The Chris Voss Show - The Chris Voss Show Podcast – The Future of GPUs: How iFrame.ai is Redefining Cost and Performance

Episode Date: July 23, 2026

The Future of GPUs: How iFrame.ai is Redefining Cost and Performance iframe.ai About the Guest(s): Vlad Panit is the CEO of iFrame.ai, a company specializing in deploying GPUs and building data centers for better compute resources, primarily targeting Neocloud providers, hyperscalers, and AI labs. Vlad’s journey in entrepreneurship began at 18, with a pivot towards IT and engineering endeavors from 2008. His significant work includes e-health projects in Europe and AI applications in medical coding automation. Originally from Ukraine, Vlad has expanded his endeavors to the United States over the past seven years, using his expertise to spearhead innovation in AI-related technologies. Episode Summary: In this insightful episode of The Chris Voss Show, host Chris Voss dives into the world of AI with guest Vlad Panit, the CEO of iFrame.ai. The conversation explores the backbone of AI infrastructure focusing on the deployment of GPUs in data centers. Vlad shares how iFrame.ai provides bare metal GPU services, offering significant cost savings over conventional cloud providers like Amazon AWS. This episode sheds light on AI’s evolving landscape and the technological advancements driving it, making it particularly relevant for business professionals and tech enthusiasts interested in AI advancements. Chris and Vlad delve into the differences between AI infrastructure services provided by iFrame.ai and traditional offerings by big players like AWS and Google. Vlad explains the company’s unique position in providing bare metal services that enhance cost and operational efficiency for customers, especially in AI training and inference. Highlighting the broader implications of AI’s rapid development, Vlad also reflects on the challenges and opportunities that come with deploying large-scale GPU networks, touching on both technological and environmental impacts. Key Takeaways: iFrame.ai offers bare metal GPU services which can be three to four times more affordable compared to AWS and Google Cloud offerings. The company’s infrastructure supports significant cost savings in AI operations, making it an attractive option for companies looking to optimize AI workloads. Vlad Panit emphasizes the importance of deploying GPUs efficiently to enhance the availability and reliability of compute resources for AI applications. The discussion highlights the evolving market landscape as tech giants like OpenAI and Anthropic move towards public offerings, with AI infrastructure playing a pivotal role. The conversation also touches on the environmental impact of AI data centers and the importance of sustainable energy practices in powering such facilities. Notable Quotes: “We deploy GPUs. We build data centers, put a lot of compute there, and then sell it to near clouds, hyperscalers, AI labs—some of which you probably use in daily life.” “We’re able to sell our services three to four times cheaper than AWS.” “The difference with AI’s GPU use versus gaming is it requires uploading and storing entire models, which is more complex and costly.” “AI data centers should have their own energy supply to be better for everyone, from local communities to the data center operations.” “It’s a huge overestimation to say there’s a bubble in AI; the infrastructure costs alone are massive to support large-scale AI operations.” Resources: iFrame.ai – Discover more about the company’s GPU deployment services. Connect with Vlad or explore further through LinkedIn and Goodreads. Tune in to the full episode of The Chris Voss Show to gain a deeper understanding of AI infrastructure and the intricacies of deploying GPUs for cutting-edge technology applications. Stay connected for more intriguing discussions with thought leaders shaping the future.

You wanted the best... You've got the best podcast. The hottest podcast in the world. In the world. The Chris Voss Show. The preeminent podcast with guests so smart you may experience serious brain bleed. The CEOs, authors, thought leaders, visionaries, and motivators. Get ready.

Get ready. Strap yourself in. Keep your hands, arms, and legs inside the vehicle at all times. Because you're about to go on a monster education roller. a rollercoaster with your brain. Now, here's your host, Chris Voss. Folks, Voss here from thecifos Show.com. Ladies and gentlemen, welcome the oldest living podcast. It's still broadcast daily out of a handful of the original podcast from the 2000s.

