Speaker 1Does Anthropic have a real moat as a public company? Do you need the very best AI models? Or can the cheaper ones already cover most daily use cases? Is there anybody out there with hard data to answer
Speaker 2questions like these? We're serving 210 trillion tokens a week. We see a cross-section of tokens, cross-lab, and that is a pretty privileged place to be.
Speaker 1That's the man with the data, Peter Walker, head of insights at OpenRouter. His company makes nearly 600 AI models available to users. So Peter has an unparalleled view of demand across models, price points, and providers. And that includes Anthropic and OpenAI.
Speaker 275% of the tokens on OpenRouter go to.
Speaker 1In today's episode, we explore whether any AI lab can realistically hold on to its customers in the long term. And if you stick around, Peter unveils some of the less obvious factors affecting AI spend in 2027.
Speaker 2If your AI usage bill grew 500% this year, thinking that it's going to grow another 500% is really scary. It's really scary. Welcome to Top Line.
Speaker 1Welcome to Top Line. It's Sam Jacobs. I'm joined by my wonderful co-hosts, A.J. Bruno, the CEO of Quotapath, leader in compensation services, and Asad Zaman, the CEO of Sales Talent Agency. We are delighted to have back on the show, longtime guest, friend of the pod, the most quoted man in artificial intelligence, former strategic VP of data and insights at Carver. No, chief insights officer. Chief insights officer, Sam. Now head of insights at OpenRouter. He has moved from the second best data set in the world to the best data set in the world. And he is demonstrating insights every day. We're getting a new insight basically by the minute, some might say, some say. It's Peter Walker. Peter, welcome back to Top Line.
Speaker 2Another flawless introduction, Sam. Both tied up this time. Yeah, really good. VP of strategic what? That's great. Always happy to join you guys. Not strategic. Why are you even around? Like, what are you doing? Insights, broadly, whatever. But it is a pleasure to be back. Good to see y'all.
Speaker 1Before we dive in, let's walk through the transition from Carta to OpenRouter. Talk about that a little bit because we're recording this, I believe, on the day that OpenRouter is formally a part of the Stripe ecosystem. We'll talk about the transition from Carta to OpenRouter and then tell people that don't know, because when you and I were hanging out in San Francisco a couple of weeks ago, it just struck me that maybe OpenRouter is the best business besides Google in terms of just sheer growth, profitability, and beautiful beauty of the business model.
Speaker 2Up and to the right, man. It is an amazing, amazing place. So I give a lot of credit to Alex and Chris and Louis, our founders. Look, the transition from Carta, I've been at Carta for five years. It was an incredible place. Have so many great relationships there and love working for Henry, but I'd kind of dug into this data set for a long time, and I felt as though a little bit, I was repeating myself somewhat on what's going on in venture. It's always interesting. There's always new information, but I felt this real pull to get back to smaller companies building fast. So Carta was the biggest place I'd ever worked before joining OpenRouter, and OpenRouter when I joined was 70 people, 80 people. So the goal was get back, get closer to AI, move a lot faster and have this incredible data set to work with. And then as you mentioned, 10 days in to my time at OpenRouter, we get acquired by Stripe. So that has been a whole other journey over the last few months. That's been amazing. And we are, as of yesterday, real stripers now. So we get to go in and see what it's like in the big, bad, I think 13,000 person organization. So got to figure that out.
Speaker 3Peter, here's what I'm hearing. You join a company and it gets acquired. So I've just opened a head of insights role at Quota, and I'm going to ask you to tell us a little bit more about your path. If you want to apply. Please do it. It's actually Chief Insights Officer.
Speaker 2I will make you Chief Insights Officer. There we go. No problem. Moving up in the low. Yes. A lot of people with data-confused correlation and causation their age,
Speaker 1and you're making a clear mistake.
Speaker 2Tell us what OpenRouter does really quickly. Sure. So OpenRouter is a single API that allows you to use 600 AI models. So instead of being locked into Anthropic or OpenAI, you can use models from across the model ecosystem. OpenRouter is the first half of the business. It's the one that we're most well-known for. And we'll help you route your requests to models that are of the right level of intelligence. You don't actually need Fable 5.2 for everything you're doing. That's the idea. Save time, save money, whether you're an indie developer or an enterprise customer. The second half of the business I think is actually even more fascinating, which is if you are using open weight models, you can be served a token to those open weight models by dozens and dozens and dozens of different providers. Neo clouds like CoreWeave, Hyperscalers, TwoGuys.io. Guys in a garage with a couple of GPUs. Like you can get this token, this next token from a bunch of people. We optimize that inference for you. So we have actual people come onto our marketplace and compete to give you the best token, whether it's cheapest, most reliable, highest quality, et cetera. So that half of the business, the actual inference provider half is, I think, the hidden part. But it is increasingly becoming more and more important as so many different people try to get into the compute game.
Speaker 4What type of inference provider is it? What type of inference provider would come there? Like, is that a. Chris, I would imagine Fireworks, Base10, these companies, like, are they also part of that equation? Or is it like the tier below them, the smaller up and coming ones? Like, who's there?
Speaker 2All of them. So, yeah, we work very closely with Fireworks and Base10. You know, their businesses are oftentimes probably primarily denominated by people who want to fine tune their own models. And so you wouldn't do that on OpenRouter. But if they have any other capacity, of course, they're going to get into the game and try to serve deep seek tokens. Or Kimmy tokens or whatever. But I mean, all of the major players, DigitalOcean, Prime Intellect, Sail Research, ItWafer, anybody on the inference side is probably on OpenRouter and is competing to serve tokens to these millions of customers. This would be the spot market, right? It's a little less than spot because we're not necessarily saying that you have to spin up a GPU and have that immediately available and, like, spin it down afterwards. This could just be reserved compute that you have access to. But the moment. As anyone will tell you in the compute trade, like, the moment you have GPUs, the number one thing that you want is to have them running at full capacity at all times, right? You need to fill them out in order to make the compute hours work. And so because we have so much demand, we're serving 210 trillion tokens a week, you can compete for those tokens.
Speaker 4Because you're paying for the GPU, like, you're paying for the GPU every hour, irrespective of whether it gets used or not. So you're just. It's like electricity. It's been, like.
Speaker 2And to be clear, we're not, right? I think this is why Sam loves our business so much. It's because we don't have any of these heavy capacity needs.
