Transcript: Brian Venturo on the Battle for Semiconductors

(Bloomberg) -- On this episode of Odd Lots, we speak to Brian Venturo, the CTO of CoreWeave, about  surging demand for semiconductors from data centers, gamers and cryptocurrency miners. You can find the episode here. Transcripts have been lightly edited for clarity.

Joe Weisenthal:
Hello, and welcome to another episode of the Odd Lots podcast. I'm Joe Weisenthal. Unfortunately, my colleague Tracy Alloway, she is off today and this is particularly unfortunate because this is going to be a big one. So, you know on this podcast, obviously we talk about chips a lot, and I think now we've done like eight different episodes on semiconductors and who makes them and the shortages and the challenges of building them and so forth. So listeners will know also that we also do a fair number of episodes about cryptocurrencies. And of course that's for obvious reasons why cryptocurrencies have been going up a lot. And it's a fascinating area. Well, this episode, I actually am going to say it actually combines the two because it's slowly becoming clear that you can't really talk about the semiconductor market right now without talking about cryptocurrencies, because you need chips to mind them.

And when this, when the cryptocurrency market is booming, people want to acquire more chips to mine them. And it's becoming a sort of more significant source of demand. If you look in the video gamer community, there's a huge frustration about their inability to buy various graphics cards from the likes of Nvidia, because an Ethereum miner has snapped them up. Recently, there have been stories out of Asia, China in particular, where people are buying hard drives because they want to mine the new cryptocurrency Chia. Bitcoin is mined using specialized chips, but ultimately that takes a chip production capacity too. So really the stories are interlocked, I think, in a way that I hadn't up until recently fully appreciated. And I would not say that crypto is the main source of reason why there is a shortage or difficulty getting semiconductors, but everything matters at the margin.

And so we are going to explore this nexus today, between what is going on in the world of semiconductors, the fight to acquire semiconductors between different players, because you have the miners, you have the gamers that people who use computers, you have data centers, which of course have tremendous chip needs, and they're all fighting it out for this semiconductor capacity. And it all comes together. And it's a really fascinating facet of all of these different stories brought together.

I'm very excited about this episode. I am going to be speaking with Brian Venturo. He is the CTO and co-founder of  CoreWeave, which is a cloud computing company. And we'll get into the details of specifically what they do. But as a cloud computing company, of course, it specifically has a lot of chip needs itself. And so he's going to walk us through what the world looks like right now. He's also in the Ethereum mining space, sees it from all the different angles. We're going to talk about that. So, Brian, thank you so much for coming on online.

Brian Venturo: 
Thank you so much for having me. This is a topic that I am near and unfortunately, dear to, so really excited to dig into it.

Joe:
So why don't you just get us started? I mean, cloud computing, you know, people think of Amazon, AWS, people think of Microsoft perhaps, but obviously they're not the only players. What is CoreWeave and how is it situated within the cloud computing ecosystem?

Brian:
We typically talk about ourselves as a specialized cloud, right? So we're primarily providing highly scalable, highly paralyzable burst-compute to companies that are in the VFX (video graphics) production and rendering space. So pretty much anything that you see that's episodic on a streaming channel or that's being produced in long form content. That's being touched by a VFX artist somewhere, whether it's on the compositing layer or it's actually on the computer-generated graphics layer, that stuff needs a significant amount of compute. And that's one of our focus areas as a business. So we're there for them. And then in addition to that, it's AI machine learning. So it's model training and serving. There's one additional one that kind of sits between the media and entertainment space and the kind of batch processing space, which is called pixel streaming, which is serving real-time experiences in your browser.

So if you ever go to configure a car online, they don't have 700 million versions of that car render. They actually do it in real time, on a GPU, somewhere in the data center. And then the last piece is kind of the legacy HPC. So it's highly, paralyzable batch processing. It's people who are doing drug discovery, or if they're doing cancer research, kind of everything under the sun. What makes us a little bit different is that we're not really competing for people to host their WordPress blogs. So we're building for, we're building scale compute for large customers of compute. And then we're kind of working with them all on the way to kind of meet their needs. And we do a lot of bespoke builds for people. So we're constantly in the chip market and in the component market just because every industry kind of has their own unique needs.

Joe:
So is there such a thing as a sort of a normal chip market? I mean right now, and we'll get, we're going to get very deep into this. And right now we know that there's incredible amount of demand for chips, for cryptocurrency mining, but, you know, go back, I don't know, 2018, it was a bear market for crypto. There wasn't a lot of enthusiasm at the time. Talk to us about what a sort of normal chip market looks like in what your needs are as a company on a semi-annual basis to acquire chips, such that you can provide cloud services for your customers. Let's start with the normal world first.

