Loading summary
Transcript51 lines
- [00:00]
Travis Cannell
We're having like a chatgpt3o moment, but it's bigger.
- [00:05]
Wes
I'm sure there's going to be other players that will try and definitely disrupt you. Whether it's like, hey, we have all our GPUs in a country where energy
- [00:13]
Travis Cannell
is really cheap, bring it on, let's compete, right? I mean, price is important, but everything else I was touching on is very important as well. The customer service, the platform experience, the way the templates work. Each and everything, each and every piece of the software kind of all has to come together for the customer to have a really good experience.
- [00:31]
Wes
So you're basically the Airbnb for renting GPUs.
- [00:34]
Travis Cannell
We are. And I think we used that tagline for a while when everyone was talking about the sharing economy. We thought that Ethereum was going to be a big driver of that. Moving to proof of stake, but it just really didn't matter because this massive thing of inference.
- [00:54]
Wes
Welcome to the product podcast. With me today I have my co host, Esben, who is entrepreneur in residence at ProductLED and who is the co founder and UserFlow and Cobalt. And today's guest is Travis Canal. So he is the COO and first employee at BAS AI, which is a marketplace for on demand, low cost GPUs that power all of your AI solutions. And so the company is on a tear. They're on the Brex 50 fastest growing companies and that's how we, we found out about them. Then we started signing up for their product. We're like, okay, this is cool. I can see why this is the case. So, Travis, it's awesome to have you here. Thanks so much for coming on.
- [01:29]
Travis Cannell
Yeah, I'm really happy to be here as well and talk a little bit more about the VAAS story.
- [01:33]
Wes
Awesome. And like we were going through some of the stats as well. I know, I think. Was it like last February as well? You had like 19,000 signups for VAAS.
- [01:41]
Travis Cannell
Yeah, it's been growing a lot. I think we're at 27x year over year growth the last time I was looking at this for signups. So, you know, 19,000 in February alone is what we put out in our blog post and that's starting to even pick up even more as we get into March. But there's a lot of people interested in GPUs now and you know, we're here to. We're here for it.
- [02:03]
Wes
As you were saying, like you've done an amazing job on. And we can dive more into this, but on like getting referrals from whether it's ChatGPT, Cloud, all these different places where if you want to rent a GPU fast, AI is doing a fantastic job of becoming that number one place where it's like, okay, great, you can learn about it and then test it out pretty easily. But what do you think is driving this crazy growth? Is it just LLMs that are referring fast FPI or is it some other need here that is now more than ever omnipresent where people are like, I need to rent a GPU for this.
- [02:35]
Travis Cannell
Yeah, I think it's inference. Over the last couple years there was this idea, you know, all this money is pouring into training and it still is. These data center builds out these massive infiniband clusters. There were 4,000, 8,000, I think this Colossus 1 is over 100,000 GPUs. So that is capex and expense that goes to training. And to get money back though, you need to take that model and you need to run inference on it. And so that comes out as a cogs and inference is really allowing the model to live. And these ways that the models now and Claude code kind of lays it out bare with its context, it really starts to feel a little bit more human, like okay, you can talk to it for this long and then it needs to go to sleep and then you get a fresh one that doesn't quite remember everything. But the inference is very important. So for a while we've been thinking about that and inference is driving this. And inference in my mind is just going to continue to explode as we not only want to run the artificial models, but as more and more we take living things and digitally recreate them. If you can imagine the three of us. I think I want to run inference on my mind and there's quite a few humans that probably want to run inference on their minds.
- [04:01]
Wes
Millions of people that want that now just to ask a dumb question here. So for some people that are listening, they're like, okay, what is inference? Before we get too much deeper, can you just like kind of give the high level overview so somebody's not lost at this point?
