Real-World AI in Private Equity

Transcript generated automatically from the episode audio and lightly cleaned up (filler words removed, misheard names corrected where clear). Speaker labels were added afterward and may contain errors.

Ryan Eppley: Well, welcome back to AI in New York City. I'm your host, Ryan Eppley.

Rob May: I'm your co-host, Rob May. I'm the CEO of Neurometric. And the AI in NYC show is focused on talking to different people all around the AI ecosystem in New York. So, startups, investors, executives, just people doing interesting things. We're trying to give you a point of view on how AI is in going to impact your life if you're a New Yorker.

Anna Kirk: Hey, I'm your other co-host, Anna Kirk. I'm sales lead at Thread AI and today we have the pleasure of talking to Sam Koch from FeedbackNow about what it takes to actually have success implementing AI in practice. So Sam maybe before we dive deeper we have some burning questions for you but why don't you give an overview of the story behind FeedbackNow. What inspired you starting it?

Sam Koch: Sure. So I started a company out of Harvard Business School where we do AI for private equity funds doing outbound to companies. Sourced over $2 billion in deals for funds ranging from small mom and pop businesses that we helped funds buy to large hund00 million plus businesses. And what we realized is that the processes that we had built in this to actually find and identify companies and reach out to them was the same kind of process that we could use to actually run these businesses more efficiently. And so, along with some partners, including Rob here, we ended up buying a business, FeedbackNow, which is, like to say the buttons in the bathroom at the airport you might see when you go travel through New York, Boston, San Francisco, Miami, any large metro. Those are our buttons. And, we've been transforming the business over the past year to actually improve and optimize it in how we run the business and also how that data is used by companies around the world.

Ryan Eppley: So, when those buttons don't work, now I know who to blame.

Sam Koch: Yeah, call me. [laughter]

Anna Kirk: I feel like that is such a hot topic these days, especially with the likes of like General Catalyst and Thrive starting their own rollup strategies, other fresh MBAs starting search funds. So, what do you think led to your success? And maybe for the people who aren't familiar with some of those terms I just mentioned, you could explain why it's such a big idea.

Sam Koch: Sure. Yeah, there's a lot of concepts right now around buying and building businesses. There's a lot that AI can do to optimize inefficient workflows and really find businesses that in the past haven't been darlings of the acquisition stratosphere. So, in the past, search funds, which is a kind of micro private equity asset, would buy a lot of SaaS businesses, so software as a service. But now the ability for you to buy into like hardware businesses allows for really big margin expansion which is the kind of bread and butter of private equity.

As for like why we've been successful at FeedbackNow I think when we were looking for a business we really thought about kind of two legs of what we wanted to do. One was yeah we can optimize a business with AI but we also want to find a business that is going to benefit from AI expansion. And what FeedbackNow does is provide a level of data. You know, it's the buttons in the bathroom, but what you don't see is how many people are going in and out. When was it last cleaned, how full is the trash can. All of that data creates a model for a physical space that explains what's happening there. And so that data is now very valuable to the facility managers, the cleaning companies, to people who are buying commercial real estate. And so that is all making FeedbackNow a much bigger business than the one that we acquired, which was for all intents and purposes buttons in the bathroom.

Rob May: Yeah. So I was an investor in Sam's company, original company, and that company was doing nicely, but he kind of said like, "Hey, I don't know that it's going to be like mega venture scale." And I said, "What do you want to do next?" And he's like, "I think you could buy companies and apply these AI techniques to improve the margins of the companies." And I was like, "Oh, we should talk about this." Because we had a at the fund I was running, like we had a similar thesis. So it was interesting because I think probably like two years before everybody started doing this, we really went out and bought a company and applied AI and found like you know where it works and where it doesn't because it's like it's not as easy as you think. I mean let's let's talk about that. Like this is a good question which is like why is it harder than it looks and why are so many people struggling? Either they get analysis paralysis around doing the deal because they can't figure out the specifics of applying AI or the people that think like, oh, it's we're going to AI this business up and down and it turns out they actually struggle with that. Like what what are some of the challenges?

Sam Koch: Yeah. I mean, at the end of the day, it's a people problem, right? Like if you buy a business and the people don't want to make change, you're going to be stuck with a business where they're unhappy people and no change has been done. And so being able to like really work with the people and get them to say, "No, I do want to make this change. I'm excited about this." And that really was the place that FeedbackNow was. The company that had owned it before was Forrester Research. Awesome business. They are kind of like Gartner. They do research for large companies, but it just wasn't the right fit. And FeedbackNow really wanted to grow on its own. And so by coming outside and saying, "Hey, we want to adopt new ideas. Like we know that the way we're doing things isn't right, they were in that mode where they were really trying to learn and grow." And I think that has helped a lot. And then I think there's a big question about like whether you centralize that kind of growth of AI or whether you disseminate it and kind of have a more democratic view. We've been kind of focused as a private equity firm on on centralizing it, simplifying the business and really growing it via like identifying really simple to fix problems and then applying the AI to that. And it's been a good way to kind of improve the business.

