Why AI Builders Need to Own Their Compute with B3

Franklin Bi sits down with Daryl Xu and Viktoriya Hying, co-founders of B3, and Yorke Rhodes (NYU) to announce the launch of B3IQ and explore why owning your own compute is becoming essential for anyone running open source AI.

Open source models are catching up fast, but the hardware to run them is gatekept. Frontier labs have locked up hyperscaler capacity years in advance. Researchers get their accounts banned for studying human trafficking. Startups get put in line for last-generation chips. B3IQ is a rent-to-own platform that lets anyone own NVIDIA hardware, run private AI on it, and monetize idle capacity so the machine pays for itself.

Key Topics:

• Why the long tail is locked out: frontier labs reserve hyperscaler capacity two to five years in advance, leaving researchers, startups, and prosumers with no place in line or only older-generation hardware
• The content filter problem: NYU students researching evacuations in war zones and human trafficking get flagged and account-banned by closed models, and running open weights locally is the only workable path
• The iPhone moment for compute: B3IQ's rent-to-own model drops the upfront cost of a 400K H200 cluster to roughly 120K, with ownership in about two years, hosted and managed in a secure Oregon facility
• Owning beats renting on the math: two years of renting an eight-GPU H200 node costs roughly the same as buying it outright, and idle capacity can be matched with offtake so the machine pays for itself
• Open source is closing the gap: Kimi K3 matches frontier performance on some agentic coding tasks, but running it takes two H200 nodes, which is exactly why hardware access is the bottleneck, not model quality

01:23 From Coinbase to Gaming to AI: The B3 Journey

05:28 The Research Wall: Content Filters and Token Limits at NYU

11:02 The Hyperscaler Long Tail Problem

16:16 The Economics: Why Frontier Labs Locked In Their Model Gains

18:19 Rent-to-Own: How the B3IQ Financing Model Works

20:10 The iPhone Moment for Compute

29:27 Why Own the Box When Neo Clouds Exist?

40:24 The Five-Year Vision: A 50-Megawatt Facility and Price Transparency

The views expressed in the podcast are those of the individual personnel quoted and are not the views of Pantera Capital Partners LP or its affiliates ("Pantera"). The podcast is provided for informational purposes only to provide market commentary and for general educational purposes, and should not be relied upon as legal, business, investment, or tax advice. The podcast is not directed at nor intended for use by any investors or prospective investors and may not under any circumstances be relied upon when making a decision to invest.

Transcript

Copy

SPEAKER_02 0:00

Owning your intelligence really makes sense, but the way to do that is a little bit muddied at the moment.

SPEAKER_04 0:05

Now the goal is to have an LLM in your pocket. This feels like the iPhone moment for compute.

SPEAKER_02 0:12

We built a platform that we're calling B3IQ to enable anyone to come in, own the video chips, run private AI on them, and own their own intelligence.

SPEAKER_04 0:24

Welcome to the latest episode of Stayful from the team at Pantera Capital. I'm Franklin B, general partner at Pantera, and I'm excited for today's episode where we'll be announcing a product launch from one of our portfolio companies, B3 Labs. And so B3 Labs is launching a new product today called B3IQ. We're excited to talk about it and also to hear from York Rhodes, who is a professor at NYU, also a co-founder of blockchain at Microsoft, and is using B3IQ's products today. So welcome to the podcast, Daryl and Victoria from B3 Labs.

SPEAKER_01 1:00

But before we begin, a quick disclaimer. This content is for educational and entertainment purposes only and does not constitute financial investment or legal advice. Please do your own research before making any investments.

SPEAKER_04 1:11

You know, first I'd love to hear a bit about how you started focusing on this part of the journey. How did you get here and what made you realize that this was the thing to build?