We are the only ones broadcasting daily shows, actually multiple daily shows two times a day. Yeah, we're kicking ass and taking names. Help us, support us because we're old and we need a little bit of help. Our backs hurt and our legs we can't feel anymore. So refer to the show to your family, friends, and realtis, go to Goodreads.com, Fortezsgrisvast, LinkedIn.com, Forteschastcchristch, Christfoss, YouTube.com, Forteschuschrist, all the crazy place. The opinions on the host and do not necessarily reflect the opinions of the host or the Chris Foss show. Some guests of the show may be a missing young man on this show. We're going to be talking about AI. Everyone's talking about AI these days. It's the thing. The end thing. We're going to be talking about a company called I-frame, dot a i and we're going to be talking about the CEO today. Vlad Panon joins us on the show. How are you doing, Vlad? Thank you. Great, great. Hello from Stanford.

Oh, Stanford. All right. Great. So welcome to the show. Give us any your dot AI or any other places you want people to find your social media, etc., etc. On the internet. Yeah, pretty much. I frame AI is the place. It's pretty easy to remember. So, yeah. Yeah. So give us a 30,000 overview, what you guys do there. We deploy GPUs. We build data centers, put a lot of compute there, and then sell it to near clouds, hyper-scalers, AI labs, some of which you probably use in daily life. So that's as simple as it can be. Now, do you guys make your own GPUs or you buy them and you source them for the companies that need them for the things? We purchase compute from mainly in NVDA, AMD. Build architecture around that.

So companies can use it for inference and also for training. And then we resell it to large companies. Oh, wow. Now, are you the reason that my 4090 that's now two years old keeps going up in price even though it's used? And my hard drives cost twice as much? or how's that work? I think NVDA is more responsible for that. Oh, is it?

Like speculators on the market, but yeah. Get a video on the phone, damn it. No, I'm just kidding. Being funny. So tell us about your product. Who's your clients? If someone's listening to this on LinkedIn right now, which they tend to do, who is your specific client out there and how do they qualify to be your client?

So pretty much it is NeoCloud that has a bunch of customers that are building applications on top of their cloud. So typically our new cloud customers, they do have some software layer that allows their customers to get as many tokens per second

as possible. Our job is provide them a reliable compute for a very long period of time with a predictable and sustainable understanding of availability. Oh, wow. And what are benefits of those sort of features? Do they make me run faster? Or how does it, how does that

different than say like Amazon Cloud or something in the other clouds? Oh, pretty much. What do we do, we provide bare metal. So for those who are not in the industry, what it means is basically we provide GPU almost as is with as minimum layers of virtualization as possible. When AWS, Google will offer you, they call it GPUs in reality. it's not really GPU as a GPU. It is something that they package as a product. Typically, it's some virtualization that consolidates their capacity over many, many, many, from many locations.

And then you buy some of it for some specific period of time. So pretty bizarre concept because you overpay like five times, sometimes higher. We provide GPUs as is. which means that you can do a lot of stuff in terms of optimization. You have access to drivers. You have access to lots of deep tech stuff that allows your team, if it's capable, to optimize and provide this additional value to your customers. From the cost standpoint, because we don't add this premium layers,

we can afford to sell it three, four times cheaper than AWS. So if you pay $500 per month for chat GPT, with us you would pay roughly 30 bucks. So that is kind of the math behind it. That's some pretty good savings then, right? Yes. Yeah. Now, I don't know if this is similar and it's probably not at the scale. You guys might be because it's in Vida.

But for, I don't know, five years or something I've subscribed to Nvidia. What are they called? NVIDIA now, I think it's called. GeForce now. And what it does, it's enabled me. I used to have really, just a really old, awful GPU on my old computer. Now I've got the 4090, but now that's obsolete. But I used to have this old, it was like a, it was old, okay?

And so it wouldn't run games on my computer very well. And my computer was kind of crappy too at the time. And so I used there in G-Force now. And it would literally give me, like, at one point it was giving me one of the latest GPUs that I was accessing through the game and the internet that hadn't even hit the market yet. I would see it and I'd be like, whoa,

that's cool. We get access to that. So is it kind of like a much bigger deal than is that same sort of concept? Yes. So for a game, it is pretty similar. However, if we talk about the game, it's pretty much running on the CPU, which is a little bit different. It is central compute unit.

