like AI, we’re going to look back, you know, ten years from now, or you know, maybe even shorter than that period and we’re going to be like, “Okay, wow, I can’t believe it made those kinds of mistakes,” you know? And but it takes time for technology to sort of get to its end game state, and we’re kind of early innings on sort of peeling back the onion and getting exact answers all the time.
But there are six or seven different tech capabilities that the industry is investing in. So, first is, like, shifting from general intelligence to verifiable intelligence, and what I mean by that is just citing sources and checking itself. So, we’re seeing some of that being rolled out. A second is sort of multi-layered architectures, where—you guys know what RAG is, it’s Retrieval Augmented Generation—RAG, you can anchor on vetted data. It’s almost like, you know, the internet when it went back and says, like, this is true, and this link is real. It’s a government-provided link, and so you know it prioritized that higher. That’s how PageRank was created for Google. And so, using RAG can be helpful.
Reinforcement learning is absolutely being tweaked, and what I mean by that is you actually get rewarded for not knowing an answer. And this is actually kind of an interesting nuance. The AI wants to give you an answer because it’s trained to say, like, let me give you an answer, but more and more it’s getting rewarded for giving when it doesn’t know, don’t make it up, you know? Let’s just say I don’t know, or I have a 70% confirmation or not, you know? And we’ll get into some of our companies that are sort of facilitating and helping this.
Another model is iterative multi-pass workflows, which basically are taking an iterative approach to giving you an answer as opposed to just, sort of, a single pass. Multi-agent consensus; we all know about agent architectures and how you can have five or six agents running concurrently to solve a problem, but you can also have multiple agents trying to solve the same problem. Not parts of the same problem, but the exact problem. And then constrained-knowledge graphs. So, you know those are some of the tools that are being worked on.
I still think it’s going to take some time for them to go through, but I do want to point out a couple of RRE portfolio companies that are helping this along. And one of them I’d love for you to talk about, Will, which is our investment in a company called Prime Intellect. So, maybe you could share with our audience on that one.
Will: Well, sure. I mean, people should think of Prime Intellect as a leader in reinforcement learning. Essentially, Prime Intellect is making available vast quantities of GPU capability to support reinforcement learning. This is that notion of the AI recognizing what it doesn’t, in fact, know what it is only perhaps 70 or 80% confident on, and exposing gaps iteratively to basically move towards a better standard of awareness of its own capabilities, and then fill in those gaps over time. So, one of the powerful approaches to solving and progressing AI, and as Raju points out, I mean, this is one of the core capabilities that’s making AI better. So, we’re proud to be investors in Prime Intellect. And they’re seeing just breakout growth as demand for GPUs and reinforcement learning skyrockets.
Raju: Yeah. I think this one’s going to be a really powerful one, Will, I agree with you. I mean, RL is, you know, if you’re rich in the AI world and you’re one of the hyperscalers, you can get access to it, but there’s a bunch of companies and enterprises that don’t have access to those models and those tool sets, and Prime Intellect exposing that and giving people accessibility to it, I mean, that’s a game changer. I do want to mention one of our other portfolio companies, called Farsight, which I think you guys have heard a podcast about earlier. It wasn’t a video podcast, but it was just the audio one.
And what Farsight—just as a reminder to our audience—is, they are effectively enabling a bunch of financial institutions, starting with M&A banks and moving to, like, hedge funds and wealth management organizations, and you know, private equity firms to basically streamline their operations, their output operations. And M&A banks, you know, we all know they create something called a CIM, which is a Confidential Information Memorandum, when they’re helping a company get sold or merged, or you know, whatever, and that’s a long exercise. It’s like, a hundred-page deck, and it has to have constraints around it, in terms of, you know, if it’s Goldman Sachs or Deutsche Bank or others, you know, Deutsche basically says it needs to look in this framework. And those decks take, I don’t know, a month or two to create because it acts as the data room from the company that you’re trying to do an M&A exercise on. And Farsight does it very, very quickly, it basically can produce something super rapidly, which is a first iteration of the results.