- [04:15]
Travis Cannell
Yeah, let's back up for a second. I mean when you train the model, you're making the model weights. It's really a production step. So training is something that takes these massive data centers. They all need to be wired together. It is a very intensive process where this cluster will run for months at a time. And the output of that is a, is a model that's kind of what you're Talking to. When you pick a model with ChatGPT, when you pick a model with Claude, everyone's talking about Opus 4.0. That's a model. You know, Sonnet is a model. ChatGPT has its 5 series, that's a model. So those were trained. And then there's a step where they're fine tuned and there's feedback with actual humans. They're worked through until they're actually get ready to talk to actual people. But the model itself, you cannot talk to it if it's just sitting at a file on your computer. It doesn't work. What you need is you need to take that model and you need to load it into the RAM of a GPU and then you can talk to it. So when you're talking to this model, every time you prompt a model, it is running on a GPU somewhere. And that's very different from training. When the model is running on this gpu, it doesn't require Infiniband. Sometimes you can compress the model and there's different optimization steps so you can get it running on a smaller footprint. That's called quantization. And that's a big area of active research. A lot of these open source models, they keep on trying to squeeze them down so you can run them on a 32 gigabyte GPU, maybe even smaller. But inference you can just think of when you're talking to the model, when you're prompting, when you're in an interactive session, this model is running somewhere, it's alive. Inference is life.
- [06:05]
Wes
Thanks for jumping in there. And just to give a little bit more context, that's always super helpful. Now when it comes to when the. Did Vast AI go into like hyper growth mode? When was that like pivotal moment for you and the team where you're like, oh my goodness, like this is taking off, things are growing every single month. Like, when did that kind of happen for you?
- [06:24]
Travis Cannell
It's interesting. I mean, vast launched in 2018. I came on in about 2022 and we've always been about doubling on a trajectory, sort of the long term kind of doubling with some deviations from that along the way. But it's, it was really around November that we started to see things kind of heat up and now it's going crazy. Just in February, March, there is something happening. And it's happening kind of right now. It's really 2026 is looking to be a lot different. I mean, on my team, from what I can see on the ground, it was Opus 4. 6 and Claude Code. There's some sort of combination where we're having like a chatgpt3o moment and. But it's bigger. It's bigger this time. And is it just that more people
- [07:11]
Wes
are like coding and they're like, hey, I need ICC more GPU power. And like, okay, great. I fast just so happens to be the tool I need in that equation.
- [07:20]
Travis Cannell
You know, I mean, this is all happening so fast that we're still trying to kind of figure out and sort through exactly what we're seeing. And we're a worldwide company, so North America isn't even really our biggest market. So we have a lot of users in APAC Europe. It's a worldwide phenomenon where people are coming and wanting to run GPUs for different things. Experimentation, fine tuning, fine tuning models on their own data. Then of course, running models in a more private way is important to a lot of people. So we do see some of the templates that are popular. There's a lot of image generation, video generation. The video generation requires a lot more computer for the given output of videos. So there's these companies out there that are fine tuning models, you know, for specific video type purposes. I think that's really exploding and taking off. But it's just one of those things where it's hard to. Like a frog boiling, like suddenly we're boiling and, you know, you're wondering how we really got here. It's hard to know for sure, but it just, it seems like it's coming. There's a lot of different things kind of all coming together. And then on the LLM front, there's a lot of people that are. There's new use cases opening up. And as things get cheaper, you know, we expect that to accelerate as well. I forget the name of the law, but as things get cheaper, they become more useful and then people use them more. So it's this kind of weird paradox that you hit.
- [08:49]
Wes
Yeah. And I also think like this goes back to what is happening in even just development too. It's like clock code and everything else here too. It's like they're. If you look at the total adjustable market, there's like the very small amount that is just builders. They. They know how to code. They. They can do everything. They understand the complexities of how to build it all. Very small part of the market. And now it's. It's like the bigger part is editors. It's everybody else who can just prompt and figure out, hey, this is how I want to describe something. And they're Like I can code now and that size of the market is 10x what the builder one was. So that's why we're seeing a lot of these AI companies just like explode because they're actually. The TAM is way bigger and way deeper than anything they've ever seen before in that same market. So it gives the illusion of like, wow, they're just like on this breakout growth like trajectory. But like the same person that uses like a lovable is not going to be using somebody like, like an ide because they're like, I don't quite get it. But now it's like that's they can because they just plug in with cloud code, they're good to go. So that's fascinating. Now why would you say AI teams choose vast instead of like AWS or other alternatives like Lambda Labs or some other pieces here?
