His $100M+ Business: Selling Real Life to AI

0:00Bobby Samuels: Can you bring up what happened with Spirit Airlines? So Spirit went bankrupt. And Spirit has this treasure trove of data that could be obtained for training models. It was compelling enough that you had multiple bidders, but then Google won with a$10 million bid. Which, to be honest, we were sort of surprised it wasn't higher.

0:41Bobby Samuels: Welcome to Media and the Machine. I'm Rob Kelly. AI is the biggest technology shift of our lifetime, and this show is about how to profit from it. I'm a three-time founder and CEO of content-driven businesses, and each week I talk with the people on the inside of AI and content, the ones building this new world and learning what works firsthand. I'm an optimist about AI, but I'm also a realist. So every show ends with me asking our guest what AI means for our jobs, our families, and our future. My guest today is Bobby Samuels. He's the co-founder and CEO of Protege. Bobby gives us a peek at AI's frontier and its data needs.

1:18Bobby Samuels: Protege reports well over$100 million in gross data transactions this year. They keep a cut of 30-plus percent. For those of you into data licensing, we get into the weeds. We explore company data deals like Google's$10 million bid for Spirit Airlines data when they went bankrupt, which they barely outbid to AI unicorn startups. Bobby explains why you expected an even higher price. I ask him what my own startup's company data is worth. We unpack who you need to talk to inside the AI labs. Partnership teams help make deals, but researchers drive the agendas. We explore a Polish soccer league's training footage.

1:54Bobby Samuels: Think six or eight synchronized camera angles. You'll hear why AI developers love this stuff. Then we turn to Bobby's two young daughters growing up in this age of AI. What should kids today learn as AI advances? Should schools ban AI? Will his girls even go to college? And my special thanks to Dave Davis, Carlos Sanchez, Sarger, and Dan Stiney for putting Protege on my radar. Now please enjoy my conversation with Bobby Samuels. Can you bring up what happened with this Google bid for Spirit Airlines? Just recap the story. What's going on there? So Spirit, a once great American institution, went bankrupt.

2:37So assets were being sold off. And one of the assets was the data. This resource that wasn't valuable or wasn't on anyone's radar even a couple months ago. But when you're talking about training AI to do business-related tasks, the best training data are business-related tasks. And Spirit has decades of history of office communication, presentations, financial reports, etc., etc. And so there was this treasure trove, really, of data that could be obtained for training models. It was compelling enough that you had multiple bidders. Some were training data-specific companies like Mercore and Micro One.

3:25But then, obviously, Google was the one who won with the$10 million bid. Which, to be honest, we were sort of surprised it wasn't higher, just given what we're seeing in the market for training data. I do think that the spirit story has really gotten the training data world, and specifically the training data world for company data, into much more of the mainstream than it had been before.

3:52Bobby Samuels: Yeah, suddenly the new phrase is company data, right? And it's worth something. Let me just rattle off some of these assets just from the bankruptcy filing just to give a sense of the scale. So communication-wise, we're talking about, this is all a list of Highlights of Spirits data. 100 million emails, 500 million Microsoft team items, 7.2 billion observations about competitor flights and pricing, 7.5 billion revenue transactions, employee records, 3.4 million payroll records, 1 million time cards. and I know we'll get into how it's de-identified, of course, for privacy reasons, 30 million lines of code.

4:36Bobby Samuels: And Google beat out two AI startups. This is no one off of Google, Micro One and Mercor. What exactly are they going to do? What answers will they now provide via Gemini that will be better than before? Right now, the biggest money-making use cases are all around the enterprise. And if you think about what data is needed, it is enterprise data. So if I want to understand how to make a set of decisions on pricing for an airline, having how Spirit has thought about that for a long time is helpful. But not just the train of thought, but what are the documents that go into that? What are the competitor information that goes into that?

5:25What are the team's messages that go into that? And so it can create this sort of agentic worker, but it only can do that by getting data and getting data that is emblematic of actual people's work. And I think the scale of what Spirit has is pretty unprecedented. And it's like a name that folks know.