Speaker 1It's a marketplace with no. A marketplace with no overhead. I mean, you know, 70 people. And Peter probably doesn't want me sharing the revenue numbers. But it was just incredible, the growth that you've seen just even in this year. And you sit at basically the nexus of all AI. Not all, but much of the AI consumption in the world. You are the bridge between people that are. Deploying intelligence into the enterprise and using it every day, and the people that build and provide it. Which is truly an incredible business, and awesome to see how you ended up there.
Speaker 2And I want to be clear, the core labs, Anthropic and OpenAI, are serving many multiples of the kinds of tokens that we're serving. Where we, I think, stand out is that we see a cross-section of tokens, cross-lab. And that is a pretty privileged place to be. Especially as OpenWeight has become actually viable over this last year. You know, if we were having this conversation last year at this time. OpenWeight versus Closed Frontier, like, the gap between those models was massive.
Speaker 1To the point, like, you know, people would say, well, how are we ever going to. We can't, we don't have the infrastructure to have an OpenWeight model. You know, there's all of this complex security and IT and all of this stuff and fiber that we need to put into place. Whereas, as you said, a single API key through OpenRouter, and you can have access to any model, any intelligence, and any inference provider. So, really, pretty special. In terms of the business that you're building. And also, I think the other thing that you mentioned is, there's no downtime. So, you know, sometimes when you're using Cloud, Anthropic goes down. You have dynamic routing to the point.
Speaker 2We would fall back to a different capacity at other places, exactly. So, you're, whatever, five, six nines of reliability there.
Speaker 3Peter, given how early we are to all of this, the growth that you're seeing, the TAM, the opportunity. I have to assume Stripe is like, oh, you guys are going to operate independently for quite some time. I mean, if I'm. I'm like, no, do your thing and take this for as long as you can.
Speaker 2I love that, AJ. I think that's a great path forward. But, yeah, we're super excited to join. We are, like, literally, you know, getting our laptops this next week. Like, it is upon us now. So, we're super pumped to actually get to do stuff together. Because it's funny, I hadn't been through a big acquisition like this on the acquisition side. But, you know, the announcement, and then there's that waiting period of, like, a couple months, figuring out regulatory or whatever. And you're just kind of, like, itching to get started. And so, now we actually can. Let's jump into the core thing.
Speaker 4So, as the head of insights, you've been posting about and getting quoted everywhere about token use. Yeah. And I think the question that we should be building towards today is that by the time this releases, a few days after, most likely, Anthropic grows public. And this is the most magical business any of us have ever seen. You know, it really is. But there's going to be this question. that everybody is asking, which is like, what is the value of Anthropic now as a public company and you know retailers want to be asking this question public investors asking this question and when you're doing that one of the things you're trying to figure out is how much is every dollar of revenue worth and so what is the type of revenue this is and it's slightly different like is this revenue durable the way that we have you know past revenue of tech companies was durable or is this a new form of revenue is the question i want to build towards and i want to use tokens as the thing to get there so can we start with like what is a token just very simply
Speaker 2yeah we can uh we can go all the way up and down the stack if you want let's do it so tokens basic unit of large language models so you can think of it as a word but more often a part of a word little bits of text whether it's spaces between words a couple of letters together etc that comprise the basic unit of these large language models that are read and write by the models themselves and all the model is doing is effectively on the back end assigning a number to that token and then predicting which number is most likely to be next again and again and again and
Speaker 4we've landed on like this is the best way to price a model now right is like tokens well
Speaker 2i think it was the only way to price a model for a long time is that tokens are the basic output unit of llms and therefore it made sense for a while to have token prices be the economic currency of ai whether that still makes sense or whether we'll see you know differences in token pricing i'm i'm skeptical in the near term like i think that every new model that comes out will have the standard input token price and output token price but what actually ends up being charged to the customer if you're using an api versus a subscription can be very very different than the listed price and we can get into why that is i think one of the things that's confusing about ai is a lot of people not enterprises but mostly individuals have really only still experienced ai through a subscription yeah so they get paid 20 bucks a month and you get you know most people that's more than enough right you're not going to hit your 20 limit enterprises vastly different and they're going to be charged by on the api and the api is direct usage how many tokens did you use and the
Speaker 1other thing is that not all tokens are the same right different yeah it's which is a fascinating thing to think about because it's a fascinating thing one of the things we were talking about when i saw you last was how in some ways like tokens are the atomic unit of machine thought and some machines think faster or
Speaker 2slower or more intelligently or it's not so we
Speaker 4use tokens as a way to like compare even models right the fable is usage across like companies were like anthropic usage went down versus open ais but if you take a task the same task and you have two models do it they'll actually use different amounts of tokens for it and there's a different cost per token as well is
Speaker 2that correct 100 and it leads to a massive amount of confusion in what stats are valuable number one and what my ultimate bill is going to be number two and i think it really resolves down to the question of what is a task tasks seem to be easy to define but they're really not they're amorphous variable things that we want to say this is an economically valuable unit of work whether it's a pr that gets merged into the main code base okay that's a task but how do you compare that task one part of your code base to the part of a code base of another company or a different part of your company when they're not working in code or they're working in slides or they're working in docs or whatever these things are very slippery and so there's been some work by a bunch of benchmarking companies on cost per task and they'll run what they say is exactly the same task over a bunch of different models and see how many tokens did it take how long did it take and how much did it cost those are useful as broad examples but they don't speak at all to your code base they don't speak at all to your usage and so benchmarks are nice to have and good i'm glad they exist but our our stance internally is like usage is the benchmark you have to be doing this on your own company usage to understand what models work
Speaker 3for you you guys are all way smarter than me but this feels like when i was in 10th grade learning what a calorie is and take the amount of joules that it takes to heat a water one degree celsius and i'm like i can't understand what that means it is it really is like difficult to understand at the atomic level like what this is and then the question i have is has the definition of token changed at all in the last three years or has it stayed pretty relatively stable