SPEAKER_02 1:23

Yeah, absolutely. I can kick us off. And Franklin, always good to chat with you. And York, really good to be on the podcast with you. Thank you guys for having us. I'll dive really quickly because I'd love to hear everyone's perspectives on AI and ownership. So I'll give the long and short story on V3's journey. We came out of Coinbase, our team. And at Coinbase, we worked on a lot of the core infrastructure around trading execution. So the problem we were always trying to solve back at Coinbase was how do you enable secure trading execution for a hundred million KYC users globally? And when we actually left about two, two and a half years ago, the question we asked ourselves was where do we take that expertise? What customer segment can we apply it to? Gaming just felt like a natural fit and a natural home. Games run tick by tick in milliseconds, blockchains, as York knows, produces blocks to the best of their ability in seconds. And so marrying the two was essentially the same class of problem that we saw back at Coinbase, which was secure fax execution for consumers. And through that, we actually had a first hand or first front row seats to what was going on in the AI space, which was AI compute demand was exponentially growing and GPUs were pretty scarce in supply. So we did what we always do at B3 Labs, which is we built vertical and we entered the AI space with two products, one on the software side and one that we're launching on the hardware side, which we're super excited to talk about today. On the software side, really briefly, the product is B3OS, and that's solving for a problem that we identified pretty early on last year, which was AI is probabilistic, as we all know, which is fine if you're drafting an email or just sending a chat or text request, but extremely risky and irresponsible when you're dealing with people's funds, especially on-chain. And so we built B3OS as this secure, reliable execution engine that lets agents operate safely with people's funds, both on-chain and off-chain, in inside enterprise systems. So that's B3OS. And then on the hardware side, we're super excited to talk more about B3IQ, which we're launching today. And the problem we're solving there is something we saw early on, which was GPUs are in short supply. Everyone wants them, either for private compute or for their own inference purposes. The challenge is Nvidia chips are often gatekeeped, they're expensive, they're hard to manage, and renting them are super expensive long term. So we built a platform that we're calling B3IQ to enable anyone from small teams to individuals to mid-sized enterprises to come in, own NVIDIA chips, run private AI on them, and monetize them when that capacity is idle so they can let the machine pay for itself. It's launching pretty soon and something we're super excited to chat through today.

SPEAKER_04 4:32

Yeah, it's something I've been hearing about pretty much every day, especially in Silicon Valley, where you know you're hearing of a lot of companies that are becoming AI native, or at least trying to be. And one of the first questions they run into is well, should we own our own compute? Right. And, you know, some are using open source models, some are trying to figure out if they need to partner with a data center or build out a whole supply chain themselves. So it makes a lot of sense to go at this problem based on the things that actually matter to them, right? It's self-sovereignty and security, being able to control your own destiny, it's the cost, it's the logistics of it all. But you know, it seems also like it's not just about these startups that are having this problem. York, I'd love to hear from you. You know, what was the pain point that you were having in your research and with your students when it came to being able to access compute as you needed?

SPEAKER_03 5:27

Yeah, great to be here and good to see you again, Franklin. It has been a it has been a while. I've been working pretty closely with Daryl and Victoria specifically on the different types of things that we do in the context of educating global affairs students in emerging technologies and how to apply emerging technologies for good, essentially, right? And so Microsoft, for example, has an AI for good lab. You see a lot of posting from them. And we actually recently, the students on a project we were doing with an external PhD research advisor on evacuations. We were basically looking at sort of case studies around the world on how you model and bring a data framework into looking at evacuations. And honestly, didn't really know where we were going. We just knew that this modeling with data and the data frameworks was an interesting thing that we could relatively easily do now that we have these amazing tools and this amazing infrastructure. And so that's sort of one story of many about what we're doing in the lab. But the cohorts this summer was kind of finally up and running in terms of how do we use AI models to actually build these types of informing frameworks and data dials where we could do all kinds of things like Monte Carlo simulations on you know different outcomes based on how you tune those, turn those knobs. And I actually knew Victoria and Daryl just as they were leaving Coinbase and so knew they were going to do some very cool stuff and stayed in touch. And they presented me with this idea where we could actually start to move some compute away from the center and look at ways that we could, with the right scale, execute open source models that we could get closer to. And that's a really strong framing when you think about where the world is today, which you know, most people who are using AI models are just scratching the surface and not really considering cost optimizations, privacy, sovereignty, and all the other things that you eventually get to. And so, you know, that's a let me pause there. There's a whole lot more we can talk about, but it's an exciting opportunity to really bring this new lens, right? That's current, very current in the market.

SPEAKER_04 7:44

Yeah, definitely. It's it's something that really shows up once you start getting into the sort of second and third innings of people using AI models within their company and realizing, wait a minute, you know, I'm using sort of the off-the-shelf sort of chatbot experience, and maybe I'm starting to code with it. But once it really starts to touch the actual private or confidential data that I'm working with when it comes to getting real stuff done, I suddenly am thinking about wait, who am I handing the data to? Am I able to do everything that I want to do with it when you know companies like Enthropic and OpenAI are themselves even just trying to figure out what are their policies for the kinds of work that they allow their customers to do with their products? So, York, from the research that you're doing, which sounds like it's really focused on just doing good with technology, where were the the pain points or the sort of walls that you were running into that made you think about actually being able to own your own compute and access your own hardware?