And they’re not all correct, but one of the things Farsight has been doing since, well, not day one, but a very, very long time ago where other models don’t do, is give you information about its sources. So, rather than sitting there and saying, “I have to recreate this chart,” it gives you a link and says, “This is where all the data came from.” And so, you can just click on the link and verify it. And it gives you the ability to sanity check the work just as fast as, you know, getting the output. So, these are tools, and companies that are playing a role in certain verticals when they do things like this in an enterprise capacity, and they provide the links, you shouldn’t trust it. You shouldn’t trust that the output is going to be perfect, but you have to have a fast way of verifying it, and by providing those links, and you know, you’re going to find that their precision is damn good, but you know, if you want to check it, it’s a very quick ask.
Will: I think you’re capturing one of the, you know, critical professional skill sets of the modern era, which is, how do we as humans check the accuracy of the result that the AI gives us? And Farsight’s way ahead in understanding what people need in 2026 to get comfortable with the output. And that quality process in how we executed and how much, as humans, we execute it over time, I think it’s going to remain one of the evolving questions in this category.
Raju: I agree with you, Will. I mean, I think it’s a really—there a lot of people are just going to take the results and go with it. In certain businesses, you cannot do that. Financial services is absolutely one, healthcare is another, and you know, legal is another. You have to be able to provide those links. And you know, when you do a hundred-page PowerPoint deck in, you know, an hour or less and you’re creating the final end product, not having those links means the work isn’t actually that valuable [laugh] because you got to go manually recreate it.
Will: Yeah, and we’ll talk about this more later, but for those AI companies that are supporting national security decision-making, you’ve got to go a step beyond because there are lives on the line, critical strategic decisions are going to be made based on the output you’re giving, and in many cases, you’re replacing the role of a human or an analyst professional whose judgment you know might have been refined over decades.
Raju: Oh, that’s a super interesting point. That’s amazing. All right, I’m going to move to the second core area, which is cost. We all have heard, you know, there’s, like, I don’t know if there’s rioting, but there’s definitely picketing: I don’t want a data center in my neighborhood, you know, power utilization, my energy costs are going to go up. This is a real issue. It is a real issue, and there are, you know, technology shifts and things that we can do, and that are being done, but we have a massive cost issue associated with AI.
And one of them is, like, we have GPU shortages and supply bottlenecks. That is a reality, you know. I think everybody sits there and says, “I can’t get the GPUs that I need.” And I know one of the great things that Prime Intellect does is it gives you access to GPUs on a rented basis when there’s spare capacity. I love that. But you know, it’s a massive issue, and we’re hearing it all across the globe.
There are things that are being done to, sort of, repair it a little bit, you know. We’re shifting away from the hyperscalers to sort of Neoclouds like CoreWeave, but just as importantly, we’re doing some math improvements, you know? Things like quantization and pruning, where you’re stripping away the non-essential neural pathways, you’re basically skinning down the problem set that you’re trying to solve. Knowledge distillation, which is training what they call student models that use the outputs of the teacher models. I think that’s, you know, we know a little bit about what happened with, kind of, the Chinese model that came out that’s freaked everybody out. That was using basically knowledge distillation.
Mixture of experts, like, designing modular architectures that only activate specific subsets of a model for a given task, and that drops the active GPU load, you know, per request. So, there are things that are being done. I don’t know. What are your thoughts on this? You think that this problem—because you know, Will, you and I are from a networking era, Will, and I once—I was at Bell Labs, and they said to me, like, “How much bandwidth is enough?” And I said, “It will never be enough.”
And people were like, “What are you talking about? Like, if you got this pipe, you could stream the entire Library of Congress.” I said, “Yeah, that’s text. What happens when you move to images? And then you move to videos? And then you move to real time? And then you move to holography? And then you move to, like, virtual reality?” There’s just going to be more and more and more. Do you think we ever get to a point where it’s like the GPU accessibility is enough, or do you feel like it’s just a long, long way away?