- [10:05]
Travis Cannell
Yeah, I think that's a really good question. You know, it's. I think a lot of it boils down to cost. So we're trying to bill ourselves as the most efficient solution for inference and basically have created this system where we have all these different hosting partners that are setting their own prices, they're listing their own equipment, they're making their own claims about what this equipment is and how good it's going to be. And then we sort of verify everything. So we verify the equipment is kind of what they say it is, and then we test the bandwidth and everything. But we're trying to keep it very minimal and bare bones and we don't own the equipment ourselves. So our platform tends to have the lowest prices for each GPU.
- [10:50]
Wes
So you're basically the Airbnb for renting GPUs.
- [10:53]
Travis Cannell
We are. And I think we used that tagline for a while when everyone was talking about the sharing economy and when the sharing economy was really cool. And a lot of those too. We've moved into data center operators, what we call data center operators. We have a blue label on some of the offers and those are run by reputable data centers that have more equipment. Many times they have the certifications and so those are kind of for more, for serious production workloads. And then we have kind of the longer tail as well. I think our median kind of host might just have a dozen GPUs. Oftentimes they put them in a data center as well. But there's kind of these. It's an entrepreneurial class that has been running and managing GPUs. And they don't want to do customer service because customer service is difficult. And they don't want to do marketing and they don't want to deal with business. And so it's like can I just buy this equipment? And then yeah, can I just buy this equipment and get some money? And we're like, yeah, you can just come to vast list it with us. You can control a lot. You can control the price that you want, you can control the terms. If you want to rent it on interruptible, if you want to rent it on demand, how long, you know, as, as the owner of the equipment, you can determine a term which is kind of important. Like do you want to list it for six months, one month, do you need it back next week? Maybe you only list it for a week. Probably can't get the highest price if you're only going to list it for a week. But you know, we, we want to create liquidity around people trading compute and coming to terms for compute. So that's always been the goal.
- [12:35]
Esben
You, you basically have a two sided marketplace. Right, but, and it sounds like you are offering your suppliers, you take away some of their, the work they don't like to do and then you have the demand which is there due to the whole AI boom. So I assume it's been harder to build the supply side up though than the demand side or has that changed over the years? Was it initially the demand side that was hard and now it's the easy part or how has that been?
- [13:04]
Travis Cannell
Yeah, you know, it switches, it kind of comes and goes is what is harder. I think at the end of the day there's only one side that's inputting money into the system. So we tend to focus on that and that tends to be sometimes the harder scenario. The way that our system works because the owners of the equipment can list offers, they can raise their price as they see the availability dry up. Like any good market system, that is a signal, It's a signal to invest, get more compute, bring on more capacity, maybe grab a couple machines that you had sitting on some other platform or you had that weren't earning as much. There's other ones out there where they don't let you. They tell you how much they'll, they'll pay you. They don't have that kind of very fast market based signaling. And so you know, our hosting partners tend to expand. They have a great retention and expansion kind of metric. So they reinvest with us when they start to earn more too. They buy more equipment. So on that side a lot of it is helping them. We're getting into that More and more helping them with potential financial partners, potential partners to buy compute where we know there is refurbished servers, things like that, at good prices. We're trying to get into that more and more with our sales team to help those suppliers. But I think both sides of the market have their complexities and difficulties, but the demand side is quite difficult and you have all these users that have AWS expectations in terms of how stuff should work and want to pay you $5. So it's quite a lift to get a 24, 7 support team running to deal with questions coming in from APAC on Friday night when you're trying to go get some dinner and pack up for the week, you know, and then over the weekend it's just really, you know, people experimenting and playing with these things all the time. So that's certainly, if I look at our team, there's a lot faced that client experience and making sure they have someone to talk to, you know, and how to set up an SSH key because a lot of people don't know how to do that. So they see step one is I need an SSH key. Okay, so do I get that from you or where is that? And you're like, we, you know, we have a lot of Linux training that we like to do at vast. For someone that's trying to do something quite complicated, typically more complicated than might be the more, something more technical than they've done before. So anyways, that's a very long winded answer to your question, but hopefully kind of covers most of the bases. But I do think over my four years it's been the customer side that's like more important and harder in a certain way because they're giving us the money. Our hosts, they're our partners, they're incredibly important, but they're not our customers. You can always ask a very simple question, who's giving you money? If somebody's like, hey, I'm your customer, you got to do what I say. And it's like, wait, am I giving you money or are you giving me money? Because who's ever paying? The customer's always right, but that's the person that you really want to help and understand deeply what problems they're having with the platform and track all that and make sure that they're having a really good experience. And so that's been quite something to set up that support team for them and try to deal with everyone from that $5 customer, that's somebody in India just trying to run a model all the way up to Somebody running a production workload at scale that's relying on us for their business.