5:48Bobby Samuels: And I've heard this phrase, digital exhaust, early on. Some folks were talking about that. But then what I'm hearing more now that it's sort of getting formalized as an actual industry here is decision data. Is that the phrase that you use? This is decision making data? I've heard both. But whether it's decision data or traces, it's really how do people make decisions? And how do they make them in the real world? A departure for sort of the historic data industry for AI has just been a move, and this is one we've believed strongly from the beginning, that the best way to train models is not through contrived examples, but through what has actually happened in the real world.

6:31And basically, for agents to learn how to build a company, they need to get a bunch of data on how companies operate. That's why the Spirit data, I think, was so impactful, and why we were surprised it wasn't more. given that you're seeing millions of dollars being thrown around pretty regularly for much smaller companies, and folks like Micro One saying that they're committing a billion dollars on data spend over the next period of time.

6:57Bobby Samuels: I'm curious how, because you probably can identify a little bit more with, in this case, Mercore and Micro One, the ones who appear to have lost on the bidding, but they bid a lot of money. They bid$7.5 million and$12.5 million. How do they value those assets, that data? How would they even come up with that calculation? The whole data world is certainly more art than science. I am sure they both did back-of-the-envelope calculations on, here's how much we think we could sell the data for over time to different prospective buyers. Let's add a margin on top of it and back that out. That's generally how I'd imagine.

7:40They may buy it wholesale and effectively sell it retail, where they cut it up into a bunch of different products that they'll then license to many different labs.

7:49Bobby Samuels: So if they had won, they actually would have gone and tried to sell it, license it to Google for Gemini and to OpenAI and Anthropic. No question. Yes. And so I think for Google, part of it was probably an exclusivity question. And this wasn't just a one-off, right? It was the largest kind of known company whose data is being bid for in this way for AI training. But where else is this happening with smaller companies? You hear about folks now incessantly getting hit up. Investment firms having people from every firm you mentioned and a whole bunch more reaching out, asking for their company data.

8:29So it's investment firms, it's companies going out of business, like Spirit, it's VC-backed companies. It really runs the gamut. We've been hit up by folks. They want your data. Oh, sure. And we've politely declined. But yeah, I think there's an insatiable appetite for company data right now. And you're seeing the dollars being spent accordingly.

8:54Bobby Samuels: So it doesn't have to be a failed startup. It could just be for an active company. And that's probably more, I think failed startups are easier because the business risk from a strategy perspective is lower. Spirit doesn't have any IP that it really cares about anymore from a strategy perspective. But all else equal, ongoing concerns are more valuable. We joke that buying this weird data is a great way to teach models how to build a failed airline. But ideally, you want to have the data in a company that is working to teach the models to act accordingly. I've got a friend who, a close friend whose friend just got paid a million for his 100-employee company's data.

9:41Bobby Samuels: and in this case, MicroOne paid a middleman. I wasn't able to get who that was. Who's the type of middleman in that transaction? There are all these companies. And I would say, in general, we're doing less and less and pretty much very little in the company data space. But there are all these companies that have popped up to sell, in some cases, to the Mercos and MicroOne's and handshakes of the world, and probably some to sell to the labs directly. But yes, there's this whole universe of companies that have popped up fairly opportunistically to do exactly that. I've got a SaaS startup around for 13 years.

10:21Bobby Samuels: On average, five employees during that time. Worth anything, data-wise? I'm sure it's worth something to one of these companies. I don't know if it's... Tens of thousands? 100K? Probably tens. It may end up being too few employees. but I suspect given what's going on, you'd be able to get something for it. Gotcha. So the emphasis is just high volume number of transactions, decisions being made at an enterprise, the more, the better basically. So the larger number of employees and the larger number of years they've been working on decisions. And if it's a successful company versus a bankrupt company, maybe there's a little premium.

11:00Yeah, yeah, that's probably right. All else equal, the model builders want. more decisions that they can then train their models off of.

11:08Bobby Samuels: If this AI data training world was a three-act play, and let's say act one was AI models scraping the internet and books, just to simplify things, are we in act two right now? It seems right. Yeah, I think more specifically, like is act two audio, video, company data, physical data, egocentric data? Am I tracking? In some ways, maybe this is Act 3. So Act 1 was scraping. Act 2 was having humans create a bunch of data. So the original scale business of circle the stoplights so I know whether this is red or yellow or green for my self-driving car. And then the third act is getting data from the real world.