Brian:
Yeah. So let's, let's start off by framing the kind of our annual needs. We're probably buying between like, let's call it seven and 10,000 servers a year. Okay. So that's seven, 10,000 servers is going to be built two different ways. One portion of that we're going to build in house, which is typically like a really dense GPU compute. And then the second piece is going to be kind of more generic CPU builds that we buy for ancillary services and other things that may not be GPU focused. And that's typically from regular OEMs, like Supermicro, HP, Dell, whoever that may be.

Joe:
What does the typical market look like?

Brian:
Well, you know, like sitting here today, I would, I would die to go back to like six months ago. The typical market is a place where you can, where things are in stock and you may have trouble finding them, but it may take you a couple hours and you can find the stock that you need and have it delivered the next day.

 It's not really like that right now. I spend most of it's a little while, but half my job here is really scouring the internet to try to find people that may have stuff. Like if I'm looking for a certain item, whether it's a CPU, whether it's RAM, whether it's, uh, a special motherboard that we're looking for, you know, I go through kind of our regular partner channels first, which is some of the largest attributors. And then typically right now I get told there's no chance to get that within 16 weeks. And then I go to eBay. And I don't go to eBay to buy the stuff I go to eBay to say, okay, is there a supply out there? How do I source it? Because it's really hard to kind of connect buyers and sellers in this market because there's so many people that provide services and our distributors and resellers.

So it's really like, okay, like get a feel for what's out there and then go hunt it down. So it's, it's pretty wild. Like last year it would be no problem for me to go out and get 10 terabytes of RAM. Right. And like 64 gig DIMMs RAM and build up 10 terabytes and get it overnighted to me if we needed it for a build. And there was a time late last year where we actually had somebody get on an airplane from California to Chicago because we needed something done in like three hours. That's just not possible today.

Joe:
Wait, did that person have a chip in, or some sort of chip set in there, uh, that they just traveled for? They just got on the plane and brought it to you.

Brian:
Yeah, we needed that. We had to deliver something for a client. There late on a Friday night and we needed, I think it may have been like 20 terabytes of RAM. And we found a supplier in California that was still open that had it in stock and we paid them extra to put them on a plane. So a lot of what we do is very much just in time. Right. So we've done dumb stuff like that before, but if I needed to do that today, like I don't even think I could source the RAM.

Joe:
So again, we'll get to what it's like today, but I want to keep diving further into the normal market. When you talk about GPU's CPU's RAM, talk to us a little bit about the difference between the two GPU and CPU and how they serve different types of applications for your

Brian:
Customers. Let's start with the VFX industry. Okay. So it's really there for two different things. The first is for virtual workstations where artists are actually now fully migrated to the cloud and working in the cloud, on a desktop that, that we've helped them set up. So we'll, you know, we have hundreds and hundreds of artists doing that today. They connect to us via a low-latency application from their laptop or from their desktop at home or whatever it may be. And then they're connected to a hardware accelerated desktop in our cloud. That's one use for the GPU, right? The GPU provides the hardware acceleration of the desktop, as well as the graphics for what they're doing. And then the second piece for GPU compute on the VFX side is for the actual rendering itself for us, we actually spend kind of two different markets in the rendering side.

One is for, for large studios where they have people rendering every day. And then we also operate a direct to consumer render farm for small freelancers and artists that has just over 25,000 users at this point on both sides, though, the demand there is bursty and crazy where, hey, they need to get a shot done. Um, they needed done in the next two hours to meet a deadline. And if they were to do it on their own computer, in their own farm, it may take them 40 days. We really become the only option at that point for some folks. Um, and you know, it's my job here to make sure that I have the GPUs or the scale of GPUs available to be able to handle a large number of concurrent users like that. I'm looking

Joe:
At your website and scrolling down. And one of the thing is, is the GPUs you need when you need them. And you say, we currently have over 45,000 Nvidia GPU. And so when we talk about this battle for chips, between gamers and miners, these, and I don't know if they're precisely the same, but generally speaking these chips that you have, or like the chips that everyone.

Brian:
No, they're not right. So they're very similar in nature,

Joe:
But they're similar, right? So we're an Nvidia cloud services provider partner. Okay. So we're restricted from offering consumer-based GPU's in our cloud. So we're actually, we're required to buy data center GPU's.