Speaker 2i think you could say it the definition of what a token is has stayed stable what's changed is that there have been improvements in the models reasoning as the biggest one where you're starting to get various types of distinctions within tokens that matter very deeply so let's talk about the biggest one which is beginning of this year agents became a real thing agents are independent uh models that can go off and accomplish a task for you you prompt it and it could spend as long as a day working on that task for you if it's a really complex task bit of work and it can do that autonomously you don't have to check in with it all the time that's a massive change in the way that ai is used agents read tokens very differently than humans so a human who's interacting with a chatbot probably a lot of the listeners have done this where you you type into the chat the model thinks for a little bit gives you an answer maybe it points you to a website whatever it's like a back and forth conversation if it's an agent it's not doing that it's adding a little bit of work to it it's adding a little bit of work to it it's adding a little bit of a you start with a question and then the agent is consistently rereading its own answers to that question as it gets longer and longer and longer and that agent reading was we what we call cashed tokens and those get charged at a much lower rate than output tokens so the entire industry is now optimizing itself around how much how many of these tokens can we get into the cash because it costs about 150 at the price of an output token
Speaker 3an open router i assume does some level of analysis on those cash tokens as
Speaker 2well yes 100 we do a ton of analysis on that so one of the reasons why routing is hard is okay if if we're uh just you know helping you save some money on your ai inference your entire shop uses claude there everybody's using fable and you ag are getting mad because you're like why are these customer success reps using fable they don't need fable this is like a very easy request great let's use kimmy instead so instead of routing to fable 5.2 you route to one of the glm models and it saves you a bunch of money so you can use fable 5.2 and you can use fable 5.2 and you can use fable 5.2 if you are routing to different models there's a chance that you mess this cash thing up and in the moment that you mess it up the next time you send that really long set of responses to the new model you're going to get charged a lot of money for those new responses because you missed the cash so a lot of what we do is try to make sure that we're keeping the cash between different routes as you go through different models so that you are not spending a ton of money that you
Speaker 3shouldn't be spending when we say cash we could also say memory is that is that fair to say
Speaker 2exactly short think of cash as short of money and you're not spending a ton of money that you short term memory so
Speaker 4isn't like when we think of market share like the whole conversation around market share is based on tokens right i think there's two
Speaker 2two market shares tokens or spend so you can talk about tokens yes how many tokens does one lab produce versus another and then the second is how much money do customers spend on anthropic versus open ai every month yeah
Speaker 4exactly so like the revenue you can be like okay anthropic has more revenue than open ai so they have more enterprise market share etc etc but the token ones is actually pretty flawed because like you could game that right like you could you could be producing a lot more tokens at a lot cheaper rate um and so and then just look like you have bigger market share in terms of usage like isn't that
Speaker 2possible yes but i don't think that model makers are gaming token market share by producing models that produce more tokens just for the sake of doing it so
Speaker 4is there some sort of desire or reason to be token efficient like to say that i use the least amount of tokens for this because it's not like i can price a token cheaper and just be like it doesn't matter how many tokens it matters how much cost in total for that task but you're saying that actually having less tokens irrespective of the cost matters
Speaker 2as well it does i mean in just in terms of the cost efficiency to the model lab themselves and the amount of compute that they're going to need to run it but you bring up a really good point and i think that's a really good point and i think that's a really good point some models are going to be less token efficient than others so generally speaking closed frontier labs are slightly more token efficient than many of these open weight players deep seek for instance or the same amount of task might use a lot more tokens they're looping back and forth it's like thinking a lot more for the answers maybe it's doing it very quickly so it doesn't actually add that much time but it is spending more tokens but the price per token is much much lower so it actually ends up like yes it might have spent more tokens but you spent a lot more dollars even though it looked either token chart and the revenue chart look different there's no perfect way to do this there's it's it's an incredibly complex thing and i think that it's one of the reasons why cfo's probably have been so frustrated um over the last like i don't know a year or so as you look at these budgets and you say one my ai spend is increasing very very very rapidly two my number of tokens used is also increasing very rapidly and three i don't know if i'm getting anything for it like what am i getting the economic roi that i want out of this really massive increased token spend and that i think has been the conversation over the last three months what do we what does
Speaker 1this group think what
Speaker 4does this group think about that question i spend six figures now you know we all probably do like you know spend a ton of money on models and undeniably they make what we do better like we are better in terms of what we do um and the business is doing really well too um we're flying but we're flying because the market's hot in terms of what we do in this moment we have great product market fit the thing that i can't say confidently is that yes it is making us better but it is not the reason we're making more money like i don't i cannot allocate this more money that we're making to it yet and i might be a user of it like you know for sure like everybody can like say what they want to say but i it's not a clean line for me i can't say that oh because we're better we're making more money like we're better at delivery but if if the ai wasn't there like would we not it's it's a hard one for me i
Speaker 3can point to that on from rpnl in terms of services we're delivering i think where i i net out on this is i i'm lucky that i have a product team and a cto and this is like these evals they're doing constantly on a daily basis for not just internal use of the team but obviously for our product and inference costs and that's something that i like appreciate because i i what i heard in the board meeting two uh two weeks ago was our cto said every day he just gets on and kicks clod in the teeth to find a couple more nickels or pennies in the cushion for uh for us and like that's a good analogy of like every day we're just looking for where we can save money using clod itself yeah i think
Speaker 2both of those stories resonate really strongly like we talk to cfos all the time or even individual developers who can have very very cloudy ideas of what apple roi is being driven by this ai usage and then others of them who are like no i can point down to the dollar or cent figure exactly what i'm getting out of these models and i think that one of the things that became clear over the last year is okay ai usage has ramped incredibly quickly prices per token have fallen fast but not nearly as quickly as ai usage is so if you're looking to 2027 and you're a cfo if you spent if your ai usage bill grew 500 this year thinking that it's going to another 500 is really scary it's scary it's really scary because then you start actively making the choice and i think that as much as people have said ai has not had an impact on layoffs yet and i think we can debate those statistics like at some point you cannot continually increase ai spend and not compare it to your payroll spend yeah because in this case you are doing intelligence work so eventually you're going to have these trade-offs and if you can't attribute roi to one of them it becomes a more
Speaker 3difficult we're doing exactly this work in commissions for some of our higher paid customers where they're actually looking at their inference their their ai cost versus the amount that they're paying for commissions and saying like we're doing all this work on our ai costs why are we not doing this work on our payroll costs it's really interesting that just happened in the last
Speaker 1two months yeah i i think something's happened i i literally aj think that's the time frame i don't know why but i do think it's like the last two months where all of a sudden agents are built into both of the big frontier labs at this point and then with the rollout of instinct in town and all of these like personal agents you can just see like i was sharing in our group chat that we're all in like instinct did something that my assistant could never do and i could never do which is close down my city bank checking account which sounds it sounds stupid except that it was like a literally impossible thing to do because of how arcane their website is you just have to assume the models are going to keep getting better and any intelligence task will probably that is not specifically meeting somebody in person or talking to them on the phone in a very interactive way displaying deep empathy like pretty much everything else probably can be accomplished agentically