Will: I think we’re going to find ourselves comfortable with smaller result sets and smaller models. And you know you’re right to point back at the networking era and the trends that you talked about are what gave us compression, they’re what gave us encoding, they gave us necessity for all those things, and yet you talk to any protocol engineer about the web and they will say, well, you know, http is really chatty [laugh]. And chatty protocols sold a lot of boxes, sold a lot of equipment for decades, and built a lot of market cap, and a lot of successful companies. I think it’s good to hearken back to that, Raju, because I believe—and I thought what you said earlier about the refinements in the architectures was so important—I feel like we’re in this moment of peak projection, and maybe I’m wrong, but I feel like we’re in a moment of peak projection about our energy and data center consumption needs. And you know, many of the forecasts have data center energy consumption today. I think it’s at about 2.5% to 3% of total US energy consumption, and it’s expected to basically double in the next three to four years with data center build-out.
So, what comes with that? You know, what I think comes with that, actually, is the implementation accelerated of a lot of the methodologies that you described, plus the acceleration of other energy capabilities to bring down that number. And you know this is where the best architects will really distinguish their companies.
Raju: I think you’re right. There’s a lot that is going to happen because there’s a lot of really smart people thinking about this problem. There’s just not enough GPU capacity. Forget about the energy for a second. Just GPU chips, you know?
And there’s people developing a whole litany of, like, very narrow-focused GPUs that are specifically around certain functions and we know the hyperscalers are all launching, but I think that there’s going to be overlay capabilities that you just have to get better at math. You got to get better at math, you got to create a different type of architecture. I’m kind of excited by a couple of companies that we’ll talk about soon in this area. But the cost, the GPU shortage is real, it’s very expensive, and I think we’re going to see some, you know, shifting in the landscape. But let’s crack all of the elements of cost, and then go to the companies that I think we should mention that are in our portfolio.
The second area of cost which is problematic is just cooling constraints. And even if you had all of the GPUs that you needed, and you had all of the energy, which is the third topic, but let’s say you had all of the energy and all of the GPUs, they’re going to burn up, man [laugh]. They’re going to burn up. And I know you were looking specifically at a company that was dealing with this problem in an innovative way for space. Because space has its own, you know, issues, and you are our resident space expert, so I’m going to let you kind of describe this one because it’s kind of cool. We’re going to have data centers in space, but there’s a cooling challenge.
Will: We are going to have data centers in space, and we’re going to have a lot of other components of economic infrastructure in space that are going to support life on earth. The beautiful thing about space as an operational domain right now is it’s not that hard to get there, it’s not that expensive, and you know, as a society, we launched 300 times last year. We will launch multiples of that in 2026. And once you get there, the network connectivity is really good, less than ten milliseconds to ground via Starlink and other things. And so, when you think about putting a computing asset up there, like, a data center, or some core set of GPU capabilities in the cooling problem, well, it’s a fascinating domain because in the right orbit you have perfect solar energy, 24 hours a day. The sun never sets in the right orbit, and you don’t have to deal with the diffusion of energy that happens as solar energy passes through the atmosphere.
And you have a near absolute zero operating environment, but it’s a vacuum. And so, to operate in a vacuum in space where you’re generating a lot of heat, you need a way to dissipate that heat, and it turns out for the companies building data centers in space, radiators actually the ability to distribute heat and move it away from the GPUs is a core area of innovation. This is fundamental thermal science at the forefront of the AI-in-space data center revolution and you’re going to see a lot of IP created around that problem. And you know, I want our listeners to hear my conviction that we’re going to see operational data center assets in space supporting important AI-related work within the next three years. So, they’re coming.
Raju: That’s cool. I love the specificity of three years, and that’s a cool, cool problem. I love it. There is—you know, just for our listeners—there’s a handful of things that people are working on. One is, you know, direct-to-chip liquid cooling versus air cooling.