12:06Our take is for the data businesses, this third act is both ongoing, but also by far the largest. Like Spirit Airlines is real world data. The exhaust is how I think about it, which is people living their lives, companies building their products, creates this trail of data. and that is the best way to train models how to behave is how the world has operated to date.

12:32Bobby Samuels: Gotcha. Why don't we talk about the way you're structuring Protege. So you've got GMs for different verticals. You've got a GM, it sounds like, for audio, right? You've got one for healthcare and then you've got one that I was calling, I don't know your name for it, but physical AI, video, spatial. physical and spatial intelligence. Yeah, exactly. Exactly. Great. So you've got those three. You don't have one for company data, it sounds like, correct? Correct. So these GMs handle both sides. They go out and help acquire content in their space. And then they also are going out and licensing the said content to AI models and buyers or licensees of the data.

13:15Yes, that's right. And we don't typically buy the content, but we do go out, strike relationships, do the right curation on top of it, embed with research groups, and then license the data out.

13:26Bobby Samuels: Do you treat them as sort of business units, the GMs? Yeah. It depends a little bit on how closely aligned they are with different sort of related verticals in the business, but more or less, yes, separate business units. And certainly the spatial and physical part of the business is pretty standalone. I'm curious, stack ranking them, what's the revenue size of the three? Who's biggest? Healthcare has been the biggest. Actually, a lot of the audio deals have been in health. There's overlap, yeah. Yeah, exactly. So that's the biggest. And then spatial and physical has grown reasonably quickly. And I think the information said earlier this year, you were tracking for 100 million in data transactions.

14:08Bobby Samuels: Is that accurate? Yeah, we've surpassed that by a comfortable margin, but yes, clearly nine figures. Okay. And for all of 2026, you'll be comfortably in the nine figures. Yeah. Excellent. Yeah. No, it's great. We feel like we're in good company for having started the business two years ago and change and now being in nine figure gross revenue. And right. So that's gross. That's excluding the payout for any data that you've acquired. And I think you also have mentioned typically you take around 25 to 30%. Yeah, and that number has been going up as we've, in part, started building more differentiated products and offerings.

14:51So it's probably 30-ish, but it's been trending upward.

14:54Bobby Samuels: And last time you and I talked, you said you'd done in excess of 100 deals. What are you up to now? Once we passed the 100, I sort of stopped tracking every single one at that level, but it's somewhere in the low triple digits. And just doing some back-of-the-envelope math with that, you know, nine figures of total transaction value, I mean, is a million-dollar sort of average deal size pretty common? Yeah. I mean, so a million dollars is sort of the mean. I think it's probably right, but you get eight-figure deals, you get low five-figure deals. You get a pretty wide range of deal types. And are the AI, let's call them buyers, the ones licensing data content from Protégé, are they the big frontier models?

15:41Bobby Samuels: Are they all of the big frontier models? Common names we all know? We work with a variety of the vertical AI builders as well as most of the frontier labs, yeah. Okay. I was curious, most would assume the frontier models would account for most of the revenue. Is that the case with you guys or is it more of these verticals? That's, I think, with everybody, that's pretty much the case. And with us as well, the majority of the business today is with the labs. Yeah. And last time we chatted, you mentioned that the AI labs, the big ones, Frontier Models, have billion-dollar budgets for data. You're talking about, you know, for pre-training, post-training.

16:21Bobby Samuels: I think you talked about RL as well. By the way, have you had a chance to look at the Anthropic prospectus to see what they're spending on data? I haven't read it cover to cover, but our CFO, who's terrific, did a sort of explainer for the company. What were the main insights? Probably the biggest one, there's like the$80 billion loss price tag. And the team's take is like, that basically is BS. Most of that's like stock-based compensation. And if you actually look at the fundamentals of the business, it's like very, very strong. Their NRR is 900%, which is, I've never heard of anything even close to that.

17:01and also the amount of cash it has on hand is insane. It rivals every public company, even the biggest.

17:09Bobby Samuels: So it's good news for you guys, good news for Protege. Yes, it's good news for us. Yeah, I haven't had a chance to read through every page verbatim yet, but I did go through and look for some mentions of data. And they don't give absolute numbers, but there were some suggestions that their spending on data increased, for instance,$165 million in 2024, same in 2025. This is just the increase. They don't say what it went from and to. And then an increase of$251 million in first half of 2026. So I guess that would track maybe to your billion dollar budget estimate for the frontier miles. And look, I mean, our take is data today is the biggest bottleneck.