Speaker 4but like if you think about this right and like stripes i think co-founder said on the podcast with sam altman that they spend about 10 percent of the budget on tokens right now uh the operational budget and i remember sam altman was like that's not enough he's like i understand you know we're trying to figure out how to spend more and so it's like you know what's the best way to think about token spend i think of it as like a ratio of like how much are you spending on wages versus ai and it's becoming a bigger and bigger percent right like 20 25 of wages you start spending on ai and so you have to really ask yourself the question that you know what is it doing and i think going to what we were discussing a few seconds earlier you know if if you have a product team like an eng team you can see clear roi like it's it's impossible not to when you get outside of that it becomes trickier and so we know it has really to a product market fit when it comes to code the rest of the stuff is you know we're like okay it's code then it's support and then there's you know some like it helps us with research and then what and so talk a little bit about token prices are going down at the same time these ai companies have to keep unlocking more and more revenue to justify all the spend that's happening in camp ads how do you view this overall thing and then there's these pressures from people like us who are like what is it doing like is it really doing anything like how do you read this situation
Speaker 2you hit on the first main point which is coding is the verifiable task category that has been most transformed by agents because there is both a massive corpus of code to train on and code itself has verifiable yes no outcomes it worked or it didn't and that set of criteria doesn't exist in all of these knowledge work fields doesn't exist for a deck doesn't exist for a doc that's why i think why the models are much worse at writing than they are at code because writing doesn't have a right answer um writing has a style and personality and whatever and this like this amorphous latent space whereas code is like this worked this didn't work and yes you can have right clean code and you can have right clean code and you can have right clean code or more efficient code but that itself is a loop that you can put into the models so code is the first and best use case this is i think something that anthropic got very right they've been very they've been incredibly focused on this only we will build the best coding models and for a while the other labs were messing around with all sorts of stuff and now it feels like everyone's like actually coding is the thing that matters the most what happens beyond coding this is why i think the diffusion of some of this intelligence might take longer than we think because there's less verifiable corpuses of information on these coding and you can see this when we're generating data for robots or having people put cameras on their head and fold laundry like those data sets even if you had a million people do that are massively smaller than the internet and massively smaller than code so i think some of this will take a lot longer to diffuse than people think which in some ways could be good because then you don't have this intelligence explosion i don't know if you spend a lot of time on the internet you don't have this intelligence explosion i don't know if you spend a lot of time on x you read like back-to-back posts about how we're all gonna die in the next six months versus like this is never gonna be economically valuable ever so you know pick your poison what's your
Speaker 1take because you you're also in the heart of where it's all happening you're in the bay you're working at open router you're gonna be on cheeky pint with the collison brothers pretty soon i would imagine
Speaker 2no invite for the guinness there i maybe i'll get to drink one though um look what's my opinion i'm hopeful that we avoid an intelligence explosion in a way that is very difficult to contain and very scary i i think intellectually i believe in the alignment problem arguments like i don't see a logical reason why alignment would just naturally happen i think it is a problem and we'd have to work on it i'm just hopeful that things will happen more slowly than some of the people when you say
Speaker 3alignment problem can you just explain that because i'm not sure everyone knows what
Speaker 2that means yeah alignment is just do the models understand what we want and act accordingly so the classic version of the alignment problem is from nick bostrom a paperclip decades ago yeah just like hey you're a model your goal is to make as many paperclips as possible go ahead and do that and what does it do it takes over the internet and takes over our energy grid and turns us all into matrix little vats to provide energy for the model that is created to make paperclips and it takes over the galaxy and creates as many paperclips like chill out yeah bostrom bostrom scared a lot of people with that but i think the core problem is a real one which is does the model understand what we want and align itself to our goals and it's increasingly this was this was all science fiction but now guys it's not as science fiction because we have agents that run that can run overnight or over a week by themselves doing stuff for us well you saw hugging face and all that hugging face was wild that stuff is that stuff is serious and like in the bay it's talked about all the time because
Speaker 4i think if you like see the magic of the ai every day it just depends on like it's interesting if you hang out with engineers who like literally have seen the magic of its fullest potential today right like they are at the edge it's like i remember growing up in pakistan it'd be like there would be these people that would go abroad they would visit america and europe and they would come back and they'd seen this other world and they would tell us about it and it was like they'd seen the edge and it was really like fascinating inspiring thing to hear about and then like i remember the first time i went and i felt it as well it was like wow like this exists like "What the hell am I doing there?" And it's like, that's kind of like when you talk to these engineers, they've seen the very edge of it and it's magic. And so then it sights and scares them in a way that for the rest of us is just not the case. That said, the way that you explained like the two choices, like, you know, we either get this intelligence explosion that's highly disruptive or we don't. But then does that mean the economy kind of hits? Because then where's the revenue growth coming from that keeps justifying the spend?
Speaker 2Is that the choice? This is the thing for me that, you know, we can put aside some of the more science fiction philosophical debates and just like focus on the commitments. So you got a big IPO coming up. I think it's probably going to do pretty well. But you made the key point. If this doesn't continue growing at the rate it's growing, will we have overbuilt on compute? Will we have over committed to building massive data centers that don't get used to the way that we think they should? That I think is the key risk. You can get a lot of- You can get a lot of opinions about whether or not we're going to do that. But if you're looking at past booms, like we, you overbuild on some key aspect of the rollout, right? Dark fiber and the dot-com boom, railway tracks.
Speaker 1Yeah, but again, but fundamentally, I think railroads is a different example, but the dot-com boom on fiber is it did take a long time to get back to the peak NASDAQ levels, but fundamentally it was correct and we needed the fiber and we will need this compute and we will need these data centers. Now, if we're going to do this, we're going to need this compute and we will need these data centers. So if you're, again, if you're like a trader or you need the money in the next five years, yeah, like you have to be thoughtful about how you're investing, but fundamentally, like we're definitely 100% going to use the compute. I always remember when, you know, I got like a Commodore 64 and then Commodore 128 came out and you're like, what do you need all this memory for? Like what do you need all this capability? Every time your phone comes out with like 256 gigs of RAM, you're like, oh, I'm going to need all this memory. And pretty soon you do.