The difference is pretty much, you have to upload the game to this local computer memory, and you're using GPU as accelerator, which helps it to render something better, do this stuff better. In terms of GPUs for AI, it's just different types of operations. What you need to do, you have to upload the entire model on that GPU. That takes time, that takes effort, and it's not very easy to change it. So, for example, if in the application you can switch between models typically, like chat GPD, Gemini, Entropic allows you to switch between models on the fly.

But that is application layer. Reality is they have to deploy this model and it takes time to some specific GPU, lock it there for a very long time, so it's cost efficient and then just provide you access to make some calculations on that. So the entire processing happens there. Another thing is if you are not just limited by inference, which is just chatting with AI, but also do stuff like training, you need to have many, many thousands of GPUs interconnected between each other. And if you have this connection between multiple internet, providers, etc, you can imagine how slow that could be. And it just cost and efficient because you still pay for GPU hour, but internet. connection in simple words. It's just messing up

with your budget. So it's not cost efficient. So I, larger new clouds, they rent space from they basically rent our DCs in the same location with internet connected GPUs. So the architecture supports

these kind of use cases. Now, I know one of the problems that companies are finding, they laid off all these people and then they are finding, it seems like AI might not be as efficient. And then one of the problems they have is the people that are left that they've tried to stack, basically the people they've fired job onto them. They're trying to make them these sole people that will run agents.

It's my understanding. You can correct me if I'm wrong. But one of the problems they're having is they're now having to, what's the right word? They're having to basically dole out or hold on to the tokens because the people that are left behind. all the tokens and blowing up their budgets. So they've had to scarcity, you know, say, hey, you can only do tokens so much. Do you guys, is that a way that you guys help maybe them reduce their cost for tokenization

and GPU requests? Oh, so that's, do you use French in your, in your podcast? No. Okay. So, yeah, I'll just, I'll just limit myself to, to, to. Yeah, you can swear in the podcast if you want. That's what I meant. Yeah, sure.

I think all this cost per token is BS because the reality is, yeah, like for example, EWS. So, okay, you have a model. Model costs nothing for you. If you use some GLM 5.5, which is open source and by any means is pretty much as equal by efficiency as impropic. And definitely better than chat GPT latest model today. And definitely better than GROT. And it's basically free. You can download it the same way you would download whatever, like application from Apple Store.

Even more free because you literally can do whatever you want with a source code. So then you can rent GPUs and run it on your GPUs. So you cannot do this on your computer highly unlikely unless you are like some very sophisticated gamer and know how to optimize stuff. And you have powerful computer. But the way to go for you is to rent GPU. So then to your question about tokens, AWS will sell you server, one machine, one computer, let's say, with 8 GPUs for $130 per GPU per hour.

So you would pay $130 per one hour of usage of that compute. And if you'll go with bare metal, you pay roughly $15 for the entire thing. Long serve. It's not hard to make simple calculation. And here we are. It's at least five to six to eight times cheaper if you would just do it in a different way. So all this price per token conversations, it's more marketing narratives from Anthropic, from chat GPT. They try to play this game between each other.

It is not actually linked to any real mathematics. Ah, that's interesting. So now you've mentioned this term metal a couple times. And I see there's VPC interconnect. Are those two, can you give us a foundation on what those are and how they work? Yeah, pretty much there is a lot of different ways how to interconnect nodes. Our specific, we started with interconnecting GPUs, IP2 IP, so-called, which means that you can run your models on your computer.

However, use some remote GPUs as accelerator. So that is pretty much similar to what you mentioned in the beginning. It works for inference, but it's not very great solution for training, especially if you want to get some. If you train underlying models. If you fine-tune something, yes. But if you do some large-scale training, it's just no way to go. Yeah.

Yeah. And so basically, instead of me as a company, let's say, do you find most of the people use you maybe medium to small? companies that they can't really afford to buy in bulk the GPUs needed, or is it usually larger companies are coming to you that are trying to offset the burn rate? So our companies, pretty much, they are applicable, like our customers, they are providing applications for customers for their end users. For example, if you generate a lot of videos, that will be probably our customer who is renting GPUs at scale from us and then resells it

to you for way better margin. And that's why the narrative actually sticks on the market about this token price. Because they love when Chad GPT comes out there and justifies why they charge you 10 times more than it really costs. So everybody kind of align on this message. Not us. We are really interested in just making it more affordable. So more and more people can do stuff with the eye. So can like, I'm a small businessman. Some people may have small teams five to ten. 20 people. Is that, are we able to, is that affordable for us if we're trying to develop AI systems? Or is it more something that's going to, you got to have some serious bank going on.