SPEAKER_03 8:53

Yeah, I mean, aside from sort of the opportunity to exercise that muscle and helps students who would not have thought to do that, right? How to do that. I've also done some local, pretty hefty laptop usage of open models and seen the benefit of that, but also the difficulty of doing it on laptop class hardware. And so the students sort of ran into, I would say, two walls. One of the students who was doing some, we do all the different types of research, including human trafficking and things like that. And so you very quickly trip the filters of most centralized model providers. And so one of the students actually had her account banned. So because of you know, tripping on these filters, which are, you know, sensitive topics, obviously, right? And so that happened a couple of months ago. And so, like her resolution was to try to figure out like, okay, well, what do I do here? And, you know, I there are normal processes to go through to try to get those unlocked, you know, with companies like Anthropic, Anthropic in this case, not faulting them like all the models would trip up on this type of language, you know, and in the context of this type of research, that's just gonna happen, right? And so, you know, getting connected to Anthropic in order to have that discussion is something important. That's another resolution path, but we have to date not been able to do that. So this is an example where you can take the open source, you know, edge infrastructure that B3 Labs provides and actually sort of go to town and not have to worry about those types of same filters. And you know, we're doing this for good and for you know, tech for good. So, you know, there's there aren't the same concerns that you would expect when someone's you know using these terms in a normal chat context. And then secondly, the same student who's actually has a ton of initiative is every day running out of running out of token quota. And so on a la, you know, uh academic lab budget and a student budget that becomes rate limiting in terms of what we can explore and what we can do. So those are two very specific examples aside from just learning, right, what you can do in this part of the field.

SPEAKER_04 11:01

Yeah, that's great. It makes me think about how you know the the hyperscalers today really just have the bandwidth to focus on really just the biggest opportunities and the biggest customers that they have. And it leaves this, you know, giant long tail of activity from people who, you know, want to use compute, want to use the models to explore, do things that, you know, are not necessarily profit-driven, but are going to create a lot of value for the world. And those, you know, can often be left behind, especially in this time when the the competition is fierce, just to access those compute resources and even just putting yourself in the queue with those hyperscalers is so hard. So, you know, being able to go direct and be able to actually own your destiny there just feels like such an important priority. And that seems like what B3IQ is meant to address. So, Daryl, Victoria would love to hear, you know, how you guys have designed this solution. You know, from what I've heard, it's just a really elegant way of attacking this problem.

SPEAKER_00 12:03

Yeah, absolutely. Happy to share just like a macro overview, and then Daryl can also go into just how we structured it financially for it to make sense for labs, researchers, institutions. At a high level, we ourselves with B3OS ran into a problem where we needed our own compute because we're processing people's funds. We're touching money. We can't have this trading strategies be fed to an anthropic or an open AI for them to train their models. Then your strategy is non-existent. And so we are our own first customer. And as we were setting things up, we kept getting inquiries from friendlies, institutions, other startups who were like, hey, I actually need my own compute as well. I need it for either private inference or I need it to actually build, create, and run large LLMs. And that's where the aha moment for us came about. We weren't necessarily talking to the Fortune 50s, and we're definitely not talking to the enterprises of the world, but it's the smaller startups, the labs that need help, institutions that are doing incredible work. A lot of what York's students were working on and they were sharing in this exposition we recently had at the Microsoft Research Lab, kind of touched on the Ukraine war and the Russian war and evacuation of Mariokal, which I'm a Ukrainian. And so the topic really touched what's so top of mind for folks in the area where when there's a bomb that's incoming, you have to make a snap decision of am I going to go to a shelter or am I just gonna wait it out in my home and hope that nothing happens? And simulating those types of responses in people help protect people and help save lives. But then as you're doing these simulations using these closed black box models, you trip up certain wires where it's like, oh, are you trying to bomb a city and that's what you're doing? Is this a terrorism prompt? Are you trying to get around our harness? And so I think for us, it was like, whoa, there's a lot here. There's other institutions that are working on cancer research. And for them, uploading pictures of people's bodies or private PII HIPAA compliant information is just it's not possible. You can't use an open AI, you can't use an anthropic for that. You have to have all of these regulations and be able to do the meaningful work required with using kind of the tools at your disposal. And so there's that market. And then another interesting one that came about through our research and just talking to folks was media and entertainment, which is one you don't hear about often because Hollywood oftentimes is shying away from AI, right? You have the writers' guild protests and everyone's like, ban AI, we can't have AI. But the simple matter is AI is helpful in a lot of cases. CGI can get improved. You're able to go live with like the next House of Dragon series on HBO so much quicker than the two years that it generally takes. And so they don't publicly talk about it, voice it, but they need, they're in desperate need of compute to be able to kind of expedite the entertainment industry. And they're getting overlooked because these giant AI labs are kind of a first-tier customer. They are a priority for anybody that's offering compute and GPUs. And so there's these interesting, like you said, long tail folks that are just getting overlooked, that cannot get access either for, you know, HIPAA or privacy reasons to the open source models or need to run and train their own models, but are just not getting the actual GPUs from from the cloud folks or don't know how to even set up their own data center and be able to run all of this. It seems crazy to think, oh, I'm gonna buy GPU and what do I do with it? How do I run it? Is it gonna be loud? Is my you know electricity gonna like go out every single night? Is my wife gonna divorce me because it's too hot in the house now? Like, well, what does that even mean? And so you're kind of at this, okay, beholden to the easiest option, right? Is I'm just going to try my best and use these closed loop models without necessarily understanding what that means long term of how it's going to impact you. And if you're going to lose all your work, if you're going to get outsourced, if you know the privacy is going to actually remain private. But that's kind of how we approached it. And it was really the aha moment for us of like, wow, there's so much unmet demand. And it's this pocket that we could really service and we could really work with. And that's how B3IQ came about. But Daryl, you had some really great insights when we were chatting with Sean or CTO through just the economics of it. Because I think that part is very interesting too.