Joe:
But from a technological standpoint, we're, they're talking, we're talking about roughly the same type of technology. 

Brian:
You know it, yes. It's like the, the data center GPU's typically have a little bit higher quality. Like they may have error correcting RAM instead of just regular video RAM on them. And then typically the frame buffer. So the actual amount of video memory on the GPU is going to be higher for the data center cards.

Joe:
Got It. But technologically, what people are lining up for in stores outside of video games, supplies, stores, and what you have a deck of 45,000, probably more by now, it's a, it's the same kind of technology. And it's the, the, the artists that the animation artists who probably worked for big studios or television shows on Netflix and so forth, they all want access to roughly that same technological capability.

Brian:
Yeah. It went from having the access to that in their office now that they may not have it staff that's sophisticated enough to set it up for them or their office may no longer exist. Right. And we've seen this huge rush to the cloud. The wild thing is we get on the phone with, with studios and they're like, by the end of the call, they're like, okay, when can we start to be honest with you? The problem that we have as a company is we're actually limited in our ability to onboard, right. Is there's just so much demand for this cloud migration. And then when I look at that from a capacity planning perspective, it's like, you're looking at this environment that we have today, where the world is in this one, big shortage of everything possible, doing the capacity planning and scaling for that, as this is kind of hard

Joe:
You said something, you said burst capacity. And again, I'm looking at your website, it says the GPUs you need when you need them. And so what I take from this is that perhaps there are some, you know, consumer startups or whatever, and they might use AWS cloud and they use a sort of predictable on a day-to-day basis, roughly stable share of cloud capacity. Well, you're offering though, or I guess part of your niche, it sounds like is here's a, someone who doesn't normally need computational power, but when they need crazy computational power, a burst of it, that's not regular. You have huge, you have huge scaling ability

Brian:
For them. Yeah. You know, that, that's the crazy thing. When we hired a director of sales last year, I had known him for like 10 years. And I was like, dude, like every customer we talk to who tells us, they just cannot get the capacity they need at the big clouds. And he was like, yeah, okay. Like, this is a bunch of BS, whatever. Okay. You're being like hyperbolic. And then he joined us in the first like 50 meetings. He did, every one said the exact same thing. So I think that the industry really struggles with that access to scale. And the reason for that is just, I don't like AWS GCP and Azure are building these they're kind of hyperscale data center regions to handle every single use case under the sun. So they've got X number of types of CPU compute.

They've got all different types of optimized compute. And you know, at the end of the day, there's only a certain portion that's available for the GPU stuff. Right. And when that's our entire business, like we can really approach it very differently. You know, what we want people to do is to come in and say, Hey, I need to get 2000 GPU's. I need them for three hours. If they were to do that in AWS, like one it may not even be possible for them, if they don't have a good relationship. And then two, they probably have to do a lot of like manual engineering and reservation planning to get there. We have people coming all the time that they'll burst. They're like two, three, four or 5,000 GPU's for like six hours. And then we won't hear from them again for three weeks. My commitment to them is that I want them to have that capacity without having to talk to me.

Joe:
And I, I guess I kind of intuitively this before, but one thing that's making this clear is that cloud computing, you know, you think of a cloud, like a literal clouds at the cloud up there, but cloud computing is just not a commodity, both in terms of the specific types of applications, but also in, uh, scaling and the, the difference between sort of like standard runtime, consistent demand versus cloud that can be turned on at high scale for a client, uh, at a moment notice

Brian:
That's a really good point. I like to think that our scale helps us break the commodity label, but for anybody who goes in and needs like one virtual machine with like four CPU cores and 16 gigs of RAM, like, all right, go get it. The cheapest place. Right. Same, same. But that's not even the people that we go after talk to. Right. I, I can't help them. I can't help them there. Right.

Joe:
Why is RAN hard to come by?

Brian :
Okay. So let's go, let's go to crypto mining,

Joe:
I guess where I was, I was trying to like, hold off the crypto part for as long as we could, but I knew eventually I would ask a question and we'd stumble into the crypto part. So let's, I guess we're there now.

Brian:
Yeah. So this, a lot of the supply chain stuff here is going to be driven back. It's going to be kind of linked back to crypto.

Joe:
Well, actually then in that case, let me, so before then you mentioned how you would love to go back like six months ago because your job was really easy. So when did you notice that things were starting to change and that sourcing compute, whether it's GPU CPU or RAM that was suddenly starting to get tighter?

Brian:
You know, every February it gets pretty rough because of Chinese new year. Okay. Right. So you kind of rushed to get all your orders in before the end of...