Ryan Eppley: The Mamdani approach to [laughter] PE and we knew it was coming up at some point. We're going to talk about that later. Yes, he won the election last night.

Well, I want to like anytime I hear a deal go down, can you like walk us through the actual terms of the deal a little bit, whatever you feel comfortable sharing, like how did the first deal really get done? Pen to paper.

Sam Koch: You want to start or you want me to go?

Rob May: So, so I'll tell a little bit of the story and Sam can tell. So, we my first startup sold to our first big customer was MoMA and the guy who ran Momma at the time, the CIO was a guy named Steve Peltzman. Steve was not Forrester and Steve was the general manager of this business and so the first thing we did was we talked to Steve about so Sam and I sort of strategized talked with Steve about this and we went to make Forrester an offer which I will tell you as two guys who had never done anything like this before to go to a publicly traded company and offered to buy a division of their business. Their first comment was like you guys are nuts. Right. But we were we were like really really serious and we started talking to investors.

Ryan Eppley: Were you wearing your Rolling Stone shirt or were you in suit and tie?

Rob May: I was not. Yeah.

Sam Koch: You wore the suit and tie and everything.

Rob May: I was in between. I was not suit and tie but I was not I was not rock and roll.

Ryan Eppley: So I want to buy your company and also have a concert later.

Rob May: Yeah. Yeah. So what I don't know what happened from there? What was your recollection?

Sam Koch: Yeah. It definitely was like a first meeting where they went, "Nah, you're you're kidding." Yeah. But we really kind of came and showed the vision of what we thought FeedbackNow could be to them and where we thought we could really help with this transformation, with investing more capital into the business, with making changes that just their business model wasn't able to do. And so they saw that synergy and they actually are part owners in FeedbackNow now FeedbackNow later [laughter] but they've really been a great partner to us in kind of the journey and the spin out. But yeah, it's been a crazy experience. If if you told me two years ago that you were going to buy a business that had international operations, worldwide deployments and I'm thinking about an appastal.

Rob May: Yeah. Sam say Sam and I spent Do you guys know what an apastil is?

Anna Kirk: No.

Rob May: It's like an international notary. Yeah. But I spent a whole half day like we had to go to like city hall to get it and like then mail it out through this very complicated system. Yeah. And everywhere we went, they're like, "No, you haven't done this yet." And we're like, "Oh, we got you."

Sam Koch: You know, you [clears throat] have to get a notary, then you have to get the state department to decide that it's a valid notary. So, I just I don't think you're an entrepreneur until you've opened a bank account in France because that is maybe one of the most difficult things I've ever [laughter] done in my life.

Rob May: Yeah. It really makes you appreciate, you know, we a lot on America and particularly this the situation that it's in right now with with all the populism on on both sides. But, it's really a remarkable place to get stuff done. Like when when you look at when when we had to they FeedbackNow had employees with Switzerland, Switzerland, France, they had employees in other countries and we had to give them like three weeks notice like from the time the deal was signed and ready to go. Then we had to wait three weeks and give them notice so that these employees could decide if they wanted to take our deal and our like it was so complicated like the rules over there and like I know it's bad to be able to like hire and fire people at will in some sense but that that lack of friction in the labor market here is just a fantastic reason that America's been so successful because you just you just keep going like it's easy to I it's hard to get stuff done but it's easier than in most c countries and I think we learned a really really painful lesson.

And I will say from yeah, I don't know if we're allowed to give specific numbers because we I we might need some other people's proof, but let's just say like we did not raise as much money as we hoped to get the deal done. And the seller provided a seller note to get to sort of finish it out. And we raised the money via syndicate from some family offices and one venture fund to get the deal done. But it was hard because it was a com it was kind of like a tweener between it wasn't growing fast enough for venture and it wasn't profitable enough for private equity. So it was a little bit of a like people didn't know people have mental buckets and they didn't know where to bucket it.

Ryan Eppley: But I mean now that's the hottest space imaginable, right? It's like you now you don't want to scale.

Rob May: Yeah. To venture scale.

Ryan Eppley: Sam, you mentioned something earlier that I thought was really interesting around this management team's openness to adopting that technology. And in my in prior lives, we've looked at like I've looked at doing something like this on the AI transformation side and where we've always come down is, you know, like we're always going to get handed the problem children. In terms of management not being open to tech change. There's something that doesn't come across during diligence that we miss. I'm curious kind of what your experience was with Service Now, why you were so confident it was a good acquisition target and why you were so confident in the culture.

Rob May: If we bought Service Now, it' be a totally different business.