Speaker 4I used to play, I don't know if you guys remember this game on, on my four, it sits, I used to play Columbo and it used to be on a floppy disk. And it was like, I remember when CD-ROMs came in the computer, I would say, what the hell do you need a CD-ROM for? And you could buy like a CD-ROM that had all the software you would ever need on a computer and like one CD-ROM. AJ, do you have a story that dates you as well? I'd like to get
Speaker 3one for AJ. Back in math class on my TI-80. No, I
Speaker 4was, I was playing even, you know, I
Speaker 3was playing snake on my TI-83 back in the day. Like all of you guys were, there we go. That thing did not have a couple of letters. I recognize, but look, there's only 20% of the U S uses AI on a daily basis. Like the room for growth is obviously there. That's not being argued. The argument is that we're five years into the future and we have all these commitments five years right now. The data, the building, the infrastructure, the politics, everything has to fall in line for that to work. And that I just, it, that's the tough thing to see right now of like how that all it's a
Speaker 2durational, it's a durational mismatch. It's not a question of, we're
Speaker 4not saying that the AI is not powerful. It won't change the world. All of those things are there. We would rather not have a market trash. I kind of want one. I kind of want, I
Speaker 2want a little one. I want like a, a, I want a financial problem.
Speaker 4I keep having the little ones, the little ones we have the five, 10%, we have every few months now.
Speaker 2Yeah. Maybe a little bit bigger than that. Um, but just to like, I think we're, uh, you know, this is, this is dipping back into my old job at Carta a little bit, but like I, the startup sector, for instance, has been running incredibly hot for the last three years. Like these, somebody needs, somebody needs to take a haircut to these valuations, but things
Speaker 4are changing. Like when you think about the state of the market, like you've got tokens and usage, um, that like show up. You know, how people are using the models, then you've got the investment in terms of, you know, building out the compute side of things to provide the compute that the market will need. And we're trying to like find a balance between these two things. Um, but what we have seen in the last bit is like how you finance these, uh, deployments. It has changed a little bit. Like before the, the bar was much lower. Now the bar is much higher. You know, some companies are not getting funded for these long-term contracts. The amount of down payment they have to put is higher. The, the debt rates are much higher. So like the bar is increasing that might create like some sort of, you know, stabilization.
Speaker 2Maybe stabilization. I mean, I just think that there's a lot of crud in this ecosystem that has been built through this hype bubble that will get inevitably haircut the way that it always does when we hit a recession or a rocky patch. I do think that one of the reasons that these companies that you talk about on X all the time about these massive, massive early stage valuations, part of the reason that's happening is because they need actual capital. They need hundreds of millions of dollars in order to build the thing that they're doing. And so if you raise hundreds of millions, you have to be worth billions basically on the day of that raise, no matter whether or not you're actually going to be worth billions. It's a little bit of a accounting fiction, but also there's just a ton of startups that are just not worth what a stated price is by
Speaker 4a lot. I think it's hard to like imagine that, you know, we know that it's always a few companies that provide the majority of returns, but like the scale of the companies that will fail to provide those massive returns is what we're kind of getting used to right now, right? Like if we look at this market right now, there will be like multi-trillion dollar outcomes over here. We can see it already in Tropic and Light the Paste. And so there will be these massive, crazy outcomes, but there will be hundreds of billions of dollars that just go down in flames. And that's a, that we've never dealt with that before.
Speaker 3That's crazy. Well, and, and I mean, Sam and I have both been at events recently where the, the venture funds are like DPI of 0.25 X, but those aren't the best venture firms. You guys are hanging out, but they're, but they're saying it's the 75th, but they're saying it's 75th percentile asset right now. Like that's you're you're who, what venture firms are giving, I saw a $4 check at some point from one.
Speaker 2I mean, Andreessen, like the returns are running to the winners, you know, like Martin Casado at Andreessen has had a hell of a summer, man. Like three or four massive acquisitions. I, I just think it's, you know, there are winners and losers obviously in all of this. I just like the power law. If there's one thing that I loved about working at Carta, it's that it made you realize that the power law drives your whole life. Yeah. Not just in startups. Like power law is the thing that matters. Um, whether it's broad economically, a couple of your relationships, whatever it is, it's like a couple of two or three of these things matter more than all the rest of them. And it sounds like a good Sundays with Sam, right there. There you go. Stealing it. Let me guess. Let me guess right at Sam. I'll write it in your voice.
Speaker 1Yeah. This episode of top line is brought to you by Thinkific. Most customer churn isn't a product problem. It's an education problem. Customers who never got to value never renew. Thinkific is the leading enterprise learning platform where courses and community come together to power business growth. Thinkific helps customer success and revenue teams build learning programs that lower churn, accelerate product adoption and scale without adding headcount, automated onboarding paths, role-based access for complex orgs, AI tools that dramatically cut admin time. Your learning platform should drive growth, not just track it. See what's possible. Book a demo at thinkific.com. That's T-H-I-N-K-I-F-I-C.com. It's like terrific, but thinking about it. Thinkific. What a great tagline. That's actually a good tagline. Well, I made that up.
Speaker 3Oh, you did?
Speaker 1Yeah. Because I was like, I wonder if the name comes from terrific, but Thinkific.
Speaker 3That needs to be their hero on their website. You crushed that. It's like terrific, but it's thinking.
Speaker 4Thinkific. Hasid, will you support us and our sponsor? Strategic narratives and positioning are your strengths.
Speaker 1All right, folks, we appreciate you. Thinkific. Amanda Malco, CRO at Thinkific. Thanks for your support, Thinkific. You're great. And Pavilion Gold, is happy to have you as a member. God bless you. This episode of Top Line is brought to you by something we're building at Pavilion. Listen, if you're leading go-to-market at a company doing anywhere between one and a hundred million in revenue, you're probably making a few decisions right now that you wish you could talk through with somebody who's made them before. Maybe somebody who's making them right now. Do I hire another rep or fix the pipeline first? Is this a marketing problem or a positioning problem? What do I tell the board when the plan is done? I'm not going to tell them. Three months ago, already needs to change. That's why we built Pavilion Scale. We bring together small groups of eight to 10 GTM executives at similar stages, non-competitive, led by a moderator. You call them small councils. You meet with the same people, get to know each other's businesses, and talk through the decisions that are actually on your desk. The part I'm most excited about is what happens a few meetings in. You can say, here's what's really going on. And the people that are actually on your desk are going to be the ones that are going to be the people in the room already know enough about your business to help you think it through. If that sounds like a group you'd want in your corner, check out Pavilion Scale today.
