Paweł Czech & Cole Crawford, NATIVX | theCUBE + NYSE Wired: AI Factories

Palo Alto Studio Connection Silicon Valley and Wall Street. I'm John B here with Dave Volante, my co-host. Welcome back to the Cube Studio here at the New York Stock Exchange. I'm Jam Allen, co-host of NYC Wired, and today we are talking AI factories. We talk all things AI is a race for models, chips, and compute. But what if compute itself is becoming a commodity? Oil is a global market because we figured out how to measure it, price it, trade it, and ultimately build financial markets around it. Native X is betting we can do something similar with compute. Joining me now to unpack that are Cole Crawford, founder of Senova Global and co-founder and CEO of NativeX, and Pavle Czech, co-founder of Native X and CEO of Starfield. Welcome folks. >> Thank you. Thank you very much. >> So, I'm very excited to unpack this company. I'm going to ask you guys some questions. But first, this is a special story for NYC Wired because the overall thesis and mission of NYC Wired is to connect great minds, builders, breakers, disruptors, build a community, get people in the right rooms, and that was Brian Bowman and John Frier's ultimate mission. You guys are actually a live test case. >> Living the dream. Living the dream. >> Talk us through how this all came to be. Um so Brian effectively was speaking to me about um how will a compute marketplace look like and uh he immediately said okay I have somebody that you have to meet there's like no question this is the best person in the world that's working on how um markets right will evolve and uh what technology will be used and how will that impact uh how we create markets um and introduced me to Call and uh Cole was uh you know got a call from Brian in a similar way. Um I flew down to meet him in Spain. He picked me up at the airport and we spent a day in a restaurant and uh drawing up the concept on a napkin literally uh which we still have and then a receipt from the restaurant and that was in December and today we're here. We're uh going live. Well, you know, Nvidia's story started on a napkin in a Denny's, I believe. Right. >> That's right. >> So, maybe we're heading the same journey. >> Cole, you've built your career in open source. That has been your wheelhouse. You have been singing that swan song for a long time. >> Native X, I mean, what? Break this down for me. Like, is this really about trying to financialize the world of open source? Like, >> no. No, it's actually the same mission as open source. Like the point the reason I was always excited about open source is number one there's a little bit of meritocracy associated with the code that gets contributed. Now AI has like really replaced a lot of that that currency if you think of code as currency and open source but really what open source is about is transparency and native X is really the culmination of work that I was doing on sort of tokenized real world assets and building the banking rails and the and and all of the software as a service capabilities. So, you could take a commodity uh a commodity or an RWA, put that on chain and and start settling against that atomically with with a two-sided market because two-sided markets are great. I for one have thought that these closedd dooror sort of sort of backroom you can argue circular financed. Uh we can maybe talk about that later, but you you could argue that this is not a transparent marketplace today and NativeX is very much in the same vein as open source about trying to increase the the transparency and build a true two-sided market. >> Okay, let's talk about why it's non-transparent though. Like let's unpack some of the norms, right? Because there is legacy thinking in place here. Some folks argue it's not funible. You cannot measure a H100 in a data center in Poland and a H100 in an Nvidia infra center in San Jose. >> We agree. We agree on that. We agree. >> So, so let help me understand how you truly make it funible. What is the unilateral benchmark that location, performance, everything can be measured against? >> Sure. And and it's pretty straightforward. We actually have a hedgeable, durable, deep bench of commodities that we know how to trade today. That is actually the entire back half of the forward curve for accelerated compute. And that's energy. Energy is something that that's truly scarce. We might be in a in a in a synthetic scarcity situation today because Micron is catching up to the demand and Samsung and Seagate and other component part manufacturers are catching up to the demand. And if you remember, I built open compute. So I was privy to some of the supply chain and and uh demand side of what was happening when when Facebook and others were were building infrastructure at a at a pretty fast pace. And guess what? they caught up, right? The the supply and the demand, they caught up. So, it's it's one thing to say on a GPU hour an H100 has gone up for a month. That's to me a little bit like weather. We weather can swing. You know, you can have cold to hot in one day, but climate is a little more consistent over a little bit longer time period. And the transparency for a two-sided market in a commodity like this needs a common durable underlying that lasts across the capital outlay for the entire thing. It can't just be on the GPU because the GPUs they do depreciate even even if today they're up by some marginal percent they depreciate over years. You look example at at at OpenAI yesterday announcing Jalapeno actually being a real thing at half the power and 1.9 to to 2x based on semi analysis based on their their testing half the power 1.9% uh excuse me 1.9x the performance what would that do what will that do to the futures of an H100 when the substation and the data center actually