- [16:45]
Wes
If we look at your website, the main moat for you kind of like were thinking as like, hey, it's definitely, it's cheaper price than the competition. That's the one you definitely beat a lot with. Is that something you'd agree with? And like, how are you planning on defending this moat in the future? Especially since it's definitely the hot market to be in and I'm sure there's going to be other players that will try and definitely disrupt you, whether it's like, hey, we have all our gpus in a country where energy is really cheap or something like that, where they can actually compete on price. And it could be a little harder to do that with a distributed network.
- [17:21]
Travis Cannell
You know, I always say, bring it on, let's compete. Right. I mean, price is important, but everything else I was touching on is very important as well. The customer service, the platform experience, the way the templates work. I mean, each and everything, each and every piece of the software kind of all has to come together for the customer to have a really good experience. There's been competition since 2018. Competition is just a sign that it's a healthy marketplace. And I think the key is moving fast and trying to stay ahead of the competition. And we will have a new version of our website out very soon that focuses on the pricing of our GPUs. It's quite interesting. We have a lot of pricing data, so we'll start publishing a lot of the 30, 90 and 180 day pricing for each GPU type, which I think is a cool thing for the layperson to kind of monitor and look at. If you're in this industry, you want to understand what's going on with B200 or H100 pricing or even 5090 pricing. We'll publish that a little bit more and move a little bit away from comparing ourselves to so many other competitors. Because I think the pricing kind of speaks for itself. And the way I think about it now is more of an efficiency angle for inference.
- [18:37]
Wes
Yeah, that's good. And when it comes to the unit economics of like a single GPU that gets hosted on bas, what does that, that look like? Obviously it will depend on what kind of gpu. Then there's like, you know, I, I don't know if it's hundreds or thousands that you have access to, which is thousands, but what does that typically look like for somebody that is, you know, hosting their GPU on fast?
- [19:00]
Travis Cannell
Yeah, we, we actually Just had a cake party because we hit 20,000 GPUs. So that's the scale. It was shaped like a 50 90. So that's, that gives you any hint of our best one. So yeah, the unit economics of owning and kind of putting something on vast is a complicated question. I do have an idea for a calculator and some different things but it's. I'll start with some of the things that maybe you know, most people don't think about. You did mention one of them though, the cost of power. So the cost of power is a pretty big input and you really have to look at the generation of chip and the type of chip to understand how big the power fraction cost is of running it. The later chips are using more and more power. The consumer cards tend to use more power as a fraction. Some of the enterprise cards are like a little bit more optimized for power consumption. So the cost of power is pretty big because if you're just going to plug these things into your garage and you're in California, that's going to be tough because you're paying, you know, 25 plus cents per kilowatt and somebody who is in Texas or in Iceland or in a low cost of power place is going to have an immediate pricing advantage over you. So the power cost is a big input. Then the price of the equipment itself, that's going to be your capex, that's an important piece. So the people that are looking at this kind of know the prices of these GPUs pretty well. We don't try to predict the prices of them. I mean I've, I've been scratching my head for four years in this industry. Like I just remember back as a quick detour story. I remember back when Ethereum was all being mined on GPUs and then they announced they were going to stop mining it on GPUs. And we were just, I was thinking, man, oh boy, here comes like a glut of GPUs and prices are going to drop. The price of the GPUs are going to drop and then the, and then the price that we can rent them all for is going to drop. So that's going to affect our bottom line. So we're kind of like preparing mentally for that. And the exact opposite thing happened. The image generation stuff, it was really just landscapes and very bizarre stuff. Couldn't even do people couldn't even make a face. But you know, that's kind of where one of those little booms started. And the price of GPUs went in the other