Ryan Eppley: Just Oh, no. FeedbackNow. FeedbackNow. Edit that out. [laughter]

Sam Koch: But no, we really I think did our homework around the team and we're also really trying I think you have to lead with value at the end of the day. Like taking taking everything about AI out off the table. When you're going into a situation like this, the first thing you have to do is be a listener and be helpful. And I think we did a really good job of coming in and first being like everything off the table. We're not going to we're not going to change anything. We're going to listen and we're going to like just make tomorrow a little bit better than today was. And that earned a lot of trust within the organization to the point where like, hey, there there are some uncomfortable changes you have to make when you're really trying to do a transformation. And if you start and rip that band-aid off right away, you can do that and that's good in some situations. But if you've built trust first and then you do it, I think the organization is much more likely to back you through good times and bad times.

Ryan Eppley: Well, walk us through that. What specific workflows did you really start with? You came in there one day, you built, you know, a bit of rapport. You probably took a couple weeks to understand how things were going. What did you start with? What was the first like workflow actually automated?

Sam Koch: Totally. So first big problem that I saw was we provide analytics to companies of how their sentiment, traffic movement, all that stuff. But data by itself is useless. What what's valuable is the insights that you glean from that. And so our team on average was spending about four hours per quarter per customer coming up with interesting insights and examples of how businesses can do better using our data. That adds up to a lot of time, hundreds of hours of labor. And all a lot of that was actually outsourced to a consultant that we were paying hundreds of dollars an hour to do that work. I built a workflow that included some SQL, included some calls to OpenAI that actually automated about 90% of that workflow to build PowerPoint slides automatically. And that reduced the total time to about 15 minutes per presentation.

And that didn't mean that the team was doing less work. That just meant that they were doing more important work and that they weren't spending their time just like digging through the platform, but instead were like really coming with with deeper insights, deeper understandings of the business, having longer conversations with the customers. And that drove value for everyone involved. And that was like a big sticking point. And when I came in, everyone was like, "We need to hire more people. We need to hire more people." And the answer was, "We actually didn't need more people. We needed a better process to improve."

Anna Kirk: If you were to have done this deal today, would you see the same outcome?

Sam Koch: It's an interesting question. What do you What do you mean by that?

Anna Kirk: What I'm alluding to is that I think the space is a little bit more crowded now in terms of people who might be following a similar strategy. There are a lot of few there are a lot of acquisition targets but probably a lot of noise also like do you think that the space has changed do you think that there's still the same opportunity set that there was a couple years ago

Rob May: so what's so people people didn't like this business right because we bought it when we bought it when we started this process almost two and a half years ago there was still this point of view that like hardware sucks and we saw this idea that in a world where software gets easier to write. Hardware is a point of defensibility and is a good thing. And so I think AI's changed a lot of your points of defensibility and in sort of a competitive strategy sense. And I think we saw that it was interesting and al also this had the this business had a really unique property which was there was a chance to go drive AI up and down the stack and improve the workflows but also there's this like bigger vision which you could talk about of like what this could become when you have data and sensors in all of these physical locations.

Sam Koch: Yeah. Because like any great AI model that you use today is built off of tons and tons of data that feeds into it to make the weights what they are. And what we got really excited about for FeedbackNow was that FeedbackNow is the weights and measures of your real time facilities. And so whether we're building that model for customers, which we are, or others are building on top of what we've built, it's a very important part of the stack. That was totally under valuated. And so that I don't know if if the deal could have been done now the same way it was done then. I think interest rates were a little different. I think people's propensity to invest in this kind of stuff was different. But I also think in some ways like the belief in AI and understanding of what it can be has also grown.

Rob May: So yeah, these models typically don't take in real-time data yet. And if there's anything I've learned as an investor and entrepreneur in the last 20 years, it's that if you can take anything and make it faster, you normally expand the TAM. You open up use cases that weren't good things happen with speed. Yeah. Exactly. And there's almost always things that'll surprise you. And so I think I think these models will incorporate real-time data going forward at some point. And I think FeedbackNow can be a big part of that. And also as you have robots that come in to do all this stuff like you need an operating system for the robots to understand where to go and what to do. That was a big part of our thesis a couple years ago.

So, why don't we talk a little bit about your first startup because one of the things I that I think helped you with this deal that you had actually done some stuff in the AI space before and tell us a little bit about that idea and like sort of how that came to be and what you had to do to build AI for your private equity outbound email business.

Sam Koch: Yeah. So, auto self auto is, deal sourcing for private equity firms, and some B2B businesses as well. We use AI to identify the right person to reach out to, generate messages that are highly personalized, and then also send them out over multiple channels. That's email, LinkedIn messages, handwritten letters, all written by machines. And that's been a really good business as as a service business, but we weren't able to really scale it beyond just like a nice growing profitable business, which is not a failure. I think that's a successful company in most parts of the world. But the way we did it was rather than go raise5 or $10 million and build a huge team and invest overinvest potentially in growth, we saw a road to profitability and me and my co-founder just went straight for that.