SPEAKER_02 16:16

Yeah, I think it goes back to what Franklin was saying earlier around hyperscalers reserving all that capacity, right? And that leaves certain folks out of the line, out of the demand. And so I think if we take a step back right now, I think compute is at this interesting inflection point where open source models are getting really good. I think Elon predicts that most will be fable level or potentially exceed fable level on certain tasks in just a couple quarters. Right. And so, how do you reserve GPU capacity to run these? So there's all these sources of demand. And these gatekeepers right now, in a way, are the hyperscalers, right? They reserve capacity years in advance. Frontier labs come in and they buy up from our understanding, you know, not three months, not six months, but two, three, four, five years worth of capacity. So if you think about it from a very maybe dystopian, grim perspective, like the Frontier Labs, if you assume scaling laws, they've already locked in their model gains from now until 2028, 2029, 2030. Right. And so how do the new AI folks, the startups, the researchers, the media and entertainment folks, like how do they compete? Or are we all just going to use their models? And I think for us, that was that light bulb moment that Victoria talked about, where maybe there's a different way where we can bring GPU capacity onto the market using what we have, which is access to NVIDIA chips, knowledge around assembling and hosting and managing these GPU systems. And then on the other side, matching that with what the industry calls like offtake, right? So matching that with demand. And that's what B3IQ is. We're a platform that enables anyone, whether you're a startup team or you're a researcher or you're someone from the media and entertainment space to come in and buy these GPUs and run private AI on them and also monetize them. And the additional turning on top, if you will, is we have this unique financing model that we call rent to up. So instead of paying 100K, 150K for a GPU system up front, you can essentially pay for it with as little as 30% down. And I think to help crystallize, it's helpful to take like a simple example here. So if you just take, let's say, a GPU cluster of eight H200s, which is the, let's say, the workhorse, like the Tonda Civic, if you will, of the AI hardware space. It's not your Ferraris because those are the black wells and the Rubens that are coming on to market. But these things, these babies are the ones that people use right now for most training runs because everything else is a short supply. That costs about 400k to purchase outright with all surrounding equipment, chassis, etc. And based on today's rates, if you're an AI startup, we just spoke to one this week. On average, if you're renting capacity for one year, you'll be paying around 200k. So in about two years, you would have paid 400k, which is the same amount as it takes to own the machine. Right. So in a way here, renting maybe makes sense in the short term because you need that capacity immediately. But ownership really makes sense from an economic perspective longer term, because for the amount that you're paying to rent it for two years, you might as well just buy the thing. So for us, we're doing this rent-to-own model where not only are we lowering the initial cost of capital for someone, so if they're trying to buy that 400k machine, it's just 120, but we're also giving them this ability to own it in just two years. And they don't have to worry about hosting a data center in their office or in their shoebox of a New York City apartment. We host it, we manage it in a secure US-based Oregon facility. So that's what we're super excited about.

SPEAKER_04 20:10

Yeah, it is almost like a callback to how Apple and smartphones got so much adoption so quickly. And it wasn't because the devices were essential to anybody, but it was that they came out with a pricing innovation that, you know, you could pay off your phone over time, you get it for free at the beginning, and you buy a plan for two years. By the end of it, you own your phone. This feels like the the iPhone moment for compute, right? I mean, that's really what we're talking about when it comes to the next stage of AI adoption, which is that yes, you know, we've all accepted that LLMs are going to be this ubiquitous tool in our lives, but now the goal is to have an LLM in your pocket for many companies because you know, you just need that direct control, you need that direct access and ability to do all the things you want to do. So in this iPhone moment, it makes total sense to say, hey, now is the time to actually come out with the right business model to actually match the demand that is coming from, you know, just this much, much bigger audience that's seeing the value from AI.