Joe:
This is where it's all being manufactured

Brian:
Before the middle of January. Right. And then if you get it in before the middle of January, they're going to do everything they can to ship it out before the start of their holiday. Um, if you send it by air, you'll have it early February, send it by boat end of February. But like, if you wait, there's nothing that gets processed in February in chunks. Okay. So that was kind of like normal operating conditions and middle of January. And, you know, crypto was kind of doing his thing and it was kind of ripping the roof off and you saw it coming. Right. But a lot of that was, you know, that was really on the consumer side. And so consumer GPU's are really tight. Didn't really impact us. And then starting in March orders that we had placed in January that were supposed to be in stock in the US um, when we're supposed to be delivered the next day, they didn't show up.

Right. And this is like, let's call it medium quality server grade CPU. So like AMD Epyc Rome platform. And we had maybe 500 or 750 of these things on order. And it was like, okay, fine. Like the other components that we need to do, our bills aren't here yet. Like they arrive in early March and we were actually waiting on GPUs at the time and I kind of, wasn't really pressing about it. And then middle of March came and I'm like, where's my stuff. And you reach out and they go, oh, we didn't foresee the semiconductor shortage. I'm like, what are you talking about? He told me these were in stock in vendor warehouse yesterday. Like that it was fine. And this is kind of what you start to see when the market gets stretched. Like this is, you have conversations like that. And people just flat out lie to you in the supply chain.

That was the first real telltale sign. The second one was when I went to buy, I think it was like 800 sticks of 64 gig DDR4,

Joe:
What's DDR4?

Brian:  
It's system RAM.  It's the most recent commodity system RAM, um, there's new versions coming out, but this is really what's used today. Okay. You reach around versus the RAM because a lot of that stuff comes out of liquidation. Right. So you can get it for this particular component for RAM. You typically buy it used, you're not going to go out and spend 60, 70% premium on this stuff. Just because the, you know, that doesn't really wear down. Okay. So I reached out to one of our main suppliers and I said, Hey, man, like I need 800 sticks. And he's like, dude, I got nothing. I'm like, what are you talking about? Like, what do you mean you give enough? And he's like, China.

Joe:
When was it? What, what month? 

Brian: 
March. Middle of March. Yeah. Okay. And he goes, he goes, dude, like I had some guy in China buy all of it last night. I'm like, what do you mean all of it? And he was like, yeah, he bought like 10,000 sticks.

Joe:
Stupid question. Stupid question. I don't know if this is proprietary, 10,000 sticks. Like, what are we talk about from a dollar standpoint,

Brian: 
Uh, two and a half million? Maybe. I think it's maybe two and a half million bucks.

Joe:
Okay. Okay.

Brian:
Right. But 10,00 sticks of RAM is not really a small order. Right. You don't just like buy that and throw it in the closet for rainy day. Right. Well, I guess maybe this year you do. And I said, okay, like what's going on? And then he was like, you know, one buyer came in from China and bought it all. And I, and you know, I kind of give them the, the expletive-laden, oh, you gotta be kidding me a couple days later, he called me and he was like, Hey, I got like 400 sticks. Like, do you want them? And I said, yes, I'll take like everything you have. So that, that was kind of my reaction to that first one. Then a week later I went back to the same guy and I said, Hey, I've got a, I've got a need for like Intel Xeon before, which is chips that probably stopped production like 2019. Like, what do you got? And he's like, dude, like, are you going to kill me? If I tell you that I sold them all to China last night. I'm like, okay. So this is at the point when, when file coin was going from like 20 bucks to 180 bucks.

Joe:
So just for listeners that don't know, and why don't you explain it further? But Filecoin is a new cryptocurrency is actually, I think, first conceived of back in like 2017, it only recently launched, but the idea is essentially cloud files... distributed file storage in the cloud, rather than hosting your documents with a company like Dropbox or box, you host it on a network of a decentralized blockchain network of computers all around the world.

Brian:
Yeah. And all around the world. There is a little bit disingenuous at this point. I'd say it's 95% centralized in China. Okay. But the ecosystem in China is they are insane for this project. And the demand for this project spans, uh, it spans every component in the supply, right? So you, you need GPS for the actual computation of proofs. You need CPU's because you need high, high core clocks for some of the processing they need RAM because, uh, because of the storage operations and actually putting stuff in cache. So they're kind of buying across the board and got it. Like if you, if you search real quickly, you'll find mentions in the news of people putting $1.3 billion into FileCoin mining. And I mean, there's real money over there. So this FileCoin driven demand has kind of sucked up beyond what the typical component demand in crypto bubble or a crypto bubble or crypto expansionary times it's called that instead of a bubble is going to drive, right?

