The Power of WEKA with Antonio Gracias and Liran Zvibel + Jam with JCal: Uptrends AI | E1986

0:00They go between 5 % and 12 % of the fund. You know, this is just at five. We would have bought more if we could have. The round wouldn't give us any more. So, you know, I'm hoping after this podcast, he decides to sell us some more shares. This round was super oversubscribed. And so we were scratching and cloning everything we get. We've offered to do some operations help, which I think, wow, we were able to get the allocation we got. But we were super impressed with these guys and with what they've done. And you'll, you know, you hear about the customers, the logos, the growth. I mean, I think you'll feel the same way.

0:24And this is a strategy that some savvy managers are pursuing now. Just if you have that winner, trying to get as much into it as possible. because of the power of law. I see Brian Singerman doing it. I saw one time Sequoia do it with WhatsApp. But I will say we invented it. You invented it, yeah. And we've even taken a page out of that book with our fund instructions. So yeah, great artists copy. What is it? Immature artists copy, mature artists steal. This Week in Startups is brought to you by Command Bar. Seamlessly integrate an AI-powered guide into your software, making navigation intuitive and interactive.

1:00Visit commandbar.com slash twist to get a custom live demo. Dot tech domains. Don't miss our Jam with Jcal contest. To apply and get more details, go to jamwithjcal.tech. Brought to you by dot tech domains. And lemon.io. Hire pre-vetted remote developers. Get 15 % off your first four weeks of developer time at lemon.io slash twist. All right, everybody. Welcome back to This Week in Startups. I'm your host, Jason Calcanis. And I'm really excited today because a friend of the pod, Antonio Grazias, is here. He is a tremendous investor, one of the greatest of this generation from Valor Capital. You know some of the amazing investments he's made over the years from SpaceX, Tesla, and now actually an investor in Athena, an investment we just did together.

1:49And we saw along the Newswire press release about this new investment you're doing in Weka. And so I thought, gosh, this seems like a very important company. And so I asked you and the CEO, Antonio, to join us. Laurent Zivbo is the CEO of Weka. It's pronounced Weka. And there's some background to the name, yeah? For people to remember, Laurent, welcome to the program. Thanks for having me. So Weka is pronounced as any Greek unit. So you probably know how to pronounce Mega, Terra, Exa. So you're just doing a Weka. Weka is 10 to the power of 30. and if EXA is a thousand peta, which is probably the biggest number people really can interact with, WECA is a trillion EXA.

2:32So it's a huge, huge, huge number. And you know, Bill Gates at once said, we'd never end up getting more than 640 kilobytes in a computer. Why would anyone need that? I'm not going to predict whether we will or will not need a WECA, but hopefully we're all going to leave to the point that someone's going to need that. And so as we start here, Antonio, we're always interested when a capital allocator who likes to tackle hard problems comes into a company and writes a big check. So maybe you could tell us a little bit about what you discovered here with Weka and what the investment thesis is. Thanks, Jason.

3:07And thank you for having us. We're really grateful to be here. It's always enjoyable to see you and talk about our companies. So, you know, I think first and foremost, we're spending a lot of time in artist intelligence. As you know, we invest in DeepMind back in 2010 and I was on the board of Tesla during the being of the eye vision systems there. So we thought about that a lot, and we have dozens of investments on our intelligence. They're focused in two areas. One is infrastructure. The other is what we call verticalized applications. These are applications that have very good proprietary data, and then a reinforcement that's very tight.

3:36Weka is in the first category, infrastructure. And what they're doing is making the data center much more productive. So the average GPU is utilized about 30 % of the time. This is like a close regard and secret, but that's what we think. Weka makes that a lot better. and they do it in a very very interesting way i'm gonna let lauren tell you about the product but i will tell you when i a funny story to lauren when i um when i first met him and we actually went a second second time with him we went to the product i asked him to kind of explain the math to me how this thing works kind of amazing how it moves it makes the back end of the gpu the memory behind the gpu much more productive he went into the math and he started diving into the math deeply on a whiteboard like a like a you know kind of plays professor kind of thing and i'm watching this i'm asking some questions and i realized that i actually remembered my math i knew what he was talking about so i was i was very impressed but he's a deep product expert so that's the the first thing is the product's amazing it fits inside of our investment thesis for the infrastructure in the data center the second is that um you know our strategy as you know is to write kind of smaller check and you know how many bigger check bigger to bigger check we wrote a small check into weka um followed them they beat every number they ever gave us and then and exceeded substantially product market fit was incredible and then i got to know leron because you know you know companies i was impressed with with leron's team and how well they execute i mean they are really product market experts they're product experts and they built something before the market knew it needed it and then the market came to them and the company just exploded but that's what we did and this uh strategy to feel their bet get to know the company and then place a second or third bet it's worked before yeah yeah it's right it's worked very well.