And so with a small nimble team, that means you have to find ways to cut costs wherever you can. And that means automation, that means support systems, that means figuring out ways to offload burdens away from yourself and onto either the customer or a service platform. And we were really success successful in that. And I think that did really help me identify how to do that in other businesses as well. And that that was kind of the playbook that I came to Rob with. I was like, we built the playbook, how to do this. I'm running a profitable business off of pennies when it should cost dollars. And I think a lot of these funds that I'm showing these deals to are trying to turn a dollar into 90 cents. And I think you can turn a dollar into 75 cents on the cost side and grow a lot faster.

So about 95% of you know money that's made off of private equity deals is either based off of margin expansion which is like I bought a business for three times revenue and I'm going to sell it for five times revenue. Debt, so I bought it with debt and then sold it and my equity is worth more but the debt hasn't expanded as fast. Acquisitions, so I bought more businesses to roll up into one business. But actual synergies and efficiencies is less than 10% of the reasons why private equity funds make more money.

Anna Kirk: Can you expand on that a little bit? Like when you say synergies and efficiencies, what do you mean by that?

Sam Koch: Like outside of just like cost cutting like actual growth levers of new products or new innovation. Coming up with ways to reduce cost that isn't just like cutting people.

Anna Kirk: So yeah. Yeah. Back to our people conversation that can be a tricky one to lead with.

Ryan Eppley: Should we hop into news?

Rob May: Yes. Actually, while you guys do that, those bourbons have run through me, so I'm gonna go to the bathroom while [laughter] you guys hear the news.

Ryan Eppley: All right, kick us off. What do we got on the first first docket?

Anna Kirk: All right. Well, this would sound like any other headline in AI except for one small detail. Cartesia AI, which is an AI texttospech model company. If you're not familiar, just announced they raised a hundred million led by Kleiner Perkins in tandem with the announcement of its new model, Sonic 3. Which Rob was very excited about because Sonic 3 is one of the first state space models to get major funding. So notable in that sense, but I'm curious, Ryan, what do you think about the future of this space? I know there's some criticism that you know foundation model companies are getting like that service is getting commoditized. They're trying to scale at a venture rate and that's just not sustainable. What do you think about that?

Ryan Eppley: Probably spend the entire podcast talking about this mostly because you know my my co-founder who is a very very gifted technical mind studied mathematics in a master's degree. She's all in on statebased models in a lot of ways because there just seems to be more opportunity in the growth side even though they're not as you know tried and true yet. There seems to be more that can happen whereas transformers are 2 plus 2 plus 2 plus two. It will get better, but there's not an exponential better. State space models seem that they they have this blackbox ability that we can tap into. So, it's something I'm very excited about.

Anna Kirk: Maybe we could take a quick step back and could you define what a state space model is for those of us who haven't heard of it before?

Ryan Eppley: Yeah. You know, and I don't know, have you spent too much time in the space at all yet?

Sam Koch: I've played with a bunch of language models more as a as a user rather than a builder necessarily. So, I've I've got some language models that we've like trained to do customer service, for example.

Anna Kirk: All right. The Bourbon Man is back. Yeah. We did hop into state space models.

Rob May: Yeah. What's interesting about space models from a technical perspective is they run a lot faster than transformer models. So, I think what a lot of people don't understand is that most of modern AI is built on what's called a transformer model, which is one type of neural network. There are a lot more. There's also lots of types of AI that aren't neural networks that you can run and ChatGPT cloud all these things are very very similar. They're they're built on various transformer models. Statebased models are different. And so the reason that I called this article out is that it's the first major funding round for non-transformer-based AI model which is which I think is really promising for the industry because you know there's a lot of other ideas out there and I think people that are working on those some of some of these new architecture ideas are going to be necessary to push AI forward.

Ryan Eppley: So yeah, I think back to your your point, you know, transformers have had a architectural problem where you can only get them better incrementally. I think the idea of a stage space model is you are having it train itself faster in a loop, allowing there to be more expansion and opportunity. You know, again, it's early. It's the first major announcement of a round, so I think we'll learn a lot over the next year or two. But what gets me excited is just that there's a new potential technical breakthrough that can change a lot of the research that isn't just hey we're going to pour more data into the old architectures. So that's something that I think is just opening up a lot of opportunity.

Anna Kirk: Yeah. That very very exciting. I'm I'm curious your thoughts on how much better it can get and also how much better we need it to get because like when I run my 11 labs voice cloning it's pretty darn good. Mhm. So like if I get five 10% better I think it would be almost indistinguishable from a human. What what do you think that exponential state base model can be?