make up roughly 60% of the capital outlay of what in at least Nvidia but you know more more to the point what the world today calls an AI factory. >> So let's just get into that for a second. You mentioned SE analysis. We had Jordan and us on the show last week. Him and I spoke a little bit about this world of GP per hour versus goodput. Very interesting the cluster mask philosophy that they those guys are developing right again create that independent verification but at the end of the day who truly verifies right if we think about how other commodities have been built oil gas they're built over time unilaterally people you know see some sort of ubiquitous value that they can measure against >> right >> we don't do we have that level of data like who is measuring this >> we sort of do and I you have it more generally and I think I think the beauty of native X is like we can actually abstract some of that away because there's a lot of independent testing that goes on from Sumi analysis and good bush and other other firms. In fact, you yourself can benchmark this. There's really good tools built into Llama and you know other other um harnesses that will give you kind of the token output. And there's nothing wrong with having a standardized benchmark in performance per watt per dollar where in accelerated compute performance is tokens tokens per watt per dollar. That's a great golf score. It's not a great hedgeable market. So an H100, which by the way there's multiple H100s, right? A GB300 can be PCIe based. It can have NVLink. It doesn't have to have NVLink. So think about how many thousands now you have Jalapeno, you've got Samanova, you've got Gro, you've got Cabbrris, you've got Posatron, you have a number of of silicon companies that are all going to introduce models and and generational change as I guess we're calling this Wong's law um right as as the performance of accelerators uh get better every nine months, right? This is not like Moore's law uh where it was 18 months. So, it's it's fast and you can actually look inside of your own harness and say, "This is how many tokens per watt per dollar I'm seeing on this card." And you can easily build standardization around this pretty transparently. The challenge comes when you try and hedge that against the overall capital outlay of the data center, the substation, the racks, the rectifiers, the networking components, the CPUs that do a lot of of the interfacing that you're missing 60% of the asset if you're just looking at GPUs. >> So Pablo, I want to go to you for a second. People want optionality, right? In a market like this that's this hot where supply isn't really necessarily able to meet demand optionality is a challenge. Okay. You have built a career around community builders all sorts of peripheral players right not necessarily the guys inside the NVIDIA data center folks that are building apps on the edge all over the world. >> When we think about the model for this who could really benefit from this >> talk me through that and I want to challenge you a little bit on the NeoCloud framework but first I want to hear from you. >> Um sure. So, very much like you compared um the to a commodity like oil, right, that we learned how to trade and create different types of contracts around. >> Um people that are going to be offtakers, they're going to be offtakers of a specific type of compute, right? So, it can be edge compute. It can be compute that has a specific um source of power. It can be compute that has a green offset, right? is producing a specific amount of tokens. So it not it's not homogeneous. You need to create that common denominator and our experience right now is that for different use cases, people want different type of compute. >> You know, you don't have to power up the sun to uh cook, you know, uh something in the microwave. And this is what we're doing today. And we're always buying the sun. We're always paying that u maximum premium. And this is because nobody knows how much is your capacity actually worth. What's the value of um of that compute. Um so number one for us as a neocloud as a community builder it was very important to bring a tool bring a system into existence into the market that will allow uh more democratic access. It is about being able to build the AI economy and make it more accessible for everybody. It's not only people that can make a $20 million purchase from Google, right? Or uh or somebody that has to pay a hundred billion like um like we know that there are deals between Nvidia and OpenAI, right? For hundred billion dollars to access specific compute. That's not the totality of the market. That's not what the average company, the average builder, the average enterprise buyer wants. Um, >> and this isn't based on like any price list, right? This isn't based on AWS or Azure. It's not based on a pricing index. It's based on essentially how much AI can I get for my token. >> I tested delivered buckets. So, it has a great comparable in act in oil. So, you know, the in in print crude the standard is a barrel. A barrel happens to be 42 gallons. I don't know why it's 42 gallons, but it's 42 gallons. If we kind of look across the industry, it made sense from a from an economic perspective, from a dollar perspective to do this based on a guaranteed delivered 20 kilowatt hours of compute. It's energy normalized. We call this energy normalization. So now it's not based on what a very good or successful salesperson sold an H100 for. That that feels a lot like liebore to me. >> Yeah, for sure. And I want to talk about the marketing side of this business because I feel like we don't talk about it enough, right? Like in the whole world of AIA factories, Nvidia is a phenomenal company, right? Like what they build, how they've lock folks