Brian:
So it went from, you know, 2018 or late 2017, early 2018, where the demand was only for GPU's to now you have Ethereum driving demand for GPUs, you have FileCoin driving demand for everything under the sun. And the amounts of capital that are being thrown at us. And China is just draining the supply chain across the world. Wow. Right. So that was really when I was like, okay, like my life just changed. And we have a risk here where we may actually be governed in our ability to grow by access to the components.

Joe:
Yeah. I just want to point out, like, I'm looking at a price chart right now of file coin the coin, even as recently in January, like say January 1st one coin one FileCoin was trading at $22 fast forward to March 31st a Filecoin. According to this chart, I'm looking at was at 192. So that's almost like a 10 bagger and the price of FileCoin over the span of about three months, which really, it seems to me then just makes the demand to acquire all of these chips. Even that much more. If, if acquiring one coin is worth 10 times more than what you're going to be willing to pay for the compute to mine a file coin. That just went up a lot.

Brian:
Yeah. You know, and one of the things that is unique or weird about five point is the staking requirements, right? So, you know, a lot of the cost of actually building the storage isn't in the storage profile point, it's the, the amount of, of coin. You actually have to stake to be able to provide your storage to the network. Oh. And at the time, like when it was like 190 bucks, like we had a lot of internal conversations about this, like, why aren't we going crazy on this? And why aren't we doing this for us to provide a one petabyte of storage, which isn't a ton of storage to the network. We would have had to stake a million dollars where the FileCoin. Okay. Right. It costs me maybe like a hundred K then build one petabyte of storage. So 10 times the hardware value I have to put up in staking risk and a coin that just went up like 10X, I think at the time, like it was this positive feedback loop where people were going into it because they want the price is going up, but to get into it, they had to buy more.

Joe:
You buy, wait, so you buy hardware to mine, but then to actually be able to plug your hardware into the network and actually, you know, use it, you also have to buy the coin itself. So as you say, there's like this like weird, like positive feedback loop that emerges where you make this capital investment in semi and chips, but then to use the chip to acquire coin, but then to actually make the chips worthwhile, you have to buy the coin itself. And so that you get this like crazy.

Brian:
Yeah. You know, it's like, I'm not going to throw the P word out there, but in terms of schemes, it's a pretty good one. It's a project that I think that internally, like, we're, we still look at it as, you know, it fits our infrastructure really well. But right now it's just so hard to get into it because you have to really have excess capacity for it and getting access to that excess capacity. It's just so hard, like kind of bringing us back to the broader conversation here.

Joe:
Right. And now another, now I want to get into some of the other sort of coins that are driving this dynamic. You and I had a chat like on the phone, like two or three weeks ago before we did the podcast and now, and I hadn't like sort of seen any articles about it, but then there's this other coin Chia that's in the mix. So what's happening with that?

Brian:
Yeah. So that one is, it's pretty crazy too. Okay. This is a, I guess a different take on proof of work. Right. So I guess they're calling it proof of space, time. Yeah. And effectively what you do is you use CPU compute to create what they call plots. And then once a plot is created, it's like a, a couple hundred gigabyte space on a hard drive. You can move that. You can move that to cold storage. And then for up to five years, you can use that as a, as a mining mechanism to gain, to gain rewards on the network. That's another one that I didn't really see coming. And we build large like multi-petabyte level storage clusters in each of our data centers. And they're, they're SEF based. So it's a software defined storage technology. And one of the things that we do there is we, we build them to have triple replication, right?

Brian:
So for every one petabyte of storage usable for my clients, I have to have three petabytes like on the ground that installed that makes me a pretty big buyer of hard drives. Yeah. This is another piece that kind of showed up in, in January that I didn't really understand yet is we were, we were trying to get access to, I guess, uh, a quasi non-public skew. And it's one that really meant for data centers. It would reduce our latency and our storage access patterns. So we went through this process of kind of qualifying with the, with the manufacturer and they were asking me all these questions, like, what are they being used for? But, and I'm like, dude, they're hard drives, what do you want? Like are you selling them to me or not like, if you want to increase the price, increase the price, I'll buy don't care.