Speaker 4Coming to this question we had at the beginning about Anthropic's revenue and, you know, other AI companies' revenue quality as well. One of the interesting things was how at the start of this year, OpenAI was lagging Anthropic, and then they released Astra. And Astra was not significantly better than every other model, but it was like maybe incrementally better than the best model that Anthropic had in that moment. But they also had other mechanisms that they were able to use. They started focusing on the enterprise and then they started using pricing in addition to a high quality model. And they carved away a significant amount of market share as a result. And they reignited their growth. I think that growth was very similar to Anthropic's during this last period. And then Anthropic now released another great model and then they became number one. Talk to us a little bit about what you're seeing at the frontier level, like how easily customers are moving around from one to the other. It feels like in the old world of software, if you got a customer, people assumed you could keep this customer, there was stickiness. Has stickiness just gone away? Yes and no.
Speaker 2So what we see from our data is a couple of things. One, it's very rare. Once you start flipping back and forth or using multiple models from multiple labs, it's very rare that you suddenly just like swear off a lab entirely. So yes, you might have said Astra is a great release. We're going to make that GPT Astra from OpenAI our like daily model. It's going to run a lot of our work for us. But that doesn't mean that you're no longer using Anthropic models. You're still using them in bits and pieces and for other things. Same thing with Anthropic comes out with Opus 5.5. Opus 5.5 is awesome. Let's smack it back. Our daily driver, but we're still going to use some OpenAI models.
Speaker 4On this, can you just explain something that I think it took me a second to realize? They're not using all of these through Cloud or ChartGPT. So they're switching between the models. So where are they using it?
Speaker 2So you can think of it in classic CFO fashion as COGS or OPEX. So you can use it in one way, which is this AI model runs a part of my application. Right? So Legora, for instance. Legora has used many models underneath their software to help lawyers make better decisions. They might be serving those lawyers with Opus 5.5 today, or they might be serving it with Astra. They're probably not exposing that to those lawyers. They're just saying you're using Legora and we're serving you tokens from individual models. So that's one. It's inside the product. The other is with your engineering teams, with your marketing teams, with your successors, we as a company consume tokens from this lab in order to build stuff in our company. So two different sides of it. And you can switch a model out of production or you can switch a model out of your engineering usage. And both, you're right, Asad, pretty easy to do so. Used to be harder. Now it's like, oh, Opus 5.5 is good. Let's try it. Let's throw it in there. Maybe it's better than what we're using today.
Speaker 3Peter, something that Asad has said many times before, Legora, Harvey, you show up and you get this PhD level when you first start using it. And then all of a sudden you're switching to like a kindergartner and then kind of a. And switch. Is that overblown or is there some truth to that?
Speaker 2There's some truth to it. I don't want to speak to Legora and Harvey in particular because I don't know. I haven't used those products. I think what, you know, I was recently at a conference where Kyle Poyard on Growth Unhinged was talking and love Kyle. And his work is like, it's very difficult sometimes now that you are in a rapper game to say how far, what layers of abstraction can we put between us and you buying the model? Because eventually, if you don't do that, what your customers are going to do is they're going to say, why am I paying for you instead of going directly to ChatGPT for this? Like, what extra value have you brought to me that the model itself cannot give me? Maybe it's the harness, maybe it's the data, maybe it's the context, whatever it is. Like, you need to predicate your value on the products you're building on top of these models so that you don't get competed away by the models themselves. Even so, I've never seen a part of tech where two things are true. One, there's very, very, very, very, very small switching costs. Almost none. So like, if a model is better, people will switch to it. And so that actually makes all this media hype that you need on X and stuff pretty important because that feeds the thinking that, oh, this model is better. So many people are switching. Like, the media aspect, the actual brand of it matters a lot. And the second is, even if your model is awesome, so if you're sitting at Anthropic right now, they have Opus 5.5, it's still everybody's favorite model across most of these metrics. They know that that time is coming to a close, and they're going to have to release a new version of Opus in the next couple months. Even if you win, even if you crush it, you
Speaker 4only get. It's six months, right, is the window. Like, in the S1, I was seeing how most of their revenue, like 80 plus percent, comes from releases within the last six months.
Speaker 275% of the tokens on OpenRouter go to models released in the last 60 days. It's planned obsolescence. Like, no matter how good a model is, it will be out of date in six months at the most.
Speaker 4How much of the. The choice as to what I'm using is model quality versus the cost. Because it seems that OpenAI really used pricing as one way to carve away attention. Like, a really good model, comparable. But maybe, like, I think we're going back to the earlier conversation we were having, where I'm starting to be like, okay, how much of revenue has been driven by this thing? So, if somebody came to me and is like, hey, at the company level, here's a model that is comparable to what you're using. It's 5% less good, but I give it to you for 30%. I would really consider that, right? And so, how much is price coming into this equation now?
Speaker 2It's a big input. People kind of bucket their work into different task categories. So, for instance, in coding, generally speaking, there is less price discrimination. Meaning, you can always find more value from higher intelligence in coding. And so, you're going to pay up for the most intelligent model most of the time. In other places, maybe intelligence is kind of saturated on that task. And you're saying, I would like to read this PDF. Like, the goal of this model is to read this PDF as good as possible. Well, if I find a model that reads the PDF 98% as well as Opus, but costs me literally 10% as much, I'm going to use that model. Because reading PDFs is not the most high-value task for me. And so, there's a bunch of distinctions that happen across the kinds of tasks you're running as a business. And what we're finding these days is that, you know, there's a lot of things that you can do that are not the most high-value tasks. And what we're finding these days is that, some of this is just inertia. Some of this is like the crud of, oh, we hear all the time that from CFOs that they go, my whole company is using Opus and Fable, and I know they don't need it. They're just using it because that's what they have access to. Or because, candidly, it is more important to use them, right? You feel cool if you're using the latest and greatest models, but you don't need all of that all the time.
Speaker 4It's like the company that doesn't give you the MacBook Pro and you're like, oh, I have the ThinkBot. Yeah, you're like, I'm on Lenovo? Like, screw that. Not for me. Not me. Except that
Speaker 3it takes 28 minutes to answer you. That's the downside of it. Have you guys ever used like MaxEffort Fable before? I used it once and I was like, no, I'm never using this
Speaker 4again. I use Lightning on Chachi, but it's crazy. Like, it's amazing.