in like it's it's unbelievably impressive, right? They're also marketing maestros though. Like if we're if we think about it frankly, so is anthropic. I mean, we're seeing that on OpenAI. They are very very good at leading with the message. When you think about the appetite and the enthusiasm around having something that has a standard commoditization, you know, I asked if a gentleman on the show last week, one of your I guess somewhat of a peer in this industry, you think Jensen wants this? And he said, "Yeah, absolutely Jensen wants this." I kind of beg to differ on that, right? I I think that there is a lot of ambiguity in this marketing model that we're in right now, and it serves the titans of industry in some respect. >> The NeoCloud, and you can challenge me on this. I'm interested to know your take. The Neocloud model though is an interesting one because it allows them to hedge capital, right? If you have predictability around cost outcomes and also gives them some sort of competitive advantage on the performance layer. What do you think about both of those things? Like how do you respond to does Jensen want you know a price index for GPUs and what is the unique advantage of this working for the world of Neoclouds broadly? not just your own but like you know the quaries of this world. >> I I would say anytime anytime you're in a position to play sort of kingmaker because you have the supply and you create the demand and that demand comes from the fact that you control the supply. >> Um that's not a two-sided market. That's a that's a one-sided market >> for sure. >> And and so you know does does any person in that position want transparency? I don't know. you know, I guess depends on your motive. Um, if your motive is purely financial, then absolutely no, you don't want that. But all of your competitors do. And this is a part of why open-source exists is because that was exactly Microsoft. >> Go easy on Microsoft. That's like my pension. Okay. >> 25 25 years ago. I mean, they had a closed sourced ecosystem. They gave, you know, they gave priority to their biggest customers. Linux was created in part as an alternative to that the transparency to make the code do what you want. You're you're now seeing competitors and I mean OpenAI became a big competitor to to Nvidia yesterday. Ser Cberus from a performance per per dollar perspective is you know one of the best chips on the planet right now. >> Um I think your competitors when when you are as big as they are people start gunning for you because they see market opportunity. This is just this is capitalism. I mean and that's great. >> And in the NeoCloud space, your core, your Nibbius, you're selling H100s, you know, at a standard price, right? What are you competing on then? You're competing on the performance like break me talk me through like 5 years from now if we have a really clear global commodity around GPUs, how you have competitive advantage in that market as a neo cloud. I think I think AI I'm sorry Pablo I want I actually would like you to talk on this too but I do believe that in this particular case >> AI coding agentic coding has largely taken away software as a moat and now it becomes execution and so from a neo cloud perspective what's your uptime what's your time to response what are your SLAs's how do you support your customer on the missions that they're on can you fine-tune LLMs or the harness for their specific use cases this is an executionbased business now not we have a SAS differentiation because you can create that minutes >> you know interestingly people say that the Neocloud bubble there are skeptics that argue it's a bit of a real estate play right so in some respects this really challenges it separates the rubber from the road if that was to be true >> how can it not be >> um yeah uh absolutely one way to think about it is um are you actually optimizing the use of your let's say real estate or the amount of capacity that you have. M >> so if you create number one a market where everybody can buy and sell right you create liquidity um and that compute becomes tradable instantly right you can settle you can have physical delivery of the compute as well right then the question becomes okay do you actually have megawwatts or bragawatts okay somebody said it I love that >> uh so uh you need effective capacity effective capacity means that it also has to be delivered ed um at a latency that that is acceptable for your use case. It has to be in the uh jurisdiction that is relevant for you to actually generate the tokens. So when we're talking about an AI factory and we're saying an sovereign AI factory is something that can be attested, delivered, right? It has to be available to you at a latency that you're willing to accept >> and to your point at a cost that is relevant for your business. M >> and this is where one of the biggest challenges of the industry today uh lies is the lack of profitability, the lack of return on that real estate. >> So our claim is twofold. Number one, our goal is to deliver that capacity instantly at um fair market value that is determined by an actual market. And then secondly to create a financial um um I would say a a fintech approach to compute where you can actually create products like uh future contracts like swaps uh things that will yeah >> markets love regulation right I mean you know they like predictability so guys a lot has happened for you in what is now eight months you're here today I know there's some exciting stuff happening I think you're also about to announce a deal with ICE. Can you talk us through that? Like, talk us through what has been happening. I mean, there's clearly an appetite here. Bring us up to speed on the progress thus far. >> We we I think that was announced. We we announced that