And it wasn't until March when like this, when Chia launched that, I was like, oh, like these guys had a run on their stock and they weren't able to provide like product to their real customers. So the demand for Chia for hard drives with Chia it's two-fold right. So the first is you really need like high-speed NBME or discs that are local to your CPU compute to run the plotting. And then once the plotting is done, you can push it off to whatever cold storage device, whether it's like spinning RUST, like hard drives or whatever it may be. But then it just sits there. And the access pattern is pretty low in that data. So this is another thing where I have conversations say, Hey, like, what do you got gotten in like says like 7,200 4k on hard drives. And they're like, would you believe that somebody bought all of these last night? I didn't expect to be getting dragged into chia. And I was only really dragged into it by the problems I was having acquiring. I was having acquiring things. And like, you know, every once in a while, like the bells go off, like, yo idiot, go look at this. Like, what is this thing? Yeah. So now we, we've got a pretty large chia mining operation.

Joe:
So this is something that I want to get into. And it's actually pretty interesting in the way you're set up core we've prior to like, you know, all this like crypto going nuts over the last, over the last several months. So particularly it seems like as soon as January and March, when all of these altcoins just started going bananas, your clients were people who actually had what most normal humans considered to be real world needs. So like animating a TV show or animating a movie, they're actually making things actually actually making things. And so you are, you'd go out into the market and buy chips and it wasn't that hard. Now you're also involved in the mining, but you do it in such a way that it doesn't, it compliments the service you provide to your clients as opposed to competes with them, I guess. So can you explain a little bit about how you use cryptocurrency mining yourself at core we've to, uh, compliment your business, uh, compliment your existing

Brian:
Business model? The permissionless revenue from cryptocurrency mining has really been, what's allowed us to build to the scale where we can provide these bursts compute services, right? Like without crypto, especially over the last like six months, we would have had to raise significant, significant amounts of money to do, to run this business. Um, I think that like life to date, we've probably sold, like, we may maybe raised like six or 7 million bucks in this whole thing. Yeah. Every time that we're, we're looking for capital or capital, we have capital requirements. Like we're always trying to protect ourselves like as founders, like you don't want to dilute it, especially you don't want to dilute to buy a depreciating asset. So we've been able to get pretty creative. And some of that spend from the crypto side, some of it's been kind of from, from just founder and partner commitment to fund things internally, what crypto does for us, like say that our clusters at a steady state or steady state, and we've got, let's call it 20,000 GPS in that cluster available for clients. Okay. If we only have like 3000 and GPU's and base-load demand in that cluster, right? What are we doing with the other 17,000? Right. And if I don't have to talk to a client and I don't have to do anything besides start a minor on those things and I can make 60% of the revenue I'd make from my cloud clients. Like that's, it's pretty good. So we, we mine on everything, everything possible when we don't have higher end workloads running.

Joe:
So this is, this is actually, this part is super fascinating to me. So if you want to have the ability to offer your clients, burst compute, then by definition, at any given time, you have to have significantly more capacity than is your sort of like normal base load and the way then from your perspective, to make the investment in that capacity worthwhile so that you can go out and buy chips that you only expect to offer to your clients that may be, you know, normal state, 30% capacity utilization is if during those times the other 70% are earning revenue for you from cryptocurrency mining?

Brian:
Yeah. You know, I think that the payback calculation is a little bit different. Sure. That's 70% with crypto mining is what allows me to aggressively expand my business. The 30% that pays the bills, right. It allows me to kind of operate, but the grand plans over here is really to kind of just continuously grow this, this compute fleet. You know, we have customers that come in all the time that are, I can't get what I need. I can't get this. I can't get that. Like, Hey, can you promise me that I can get 150 GPS? I'm like, dude, it's fine. A hundred or whatever. You don't have to talk to me.

Joe:
So another way to think about it is like, if all these people are playing the game of buying all different kinds of chips to mine, you're essentially forced to play the game too.

Brian:
Yeah. You know, we try to operate in different segments. So typically when mining, Ethereum mining rigs are being built, they're being built for the lowest cost possible. Right. Right. So the GPS are really the only expensive thing in the setup. Like the host device may have, whether it's somebody building it out of like an egg carton burden, like an eggshell carton from, in their garage, or putting it on like an open-air rack or putting it on a wood frame, or actually putting it in like a, in a closed chassies with forced air flow, the motherboard there is going to be super cheap. It's going to have a really like low end CPU. It's going to have like four, eight gigs of RAM that these aren't the types of systems that we're, um, competing to build. Right. Which is why it's without the last five months here, like it's always been kind of reasonable.