5:11I mean, our portfolios, typically our biggest winners are largest investments, very different than many managers that kind of spray out lots of positions. We like very concentrated positions. We'll write lots of small checks, but our big checks, and this is a$100 million investment in our fund. That's a large position for us. They are typically our biggest winners. And those tend to be 10 % of the fund, 5 % of the fund size, 15 when you make a gigantic... They go between 5 % and 12 % of the fund. This is just at 5%, and we would have bought more if we could have. The round wouldn't give us any more.

5:43So I'm hoping after this podcast, he decides to sell us some more shares. This round was super oversubscribed, and so we were scratching and cloning everything we get. And we've offered to do some operations help for Weco, which is why we were able to get the allocation we got. But we were super impressed with these guys and with what they've done. And you hear about the customers, the logos, the growth. I mean, I think you'll feel the same way. And this is a strategy that some savvy managers are pursuing now. Just if you have that winner, trying to get as much into it as possible because of the power of law.

6:14But I will say we invented it. You invented it, yeah. I see Brian Singerman doing it. And I saw one time Sequoia deal with WhatsApp. So I think collectively people are starting to become aware of this, Antonio. And I do give you some credit there for pioneering it a decade ago. And we've even taken a page out of that book with, you know, our fund construction. So yeah, great artists copy. What is it? Great artists copy, some of the steel, something like that. Immature artists copy, mature artists steal is the strategy here for us and for maybe some of our contemporaries. All right. So let's get into the product.

6:49The company's existed for a decade. I know you've been moving large data sets around. So explain to us, Laurent, what happened in the history of the company where you realized, hey, managing large data sets, moving them around, and then all of a sudden these H100s and AI sort of hit a tipping point. I think we'd all agree about two years ago, maybe you could tell us a little bit about the strategy of the company here and getting that extra 70 % out of every H100 for people who don't know. People call it a chip. It's more like a platform made by NVIDIA. And Zuckerberg just announced they did their latest LLM on 16 ,000 of them.

7:24So to get an extra 70 % from those 16 ,000 is, they're not cheap, right? 30,$40 ,000 each, I believe, or on average. So yeah, tell us about the product and how you got here. We've been doing data management. We've been doing storage for a long, long, long time. Work as our second company, the previous one called Exiby Storage and IBM Acquarda. So anyone you know that's been buying high-end IBM Storage in the last 15 or so years in previous company we've learned a lot about the enterprise about big customers and about why people buy storage while being at ibm we left ibm at 11 at that point we basically took an oath and said hey never again this market just doesn't make any sense and it didn't make sense for two main reasons one the delivery mechanism for this product is flawed while most of the engineers are software engineers and when i'm saying most is like 99 not 51 the delivery mostly comes in some proprietary box uh it makes it hard on premises it makes it impossible to run the product well on the public cloud the second problem is that the market is fragmented like crazy so it's a 150 billion dollar market but there are hundreds of products that actually sell and when you're looking at it from the large enterprises, the big corporations, they're using anything between 10 products, if they're really lucky, to maybe 40 to 50 products across hundreds and thousands of silos.

8:54So it's hard for the customers. It's incredibly difficult for the vendors. So that's the reason we said we want it out. Then back in 2013, when we started three huge, huge, huge changes in what's possible happened. Containers happened and they led to completely different thinking about software, microservices that let you take a big monolithic system, chop it down to tiny little pieces and run them on many, many, many machines. So microservices changed how you can think about infrastructure. NVMe happened, which allowed you to take flash, connect it directly to the CPU and control it very effectively in tiny little fragments.

9:40And then the third change, which was the most important one, is network finally caught up to the speed of the servers. So throughout history, the network may have been 100 times lower if you're looking at the 80s and 90s, maybe 10 times lower if you look at the early 2Ks. In 2013, network has caught up. And by the way, by now, the network is twice as fast as the servers can create frames at 100 gig, but the network switches them at 800 gig. So this is incredible. Founders, I know a lot of you listening to this podcast, you build software for a living, right? We love doing it, but we all know it's hard.