Yeah, the question of affordability is a great one, but it more like lies. It narrows down to one, one question. Like, what is your really, what's the goal? Because to run your own model on your own GPU, it would cost you at least $1530 per hour. Oh, wow. If that is something that justifies your expense, for example, if you generate tons of data, or you do, for example, you do stock prediction, that is definitely, that's something that makes sense. You would make way more money.

Again, it's not financial advice or stock purchase advice. However, if you do some high frequency stuff, that requires this intelligence, that is the way to go. And it's a pretty cheap way to go. If you do occasional request, if you make occasional requests to AI and just ask it for some stuff to help with daily work, it does make sense. Oh, this tokenization and people having enough tokens and making sure they don't waste them and stuff.

I don't know. Maybe some of these people are using tokens in the big companies for OnlyFans or something. So how did you, how did you, how did you, get into this business and how did, did you begin your entrepreneur journey here? Kind of give us your entrepreneur, hero, entrepreneur, what is that? Your entrepreneurs. I've been entrepreneurs when I was 18, I just made up a word. What was your journey and your entrepreneur hero's journey? Oh, like in IT, it actually started 2008 from software company. Actually, it started way earlier with e-commerce business, but like serious IT, which is like more,

dive into engineering. The company in Europe, we scaled to quite large company, worked with the United Nations projects, and we're pretty successful there. I was born in Ukraine, so we had a huge team in Ukraine. We worked on e-health project, which is now, it's something like in the U.S. They try to do this same consolidation of healthcare data,

but here in the U.S., where I live for, the last seven years, I think. It's a complete mess. Nobody wants to share data with no one, especially health care. It was actually cultural, if you don't mind, it was cultural shock to me when you do some blood work or stuff like that. And you go to the hospital. And the hospital is rejecting you, patient, whose blood they took to provide you your data because they are afraid that you would go to their some competitor or whatever.

Yeah. and only shares your data with your doctor, which is within their network, which is or who is within their insurance coverage. It's like it's not healthcare. It's marketplace. It's medical marketplace where insurance companies have pretty much, they own the marketplace and everybody just play along. Yeah.

So when I came here, one of the first projects with the AI was medical coding automation. And we still, we actually grown to. Over a million users. For those who don't know, rarely there are people who never heard about medical coding. However, for those who are still lucky ones, brief intro, every time you discharge from the hospital, they are assigning codes to your case, to your notes. And these codes are then sent to insurance providers. So they are reimbursing your case based on this codes.

Oh, wow. Shocking information that to pass certification to be certified medical coder, person who is doing that, you can pretty much do, you can pretty much make 30% of mistakes on exam, and you're good to go. Serious. And the problem, the world, okay, if it's not enough, most of medical coding is done abroad. So if your data is pretty much, yeah, if you work with, okay, I'll not disclose brands, because in the U.S., another thing that people love to do is to each other.

So I'll have to... Welcome to America. Yeah, oh, absolutely. I love this country, but yeah, there are some interesting things happening. So, yeah, like a lot of these data, they are ending up in India. So that's why a lot of people lost their privacy. Basically, in the beginning of last year, I think it was first quarter of last year or year before, over 200 million patients lost their data because of United Health Group leak.

And it was before, it was prior this horrific situation in New York. So, yeah, 200 million. So if you received call from some very, from a nice person with a very specific accent about something, some help with your computer or stuff like that, it's highly likely that they got your information from this, from this leak. So we started using AI in this industry

because it's obvious. It's just obvious. You don't send data anywhere. It's basically processed in the facility. It never leaves a healthcare provider. It uses a bunch of optimizations to make sure that you will get your reimbursement.