Speaker 2Sam, you and I, I think, spoke about this the last time we were together. I personally don't think I have tasks that touch the frontier edge of these models anymore. I agree. I don't do anything day-to-day that proves to me that Opus 5.5 is vastly better than the other models. I just don't
Speaker 1know. I needed Opus 5.5 to cancel my Citibank checking account. That is for sure. Right. So that's a hard task, evidently.
Speaker 3Wait, so Peter, what would you recommend on a daily basis? Like, just what tasks? What do you use?
Speaker 4It's interesting. Peter is saying he doesn't know. Like, you know, and that's, I think, part of the challenge is like, I don't know. Like, maybe I do. Like, I know that sometimes I use Sol and not Astra. And it makes more mistakes. You seem like a big Sonnet guy, Asad.
Speaker 3I think you're a Sonnet guy.
Speaker 4And I find it super annoying when it makes mistakes. I'm like, man, I went like, I've used the cheap model and it's making all this mistake. Crossed me two seconds of my life. I think I mostly use Sol.
Speaker 1I don't think I use Astra very much at all. And I use ChachiPT as my main. I can see like the
Speaker 4errors are more and it like it catches me, but I don't know how to compare it. Like, I just don't know.
Speaker 2I think this is the like in-house skill set that companies must have moving forward, which is you need to be able to evaluate your usage. You can either have someone like OpenRider help you do that, or your engineers can do that themselves. But either way, you have to be consistently evaluating is, am I using the right model for this task? Because it's very easy just to use what you've been using and then start paying more. And I think that's a great way to do that. I think that's a great way to do that. I think that's a great
Speaker 3Do you have a breakdown of what you personally have used the last three months and what's come to the top? Yeah.
Speaker 2So my personal usage is probably if you looked at the percentages, I'd say 40 or 50% of it is Opus 5.5. I don't think I probably need that, but it is actually very good for data analysis and design. I'm starting to really use Gemini models like Carbon and Argon for writing. I think they're better writers than Claude in a lot of ways. And then I have started using. Sorry to interrupt.
Speaker 1What's the mechanism through which you interact with those Gemini models? Is it some writing application that's using OpenRouter or are you just logging into Google?
Speaker 2No. So OpenRouter works across all harnesses. So if you're in Claude code, or if you're in Codex, or if you're in OpenCode or anywhere, you can use OpenRouter and have your models available to you. You can also just go to it directly and. the terminal if you want to, you know, whatever is the workflow that you're using most definitely. But I can, like right now I can pull up my harness and say, I'd like you to write this, have Gemini write this, have Opus write this and have Kimi write this and I'd compare the outputs.
Speaker 3So, but you, you chose Gemini yourself. You think it's personally a better writing.
Speaker 2I'm getting a lot of this from my Slack where people are like, Hey, I've been messing around with Gemini models and I think they're really good writers. People should try them out. It's like a lot of this is coming from input from other folks. If you are using open router and you just want us to handle that for you and to pick the right model, the best model for the cheapest price, we can do that. It doesn't always get it right, but it is pretty damn good. And you know what? It turns out a lot of the tokens still flow to ChatGPT and OpenAI or excuse me, ChatGPT and Entrofit. Like they build really amazing models.
Speaker 4So you've got these companies that are at the frontier and these are companies that both are going to go public one very soon, one in the very near future. And so when we look at their revenue, what are we betting on? Are we betting on that? There's this six month window in which if you have the best model, you will capture the most revenue. And this is the team that will do this consistently for the next five years. Won't get exhausted, won't lose its people. Won't, you know, have this group of people that are like, Oh, we all have made a hundred million. So let's throw balance. And now they can't maintain that six monthly. Like what are we betting on?
Speaker 2You're betting on that. The, the best model has within it pricing power that is close to permanent and the labs that can create the best model are very few and that they'll retain their best people and that they won't stumble and that regulation won't slow them down or the state of the world or the number of data centers. You're making all of those bets. And to date it's worked really well, but will it work in the future? I think it's a speculative bet
Speaker 4for sure. Liquidity is this interesting aspect of this batch, right? That there's two places where I think of liquidity. One is I'm betting on this company being able to maintain its position at the frontier for the next six months and the next six months after that. And I really do think like when you look at, I think the stat is one in 17 people that will have more than a hundred million dollars in the U S will be entropic employees, like some crazy number like that. So there's like so many people that will make so much money. Like, how does that not change? Like what happened with space X, a bunch of people were like, I'm, I'm out, I'm retired. I'm going to teach yoga. I'm going to chill. I'm going to do some. So how does that not affect this company's ability to maintain its position? One number two, there's only so much money in the public market. So you've had space at school public this year, you have entropic grow public, you have open AI grow public. How much more room is there left for other companies? Like, is it just going to be like every day? Like, and then we get one of these massive ones and that's the bar now.
Speaker 2I am worried about what the state of going public looks like over the next three years for not the labs. I really am. I look, it, there's no such thing as a window being closed or open. Like you can go public at any time, but like culturally, will you, if you are metrics and other things are vastly different than the labs. If you look at the corpus of private companies, that are not a labs, but are massive, you know, I think speculating on when those, that set of companies will go public is kind of a fool's errand. Like I'd imagine most of them don't because they don't have like ramp does ramp go public in the next year. It isn't, it isn't,
Speaker 3it's not different.
Speaker 2Like stripe stripe isn't gone public. Like, yeah, but the reason I think this is increasingly worrisome, or I don't know exactly how you want to phrase it is at some point. I don't know if it's now, but at some point, people who are not in private markets get pissed because the growth is not available to them. The retail investor in America does not get to access most of the best companies in America, or at least the ones that are growing more quickly. We talked about this a lot. We talked about this when you were at Carta. We have to agree.
Speaker 3I definitely agree with you. It's bad for the
Speaker 2American public and the American economy for some of these companies not to be public. Like, I'm glad Anthropic is pushing out there fast enough because at least you have one AI lab champion out there. But you know, this 2025 year private market period means that a lot of the growth will be done by the time they hit public markets. And so like, if you are not in the flow, if you're not an insider, I can see getting annoyed by that.
Speaker 3That's why SpaceX, they're saying like, oh, this is going to go to 10 trillion. Like there's a lot of momentum to say, like, there's a lot. Of growth that's not baked into the price right now.