I remember this, I know this super well because I was a gamer—I’m still a little bit of a gamer, not as much as I used to be—and liquid-cooling GPUs that went into the computer, not only did they look cool, but they really kept your GPU your computer from blowing up [laugh], you know, literally blowing up. And man, I would grind a lot of stuff back in the day. So, that’s happening. There is this immersion cooling, which is pretty cool. I read about this, where you’re submerging the entire server blades into special dielectric fluids that conduct heat away from the electronics thousands of times more efficiently than air without causing electrical shorts, so it’s got to be sort of an oil-based emulsification.
And then the natural, you know, one is a geographic relocation. Go to Iceland, you know? Go to Greenland, use cold weather and cold water, right? So, that’s going to happen. The energy constraints are known, right? We know we need more grid.
There’s answers to that. I don’t think we should spend a lot of time on this one, but nuclear, natural gas power, solar, that you mentioned are all natural candidates for this to, sort of, alleviate the log jam. What I’d love to just sort of move on to is just a couple of our companies that have cracked a little bit of the code in terms of this cost model. And I’d love for you to talk about Lovelace, if you don’t mind.
Will: Oh, sure, I’d love to. So, you know, Lovelace is really bringing forward enterprise AI engines to power decision-making at an absolute massive scale. And this is an example of one of those companies that architecturally is taking an approach that’s giving it just an incredible efficiency advantage. The founders of Lovelace built a lot of massive-scale infrastructure at Google. They are also some of the originators of knowledge graph technology, and Lovelace is elemental, is essentially an enterprise context engine that was built to enable joining millions of data points across petabytes of data in mission-critical environments like national security and financial services.
And it does that, it does that, frankly, delivering about 1,000x the investigative capability of the large language models for 1/1000th of the token cost. And you know, architecturally it was designed from the ground up around those, kind of, efficiency parameters. And its focus is on very, very specific problems where mission-critical decision-making is often associated with lives that are on the line or critical pieces of global infrastructure.
Raju: Yeah, I love that one. What a great company. And the CEO is just a marvelous human being. Actually, the entire team. We have one that I—we really can’t talk about quite yet. You know, we’re in the midst of closing it, but it deals with a lot of this challenges by pre-indexing a bunch of stuff in a data set, a large data lake, which creates enormous efficiencies in token utilization, like, a 90% reduction in token utilization. It creates incredible accuracy levels because the index allows you to look at data that sits in disparate parts of your data set, but creates linkages to one another so that the answers that you’re getting are incredibly accurate.
And I’m going to be super—actually, we’re all going to be super excited to have them on this podcast to, kind of, talk a little bit. But I don’t want to get too ahead of our curve on this one, but there are companies out there that are thinking about this mission, which is to solve a lot of the constraints that we have in terms of cost, which will be a big limiting factor. And you know it may not be—if you use current cost structures, I think people are really kind of a little bit delusional in certain use cases. And they’re just sitting there saying, like, “Oh yeah, it’ll just get faster and cheaper over time.” And if it doesn’t, you know, you probably don’t want to fire those 30 or 40 people or 10% of your workforce because the AI is not efficient yet at solving those issues. But—
Will: I think you’re making a really important point, and I want to expand on it because our listeners should think about that level of uncertainty, and not just decision-making about retaining teams or not, but also decision-making in the financial models that are underwriting a lot of the physical world build-out to support AI. I mean, I would like to know what the assumption set looks like on the part of real estate developers and others as they think about the efficiency of AI 5, 10, 15 years out from now, as they wade through permitting processes and construction processes and that sort of thing. Because those assets are going to come online in a different environment than the one we’re in right now.
Raju: I agree. We’re going to touch on this in enterprise adoption, which is the fifth topic. It is important. It is important. Like, the ROI that you are predicting has got to be tangible, and we’ll kind of get there.