SPEAKER_02 21:21

Just to take that one step further, because I think that iPhone analogy is so good. Right now, I think hardware I think is uh opaque space for most folks. So I always like to try to simplify it. And the iPhone analogy is great. I think it's like right now, if you're using a laptop class piece of hardware, you're getting like the first or second iPhone in terms of ability and what kind of LLMs you can run on it. I think what we're trying to bring to the market is iPhone 10 or 11, right? So that you can actually run some of these really effective open source models. And just to take a simple example, like the Kimi K3 came out, I think we went by. Two weeks ago, I think Frontier Labs were panicking a bit. The Kimmy K3 is just as good as Fable on certain agented coding tasks. But the piece of hardware that you need to run it is actually not just one H200 node of AGPs, but two. So you need two of these things. And the cost for two of those things is 800k plus annual hosting cost, plus the expertise, plus the supply chain access to procure them. So I think, yes, open source models are getting a lot better, but right now no one has the iPhone 10 to actually run those things, actually run that Uber app or that Lyft app. That's really good. So that's I think why GPs are still a missing piece of the puzzle.

SPEAKER_04 22:44

Yeah, absolutely. York, you've been uh a veteran of technology for so many years. You know, what do you take away from this? What would you unpack from it?

SPEAKER_03 22:53

I think honestly, uh, we have no idea what's to come in some ways. This, you know, as a side data point, I was at Microsoft in the 90s when it was a much smaller company, started with 6,000 people in 1990 when I started, five years after the RPO, and had no idea that I would be there for five stock splits during that time period, which was actually a nice bonus. But I left Microsoft in 2000 because I was just tracking sort of what's going on in the internet, what's happening, what's what could be coming. And the reason I left Microsoft was because in late 1999, around November, December, I was like, okay, I'm gonna go do wireless internet. That's the next thing. That's gonna be the next 10 10-year wave of the internet. And I looked around Microsoft and I was like, are we doing anything like that? And the answer was no. But there were wireless handheld devices around that were not iPhones, obviously, in 2000, things like these clunky handheld PDAs that you might remember. And one of those companies is not even around anymore, or a couple of those are not around anymore. But Wi-Fi was available, different bands of Wi-Fi were available, different broadbands, cellular broadbands was available. And so I left to go spend time in that category. And I wound up in two different wireless startups doing that work. I worked for the CTO of Goldman Sachs doing wireless work and built five products for our bankers while I was there. And the first product that I built, actually, so this would have been in 2000, was an alternative for our Tokyo-based managing directors who were demanding the equivalent of BlackBerry, which we had in the US. And so I had to figure out how to actually recreate that. And I had done some research and span a couple of months at a startup in that space. And so I kind of knew what version three of that was from a software perspective. And so I basically just went and rebuilt, actually worked with a software company, but I knew what version three was. So I knew how to get there really quickly. And we wound up deploying effectively a uh BlackBerry equivalent on NTT Dokuma phones in Tokyo in 2000 that gave you BlackBerry capabilities in the first time in Asia, essentially. And, you know, to your point, seven years later, the iPhone launched, right, with a lot of fanfare as a viable consumer device that, you know, to this day, the thing that I don't like about Apple devices is they're very proprietary and locked, you know, for technologists, first people. So it fits the consumer model. It did not fit the developer model well, which is why I started in PCs in 1983 when I was in when I was at MYU. So, but I think so. Where are we now as compared to that? I my sort of retort a little bit on that is we are at a place that is very much more like the late 90s internet era when people didn't really understand what was coming for the next 10 to 15 years. And we are doing that on top of all of the knowledge that we have and all of the capabilities and infrastructure that we have, both on mobile devices and on the internet. An Apple today, in my opinion, would not be possible because of the proprietariness of Apple, right? And if you just look at what the AI models allow you to build on your own in a data center, that very much disrupts highly proprietary locked-in ecosystems. And I think we're seeing that more and more every day, where there's just an amazing amount of open source stuff on top of the sort of you know, open source journey that has gained momentum since the mid-2000, 2014 era. We're just seeing it every day, right? Like amazing replacements for GIS capabilities, right? And for Earth modeling. And so that's also in a very powerful moment to be able to have access to this type of infrastructure in a lab, right? And in a context of being able to do things that help the green field of the humanitarian space without the same cost structures that you would have had a year ago, right? Like literary, right? And that's massively powerful. And that's what excites me most about this particular moment is we can accelerate the good that we're doing in the world with the power of these tools, as long as we are directed and thoughtful and guardrailed, et cetera, et cetera.