10:21You know the pain of trying to onboard new users and get them up to speed quickly. Worse, most chatbots and guides built for the task are annoying as heck. Users tune them out because we all hate random pop-ups. Thankfully, there's one company that uses generative AI to help users onboard without annoying them, and it's called Command Bar. It has a chatbot that gives personalized responses to user questions instead of a basic Q &A. And it shows users around your product like a live guide, cursor included. Even more, Command Bar can detect when a user needs a nudge, giving them a product hint or special offer to close that important sale.

10:59Command Bar is used by unicorns you know, HashiCorp, Gusto, Sixth Sense, AngelList, and others. So here's a simple call to action. You got to check out this product. Integrate an AI power guide into your software so your customers can navigate your product intuitively and quickly. Visit commandbar.com slash twist to get a custom live demo. Wow. All of our competitors, the reason there are so many products out there, they are basically compensating for slow network by optimizing for locality. They're creating a storage array. I'm sure you've heard about the storage arrays before. At the storage array, you have a bunch of controllers.

11:39Each controller owns a portion of the namespace, and you're picking whatever storage you want to buy, but how the application basically spreads their data across the day, across whatever, and you want to break into controllers in a uniform way. The problem is, once you've picked that, your solution now becomes a victim to the clients, because if all the clients would approach a single controller, it's the bottleneck. So you may have a hundred controller system, you're getting 1%. On the cloud, it's even worse. Let's say 1 % now sits on an instance that has noisy neighbors. You're not even getting your 1%.

12:19What we have done between microservices, NVMe, and fast networking, we've started as a clean sleigh, and we have created a system. It took us years, by the way, about eight years to get it to really run well, and then few more words to mature it, where what we're doing, we're looking at what the server sent, the client sent every split second, and we run the perfect load balancing of IO throughout all the NVMe devices and metadata throughout all of the CPUs. And we're showing that even on cloud with noisy neighbors, we're getting perfect linear scalability. So if you give us twice the big WACA, you're getting twice the results.

13:06And that's the reason, since what we're controlling is latency, what's keeping the GPU servers starved for data is high latency. Since we're controlling the one thing that actually matters, we're the only product that optimizes for the one thing that actually matters, we can get to the point that it doesn't matter how many GPUs you have, what is their access pattern we can get them filled up to the 90th percent so this becomes an acute problem when you're trying to solve major problems in the world um maybe self-driving or uh you know running a language model people really didn't face this problem all that often previously yeah no they have not so now comes the big question of product market fit and how do you take it to market.

13:56So we predicted customers would need bigger scale, would need bigger performance. We started this thing through simplicity saying, hey, you don't need to buy these 50 different products to run all of your workloads because we're just going to run them well. Had we had to go through that phase, it would have been a much longer journey. What are the biggest workloads today if you were looking at like categories i mentioned self-driving i mentioned language models those seem like super obvious ones are am i correct that those are in the top three or four for sure so we're now seeing the explosion of gpus you've mentioned the h100s previously there was the a's and the p's and the v's and the b the b's are coming in uh in the next and they're used a lot in AI.

14:48We are showing that we can get AI projects running on, if you're looking at the cloud, we have Stability AI running us. When they have switched to us on the AWS cloud, we're able to get their storage bill down from 5 million to 1 million, so five times cheaper. And as Antonio said, before they switched to us, their GPUs are running at 30%. The biggest problem was that they needed to get more job done out of the existing GPUs that they had, they couldn't get any more because there is a supply chain crunch. After they switched to WACA, we got their utilization up to the high 90s. So over three times more output with five times cheaper, so 15 times better.

15:32So for AI, we're bringing big, big help. We have big car makers that are using us exclusively for many years and we got there time to epoch from a couple of weeks to four hours. Explain time to epoch for people who haven't heard that term before. So time to epoch basically means, hey, we're going through another round of training and you need to go through, hey, I'm ingesting the new data or it's already there. You're running a bunch of ETL. ETL is not sexy enough. You cannot charge for it. So now it's called ML Ops or LLM Ops. But the ETL portion of it basically takes your tons of tiny files and prepping them for the next job.