But let’s focus on talent bottlenecks, which is the third issue. You know, limited resources, limited number of people that really totally grok this, huge comp packages. You know, it’s insane, kind of like what we see certain people getting in this ecosystem. And you know, like everyone in the world—well, actually, like, you know, the startup mentality, the venture mentality, the, you know, technology disruption model has always, you know, basically said, like, certain companies, what you thought was an incredibly high valuation, you could have paid a hundred times more and still gotten value out of that. And that mindset, unfortunately, there’s a lot in the lot of people in the venture industry, and a lot of people in this technology ecosystem that say, you know, it’s okay to pay somebody $100 million signing bonus, it’s okay to pay, you know, a trillion dollar valuation for a particular company because it might be, you know, $100 trillion down the road.
At some point… that math doesn’t work. And, Will, you and I talk all the time, and we do as a partnership, about what is an acceptable valuation, what’s an acceptable ownership threshold in this new era, and we have been through a handful of technology inflection points, and seen, like, the hype curve and the J curve, and you know, like, things tanking, and I got to be honest with you, we are very disciplined investors because we know it’s a long game. It’s not going to be one in inning… 0.1, like, the first out of the first inning, like, and let’s just put all the chips on red, you know? Like that’s not the way it works. So, we’re very, very disciplined, but like, it doesn’t feel like the rest of the world is being very disciplined here. They’re just rolling infinite amounts of money on this. And I’m sure you see it in some of your companies. I definitely see it in some of mine.
Will: No, we do. I mean, your point earlier about, kind of, the four technology inflection points in our lifetime, well, in each one of them, we witnessed asymptotic expectations around a handful of companies. And we’re seeing that right now, and we’re seeing that in the flows of capital to certain companies. And when capital wildly outruns the growth and performance of a business, you get this kind of dissonance that can be really disorienting about what actual value creation looks like, disorienting for employees and for entrepreneurs and other would-be entrants. You know, the antidote to all of that is it turns out to be sales [laugh] and growing revenue and generating profits, and we’re certainly all for growth and deferring earnings in favor of growth over time in our portfolio. But you have to believe that you’re actually building a business, and there is money to be made by owning a portion of market leaders, but at RRE, care an awful lot about the price we pay going in.
Raju: Yeah, we have to be because you can’t bet on all of them. You may as well just throw money at every AI company in the planet and hope that—you know, this is [laugh] this is akin to, you know—you and I talk about this all the time—but like, internet, you know, people poured billions of dollars into Excite and Lycos and Alta Vista and Ask Jeeves and Yahoo and Netscape and Mosaic and Firefox, and the paper that was out there that said, you know, “Search engine wars are over. Alta Vista has won,” a year before Google came out. And so, people are gobbling up talent, paying infinite money for companies and valuations, but we’re early innings. And I think our listeners should understand that there needs to be a little bit of patience here.
I have a sort of mental model on how technology inflection points actually materialize, and what typically happens—and you saw this in PC and internet and mobile—is the first wave of, sort of, scaling companies and revenue enhancing companies are services companies. They’re like, you don’t understand the web. We’ll build a website for you. And we saw [IXL and Science and Viant 00:32:46], and for mobile it was Zephyr, and PC had its own, you know, sort of—databases had their own sort of contracting companies. The second wave—and those services companies kind of flatten out when they get some maturity in the talent bottleneck, and you’re able to get more people.
And the second types of companies are tools companies that say, “Hey, look, there’s a bunch of data migration, data cleansing, you know, API access, models,” not sort of LLMs, but like, even smaller models that are going to clean up the environment, you know? Get all your data into a data lake, get it normalized. All of that is going to be required. There’s a bunch of tools companies that get created, and the third wave is, you know, verticalized applications, things that you take a narrow slice, you know the ROI is there, it is really clean cut, and those companies scale. And the last companies to actually, like, hit that threshold are platform companies that kind of incorporate everything. Think Oracle, think Microsoft, think you know, sort of like Apple, where they own the entire ecosystem, they have services, and they’ve got tools, and they’ve got apps, and they’ve got a platform.