17:57So I am bullish that those budgets are going to continue to go up.

18:01Bobby Samuels: What makes you confident that data is the biggest bottleneck? I was at some conference and they were like, here are the 1 ,200 things that happen in a hospital or that a doctor does. And to fine-tune a model, you will need data to some extent from all of those. And that's just one small part of the economy. And we can have the most powerful models, but like the brain is still a total black box. And that's not a question of model compute. That is a question of data. And so I think there's going to be more and more niche areas that are not so niche where what's going to unlock our ability to understand it is just more data.

18:41Bobby Samuels: Who at the AI models, and I know it's different depending on whether it's the frontier models or maybe the verticals, but who is it that you work with? Is it a biz dev person? Is it a researcher? who is telling you what data they need? It's both. It's an industry where building relationships with partnerships folks and researchers is really critical. And I think part of the reason we've been able to have the success we've had, this is, I'd say, a new type of buyer that hasn't existed historically, the researchers. And they're not the ones signing the checks, but they're driving the agendas. And it's an incredibly mission-driven group.

19:21we've built a very mission-driven company and that comes through and i think deeply resonates

19:30Bobby Samuels: you mentioned researchers partnership folks which one is more common by the way so researcher or a partnership person reaching out to defining what it is they need data wise when when we when we do it right it's almost always both i know it's in the weeds but i love with stuff like you, is it the researcher identifying that they need some new category and then a partnership person pinging you guys? Or is it the reverse? Or are they both on? I'm fascinated because like you said, it's really a new type of buyer, right? It's never happened before. Yeah. I mean, look, I think when we do it right, deal number one is partnerships person comes forward and says, hey, based on the research agenda, here's the data we need.

20:17if we do it right, deal two is, hey, you guys did a great job. What else should we be thinking about? And so I think it really is going deep in those and becoming thought partners versus just vendors that are dealing transactionally with data. That's where we really shine.

20:33Bobby Samuels: So maybe it starts with the partnership person as you get in more entrenched, then you're getting to know the researchers? Yeah, and then it's both. It really is both. And then might a researcher just ping you directly without the partnership person in the future once you've done business? In certain cases. But then it still requires and is accelerated by relationships with the folks who are leading data partnerships. And what you just described just now, is that only at the frontier models or does that apply also to, do you call verticals the sort of non - Yeah, it depends on the size of the company.

21:07You know, if it's really, really small, you might get the CEO. If it's mid-size, you may get the head of product. If you get to a larger size, it usually is head of partnerships. If you get even bigger, it's someone who reports the head of partnerships. So it's on this spectrum, depending on size of company. And then obviously, the labs tend to be pretty large. So it sounds like what I'm hearing you say is

21:28Bobby Samuels: you're not seeing any ceiling for AI data training needs. No, I don't see any immediate ceiling. Now, there's going to be some correction at some point, some magnitude and data budgets will probably get cut. But if I think about the long term, I think it's going to be a growth area for a while. Is there a part of the industry that you think will get shaken up? Sounds like you feel confident about Protege, but is there a part of all this ecosystem, call it data training, data licensing, that you feel like would get shaken up first? I'm a believer that the verticals, there's going to need to be a whole ecosystem outside the lab.

22:08I'm not somebody who thinks that the labs are going to rule every possible use case. But I do think there's going to be corrections there. I think the companies that have raised a kajillion dollars pre-revenue, maybe a couple survived. But I think that's going to get hit really hard.

22:24Bobby Samuels: And the ones raising the billions, are you talking about companies like Mercore or Micro One? You see consolidation there? More like some of the Neo Labs. I mean, Mercury and Micro One, they obviously have pretty significant businesses and I think are quite profitable. I think it's like you mentioned earlier, the company that was working with Micro One that hit up your friend's company. Yeah. I don't see how that evolves to something big. I think that's just, you're a middleman to a middleman in an area where I don't think anyone really has differentiation to date. So I think that's just a hard spot to be in.

23:01And I think that ecosystem probably doesn't last for most of the companies that are there.