direction. I believe those were all like 4090. So then suddenly you couldn't even buy a 4090. And then the prices to rent them actually went opposite. And we thought that Ethereum was going to be a big driver of that, moving to proof of stake. But it just really didn't matter because this massive thing of inference, you know, that was coming, that was the first little, little wave of that. So, and then to get back kind of the unit economics or the price of the equipment, then you also have another little hidden thing is the taxes. So, you know, in America there's a great way to accelerate the taxes on equipment based on this, the latest tax laws. And that gives you a little bit of an advantage as well. If you have some income and you want to buy a bunch of GPU servers and accelerate their depreciation, you can do that in inside of a year. So we've seen a lot of people do that. And then, so those three things that I'm mentioning, the CapEx, the price of the GPU, the price of your power, those are things that a lot of people don't even see. But then that sets sort of the foundation of what price you can offer. So then when you get into, okay, well now there's these other competitors and it seems like my price is kind of high. It's like, well, maybe you didn't think through the power thing or when you bought the GPUs. So a lot of those factors kind of go into it. And then what you're really trying to do is forecast the demand into the future. And I tried to not do that with our hosting partners. That's kind of their job. We can show sort of the prior prices, we can show the prior availability on our marketplace. But can really anyone predict the future? We can all try. And my prediction is that inference is going to continue to be important and that as new uses for inference kind of continue and we start to understand it a little bit more. And I think I saw something on X where somebody would put a fruit fly into a digital world, you know, and then a mouse, a cat, a rabbit, a dog, and then we all know where it's going.
- [23:38]
Esben
So the demand is going to keep increasing, you can say, and the supply is not going to. What do you think? Is that going to increase at the same pace or is it not possible at all?
- [23:51]
Travis Cannell
I mean, Musk, look at what he's saying. If the demand continues and we all want to live in the digital future, maybe we turn the world into a computer. That's the kind of bull case, I guess. Or if we don't want to turn the world into computer, then we start building in space and you know, feels like we're living in some sort of sci fi book that's all kind of coming true, including robots and stuff as well. Right. So it's all kind of all happening. But I do think that for inference, you know, we're kind of now just teasing out the artificial minds that want compute and they want to inference. Then we have this whole other category that really hasn't. It's going to be quite complicated to get there. But you start with a fruit fly, you maybe do. Soon you'll do a mouse and then you got mice running around, some sort of virtual simulation and everyone's wondering like Philip K. Dick did, is this real? Is it dreaming? Is that a real mouse? Am I a real human? And. And then away we go.
- [25:03]
Wes
Would you. I'm curious to hear your take on this. Do you think there's any like really massive network effects that exist in your platform that you're building today? Like what is something like if you look at meta, okay, great, like with Instagram, all these other tools, it's like when some people sign up, it gets more valuable as time gets on. It's like, well of course, here's my Instagram link or something like that. People are like, okay, now I have to join it. It gets more valuable as more people join. So I can see some pieces of this where as you get more and more supply, there is some network effects that do become very hard to compete with. But is there any other kind of big network effects where you're like, as this business gets to scale, it becomes indestructible or is there anything there where you're like, yeah, you know what, it doesn't quite work that way. Like if, you know, I'm thinking of like other competitors, like maybe like I don't know if iron you would consider them like a direct competitor. But like they have data centers in places with like really low costs of power. So there's like different places like that where like, okay, that could still get disrupted if they can lower the price and kind of erode the initial value prop you have. But yeah, what are those kind of initial, you know, network effects that you can see are getting stronger and stronger as you add more users and customers? Yeah.