16:15So you're reading and rewriting them in a good format. Now you're retraining one form of your data. So you're picking a subset of the data. You're sending it to tons of GPUs to retrain. So you need to copy the data into the GPUs. this operation takes a long long time and it's really great that now you can parallelize these on so many GPUs while you're doing that you actually need to save checkpoints so you're writing a lot because if it fails you actually want to go back and you don't want to miss so much time and at that point at some point you're done with the next retrain you have now optimized for another portion and you would want to go and run regression tests and see, hey, I've optimized the model to slightly better, if it's a car, let's say, junctions with always stop when it's raining and there is a crowd that casts a shadow over the sides.

17:19And now we want to see, hey, with these crowds casting shadows, did we get anything else worse? So now you're running a huge amount of regression. you basically would drive your cars through all of the rest of your data so you want to see that this has worked well and finally this is one one one circle and now you have a slightly better model and you start all over because it's crazy antonio to think like we actually went full circle in the history of computing from there's way too much compute way too much storage tons of bandwidth and there's no application for it and you know it's just youtube is chugging along facebook is chugging along massive amounts of data your iPhoto going to 4k storing it all sorting it all and just nobody seemed to have a use for us and now here we are Antonio with a moment in time when we have supply chain you can't get enough GPU CPU power you can't get enough storage you can't get enough bandwidth to move it around and we're doing technology runs like it's pro systems in the seventies or something where you're time sharing and you have to wait to release your product and run another job and coming out with versions of it.

18:34I mean, is this going to last or do you think all of this incredible effort to make better chips, better storage, better operating systems and processes will result in us kind of catching up to the jobs as it were? It's funny. When I was 12 years old, I went to computer camp at Michigan State University and I had to hand in punch cards. That's how old I am. And come back overnight after the run. If you made a mistake on it, it just didn't run, right? And you had to go back and like - Yep, get in line. Yeah, exactly. You write the code out by hand and then use the binary punch code thing. This is like air traffic or actually traffic control systems for the GPUs, right?

19:08So filling up the CPUs are always fully utilized GPUs. And I think the answer to your question is, I think this is always a problem because the technology is moving and the software side, the LLM side, that the model technology is moving faster than the compute. So absent a large breakthrough, like a quantum computer or something, which, you know, it's always like, it's always one year away or two years away. It's perpetual. It's almost ready. As long as you're on Silicon, I think you're always going to have this problem because the model seems to be moving faster than the ability to create compute.

19:38Now, it doesn't mean you won't go through some boom and bust cycles. You know, you and I were both running the internet, right? So I was building infrastructure back then, the connector business, right? So overcapacity, I built, blew up. Then we were short capacity, went back up again, and you'll continue to have these kinds of cycles. But artist intelligence, I think it's kind of a super, super cycle in the sense that it is much bigger than the internet. And the need for trained data is huge. It's global. And it's not going to stop. We're just at the very beginning of this. So I think it's a long-term trend that will have a sine wave with a very steep upward slope.

20:12Yeah, if we were to compare it to, say, building out bandwidth, that was a 20-year story. and yeah it we had a boom bust cycle was worldcom and all these places where they overbuilt the fiber and google bought it for pennies on the dollar people don't remember this it was the early 2000s yeah they went for it laid so much fiber and that actually became i think in some ways the precursor for things like netflix and youtube and your photo library not caring about bandwidth anymore i mean it used to be putting any viral video on the internet would result in a 10 000 or a hundred thousand dollar bandwidth upcharge from your provider and then bandwidth suddenly became free and unlimited and the storage that that was another 20 30 year story so if we were to parallel to those this could be a decade or two decade chase to get more gp at least and with with the added difference that the in that case the end use which is sort of like you know moving packets on the internet is kind of a fixed use and it didn't change very much in this case the end use is changing a lot so the applications of our intelligence are going exploding right so right now we're at the very early stages like it's it's an app on your phone it's grok and xai right it's not um a robot yet but it's washing your dishes so imagine the amount of data required just the things these things we can imagine forget things we can't imagine what's going to happen but going from um you know the at the grok app on your phone to the robot washing your dishes is going to take a lot of compute.

21:40We don't have it today. It doesn't exist. And building it is an enormous lift. And anything that makes the dentist more productive is going to be very valuable. Founders, these jam sessions with J-Cal we've been doing are a huge success so far. We've seen so many great submissions. We picked two winners so far, and we interviewed them on the show. And we're still looking for three more amazing founders. Here's all you need. If you want to come on this program, This Week in Startups, and Jam with Jcal, where you pitch me your company and I give you a bunch of feedback. Just two qualifiers. We want you to have less than$2 million in funding, so we want it to be a new startup, you know, one that needs help.