SPEAKER_04 27:33

Yeah, it's such a good point. Technology so often is just path dependent. And, you know, it seems like such a special setup right now that closed source and open source are just competing so hard with each other in parallel rather than, you know, sort of sequentially for once. And that means that the open source approach is incredibly viable today. And that forces, you know, the closed models to actually defend against that and hopefully act in the better interest of their customers because of that threat. You know, in Silicon Valley, you know, all these sort of AI native companies, when you ask them what they're actually doing in-house, you know, where 80-90% of their code is now being written by agents, I think nine out of 10 of the ones that I've talked to at least are saying that, you know, they're using open source models almost exclusively because they just can't rely on the Frontier labs and the platform risk that comes with them if they're also trying to build their product roadmap, right? I mean, they just don't know if the next fable gets shut off just as they're about to launch their product. That's not acceptable. And so, you know, they've looked at what the open source models can do and said, this actually does everything we need. We don't need all of the frontier stuff. We don't need the cutting-edge benchmark breaking version of these models to create the value and create the applications we want to build. We actually need the thing that, you know, is maybe three to six months behind those things, but absolutely serves our needs. And once they reach that conclusion, then you know, they start thinking about okay, well, is it time then to own our own boxes, right? Own our own hardware, because that's another piece of the sovereignty question and the platform risk. So, you know, coming back to B3IQ, Daryl, Victoria, you know, maybe that's the question to dive into is, you know, if I were to play the devil's advocate, there are neo clouds out there, right? AWS, Core Weave, etc., you know, they have things to offer without ownership. Why would someone want to own the box?

SPEAKER_02 29:48

Yeah, I think that's a great question. So I think it depends on the customer segment. So if you're a Frontier lab and you're raising billions of dollars, or you just went public, you have all this capital, and you've locked in your model gains. I think you're training very specialized models for newer domains that potentially can produce the same outcomes as agentic coding, right? So I think we're seeing that. And those are primarily the customers that are going to hyperscalers in neo clouds. But I think there's a certain class in the market that, at least on the commercial side, outside of academia, is similarly underserved, where they're looking for the ability to own their own box, run hardware on these, or run software, run open way models on these boxes that are cost effective and make sense from an economic, economical perspective. So, you know, owning is better than renting. And these folks are also in direct competition, I think, with the Frontier Labs, but in different domains. So I think one AI startup we just spoke to this week, they raise about 200 million in capital. They have a super unique take on simulation and generative agents, and they need to deploy that capital immediately. But I don't think they can pay the prices that the Frontier Labs are paying. And so they have two options. One, when they go to a hyperscaler NeoCloud, they're either going to not have a place in line or they're going to be in line for older generations of hardware. So H100, A100, maybe they have to stitch together consumer grade hardware like 300 or 3000, 5090s, right? All of this is infeasible if the capital they raise is meant to compete with some of the best models out there in specific domains. So like they either don't have access to the best hardware at all, or they're not getting access, or they're getting access to older generations of hardware. And I think we offer an interesting alternative. I don't think we can fulfill all the capacity they need immediately, like next week. But the idea is as we bring in, you know, investors and more folks onto the platform, we're going to be a viable alternative to some of these folks that are looking for more capacity over time.

SPEAKER_00 32:14

And I think the the other cohort here is outside of just like these, even an AI lab or an AI startup, like they're well funded. $300, $200 million is a lot of money. There's these individual AI enthusiasts and prosumers that have really taken up vibe coding. Like we see this over social media. Everybody's so excited. You're now creating your own version of Twitter. Everyone's launching their own version of a website they've only ever dreamed of. And these individuals want private inference. They want to be able to vibe code without the restrictions, without token, without paying all these crazy costs for tokens or being running out of tokens with these labs. And then also not having to worry about that idea, whatever it may be, being kind of taken into account by these large labs and that idea being taken away from them. And so they would love to have their own machine. They don't necessarily need an H200. They're a 5090 is just as good for their purposes as like any 200s of Ferrari, effectively. Like they're the Honda Civic is great for these folks. And they're, you know, when you think about how what their behavior is, they're vibe coding, they're excited about a project, they're using, you know, the machine and their LLMs open source for like three, four days at a time, but then you run out of steam, you're done, or maybe you need to take a break. And for two, three weeks, the machine's going to be idle. And in that sense, what do you do with it? You just spent 10 grand on this thing. You don't want it to be a sunk cost. Like with what we do with B3IQ, we connect you to offtake. So maybe the next vibe coder who isn't ready to necessarily purchase his own machine, but wants to get preferred rates, lock that machine in for three months at a discounted rate, cheaper than what a core weave would offer, and effectively be able to use it. And so you're able to kind of make up the cost through renting that idle compute to either individuals in kind of this marketplace software that we've built. So that's a really interesting segment through talking to actually, and it's specifically universities and researchers. That's where we met this whole cohort. And they're all in Telegram groups and they're all like, hey, can I borrow your RTX 6000 for the weekend? I promise I'll only use it for two days. And so it's this really and it's like the pre-Airbnb model where, like, if you needed a cabin up north, you would reach out to all your friends and say, hey, anybody have a cabin available? I'm willing to pay for it. Just let me use it. And so this doesn't exist right now for that cohort for that market. So even just like larger or medium-sized labs aside and the commercial folks aside, there's this demand from just an individual themselves. And I know York himself is actually working on so many really cool projects that he keeps showing us whenever we go to the Microsoft Research Lab. And it's just like, first of all, how does he have the time? Second of all, this is incredible stuff. I'd love for him to just talk about some of it because I I think it just really resonates with the point of that individual contribution that you want to make to the world and do something meaningful, but you also don't want that taken away from you or pay crazy amounts of money for the work that you're doing.