- [26:21]
Travis Cannell
With the two sided marketplace we, you know, it is a big moat because you start to get the network effects on both sides and once you're really well known for great place to go run a gpu. It's really simple. You put in five bucks and you get something working within a few minutes. You know, then that's, that's what you get known for. And so then you have that consumer pull. And so then on the supply side they know that that Vast has that consumer pull, lots of pull in the marketplace. And for them to try to compete with that, then they've got to compete with the whole of Vast and just, you know, not just their few data centers that they might have. And I would say for the companies in the category that you're mentioning, you know, they're not typically software companies. Irin is. Can you say the name again? I don't know how to pronounce right. Yes, Irin. You know, that would strike me pretty typically like they're hardware guys. They know how to rack and stack GPUs, they know how to put networking cables together. They really know how to get cheap sources of power. They probably know the ins and outs of power contracts. I have no idea how to buy power in British Columbia. We're a software company. That's what we do. We have software engineers led by a CTO CEO who's a software engineer himself. And so we're working on the stack, we're working on the way that makes it easy. And so that gets magnetic to the customer base. And then, you know, if you're a supplier, you probably even want to put your spare capacity on Vast rather than try to recreate that whole ecosystem that we've kind of created and try to recreate that pull and that magic that we're providing to our users. And then on the flip side, it just works on the other direction as well where because we attract all that supply, we have over a thousand different independent suppliers right now. We're the place to come and shop and we're the place to kind of look at the pricing and try to understand what's really going on. I mean, I think that our pricing feeds, it's like the open secret. People have been scraping our pricing data for years and using it to make all sorts of decisions. So that's why I'm kind of excited we're going to start publishing a little bit more on our own website and taking some credit for that. But that competition between the different suppliers creates a little bit of magic for then the buyers.
- [28:41]
Wes
But I could see how that would absolutely feed the network effect here. Even like iron, it's like, hey, great, we got all these amazing Enterprise customers, but we still have spare capacity so why don't we just rent it out?
- [28:56]
Travis Cannell
At the end of the day it's like, great. Glad you're trying to compete. Yeah, cool platform. That's cool. But you want to just list that on Vast. You know, I saw somebody on X was trying to follow one of my tweets and was saying, oh, I got each 100 for half the price of what they're offering. And I said, well, why don't you just list that on Vast because you're going to make more money trying to undercut our price like that. So it's just one of those things. There's a bit of partnership and competition together with some of these platforms but at the end of the day if we can provide value to them, and we do, you know, and then they can list stuff too anonymously. So that's another benefit to them is they can do pricing experimentation. They can price stuff at a bigger discount than they would ever price on their own website. It's just a different system. And then I think in the future we'll get into handling bare metal and cluster contracts better and future contracts better and things like bids. So right now our entire system is driven by asks on the supplier side. So when you go to Vast AI and you go into the console at Cloud Vast AI, what you're looking at is the asks and there's no way to like put in a bid as a buyer and saying, okay, nice, I see your $2 per H100 hour and I'm only going to offer you $1.50.
- [30:28]
Wes
But then it kind of becomes like stocks. It's like that is what it is like if you don't want to pay, you know, a certain price, you can put in that limit buy and you're good.
- [30:38]
Travis Cannell
Yeah, exactly. And it creates that little spread. And so there's some extra complexities in our system with that. But I think in the future, you know, that's, we're kind of working on a V2 that will solve a lot of those issues. So you can rent kind of independent of time and space and we price those things accordingly. And so yeah, we're working on quite a few things like that.
- [30:58]
Esben
I love how you, you think about your company as a software company. So I, as Wes mentioned in the beginning, I co founded Cobalt which was also a two sided, some would see it as a two sided marketplace but we always saw ourselves as a software company. And I think it's, it's very much about the mindset about how you Enhance your product continuously and make it easier for the customers and in the marketplace. It's just you have two sides of the market that you need to build software for. And over time you build more and more IP into the actual software, which is fantastic. Maybe shifting gears a bit, it would be you're the CEO of the company, right? And you've had amazing growth a lot due to this whole AI boom. How does one actually build an organization when you're growing at that, that level of pace? And how do you think about that?