And so, we—there’s a lot of belief in this world that in AI it’s backwards: that the platform companies are going to be the winners. And I disagree. I actually disagree. I think that it’s going to follow exactly the same trajectory that the other inflection points went through. And people who are sitting there saying, like, “I’m going to put a lot of money in the winning platform,” there’s Jeopardy conditions there, right?
Like, Deep Seek was a Jeopardy condition for a little while when they said we could train the model for 8 million bucks. And we know they used one of the tools that we were talking about before. But you know there could be another one out there that you know sort of obviates some of this. And so, there’s a rush for these companies to create an IPO environment to justify a trillion-dollar valuation, but man, it’s hard to justify a trillion-dollar valuation at the revenues that some of these companies are having. So, I think, you know, that at net-net is, you know, as an enterprise or as a startup that is saying, like, “I got to pay a trillion dollars for this engineer that is an AI native engineer.”
And you know, an enterprise that’s saying, like, I need somebody, you know, I can’t afford to get the right talent in place, I think services is actually going to be a very, very interesting stopgap where you can get services talent to fulfill the goals, much like the IXLs and the Zephyrs and the Science and Viants of the world did in the prior technology inflection points. And actually Sequoia wrote an interesting article, a blog post, about how you know, for every dollar spent in enterprise software, six is spent on services, and they want to find services companies. So, anyway, that’s interesting.
Let’s move to the fourth, and then—you know, quickly fourth—and then after that, kind of end on the fifth. But the fourth is copyright and legal bottlenecks, you know? So, I’d love for you to talk a little bit about some of the challenges here and what needs to get, sort of, solved because we haven’t put a stake in the ground. No government has yet. But I’d love to hear your thoughts on it, Will.
Will: Well, it’s an understatement to say that this is a dynamic area. You know, I think those who follow the chatter coming out of Hollywood are very aware of the concerns about ownership and copyright and likeness that the generative AI world has spawned. But I think that that actually represents only a narrow set of the critical questions here. I think data governance, data ownership are increasingly an area of strategy for both individuals and for companies. And rolling forward, I think we’re going to see a tremendous amount of emphasis placed on this topic.
If you just think about how large enterprises will rely on AI to make critical decisions in the future, they’ll need a very coherent framework for external data that they rely on, internal data that is proprietary in some way that they consider a strategic asset, even a trade secret, internal data that they are willing to license and share externally on certain terms, and that which, of course, they won’t. Data governance and sophistication in this area, I think, is really in its infancy compared to where things are going in the future. And as we at RRE look upward through our portfolio companies, some of which rely on massive data sets that they own outright that are totally proprietary, we see the incredible competitive advantage that they’re getting as they scale their businesses. And this topic is going to become more and more visible, it’s going to move higher and higher on the agenda of corporate leadership and boards because data governance is ultimately going to be corporate governance going forward.
Raju: I couldn’t agree more. That was well said, by the way. And our last topic is really enterprise adoption is harder than expected. We got security issues, we got poor ROI measurements, we’ve got pilot projects that don’t scale. And I will tell you this is indicative of every inflection point we’ve ever had, right?
What is my justification for my website? What is my justification for my mobile app? What is my justification to move everything into a digital PC environment or mainframe environment versus, you know, human-driven behaviors? And that takes longer than people think. It does.
And so, I think we’re going—we have not seen the trough, right? There’s still so much hype around AI, but we’re going to see it because there’s a lot of people sitting there saying, “Well, like, listen, I can’t have my data leave the enterprise. I can’t have vulnerability put in place because I have AI. How do I deal with AI making decisions, you know, that are improper, that put me into legal, you know, battle?” Or whatever it may be.