SPEAKER_03 35:09

Yeah, I'd love to talk about it. Let me take just before I do that, there's also an inflection point, Franklin, in the market that is really, I think, the unlock that actually makes distribution of compute and models possible. And that's not all the market, but you know, because consumers are going to go to whatever's available to them, right? And they're going to use the tools that are put in front of them generally. But if you look at the people who are thoughtful about the harness and the choice about the harness, that's the unlock. And what you're seeing in the maturity of the market, again, beyond the casual users, is the unlock. And so, as an example, like my harness is VS Code, and VS Code is open source, number one, right? Number two, my repository for everything is GitHub, right? Which was open source. Microsoft purchase it to make sure it didn't go out of business. The the those two things actually unlock your choice. So now when I go into my GitHub enterprise license, I see 20 models already and I see the Microsoft versions of models. And I can just add new models that are my open source models running on my B3IQ capacity. It is natural, right? It is an unlock of an ecosystem that was that is harder when you're locked inside of a specific provider and their interfaces, right? So that actually then allows you to make very economical choices of when do you use what, right? And that should absolutely be something that people are thinking about. We're sort of in this optimization moment, right, where you can make those choices because the underlying unlocks are there, right? And those choices to have, those types of harnesses. So that, you know, it was very easy for us to bolt on this infrastructure, right? Because our tools are that I'm training my students on. And they're grad students in global affairs, by the way. So their backgrounds are lawyers, policy people. They've never touched a VS Code or GitHub before. So it's an unbelievable opportunity to be able to educate this generation of global affairs and folks who are not technologists first, right? But now they can actually, with a bit of guidance, produce just absolutely amazing work. Victoria talked about the Mariupol siege example. It's the reason we studied that with the advisor is because it has a lot of data. And so, as a foundation for building data frameworks around evacuations in war zones, actually a really good starting point. And that was one of the things that tripped up the content filters. The other was we do work on human trafficking and forced labor. That also trips up to content filters. And then what I was showing Victoria that she mentioned yesterday when I saw them is I live near Dumbo and I spent a lot of time walking in Dumbo. And for years, the Manhattan Bridge, which is a gorgeous piece of infrastructure, has been raining down 90 decibens decibels on top of people's heads because of the subways crossing the metal structure. And it's this weekend, I was like, oh, I have these amazing tools. I'll go do a research study, right? Like, what's been the history over the last 20 years of the remediations or attempted remediations on that noise pollution that's happening? How has it been handled? And so I it's public, but I won't share it with you yet. I will share it with you shortly to look at. And I database, I've done basically not only that research, but also borrowed an excellent microphone from Victoria so I can go do some more audio researching. It doesn't get clipped by my phone microphone. And there's a ton of information out there, uh, open source data from the NTA, open source data from you know different organizations that tell you about ridership, people entering turnstiles, which lets you infer how many people are in a neighborhood. So how many people are actually affected by this? And you can do a full-blown study. And I've gotten far enough that I'm confident with the data, again, backing it up with some more first party recording data. But what I've noticed since I started measuring it, and this is actually astounding to me, is the noise on the Manhattan Bridge from trains, it does not cease for more than 10 seconds at a time during the day. It is almost incessant. And that to and it gets up to 90 decibels, like which just for context, I'm sorry, 90, 98 decibels. For context, a rock concert is 120 decibels. So it is ear piercing and worthy of study. And anybody who's dealing with noise pollution in New York City, like helicopters, Mayor Mamdanny, is absolutely going to be interested in this particular study.