Ryan Eppley: Yeah. Well, are you talking about AI or are you talking about voice models?

Anna Kirk: Cuz I'm talking about just voice. Yeah.

Ryan Eppley: Yeah. I don't know. I don't know how much better you need voice to get, you know? But I think like it's the first battle for it to like be like, "Okay, this is clear." It's more so like how much faster can it get? I think I think it can get way faster. I think voice is a good use case that you're bringing up as like almost an antagonist of it. Like how much better does it need to get? I don't know. Transformers, I've used plenty of them. They're they're pretty good for voice, but I think I've already seen Sonic be better and it's very early in it, right? Sonic is as good as I can't tell if I'm calling someone live or not. Whereas some Transformers you sometimes can't tell and then like halfway through sentence two or three, you're like, "Okay, great. I have recognized that this is now an LLM behind the scene." But with Sonic, at least from what I've seen, I could probably go the whole conversation and not know.

Rob May: And there's there's engineering challenges, right? Which is like the longer that you need it to talk, the more likely you are to have an error. You know, you can always make these models, at least for quite some time, cheaper or smaller to run. I a lot of people forget models are very big by computer program standards. And so they have a lot of errors. They're they're much more likely probabilistically to have an error when they run. They they require more memory and all that. So making those smaller can always be good things. We know we're still really far away from the efficiency of the human brain from a sort of power and computational perspective. So that's cool.

Before we go into the next article, you guys want to do our Mamdani jokes. So, all right. We have a new mayor elect in New York City, and I ask three models, Claude, Gemini, and ChatGPT, to write jokes about AI and Mamdani. They're terrible. I'm just going to tell you none of them are good. But what we're going to do is I'm going to read three. And if you've played with these models and you kind of know what's different about them, and we'll we'll let the listeners do this as well. See if you can guess which model wrote which one. Okay, so the first so the first joke, Zohran went so viral on TikTok that AI training models now think democratic socialism is just really good engagement optimization. So GPT5 keeps asking, "Have you considered freezing the rent on my inference costs?"

Anna Kirk: Not a good joke. Not a great joke. Three out of 10.

Ryan Eppley: Three out of 10. Yeah, exactly.

Rob May: Yeah. We're going to we've got some comedians that might come on at some point, so we're going to we'll face them off and see if they can write better jokes. Second joke. When the AI Mamdani, it asked, "Are you running for mayor to change the city or did you just glitch into this election cycle?" Mani Grinden said, "Why not both?"

Anna Kirk: Does is glitch like they just mean like a technical glitch or is that like a term the kids? Is that like 67 or something like

Ryan Eppley: No, it's still below a five. Maybe incrementally of four mostly because I think it's getting to the fact that he hasn't really had a job. So yeah, he glitched into it.

Anna Kirk: Okay. Oh, that's a good point. Okay.

Rob May: And the third is I asked an AI chatbot to generate a slogan for Zohran Mamdani's campaign. It came back with quote a $30 minimum wage by 2030 or the AI will unionize your Roomba. [laughter]

Ryan Eppley: All right, I'll give that a little better. Okay, you know, throwing a Roomba in there.

Rob May: A little any vote. Okay, so again the jokes were TikTok viral blah blah blah glitch and Roomba unionization. Which one was Claude? Which one was Gemini? And which one was ChatGPT?

Anna Kirk: I feel like one was ChatGPT. It was so wordy.

Rob May: Went viral. It was long. All right, everybody at home listening, you got your answers. Lock in your vote. The TikTok one, the one that mentioned GPT-5, the TikTok going viral was actually Claude, which I will say I think had the worst of all the jokes altogether, which makes sense. It's a business focused app. The glitch one was ChatGPT. So that makes sense that it might be more up with the zeitgeist. That's kind of its presentation. And Unionize your Roomba was Google Gemini. Gemini's got the jokes.

Ryan Eppley: Yeah, Gemini's got the jokes. Got the jokes. That could be a good advertisement for them.

Anna Kirk: Yeah, you got to do this again with Grok.

Rob May: I mean, they did make Bard. So, that was Yeah, we should see [laughter] what happens.

Ryan Eppley: Grok because it's trained on the Twitter data probably has a lot that's got to be spicy.

Rob May: Okay, we'll do that one next time. That's going to be a joke. Yeah.

Ryan Eppley: All right. Let's jump into this article about data centers and the fact that people are people are taking use jet engines. Did you guys know this? Use jet engines to power data centers.

Anna Kirk: I don't know how ubiquitous it is though. It sounds ludicrous.

Ryan Eppley: Yeah, there was like a lot of Bitcoin mining stuff where people were like, "We're going to use geothermal to totally change."

Anna Kirk: Like, you know, the article does make specific references. I can't imagine it's too many data centers are being powered by old jet.