Brian:
Right. I never had to compete for high-end CPU's before and never had to compete for server chasses before and never had to compete for like network cards and stuff. And this is where FileCoin really throws things into the, into a loop. But on the mining side, you know, everybody talks about how miners and gamers are opposed to each other. And, you know, I think there's been some vilification of some of the larger mining operations around the world that they're taking all these GPU's. And I'll tell you that even on our mining stuff, when the Nvidia 30, 80 came out, we looked at it and was like, oh my God, this thing's a three slot. Like we can't like, we can't deploy that.

Joe: 
What did that mean at three slots?

Brian: 
So it takes up three slots on a motherboard, right. So it's just different spacing than what they had done previously.

And typically it would kind of the, the mining GPS and mining variants, they had released in the past where we're dual slot. So it would take up two slots on the motherboard, then these crazy things that were three. And they were like, okay, like we can't deploy that at scale. And our focus at the time was just so different that we didn't even buy any, there are some distributors in the U S that show like real-time availability and real time, how many back orders they have in their queue and you see them. And they have, like, we expect 500 Nvidia RTX, 30 eighties on September 1st. And like, pre-orders like 47,000. I'll tell you like the 47,000 pre-orders, I don't think that's sophisticated mining operations. I think that's like, like you and me personally going and doing it and putting it in our garage.

Joe:
Where do gamers fit into the mix? Because I'm in like a discord group. And I don't really understand half the jargon in it, but it's basically a group for gamers who are looking to acquire Nvidia GPU's cause they just want state-of-the-art graphics for their systems. How do they even stand a chance in this world? And how does, uh, how does the industry think about sort of making sure that they can get access to this technology?

Brian:
There's two steps here that there's two pieces here that are really a problem.

Joe:
Okay.

Brian:
Yes. The mining demand is a problem. Um, but the mining demand is a problem because the economics warrant it. And you know, one of the crazy things about proof of work cryptocurrencies like Bitcoin and Ethereum in its current life is that the early on mining component of it is like the greatest user acquisition strategy in the history of the world. Right? It's like, Hey, you have a computer, your computer may already have this device. If you turn it on, we're going to pay you $10 a day. And that $10 is now denominated in Ethereum. And now you say, okay, I've got all this Ethereum, what do I do with it? And now you, as a user are now researching everything in Ethereum because you now have this financial asset.

The hundreds and hundreds of thousands of people that have gotten into cryptocurrency from mining it's crazy. And I think it's one of the things that's like super underlooked or super overlooked, excuse me, the mining economics definitely drive a lot of this. Right. And video's taking steps to put their hash rate limiters on and to, to kind of enforce that in a certain, in such a way for the Ethereum mining to say anybody, who's going to buy this stuff at real size, like by the crypto mining focus GPUs.

To me, the bigger problem is really the scalpers. It's the guys who are running bots on Amazon and they're buying the GPUs as soon as they list. And they're selling them on eBay, right. The supply isn't making it to the end user. And I think that that supply is probably artificially overpriced right now. And you know, the crazy thing is that the Ethereum mining economics kind of warranted. So maybe it's not artificial, but for the gamers, they're going, what the heck? Like this thing's supposed to be 600 bucks and it's 2,600 bucks. But between like retail and the gamers actually getting it, it's almost, there's two huge obstacles for them. I don't see a way to solve that.

Joe :
I guess if the cryptocurrency, I mean the crypto market crashes every few years or booms every few years

Brian:
Hey, it's different this time, man. Come on.

Joe:
But in theory, it could crash. From your view, would that ease up capacity for all these things that we're talking about?

Brian:
Probably not. I think what you would see is you'd see a lot of GPUs that come back to market as kind of like distressed assets. Okay. Right. But I don't think that those are going to go in large blocks. They're not going to want to be piecemealed out to end users. And you know, there's some services and some companies in the U.S. that actually do specifically that. They buy, like used crypto mines and they piecemeal them out, or the guest one GPU lots on eBay. Yeah.

Joe:
But what I mean is, setting aside the selling of used, I mean, new chips of all of these different stripes are always being manufactured. So theoretically like Nvidia and dozens of other companies that make hard drives and so forth in 2022, if let's say we're in a crypto bear market, does that free up capacity or does that free up...

Brian:
I don't think so. I think this is where you have to look at the structural change that we're seeing in the compute market. Okay. Everything is more compute intensive. Everyone's going to the cloud. Data center demand for GPU is just exploding. Okay. Even if crypto were to disappear. The demand, the growth pattern of this market, as well as kind of home gaming and everything like the global foundry space for semiconductors is just too small for all the industries that are relying or reliant upon it. Ford's going to have to cancel how, like, not deliver how many trucks this year because of semiconductor shortages? Wild. And it's impacting everything. And when you get kind of in that structural dislocation like this, like the whole, the saying in commodity trading is like high, high prices cure high prices. The timeline to cure these high prices may be several years.