And so, you know what I say to that is, you don’t drop your AI projects, don’t drop them. There’s a reason you picked them in the first place, right? But think about the model. Think about the model that you know I was portraying a second ago, which is services, sort of, are the leader in a technology inflection point, followed by tools and verticalized applications, and then lastly it’s the platform. Betting on the platform, sitting there saying, “I’m going to drop Anthropic into our environment—you know, like, Claude—and I’m going to drop ChatGPT and OpenAI, I’m going to drop Gemini into my platform and then we’re going to figure out what to do with it,” that is not really a great model.
If you look at the past and you believe the past, which I do. I happen to be old enough to see all the good and the bad—and you too, Will—the model really needs to be, you might have picked a couple of target zones. They were appropriate, not, like, inappropriate to think about, like, my call center and my customer service piece, and you know, maybe some sort of like automated billing. You think about services. We have a bunch of verticalized AI companies.
We’ve invested in those companies because the ROI is real, it has a very fast onboarding process, and you know, basically the time to value is near immediate, they are experts in their domain, they don’t make a lot of errors, and if they do, they give you visibility into it. You know, Open Envoy, which is a verticalized, you know, accounts payable-accounts receivable platform, finance platform, Farsight, which deals with the M&A and financial aspects, a company called VoiceRun, which deals with enabling your voice agents in your enterprise. You know, you want to swap over to a voice AI agent as opposed to a call center agent, you want to do this in a way that is more sophisticated. And all of those companies, not only have ROI, but like, in the case of VoiceRun, it is the services element that’s important. They give you ability to pick the LLM, pick the APIs, pick the various, you know, voice synthesis tools, voice AI agents, voice creation agents, whatever you want throughout the stack, but give you the comfort of knowing that their framework has vetted it all, and it’s a working solution. It’s the hybrid—perfect hybrid—in terms of, you know, your voice agent strategy.
And Farsight, and you know, all the ones that I’ve talked about, Open Envoy, and others, you know, are very similar. So, I really think, like, we’re going to see a world where, like, we don’t have the talent, the energy is expensive, you know, we’ve got, you know, issues associated with, like, cost of creating AI, and it hallucinates. And so, you know, I could very easily see a bunch of enterprises sitting there saying, like, this project doesn’t scale. It doesn’t work. You know, we put a bunch of money into it, it doesn’t—and I don’t think people should abandon the end game. I just think people need to be very, very realistic about the curve that actually happens, and the way that you do it.
You don’t have to sit there and be like, we’re going to build our own app, and you know, the world is simple, and Claude makes it work, and everything like that. I think you’re going to need a bunch of this stuff that we talked about, this verticalized AI applications, and then services that are going to support you. And it’s no coincidence that both Anthropic and OpenAI have poured a lot of money into forward-looking engineers and services organization because they know. And I think people thinking that, you know, the platforms have just sort of locked and loaded, it’s like the same thing as saying, you know, search engine wars are over, Alta Vista has won. And this is early innings, my friend. Early innings.
Will: That’s very well said, Raju. And I think it probably is reassuring to some of our listeners that they can look back on past transitions that they’ve been a part of and remember that moment where the emergent companies seemed like they were going to dominate every aspect of the new application, and yet they struggled to apply the technology to deliver value. We’re right in that moment right now, as you said, and it will be services and much more focused applications that—and applications delivered by application-specific companies that provide the breakthrough for enterprises to really begin to get value.
Raju: Yeah, I couldn’t agree more. So, anyway, I’ll let you close it out for our listeners and viewers, Will. But this was a fun one.
Will: Thanks, Raju. Thanks for your leadership on this topic. You know, listeners, at RRE POV, we strive to give you a sense of the kind of conversations we’re having internally at RRE all the time. It’s this kind of candid assessment of where are we really in the AI cycle that we really are talking about here, and we’ve given you a taste of it in the last 45 minutes. And look forward to our next episode, and hope that you’ll tune in with us again in the future. Thanks.