ICE was going to be building a futures contract based on the um NXCI, which is the native X compute index, >> uh that is backed by coil, uh compute oil. >> Love that. And actually, you know what? I have it here. I think I have one of the tokens >> there. There is there is your uh >> there is your coil that is that is that is worth 20 kilowatt hours of attested and delivered compute. Yeah. So, >> so as you see it, it does say coil, compute oil on it, right? Um, and it's trademark pending. >> Trademark pending. It is a representation, right? It is for people to have an easier understanding that it is physical delivery of a commodity. Much like you can take delivery of a barrel of oil, you can take delivery of this. >> And and again, this is you said something really important, Pavle. This is transparent settled settled because there was a fair market value. There was a two-sided market and a fair market value because someone put this up as you know better. You know, we're sitting here at the at the epicenter of capitalism and you and NYC what makes the money move a matching engine. >> So you put up an order and that order crosses. This is the same thing. So we do run a spot exchange where you can buy that capacity. those orders can cross and then that becomes instantly deliverable as a spot. What's great about this model specifically is everything is on chain. So you publicly and cryptographically sign that you've made that available and anybody any any any of the big four audit firms, any private equity bank, any hedge fund can go replay that publicly. It's all there. So there's there's no liebore like uh qualitative deal done here. There's compiled software that sits and cryptographically signs the energybased accelerated compute that gets delivered on something that people know how to hedge and trade today that that will survive the depreciation cycle of silicon which is energy. >> So it's all on chain. Okay. So you guys are doing something very interesting in that you're selling a commodity in one space but you're also convincing the world of the value of that commodity in another right so it's both a marketplace and a movement >> and a movement largely based on the fact sorry go ahead >> yeah and and movements marketplaces need customers and movements need believers right >> how are you attacking this like talk me through what you guys have been doing it's a fascinating challenge So number one today right we're here inviting um the industry uh and I mean anybody that is dealing with with uh the financial industry and the neo clouds coming together and effectively being able to start trading um trading the exchange. So this is today this why we're here. Uh we're going to have an amazing event thanks to ICE, thanks to the NYC uh the team here. Um so people are uh so interested, the demand is so high. We've heard that it very likely is the highest demand for any uh type of financial asset in the history. Um so I believe that the market in itself wants to uh participate. Yes. So we've wherever we speak to anybody in the world so far for the last year nobody has said I don't need this I don't want this this is not something that is going to work and function on a day-to-day basis so we're extremely confident in that at the same time um it is a question of now the industry coming together and agreeing on a standard right and the closer we are to that energy normalized standard um the better the outcome. So one of the things that we're doing as um as Starfield is that we are also taking the responsibility to deliver compute right based on um what people buy and sell. So uh we're making sure that you can actually take physical delivery of this. Um so it's a question of confidence. It's a question of managing the the risk right as well and it becoming a tradable uh asset. And above all, I'm sorry, could I let you? >> Can I just just 10 seconds, you know, there all of the Neoclouds. The reason why we're all in this kind of boat, in this movement, I think the movement built itself, right? Largely because everyone's in this race. I think the frontier folks would tell you we're in this race to AGI, but I think the bankers would tell you we're just in a race to like capitalize on the opportunity. uh and you know those are probably mutually exclusive. So noble goals on one side and and financial goals on the other but nothing wrong with that. The I think I think the point is that in in order for this to succeed as a movement you need confidence on both sides. So the supply side needs to go back to their LPs, to go back to their investors, go back to the ETFs and the mutual funds and the sovereign wealth funds that are saying, "Okay, we're going to put money into your fund." And then on the Neocloud side or any on the hypers scale cloud side, there is $4.6 trillion sitting in SPVS off balance sheets today. >> Like that number is going to grow. U so you know, I think it's solving both sides. supply side gets better transparency, neoclouds get better interest rates and better deals, maybe supply chain advantage. And I, you know, I think at the end of the day, this very quickly, and I mean like over the next 12 to 18 months, we'll start looking like a revenuebased business and not a growthbased business. And that's that's what we intend to help with. >> Well, folks, it's certainly about commodification and confidence, but it's also about community. And again, shameless plug for NYC Wired, but what a great story. Delighted to have you guys here. Looks like great events happening upstairs. Hopefully, I'll make it up. Thanks so much for joining us at NYC Wired. >> Thank you. It was a pleasure. Thank you very much. >> I'm Gemma Allen here at the Cube Studio at the New York Stock Exchange. This is NYC Wire's AI factories. Thanks for watching.