- [31:48]
Travis Cannell
Yeah, it's quite recent that, you know, we've kind of really taken off. Our growth has always been on a great trajectory at vast. We do a lot of stuff ourselves. So as the coo, I do a lot of like independent contribution myself. The CEO is constantly writing and shipping code himself. So we're very much kind of a hands on leadership team. That is the way I like to think about it, you know, kind of right there with our, our employees putting in the long hours and, and actually contributing. But it's been a journey we decided to do in office in 2024, which was a bit of a gamble and a roll of the dice. But I think we've found people that want to be collaborative, want to be in that kind of environment. It was such a weird place sort of after Covid and a lot of people like that and that's fine. But a lot of people were kind of like, hey, I miss the camaraderie and coming in. So I guess one thing that we, we did was we, we have a full like 5 day in person office environment and we opened an office in SF to tap into the talent up there. A lot of, you know, obviously great software engineers and AI engineers up there. So that's been a little bit different from a lot of other successful companies that have kind of more, maybe more leaned into the remote working.
- [33:12]
Esben
Do you think that's good for you when you're doing high growth, that you're in the office and kind of together and you can quick more, maybe more quickly tackle the fast pace.
- [33:23]
Travis Cannell
All I can say is that it worked. It's working. It worked for us. Obviously. I don't know the worlds where we stayed remote when we were remote. Prior to mid 2024, we did do this thing that. I don't know how normal this is and it might sound a little crazy, but we had kind of a continual zoom call running. Actually it was Google Meet or a Google shop, sorry, zoom, but we had a continually Google Meet running called the Open Office. It's kind of like, hey, if you're working, join. And thankfully it never led to any weird things, but everyone just joined. You join. And most of the time it'd just be silent. You know, it is a little weird. But then if you needed something, you would, you could just say, oh, hey, you know, Vince, what's, what's going on with this thing? And when we were like five to 10 people, that was kind of great because we were spread out pretty far across the US different time zones and everything. Had somebody kind of in New Jersey, Texas, Michigan, Boston, couple people in California, but we could all get together every day. And even if you never said anything, just seeing everyone there, you're like, okay, well we're working now. You know, it just kind of, it solves, I think some of that problem of, of the seriousness of remote work and certainly was motivating for myself. It was just a motivation thing. It's like, great, okay, well, Vince is, Geez, Vince is on. I might as well show up, you know, like. And so we were doing that prior to 2024 and opening our, our headquarters in Los Angeles. And then the only other thing I'll say about the office is we wanted it to be cool for ourselves and for our employees. You know, if we're going to make everyone go to the office every day, it should be cool, should have a good view, should be in a nice area. Like, let's spend a little bit of extra money so we're not all, it's like, hey, it's a startup, you know, like, so we're in the back of this, you know, factory thing and there's no windows. It's like, no, let's just, let's find a good location that we're going to like into the future and so people can come here and be like, yeah, this is a cool experience. So those are a few things that we did that I think helped and that might be a little bit different from the kind of regular stuff I
- [35:36]
Esben
think we're in San Francisco. You're definitely seeing a lot of people going back to office, like especially the high growth companies going back to office. And it's becoming more the old San Francisco again with high energy in office. And I think it is due to the fast paced level of innovation happening at the moment where people feel the need to be together. But, but also what's going on with AI is a lot of efficiency, right? So how in a, in a fast paced growth environment do you think about hiring and, and using AI to maybe reduce the need to hiring and so on. Is that part of your strategy or how do you look at that?