But again, coming back to monopolies, you can't do much because there is a thing in the US called over overcoding. It's not, again, it's not US problem. It's it's a monopoly's problem. So I just want to, I just want to make sure that people have this differentiation. The monopolies pretty much, if insurance don't like how well you do in your insurance reimbursement,

they can kick you off as a doctor, kick you off from their network, and then you're out of the marketplace. And good luck to advertise your services outside of the grid, out of the matrix. So, but anyway, we outsource right now. We made this project pretty much free for everybody. So it's pretty much automated Wikipedia for anyone who struggles with getting reimbursement. The website, I believe it's still live. It's met.comport and we sponsor it. So it's completely free.

No advertisement. If you have issues with insurance, please welcome. Yeah, but then we move towards something that really makes more sense in terms of long-term business and investment, and it is GPUs. Right now, we are in the age of this underlying technology. It's pretty much similar to building Internet in early 2000s. So that's where we are today, building data centers, and hopefully making AI cheaper so everybody can afford that.

Now, these data centers are building, I've been visiting my mom up in Utah, and they're not too happy with some of the, who's that one guy, the Mr. Wonderful from Shark Tank and stuff. And I was concerned about how much water and energy these things consume. Does your guys set up make it so that there can be less data centers because they're kind of consolidated and everyone's tapping into using them instead of starting their own maybe?

I don't know. Yeah, absolutely. So it's legit concern. And the thing is that, again, that's a good thing that you mentioned. There is another business, which is like co-location. That's what they do. Because data center in our understanding, it's the architecture of the compute and deployment of that compute. The things that Mr. Wonderful does, if I remember, yeah, the Shark Tank, right?

Yeah. So what they do is mainly real estate business. They build a bunch of facilities, connect to the grid, and here we are. So we are not very in favor, very much in favor of this approach. we think that data centers, if there is a high-quality collocation center, they have to have own energy supply, power supply. It is just by default. It's just better for everybody. It's better for local people, for local community. It's better for data center because you don't rely on some grid that might be affected by something else. And again, it seems like data center as a business

will struggle, but in reality, a lot of people who are using it. For example, medical coding folks or people doctors who are using AI, they will just be not able to use that AI in their work. So we need reliable power supply and the only way to go, I believe

is deploying like own mobile power plants. Again, we don't really do that. We prefer to lease this stuff from folks like you mentioned. And typically right now it's real estate guys who

tired to purchase apartments and now move to building this co-location centers. Oh, that's got to be pretty good for them. Yeah, it is good because there is a huge demand. Another thing, if they have enough experience and they know. So the problem is, and I think that is what create tension between local communities and these guys, is they over-optimized on everything because their incentive is clear. You need to build as much square meters, as much space as you can, connect as much energy, as cheap as you can, and sell it to some EWS, for as high as you can.

And the problem with that, over-optimization is how to build cheap. Okay, let's utilize whatever we purchased, some free land that we purchased 20 years ago because it was cheap then. And instead of building new apartments that we planned 20 years ago, we'll just build it as a center, why not? That makes a lot of, that creates a lot of problem and obviously, like, friction. Another thing, what is the cheapest way to get energy as soon as possible? Because we can lose data, like, we can lose a deal with Google or AWS or wherever is there. There is community here. Let's borrow from them until we are building something. So that is another friction point. And the third one is pretty much obvious. Like, politicians try to throw more oil. into this or gas or whatever to this fire. So it sometimes sounds even more horrible than it is. But what I agree on is whoever deploys collocation centers, whoever builds these facilities,

they have to think about more sustainable sources of energy. That is a long term, it is better for everybody. We are ready to wait. As there are customers, we are ready to wait. and our customers also, even if they don't realize that, it's our job to explain to them, that please wait six months, we'll build a power plant, will invest in it. It will be a little bit more expensive, but you know that tomorrow you will not be taxed to death because local government will have no way to go.

Somebody local will win elections on a narrative that they should kill you. It's a huge risk for us, so it's better for us to actually build something And even more, if it's more cost effective to build power plan that will sell power to data center, but also sell power cheap power to local community, why in the world wouldn't we do that? And everybody benefits from it. And it's plan B for local communities as well, as well as for data center. You want to scale, please. I think we should be just more open, more transparent and have more just more transparent. I think that's a right word here. Yeah, yeah, definitely. You have to explain people. You just can't try and roll them. And, you know, I mean, Utah is going through this huge. We had the worst, what was it? We only had three snowstorms all year. The ski slopes were just mud. It's unreal. I've never seen that in 40 years of my mom living up here and visiting. Even when I spent my teens here, it was, I've never seen no snow. It was just wild. And so, you know, there's a big concern about water. Now we're having wildfires and all sorts of. problems because of lack of water. And so it's kind of scary. I know they're moving the docks down again at Lake Need in Vegas.