SPEAKER_04 39:49

So yeah, it's fantastic that you can be doing that work. You know, as someone who lived in New York a long time, I feel like at some point you build an immunity to all of the street noise, but that sounds a bit out of bounds. So I'm glad someone's working on that. And that B3IQ and other great tools have just helped make that make that happen. You know, if we zoom out a bit and look at E3IQ five years from now, you know, the long-term vision, what it means to have an owner-operated compute network, what does that look like for you guys?

SPEAKER_02 40:23

Yeah, so I think it's a few phases. I think for B3IQ, initially, I think we want to solve the problem that's in front of us. And right now it goes back to the fact that compute and GPUs are in very short demand, scarce supply. I think there are researchers, amazing faculty members, amazing students like York and his lab that all need compute. And it's not easy to access the best compute at the moment. So we want to provide a platform that lets anyone access those instead of just folks at Frontier Labs or the folks that have the money to pay for them. So I think that's first and foremost the objective. I think as we move and grow further over the quarters, I think a few trends are happening, like we've discussed today. So open way models, I think, are getting better. And I think the economics of renting versus owning, like to me, these two trends of open way models getting better, owning your own box makes more sense than renting. All of that converges onto like the logical destination, which I think Palantir and Microsoft and some of these other folks are talking about. But I think they're speaking to enterprises. We're speaking to kind of like the masses where owning your intelligence really makes sense, right? But the way to do that is a little bit blurry, a little bit muddied at the moment for the average consumer or even the average YC founder. And we want to provide that one-stop shop, one one, you know, that kind of almost turnkey platform for them to come to us and own their own intelligence, right? And they can own the box, they can host the box with us, we can assemble them, we can provide this software matchmaking piece to not only optimize performance from these boxes and these clusters, but also match them with offtake, right? So then these boxes are essentially paying for themselves and these machines are paying for themselves. So we lower the cost of capital. So, in a way, it's both a tech and a fintech solution that I think meets consumers where they are right now, and I think is very scalable as we empower a lot of these startups and these individuals to do a lot of cool shit for I have a better word that New York is doing and fighting the good fight. York for mayor, next mayor of New York City. But, you know, empower kind of you know individuals to do really cool good things for the world, and then also empower on the other side AI startups to create better, I think, more purpose-built models than maybe what the Frontier Labs can do.

SPEAKER_00 43:05

Yeah, plus one on York for mayor. You have my vote. I yeah, and my support. But and then to just like really zoom out right now with B3IQ, we have a warehouse in Oregon. It's 27,000 square feet. We also have four adjacent warehouses that we are able to take over. And through this, the B3IQ, we want to scale into those warehouses. We want to grow. And hopefully in five years, we have a 50 megawatt facility running machines for individuals, institutions, researchers, and we're really providing transparency into the industry. So, right now, if you go onto like a VAST or any of these marketplaces where you're able to rent people's compute, the prices are so variable. You log in today and it's $3.56 per hour for an RTX $6,000. You log in just 11 hours and it might be $4.12. Those variable costs are very meaningful and impactful, specifically when you think about researchers where they have finite budgets, they can only spend X amount. They can't be beholden to these exorbitant costs that rise out of nowhere, to be quite frank. And so with the B through IQ marketplace and the matchmaking, we want to provide transparency so that you know if you go on and you lock in a price of $3 or $2, whatever it may be, you're only going to be ever charged that for the duration of the project without having to worry about the variability of it all. So we're super excited. And hopefully with Fantera's help, we can scale to that 50 megawatt facility and provide the transparency through the B3IQ marketplace.

SPEAKER_04 44:40

Where can builders and investors and future GPU renters and landlords come find you guys?

SPEAKER_00 44:47

Yeah, so b3iq.org. That's the website. You can come in, you can schedule a call with us. We actually have a very handy calendarly pop-up that comes up. So you don't write us an email and request a call. You just select a date and time that works for you, and you literally get matched with myself, Daryl, Sean, or the three of us, actually, to make sure that you have time with us so we can chat through the options, set you up with the right machine for your specific needs, make sure that you feel good about the entire experience, help you get into the dashboard. So b3aq.org, check it out, look at the machines, give us a call or set up a call so we can chat with you.

SPEAKER_02 45:22

It's also very agent-friendly. So if you're navigating the web through agents these days, your agent can immediately talk to us through Markdown and then, you know, we'll probably have an agent-to-agent conversation or something, but whatever you prefer.

SPEAKER_04 45:35

My my agent will call your agent.

SPEAKER_02 45:38

Exactly.

SPEAKER_04 45:40

Fantastic. Well, Daryl, Victoria, congratulations on the launch of B3IQ. I'm excited for uh all the rest of us to be able to own our intelligence and do amazing, cool projects. Uh and York, thanks so much for joining us and sharing that perspective. Uh, it's been awesome just learning about it as well. Great talking to all of you.