Ryan Eppley: I think it's actually this many. It's like it's like two data centers. But to be fair, I think the article mentioned they're powering like a gigawatt. Which is pretty substantial.

Rob May: And I think they can provide bridging power for like six or seven. The fact that now there's there's a market for about a thousand used jet engines a year. And now the prices are being bit up because you can relatively quickly and easily convert them into running with natural gas, being stationary, and you can fire them up in five minutes, right, to get them to run, right? Yeah. And so you don't have to wait on because the problem is you have to go get power lines somewhere. You need permits. Takes a long time.

Anna Kirk: I guess my my only push back there is like what is a limiting factor supply chain wise for getting a data center up and running? I don't think it's gas turbines usually like

Rob May: but isn't it power overall?

Anna Kirk: It's transformers and it's grid power. Hyperscalers are turning back to gas and they don't want to because technically they're abandoning a lot of their 2030 climate commitments. But that's just the reality of where we are now. It's about it's power but it's also time to power. So I'm curious like what kind of market even is there for this? Like it's a cool headline.

Ryan Eppley: Billion dollar business idea. People pedaling paddle boats that power AI data systems.

Rob May: That's just the matrix. [laughter] The humans are still relevant.

Ryan Eppley: We take one of those beer trolleys. The four of us hop on it and we power a data center in downtown [laughter] little mobile data center running around. Everybody's drinking. Cool way to do the podcast. Yeah, honestly. Next podcast. That's what we're doing. Nashville Mobile Party podcast.

Anna Kirk: Oh, I don't know. You guys have to do that one without me.

Rob May: We could probably all power one Raspberry Pi. That's it. That's all we got.

Ryan Eppley: Yeah. Sorry, Rob. You can't do it with bourbon. Doesn't work that way.

Rob May: Maybe you could, but All right. Some of us aren't built that way. What else do we have?

Anna Kirk: Going back to the jet engine question, like sounds great in theory, but like what there's got to be some issues around like you have to have some technician who knows how a jet engine works around. Yeah. Like there's got to be some other downsides to this. Why there's only two people, two dataers

Rob May: doing a big problem with the whole AI movement in general, right? Which is you're you're you're you're shifting around you're you're taking a bunch of jobs out of the workforce. You're introducing a whole bunch of new ones and it's happening so fast. So there's there's there's data center people we need. There's there's there's a shortage of electrical engineer. I mean, this part of why your company exists, right? There's a shortage of electrical engineers. There's a shortage of chip designers, firmware people who know how far like everybody now who's in tech is just a application designer. They know how to write Python code and it's causing a lot of problems and so to your point I definitely think there's a lack of technicians that these data centers are like oh man like we didn't realize how much maintenance went into these things.

Ryan Eppley: I'd also just say just like, you know, I have a ton of friends building all types of aerospace engineering aircrafts, things that are their whole kind of, you know, prototyping is go buy cheap aftermarket engines and then, you know, they'll scale up and get their own production. But I wonder if with this article they're like, man, if these hyperscalers take my aftermarket engines, like someone's going to be really upset and it's going to be these startup founders building really cool machines. So I'd be I'd be interested to ask some of my friends about that in the Twitter DMs.

Anna Kirk: I mean, if if anything though, like the hyperscalers aren't the guys who need or want this. Like all of the major oil and gas providers, I'm sure, have partnered with them in a heartbeat. So who is the real type of buyer for something like this? And maybe it's like smaller kind of inference driven data centers. But like if if one of these things can power like what like 50 megawatts, even less than that, you're talking like 20 of them for a gigawatt, like that's a lot of kind of added operational complexity that you didn't have before.

Ryan Eppley: It makes me think of like behind enemy lines or something and the guy's like there's a old jet that crashed and like I need to like spin up this like radio and like he's like tinkering behind the scenes passed away. Like you're you're putting together this technology like Frankenstein machine to just figure out how to do something, but you know, maybe maybe data centers will be powered by jet engines. I think it's a low probability.

Rob May: Well, I mean there's the some of these places there's apparently a 6 to 8 year wait to get power lines distributed to your data center that you want to build, which is like who knows what the demand's going to be at that point or when it's going to where all the nuclear companies are saying like, oh, it's, you know, within our reach to do six to eight years, right? It's very hard for the next, you know, 6 to 8 months, but they're like, "Oh, 2032, I could be the power source for that."

Ryan Eppley: so I've heard a lot of nuclear companies use that exact quote, like, "Wow, 6 to8 years, like that's in our now pitch deck."

Anna Kirk: Yeah, I think there's going to be overbuild for sure. Like, I think there's going to be a short-term time crunch where a lot of these hyperscalers are going to need to adapt using bridging power, but I think the market's going to catch up. And I think all of these nuclearpowered data centers, I think all the kind of repurposed offshore oil rigs, yeah, it's really interesting interesting and I'm sure there is a use case for it. I just don't know if it's like a macrolevel sector shift that we're going to see.