- [36:12]
Travis Cannell
That's changing pretty quickly. And right now we put a pause on hiring anyone and then the type of person that we would be willing to hire, you know, would be, would be different. It's a bit of agent manager as opposed to software developer. And it's kind of changing everything too. I mean I remember looking at our JIRA boards and I was just thinking, why do we, you know, we have so many states in this JIRA board, like planning and then, okay, then we're going to like look at the design and then we're going to start writing the code and then we're going to review. I'm just like, everything just goes right into review. Like for you the code's already written, it's done. Like, yeah, we do need a plan, but then we have the code, then we just need to test it and review it and then we need to kick it back into planning or we need to push it out. I think all teams are trying to deal with that. This agile process is definitely sort of built around the idea that you need to write code. Right? So, but to get back to hiring, you know, we're a bit on a hold right now and we'll see how that kind of evolves. But the type of person that we're looking for has more of the skill of complex technical design management, like time management, motivation, those things all still matter. The kind of core types of things that you really want on a good person still matter, but the particular skills are a little bit different because being able to write great code is no longer a big differentiator. You know, what we're moving to is that's kind of hard to quantify and it's, and it's changing, you know, in a somewhat concerning way as sort of these things become more and more intelligent. But it's like infrastructure management. It's all the things that these LLMs really can't do at all. Like, you know, which is really understanding how everything is sort of set up on the infrastructure side. Different security models, best practices, how these things are all working together, you know, ways to automate different parts of the process. It's kicking things from out of, you know, from doing to more management and systems level thinking.
- [38:15]
Wes
Awesome. Yeah, that makes sense. And I guess that's very industry specific to what you're doing in that part, which makes sense for the big client Shader. And before we wrap up here too, is there any particular like piece of advice you would have for any founder where it's like, hey, listen, if you're, you want to kind of either position your company to grow faster in this AI era or here's some things to watch out for. What would be like your one or two kind of tips you would recommend to any product that battery your kid talk to?
- [38:46]
Travis Cannell
Nothing new that is coming to mind, but some things that we touched on during this kind of hour would be the agent economy and agents and building things for agents. And it's not a very novel thing to say. I see a lot of people talking about that. But of course, you know, the idea is always cheap. It's the execution that matters and so building things for agents. So I think in the, in the future, I don't think it's even a bad thing. Like if I. Let's say your agent writes an email saying, hey Travis, I want you to be on the podcast. And then my agent reads it. You know, that's still, I'm still gonna say yes. Like your agent's gonna be like, Travis would be a good fit for this podcast. And my agent's gonna be like, hey, this is interesting podcast, you know.
- [39:26]
Esben
So did we really speak with an agent, Travis? Not yourself.
- [39:29]
Wes
I didn't know it was the Esbin agent.
- [39:33]
Travis Cannell
Ooh, what is reality? I guess that's. I guess we'll find out. I, I think that's a good place to look at. And then inference, it's going to explode. I think it's, it's life, you know, and it's, it's going to be, there's going to be more and more use cases for things as prices continue to fall for it. And you know, that's another area for, for everyone to watch. Cool.
- [39:54]
Wes
Awesome. And where can people find out more about what you're up to? So if there's vast AI easy domain, you'd ask where to rent GPUs on any LLM and probably find out the same thing there. If you don't get the URL right and you're Also active on LinkedIn as well, we can put a link there. But is there any other places or things where you want people to kind of interact with you if they got a question or they want to learn more?
- [40:18]
Travis Cannell
Yeah, Vast AI is kind of the place. Like I said, we'll publish some pricing stuff there. We're doing a little bit more stuff on X. So we're vast underscore AI on X and we're growing there. So give us a follow. We do push out little pricing snippets and things that we're noticing in the market in terms of like, oh, B200, availability is low and any price motions and things like that. So that's kind of a fun little GPU price ticker, if you're curious. And I'm Travis Cannell on X. So there's, there's my plugs.
- [40:51]
Wes
Awesome. Well, thanks so much for coming on. It's been a blast. I definitely learned a lot about GPUs and I've known before, so this is great. Thanks so much, Travis.
- [41:00]
Travis Cannell
Thank you. Thanks, Wes. Thanks, Esteban. Big success.
- [41:07]
Wes
And to wrap things up, thank you everybody for listening to this version of the product LED podcast. Make sure to rate review this on wherever you listen to podcasts, whether it's Apple, Google, you name it, Spotify. I'm going to read every single one of those reviews and that's how I know how to improve this. Also, if you want to stay in contact with Bean and learn what is going on in the world of plg and every single week get the best actionable deep dives on product led growth. Make sure to head on over to product led.com forward slash newsletter. I am personally writing each of these deep dives every single week and you're going to get a ton of it, so make sure to head on over there to product led.com forward slash newsletter.