22:16And we want you to have an awesome.tech domain name, so head to jamwithjcal.tech. Jam, J-A-M, with jcal.tech. It's so much fun. The winners come on the podcast. I have to tell you, this is the greatest editorial segment on the pod for me. I love doing it because you pitch me what you're working on. I give you some unfiltered feedback. We do it in real time. We ping pong back and forth. We pickleball. I ask you a question. You give me a response. You ask me another question. And you know what? We both get smarter and we understand your vision for your startup. And hey, it never hurts to get your company on the number one startup podcast in the world.

22:46And we're working with our friends at.tech domains because it's super cool to have a.tech, rabbit.tech, aurora.tech, 1x.tech. Everybody's using them. Heck, I use it for founderfridays.tech. So here's your call to action. Tell me about your awesome.tech domain and startup and apply for the Jam Session with JCal contest today at jamwithjcal.tech. We're picking those last three winners soon. And so take us through some of the logos, Lauren, that are doing the most interesting things. When you look at your Lighthouse customers, people who are doing the biggest, most ambitious things that you can talk about.

23:19Maybe you can give us a little highlight. Either your big Lighthouse customers or the type of customer, the type of job, if you want to abstract it a bit. Yeah, so I've already mentioned some huge AI project between deep learning, machine learning, the generative AIs. I mentioned stability. We have mid-journey and a bunch of the other very, very exciting forefront of AI. But, you know, GPUs are really useful for other means as well. So if you've been to Las Vegas, you've seen the sphere and you went to the U2 concert. You two basically were the first show on the Sphere when they got the residency.

24:02It was June. They had to start showing up and performing in September. Their initial plan was to use their own infrastructure that owned tons of GPUs, but no previous way of thinking about all flash arrays. They started re-rendering their thing. It would have taken them six months. at that rate obviously they didn't have the six month from june to september they've switched to us to switch switch to waka from the moment they got it on the floor to the end of the first render it took them under four weeks so six times faster but then what they've realized hey if you're taking a hd square thing that makes sense in a stadium you're reporting it to the sphere which this magnificent screen it just doesn't look good so the fact that we were six times faster the fact that they got the first render and they still had two more months actually allowed their artists to come in and reiterate and change and reiterate until this magnificent experience happened i don't know if you've been there but i haven't been to it but yeah i've had some friends who went to see the dead and the and the youtube stuff yeah yeah so the dead already dead and calls to use those and they have these experience where like the whole thing closes on you like a box or you're out there on the desert and now that experience is like magnificent it's so i i better say a lot of times awe-inspiring it's like being a medieval european and walking to a cathedral right that's how it feels but i want to say this it wouldn't exist without weka it wouldn't work and um it was so important that loran got invited to the opening night because he his technology enabled it to work it wouldn't work otherwise and you know the dual infilling made a bunch of bets here of technology that didn't exist before they built the sphere and one of them was this back-end gpu problem that weka solved yeah i mean if you think about it a 4k movie is probably 50 gigabit if you were to download it and i don't know if those are 8k or 4k i'm assuming 4k is enough but how many screens are there like if you were to actually make it into a television screen like is it 100 000 screens 10 000 screens it's some large number yeah huge i think they said like 150 ,000 screens and every second of the show it's over 400 gigabytes and they did some really really impressive stuff if you go there you look the YouTube videos you can see Bono and the edge singing inside of bubbles and if you're watching it there is no lag so there's not even a single frame of lag so it's all going through the through the waka being gpu processed through the waka projected all in 120 frames where no not to spill like yeah i was about to say if you're doing i was going to say 30 or 60 frames but if you're doing 120 frames it's obviously four times as much as that now you're talking about just an impossible task and they're doing some things in real time like you're saying they're stitching in the images of people which is just unbelievable and if you think about pixar which now has like i think maybe half of the all-time uh highest grossing films their biggest problem when steve jobs was starting that company people don't know this was just waiting for the render of toy story and it would you know they would make story changes and they'd have to wait 10 days to have it output and that was you know according to ed catmull their their biggest challenge was actually the rendering of it i mean what is it what would it take a pixar film to render today a full-length film do you think using your platform and everything we're seeing a lot of these things go through and it's unlikely that you're watching something enticing late at night that didn't go at some some stage through a workup so a lot of the big ones are isicus but just just benchmarking it to like make these films now to actually render a film what do you think it takes now to to render so they know a james cameron what was james cameron one that he that was super intense.