Rob May: So Nassim Taleb, who wrote the Black Swan, tweeted one time a couple years ago, he tweeted during COVID and I thought about this. He it was with respect to the shipping container issues. He said, I've never seen a supply constraint not turn into a glut on the other side because everybody overbuilds. So, I do agree with you. I mean, this happened obviously with web 1.0 and fiber buildout. We've talked about before how this is different because these things don't last 20 years and they depreciate faster, but still I think I think it'll be on a smaller cycle, but I think there will be a glut invest and buy up a bunch of the capacity probably.

Anna Kirk: It's interesting too because I think it's localized in some ways like China has really great energy capacity compared to the US and like really good builds of I it's like some insane stats about how many what what percentage of solar panels are produced in China. I think it's like more than they also built like more nuclear plants last year than we have or something kind of stat like that. Like they just have such Yeah. Well set up power and I wonder like if the US power imbalance is so bad like does more inference go to China does more you know investment there

Rob May: inference tariffs so [laughter] it'll be tariffs yeah I know I know make jokes about that forever

Anna Kirk: all right what's next Amazon is suing Perplexity to stop the to stop the model provider from allowing users to buy its items through its agent, its browser agent, Comet. What do you guys think about this? Maybe it'll set a precedent. Really, your favorite?

Ryan Eppley: My favorite article of the week. I thought for sure that this is it's probably not going to get upheld. I don't feel like Amazon will win, but it makes me feel excited to understand how does a governing body look at an agent with all the old contracts being from a human entity because it's basically saying our user agreement with a user now you have to define a user. So there's a lot of a lot of nouns that have to be defined in the world of agents and all these contracts have never even in the spirit of the contract thought about an agent before. So I'm most excited about how do you define a user in all these service agreements that we've all just clicked a button on and said accept. There are they're decades old now and all these agents are like well we get we're a user if it's doing it on your behalf. Is it your proxy? Is it a user? So to me, I think just from like a legal kind of nerd out for a moment, like how are we going to define a user in the age of agentic platforms is probably one of the more exciting articles of the week for sure for me.

Rob May: I took a more strategic thought from it. I do think that's a great point that you made, but I looked at this from the perspective of like why doesn't Amazon want this? So, if you wanted to take a negative point of view on Amazon's business, you could argue that they sell a lot of cheaply that people buy and they have a bunch of algorithms to know what you're not price sensitive for. And their business is about like their ad business keeps growing because people pay to promote on Amazon. So they know the laziness of humans and they know we're not going to dig three or four or five or 25 pages deep to find the best deal particularly on a bunch of commoditized things that lots of people sell on there. But an agent would be happy to go and look at everything in their database and find you the best deal. And I'm sure Amazon wants to prevent that.

So, I think you're going to get this weird strategic battle between suddenly these agents give humans effectively the wherewithal, like the patience and the ability to go like really fast and like search through everything and figure out everything and like stuff that you and I would never waste the time to click on and figure out the best deal and negotiate forever relentlessly. And no e-commerce player actually wants that, right?

Anna Kirk: Well, especially once if you're if you're not just on Amazon, but like if you're on Amazon and eBay and Alibaba, like, yeah, that's terrifying for someone who feels like they have a captured market. Like, I would totally make a trade sometimes of like, yeah, I don't need this in one day, especially if my agent can tell me that I'm going to need Drano in four weeks.

Rob May: Yeah. You're tra you're in San Francisco this week. You're not even like you don't care when this thing gets shipped here because you're not even going to be back for five days and your agent knows that.

Anna Kirk: I mean is Amazon just trying to build their own like if if I were them that's exactly what I would be doing. I would be blocking everyone else's and trying to build my own. Either that or you have to alter what your kind of optimization looks like for different providers on your platform. But agents could really give power back to the people over a lot of the tech right control.

Ryan Eppley: I don't Yeah. So will consumer protection come in on this one as well because right now they're they're doing it from again the dejour we have a service agreement with a user you've accepted this we get to define that now I totally agree in like the game theory of like why wouldn't a business want this to be able to be you know agentically bought through because now you have the user having more purchasing power in theory. But that's not what they're at least legally saying. And they're saying like, "No, it's our service agreement." So whether you look at it from a business angle or from a legal angle, both very exciting. This is probably going to be as important of an outcome as the OpenAI versus New York Times. I thought that was a massive massive moment in this AI era. I think this is going to be maybe just as big.