23:06Bobby Samuels: What about someone just focusing on one type of data? So for instance, I've got a friend getting into just egocentric data. That's it. Like that's their whole bet, egocentric data, or just video for that matter. Yeah, I mean, I think, I mean, look, there could be plenty of successful businesses there. But if you want to maximize the likelihood of success, having some sort of theory of what is the defensibility. If you have someone who's just doing egocentric data, that's not enough. From my perspective, to be like some durable business, like we've seen the market be flooded with egocentric data and prices go down accordingly.

23:49I think there's very little barrier to entry for that. And so I think like moats still matter.

23:55Bobby Samuels: I'm curious, just because I've had the CEOs from both companies on my podcast, Clint over a CuriosityStream and Marty at Trovio. How do you explain the difference between Protege and those two companies? I mean, I know Trovio and don't know CuriosityStream as well, but I think we take a very research heavy perspective on what we do. So we have an organization called Data Lab. The way I think about it is we have this massive content library and then increasingly a world-class research group mashed together. You've got a bunch of PhDs on your team, don't you? Yeah, exactly. How many? I don't know how many PhDs.

24:39The research group is probably 12 people or something at this point. We'll continue to grow pretty rapidly. But I think we, as a result, speak the researcher's language.

24:50Bobby Samuels: Tell me about this extra classa, if I'm getting the pronunciation right. Polish company you did a data partnership with. What's going on there? Yes. So it's a Polish soccer league. And at our behest, they started filming their games from either six or eight different angles, all in high definition. What that means is for any moment, you're able to capture all sorts of different reactions simultaneously. You're able to see the goal, for instance, see the scorer's reaction, see the team's reaction, see the frustration on the goalkeeper's face, see how the crowd responds. This would be more angles than you would see in an ordinary game that you're watching on television.

25:38Yes. But also having all of them synced together means that it's great data for building models of the world. And as someone who's not played a ton of FIFA, but enough to like, you could say like, oh, this data would make any video game better.

25:53Bobby Samuels: And what's the value, who can use that data out there in this new world? Who loves that kind of data? I think anybody building a world model to simulate what is going on in the sort of physical environment. So it's not limited to soccer. It's really humans moving and their emotional reactions. Yes, totally. And the physics of what does it look like when a ball is on a foot? And what does it look like when a crowd has a reaction together? Any video model would find this incredibly interesting. And we've licensed it to video models. If you're building a video game, this, I think, is catnip. So it really runs the game.

26:34It could be video models. It could be world models. It could be games. I think it can be a pretty broad range.

26:38Bobby Samuels: How much of your content is archived licensing versus, call it like commission capture? In other words, where you know what you need and you will put it out there that Protégé will pay if someone can generate this kind of content? Most of it to date has been archived. I think over time we certainly could move into a sort of commissioned capture. I do think that is a growing trend. folks will want to know what a specific thing looks like. And if it's not a high enough volume in archival footage, folks will want to likely create it in some way. So I do think it's a growth area within the industry.

27:19Bobby Samuels: And that was from Joseph over at Stream Genie. Give me a wild use case of the types of data that AI models want, like specifically with number of hours of type of data. What's some wild stuff you've heard? When you talk about pre-training, it's basically give me as much as you possibly can. So, you know, it's billions of de-identified clinical notes and millions and millions of de-identified images and hundreds of thousands of hours of cinematic content. It's human chores. I think as you get into robotics, we're seeing a massive amount of demand there. You know, multi-petabyte deliveries, not infrequently.

28:04Bobby Samuels: Well, what's a type of data that not just volume, but a type of data where you were surprised? One of my favorite examples was a while ago, we had multiple labs ask about dragons. They like really wanted to have high quality content of dragons. And there actually aren't that many Game of Thrones where you have high quality. And so I think we and everyone else were like, I can't help you there. I think there were a couple of researchers who were just super excited about dragons and wanted to make the models better at it. Who's got the biggest data moat in the world? If it was a set of data you could own, protege could own, what company has it?

28:43So I'll put restrictions aside about what can and can't be done with the data. It's probably Google, Meta, or Microsoft. You know, if you're thinking about how do you train sort of enterprise flows, nobody has more data than Microsoft. My guess is like Microsoft has the lead for enterprise. Meta and Google have probably the lead for like consumer.