They're moving them all around because they're trying to keep them all floating as the water drops. But it's kind of wild. So, you know, these people are kind of top of mind to it right now. And it's really important. What are some other aspects of your business, business model and what you do for customers that I maybe should have asked you about that maybe we need to get clear for people to know? So I think that what's really interesting is today's conversation about SpaceX, because the huge portion of their business is X, A.N, so-called, which is interconnected with Entropic.

We are expecting IPOs, Open AI and Entropic, I believe. And Tropic already mentioned S-1 filing, which is like pre-IPO kind of moves. Actually, that's not pre-IPO. It's literally like intention to file for IPO, for listing. Oh, really? So, yeah, I think that will affect the market massively. But again, like, it's very hard to say something in this condition without being sued. So I'll just be careful here.

So anything more we want to tell people about who you guys are and what you do? Yeah, I'll just summarize that we deploy GPUs and sell it to Neo-Clouds or three to four times more affordable than AWS or Google will do for you. So that is pretty much our business. Any future things that people will be needing to gauge in as you bring it up or future services maybe you can tease out or we don't want to give away anything that may be a secret. No, absolutely not. the only thing that might be interested for more general public is, yeah, like, this thing is pretty much capital intensive. And again, I think that like all this rumors about like how big the bubble in AI is, I think it's a huge overestimation. And again, like the numbers are showing just a brief example.

So to deploy to deploy roughly 64 servers, like 64 computers in simple language. We like typically, NVDA will sell it to you for about 700, 800K per computer, which is a lot. So you end up with a budget for this 64, we call them nodes, 64 nodes, 64 computers. We'll end up with a budget roughly 100, 150, 100, oh, actually 200, 240 million. And just for understanding, this 64 GPUs, they can serve only roughly 64,000 people simultaneously. 64,000, not people, 64,000 requests simultaneously. It's a very low amount of them. So you spend $200 million and you serve only 64,000 people simultaneously.

We can imagine that having millions and millions and potential billions of people utilizing GPU today, it's not even close to enough just to serve questions like, you know. The basic stuff. I think we're losing a little bit, Vlad. Our in compute capacity. All right. I think we got you back. Dip for a little bit on the internet there. The Stanford must be using their AI systems.

Yeah. Soaking up all the internet. We'll round out the show here. Thank you, Vlad, for coming the show. We really appreciate it. Thank you so much. Thank you. And thank you to our audience for tuning in.

Check him out. Give us the AI and any other social media website you want people to check you out on, Vlad. Yeah, just eye frame AI. And yeah, pretty much that's it. And Viva, Viva, Ukraine. Oh, thank you. I was trying to say Viva La, Ukraine, because they say Viva La,

But I don't think there's a law in the Viva Ukraine, right? You just say Viva, Ukraine. Yeah, I'm not in that club. I'm not in that club. Okay. The one thing about Ukraine is they had so many beautiful women. They were like the capital of, what do they call that? A surrogate pregnancy.

And I've photographed plenty of you through my photography stuff. And they're just so beautiful. They have that wonderful Eastern block look and probably some of the best women on Earth. I don't know. Anyway, thank you very much, Vlad, for coming to the show. Thanks for much for tuning. Go to Goodrease.com, Fortunus Christchristchristvost, LinkedIn.com, fortuneschischristvost,

YouTube.com, Forteouschiscusvoss, and all those crazy places in it. Be good to each other. Stay safe. We'll see you next. You've been listening to the most amazing, intelligent podcast ever made to improve your brain and your life. Warning, consuming too much of the Chris Walsh Show podcast can lead to people thinking you're smarter, younger, and irresistible sexy. Consume in regularly moderated amounts. Consult a doctor for any resulting brain bleed.