Speaker 4Well, that's definitely true. Yeah, but like so much has already happened. Like, that's the thing is that the size at which they grow public, they get there faster now. Like Anthropic got there so quickly. You know, maybe there's an argument where like companies will get there faster than before, but they're still like Anthropic's growing public at a trillion and a half. Like it's the second company in history that goes public at above a trillion. The other one was earlier this year. Like, you know, this is a very different world. It really comes down to like this token situation, which is like, as we're talking and we're like, it's starting right now where people are like, hey, man, like, so this is becoming a big part of my budget now. Like a big part of the budget for every company, even like a recruitment company like ours is spending six figures a year on Anthropic. Like that to me is like wild. And so at some point that number keeps getting bigger and bigger. And you know, it does start eating into labor budgets. And then you have more and more people sitting on the sidelines. It's like income inequality and also the lack of job opportunities creates a lot of friction in society. Who knows where that goes? What do you think happens when all of these employees make all this money at OpenAI and Anthropic? Like what's your guess that in the next six to nine months, what's the impact on their business performance of all of this liquidity that their employee base gets? And especially the more like senior people that make lots of money, like what happens? We've never seen it.
Speaker 2Smaller versions of this when Google went public, et cetera, different scales didn't seem to slow the companies down. You had a version of this clearly with SpaceX that minted a ton of millionaires. I think one of the differences at the major labs is that, yes, of course, you have a bunch of employees who are there because it's the place to be in tech. They're really smart and they want to work on hard problems. It's a great place to do that. You have another cohort of employees who are very, very missionary driven. They're saying, not only is this, quote unquote, the place to be because it's high prestige, but this is literally the only thing that I want to be spending my time on. And there's no better place in the world to be spending my time on it. So like, yeah, I might be wealthy now, but this is like humanity's last invention.
Speaker 1I had a friend I met up with at Anthropic's offices last time I was in San Francisco, and this guy, he was a missionary. He's like, you know, I don't know how to describe it, but Dario really is the most inspiring. Leader I've ever encountered in my life.
Speaker 4I've seen a lot of people, you know, speak very nicely about someone who's made them a hundred million dollars. Everyone becomes a follower.
Speaker 2There's a little bit of that, like religious zealotry in it, you know, he's so good looking, he's so
Speaker 4dashing, he's so charming. He's like the best at everything. He's got a great cardigan, really strong cardigan game.
Speaker 1I was like, how often are you guys thinking about, are you guys worried about the IPO? Honestly. We never talk about it. We never think about it. In Dario we trust.
Speaker 2In Dario we trust. Come on. Come on, man. Come on, man. Yeah. I mean, I'm assuming some of them will leave. Hopefully they become angels and like fund some startups instead of just taking a do stuff
Speaker 4with, but I have a feeling that almost no one in tech or no one in the world is really a missionary, but it was like when you test it against money and that's what this AI thing has done is like when, when the upside is enough, no one's a missionary. Like when the upside's enough, no one's a missionary. When the upside's enough, the builder doesn't want to be at the series A company. They want to be a director of sales at, you know, Sierra or at open AI. Like, I think there's very few actual missionaries in the economy and that's not necessarily a bad thing, but there's very few. I think there's a lot more of them.
Speaker 1They're all post wealth amongst founders.
Speaker 4And there are founders are by nature are missionaries, right? Like by nature, but I think the non-founder base, like within the employee base, like very, very few are true missionaries and maybe anthropage has a bunch of them. And they might have a bunch of them.
Speaker 2Yeah. If anybody does it, that they do.
Speaker 4Hell yeah. In Dario, we trust.
Speaker 1So, uh, before we, before we wrap Peter, it's always great to see you buy or sell the anthropic IPO at 1.4 trillion.
Speaker 2Uh, sure. I'll buy day one. Why not? You know, what am I doing with my money anyways? All that Stripe money, you got that Stripe money, you got that friends money. No, no, no, no, man. We just got acquired. There's many years left of building to before that happens. But I think it's, uh, I don't know. I think it's, it's kind of, I'd almost want to just buy, just have skin in the game. Just to be like, look, this is the thing that I'm thinking about all the time. Anyways, like might as well, might as well be in economically entangled as well, not be a bias and like, I'm a buyer.
Speaker 4I'm not saying you, I think like as anyone, like you have to be a buyer of that, like that makes no sense. Just, I think, I think
Speaker 2excite what I want, what I'm dreaming of is for all All the back and forth debates to like at least now have a little bit of skin in the game. Like, cool, if you've been hating on Anthropic for the last two years, then short it. Short it and show me that you shorted it. I will respect you.
Speaker 1Shorting is hard.
Speaker 2Exactly. I agree. There's a lot of people out there who are like, this company is a fake Fugazi marketing scheme. None of these models do anything, blah, blah, blah, blah, blah. Like, put your money where your mouth is. Sam, you were about
Speaker 3to say the big loser from the elections was? Keith Raboy. Oh.
Speaker 1Maybe. All of the receipts. All of the screenshots where he's like, the Dems will never take Congress. Take it to the bank. Bet somebody $50,000.
Speaker 4But he'll keep doing this and next time he'll be completely right and people will be like, genius. That's how like Ray Dalio became so famous. He just kept saying the market's about to crash, market's about to crash. And in 20 years, he was right three times. And they're like, this guy can see. He's the future. He's still right. It was like, dude, he was wrong 17 out of 20 years. It's like, it doesn't matter. You were right three times.
Speaker 2I love when we get the like patented, hey, this is a person that hasn't come up yet in the pod. Let's take shots at them from outside. Yeah, Ray Dalio. You're 75 years old, but who cares?
Speaker 4There's a list on your iPhone where you're like, I hate that guy. I hate that guy. Dude, episode 437, they're still alive. I mean, Principles came out 15 years ago.
Speaker 3And you're just. Just hating on Ray Dalio.
Speaker 2Look, the holiday, the holidays are coming up outside. If you want to make a I hate calendar and just send it over to me, like I'm more than happy to do that. I think you can do well. Yeah, I think you can do well.
Speaker 1Peter, thanks so much. Great to see you. Congrats on the new gig. We love consuming your content. We love talking to you more regularly. And it's always great to see you. Thank you for joining us. Next week, folks, we're going to have Sam Blond from Monaco on the pod. So make sure you tune in and thanks for listening. If you are looking for more Topline,
Speaker 5check out the Topline newsletter at toplinemedia.substack.com. That's toplinemedia.substack.com. Or if you're a video person, check us out on YouTube. Topline-media is what you want there. Have a delightful day, everybody. We'll see you next week.
PodcastTranscript AI