29:06Bobby Samuels: What about SpaceX? If you count Tesla being merged into SpaceX, let's just call it Elon Musk Enterprises. Yeah, they probably are up there from like a physical world perspective. But there are other car companies who could film a bunch of road interactions. You see that evening out then? Yeah, I don't think Tesla's data advantage there is massive. Because you can recreate a lot of that data. You can't really recreate the enterprise data that a Microsoft has or the consumer data that a Meta has. Let's talk about healthcare. So Dave Davis on your team said, really, your one of one is healthcare data.

29:52Bobby Samuels: He would argue that's where you are perhaps the best person in the world or most knowledgeable about it. And I want to ask you specifically, what part of your world can help cure cancer? You know, there's a whole bunch of data, obviously. How do you put a dent in that world, which is often very much a silo? I think if you want to talk about purely curing cancer, then it really is you get into a wet lab, you create different molecules that accelerate drug discovery within both foundation models, but then also probably vertical applications that are focused specifically on this. I also was at a conference once where early days of Gen AI, and I heard someone in the healthcare field talk about how hard it is to get FDA approval for something and how sometimes the best way to get FDA approval for something is to take an existing approved drug and then test it.

30:53Bobby Samuels: Yeah, just the off-label. Yeah, test it on new things because it's already passed. Yeah, and that's where clinical data can be incredibly useful. If you're able to say, hey, we use this drug, it was used off-label, but it had a huge effect. That's a very fast way to see potential benefits beyond waiting for a drug to go through FDA approval. What's your advice for kids and school in this new age of AI? I know you've got a couple girls who are not ready for preschool even yet, but what are you thinking about in terms of the future of school for kids? Will they even go to college, for instance? Yeah, I mean, unless they have some sort of particularly offbeat passion and interest, like my expectation is that, yes, both will go to college.

31:45Bobby Samuels: How different will college be at that point? Talking 15 years from now, 16 years? I think a lot of college is an exercise in meeting people and just getting surrounded with folks who you get excited by. I think my general recommendation and all this stuff is just surround yourself with great people and when in doubt, follow the great people you know. And so I don't think that will change. I do think that developing creativity becomes all the more important when the ability to obtain fact-based information is more readily available than ever. But I think there will be an increased premium placed on creativity, on EQ, on those types of software things.

32:35Is it that what I'm thinking about? And just trying to sort of have them have this sort of like... joy of those kinds of things, I think I benefited from off the beaten path kind of stuff over time.

32:49Bobby Samuels: Like what? You know, I got much more out of like my own moves, not off the beaten path, but like, I learned much more doing my like newspaper in college than I did from classes. College newspaper? Yeah. Same thing in probably in high school. I like in fourth grade, like started a football league with my class. It's just stuff like that where the things that no one's telling you to do, but you just let your passion bring you there. I always felt like I learned the most. What do you think about schools that are banning AI? New York State is talking about that a lot, your backyard. It's a balance.

33:26These are unbelievably powerful tools, and you want to make sure folks learn how to use them. at the same time, you don't want to sacrifice that for losing critical thinking. And so I think if I were a teacher, I think it's very hard to police not using AI in take-home assignments. I'd probably be like, you do what you want, but probably put more of a premium on grading and participation and in-class exams. So in high school, I had a class where the homework was a part of your grade, but you had the answer book. And so like you could just copy it verbatim and get a perfect score on all your homework.

34:08But then when the test came, you'd be hosed. And so I think that's like probably how I'd think about AI if I were a teacher.

34:14Bobby Samuels: Hey, Bobby, thanks for all the time today. Yeah, no, my pleasure. Thanks for having me on, Rob. Right on. Well, that's it for this episode of Media and the Machine. Before you go, a few quick things. If you want to hear the next episode, hit follow on the show in your podcast app. If you like something about this episode, the guest, the topic, even one part of the conversation, please leave a rating or review and tell me what. And if something didn't land, email me. I want to hear that too. Either way, it helps me make the show better for you. If you want to go deeper, head to mediaandthemachine.com and subscribe to my newsletter.

34:48Bobby Samuels: You'll get my favorite highlights from the show, plus essays I write about AI. That's also where you can email me or recommend a guest. If they make it on the show, I'll give you a shout out. And if this episode made you think of someone, please pass it along. That's how people find us, and I'm genuinely grateful. Thanks for listening, and see you next time.