28:01So they had many, many dozens of thousands of CPUs. They're now switching to GPUs. They have huge networks. So when you're looking at these impressive AI infrastructure, you know, AI is all the rage. It doesn't matter if you're running AI, if you're creating a new movie, that the data center basically runs of the triangle of compute, network, and data. And now that everything looks so good because compute is four or five orders of magnet faster the network is four or five orders of magnet faster and all of these folks that were when they're taking compute the network to the extremes they also need to do it with the data and this is where weka comes in going back to what antonio said earlier hey quantum computing or neural networks i i went to school in the 90s back then neural networks and quantum computing both failed too far up what has happened with GPUs is basically we won over Moors Law.

29:00So we couldn't create a single chip that has twice the transistors every 18 months. But with advances in networking, we can now create bigger and bigger and bigger computers that are disaggregated and scale. And it took us another 15 years, but this is what we're doing. So the world stops scaling up, we're scaling out. which requires that bandwidth and storage and the sharing of the jobs. You're basically chunking all that data and then managing it. And then in between those GPUs, I mean, we're going to be seeing optics between them soon. Yeah. In some cases, it's already out there. And so that's how large self-driving car data sets or protein folding is getting moved around.

29:46Let's end on this, Antonio. When you're looking at and just sort of zooming back, okay, you got to manage the data loads. So you got vertical applications here. There's another big piece people are starting to talk about, energy and data centers. These data centers need to be at a different scale, Antonio, than what we traditionally thought of, and they need a different energy footprint. Maybe you could speak to what your team and you are looking at there in terms of opportunity for capital investment, opportunity for capital allocation, and just generally what these footprints will look like because it's different, yeah?

30:20Yeah, look, we're going to, you know, 100 megawatt and gatewatt scale plants. And as you know, around the world, everyone's thinking about this problem. And you did that, Jason, we're going to cross to the Middle East talking about this. And our friends in the UAE and Abu Dhabi built a four gatewatt nuclear power plant to power data centers here in the US. You know, we have a couple of companies, and I won't name them all, that are working on renewable energy solutions for the data centers. And I think this is very interesting. you know we have a power problem in the us we were not going to have power to make this work in america and so we have to work both on generation creating more power and actually getting it to get in the grid stable so we can both have power and the power to use and have it where we need it when we need it now weka's part in this is it if the data center gets a lot more productive you need less data centers right right so like three times more productive means two there's less hardware exactly so it basically the return on capital we you know i'm investing miles but roic right so the return capital data set goes up a lot if you are using um your gpu more and very simply put so if you if you invest in things like weka that make the data data set more productive you're also making the energy footprint much lower because you just need less data setters how are people solving the problem today because you know creating energy and nuclear power plants Even in China, you're talking about five, six, seven years.

31:46Maybe they could do it in three or four, but it seems like it's five, six, seven years. So we're not going to have nuclear power in time. So what's going to happen if you have to take out the crystal ball here? Are we going to have data centers being moved to where nuclear reactors are, putting them next to the existing ones and upgrading them? How is this going to get solved? I think there's two kinds of companies in the world. There are those that go slowly and kind of do things straight by the rules and book, and they will take a long time to find power and get it right. And there are those that move fast and figure it out.

32:17And we have a couple of companies in our portfolio that move fast and figure it out. One of them is doing it in a very, I'd say, creative and unique way. The other has been executing a strategy for some time of using stranded energy to build data centers. So we have a company called Caruso that's building a data center now next to a trapped wind farm in Texas that's built because of subsidies. All the power can get used. And they're putting their data center kind of right next to it. Use that power and then using battery backup as well. And they have backup behind that for cogen. But I think we're underestimating in America how much energy we have around in hydro, in wind that actually isn't being properly utilized.

32:59it's feeding old industrial assets can be repurposed which we're seeing in the smartest players and this is i think one of the great things about this country is you you have such a fluid capital allocation strategy sometimes people feel overbuilt we talked about fiber earlier so somebody gets a subsidy overbuilt solar wind whatever hydro and then that energy is going to find a customer eventually and it's actually worth it to ship the gpus to the energy as opposed to building the energy where the GPU is or where you want the GPU is, correct? Yeah, that's correct. Because you can get network built.