Rob May: Yeah, we'll definitely be some defining moments. And related to consumer protection, I don't know if you know this, but the very first consumer protection law originated from bourbon, which I know for because I'm from Kentucky. So, people were putting all kinds of stuff in water and other I see liquids to make it look brown and whatever. People were getting really sick and that was the first

Anna Kirk: All right. Well, what was the first thing bought on the internet then? A book? I don't know. No. Does anybody know?

Ryan Eppley: No idea.

Rob May: I'm pretty sure it was a pizza. I'm pretty sure it was Domino's. Okay. Wow. We'll have to look that up. Someone in the audience, let us know. But I'm pretty sure pizza was like one of the first things bought on the internet.

Anna Kirk: That's pretty awesome, actually.

Rob May: Good. Well, that leads us into our question of the night. So, we always ask one question about New York to sort of end up the show. Favorite New York burger. Who wants to go first? I mean, lots of good options. This is a city that has really good food. There are a lot of good burgers at a lot of good places, but does anybody have a favorite?

Anna Kirk: A lot of different ways you could go too with this one.

Anna Kirk: Yeah. Smithfield Hall. Pub Burger.

Ryan Eppley: Love a good Pub Burger. You're going United. Come on. [laughter]

Sam Koch: rooting for Liverpool. Sorry.

Ryan Eppley: Favorite. I don't think it's a burger per se, but like, you know, if you got a Chick-fil-A, I'm eating it. I love I love Chick-fil-A. I just b I had it tonight. I had it tonight. Otherwise, I don't know. Like I don't go out of my way for like a burger in the city, you know. Do you have a spot in mind?

Anna Kirk: Just it's like a whole vibe because I was just there, but I don't know if you guys have ever been to Manhatta, but it's on the 60th floor in Fey. You can see the whole city. We went this weekend for lunch and it's just like a great burger but also like great view. So, it's good.

Rob May: Yeah. I mean, I would say and this this this place will be my answer to many of the best in New York questions. But, I think the Smashburger at Flatiron Room, which you guys know is one of my favorite hangouts. I think it's the best whiskey list. I think it's the best live music. It's like all kinds of things. Did I have a burger with you the last time we were there?

Ryan Eppley: You did. Maybe I did.

Rob May: And this podcast sponsored by Yeah. [laughter]

Ryan Eppley: Sponsored by Flatiron Room.

Rob May: Well, I mean, when we kicked this off, the three of us with some other people. So, so part of the way this group came together was we had a conversation at dinner at Flatiron Room that went totally off the rails and we talked about all the things you're not supposed to talk about, religion and politics and a whole bunch of things. And like we were still friends at the end of that conversation. We were like, "Oh yeah, we would be good to do this together because [laughter] we should do a podcast." Yeah. So, welcome to the cast.

Anna Kirk: Yeah. It's totally the like 2020s version of We Should Buy a Bar. Yeah. Like, oh god. Yeah. It's like It reminds me of that one episode of How I Met Your Mother where Marshall's like on a search for like New York's best burger and there's this like one burger spot that he couldn't possibly find. He thought it was Corner Bistro. Spoiler alert, it wasn't Corner Bistro, but I it's on my list to try their burger. I don't know if you guys have tried it.

Rob May: Well, this is another problem in New York, which is sometimes you've had a few too many drinks and you stumble into a place and it's late and you have a great burger and then you can't remember where you were. Totally were. So, I've had that happen as well. Good.

Ryan Eppley: Any any closing thoughts? Anything you guys want to talk about that we didn't cover?

Anna Kirk: Send us your swag. That's right. We forgot that in the beginning. Send us your swag.

Rob May: Absolutely. Send us your swag. We'll wear it on the show. Yeah, we need to start doing that. Also I'll say that as part of our end of year show, we are going to shoot here in the studio on December 18th, I think, is our is our last show that we're going to record. And we're going to we're going to profile 10 companies. So, if you want to come in, be on the show, have a few thousand people see your demo. We'll put up a form shortly to take some applications for that. So, it'll be fun. All right. Thanks so much.

Anna Kirk: Is that the first item bought online using the internet was a CD of Sting's album 10 Sting?

Rob May: Really? A Sting album was the first Really? That's disappointing. We could do better. Yeah. Should have been a pizza.

Ryan Eppley: Yeah. Yeah. Pizza was the first thing about with Bitcoin, I think. Yeah. That sounds one of the first. Yeah. Yeah. Yeah.

Rob May: So, cool. Well, Sam, thanks for being on today. Hopefully if you're watching next time you are in a airport bathroom and you see the FeedbackNow buttons you will push them now that you know. Hammer that. Yeah. Now that you know what happens. Hammer the like button. Hammer the smileys [laughter] and if it breaks you can you can call Sam. Yeah. And yeah, thanks for listening. Please please rate us, like us, leave a comment. Watch the other episodes. And if you know somebody that should be on the show, please send them our way. And yeah, thanks for listening. Please go watch the other podcasts. Thanks.