Joe:
Interesting. How many people are essentially in this game of, I mean, it's interesting. Like I can go on StockX, like the sneaker website and they actually have a few categories of chips and you can buy Nvidia chips on there and they sell for like 300% above MSRP. And you could see the price chart, how many different players, whether they're sort of large brokers or individual scalpers have essentially like gotten in this game of chip price arbitrage, and trying to get themselves in the middle, between in the middle, between the chip manufacturers or the chip distributors, or somewhere between Amazon and eBay and gotten into this game of sort of like chippery selling as if it's a commodity or a stock.

Brian:
Yeah. You know, it's not just the, it's not just the  scalpers though either. What's unique about the GPU supply chain with Nvidia in particular is Nvidia that they make the GPU chips and they actually, they sell them to, I guess they call them add-in board (AIB) partners. Right. And this is people like it's Gigabyte, it's MSI, EVGA, it's Zotac, which is like a PC partner to brand. And they actually take those boards. They fabricate them, they put the cooling devices on them and they sell them to end users. And, you know, I think that the ultimate pricing on that stuff is really controlled by the AIBs. So this is like the kind of fabricator or distributor at the end of the line there. And they're no dummies. Right. They see crypto demand, they see where it is.

And price definitely changes with where crypto is, right?They know that they have a limited supply. It's virtually unlimited demand. So it's like, this is really just supply and demand. And it's mostly laissez faire state. And, you know, I think that one of the things that the U.S. has, like one of the problems in the west here, is that I don't think a lot of this stuff even gets out of China.

Joe:
What do you mean by that?

Brian:
Oh, if you know a guy at the AIB in Schengzn and you're local and you want to come in and pick up a thousand GPUs, that's a much easier transaction. And like, in terms of like political capital, just on a personal basis, like, I think it's a lot easier for them to do than it is for them to ship something to Brian Venturo in the United States.

Joe:
Wrapping up here a little bit, you know, you've explained how your life kind of has changed in the last six months in terms of the difficulty of acquiring compute, whether it's GPUs, CPUs, RAM and so forth, what is it like now on a day-to-day basis and how have you had to change your business strategy so that you can grow your cloud capacity alongside customer demands?

Brian:
You know, it used to be that GPUs were the problem for us to get. And GPUs right now, like the data center card is like, I've got shelves and shelves and shelves packed with data center cards, waiting for builds. And they're all blocked by CPU and RAM. CPUs in particular. I had a conversation with my CEO, I've been with for 17 years now in kind of different businesses. And I called them. I said, hey, I just found a bunch of AMD, ABIC, Milan CPUs in stock, ready to ship tomorrow. I've bought a hundred of them. And he was like, well, how many do they have? And I was like, I don't know, like 250. He's like, why didn't you buy the rest? And I'm like, well, I didn't really want to spend the money, I didn't have a place to put them right now. He's like, dude, this is literally all you complain about every day is not getting access to stuff, go back and buy the rest. So I went back to buy the rest. And in that 15 minute period, the rest had gone.

Joe:
So, you know, this is a theme. It's not just chips, obviously. And we talk about this with lumber. We talk about this with the industrial commodities, this idea that you buy more when there's a sort of sense of scarcity. And so I'm curious about like how much of the tightness of the chip market is essentially, it's like, well, I don't know when the next shipment is going to be available. I don't know what I'm going to be able to buy these AMD chips at the size. I don't know next time I'm going to buy RAM in size. So how much does that have the effect of causing people to then over order so that they can at least get some assurance of the chips that they're going to need, which then further exacerbates the shortage causing more over-ordering.

Brian:
 I can answer for us. Like I mentioned earlier, that we're a pretty just in time operation on the inventory side. So what I'm talking about, if I go out and buy like 250 CPUs that gets me like one month of supply. You know the CPUs that I did buy that were delivered the next day, which was even more of a surprise, like they're already in systems are already online. And that this is a couple of weeks ago. When I'm talking about buying things and stockpiling them, like my usage timeframe is much that there's no, like, there's no speculation on component values right here. It's like, hey, we're going to need them in the next 30 days. I typically wouldn't buy it until I needed it in the next seven. But here we are and the world is on fire. 

Joe:
So six months ago you might not have made an order until seven days before you needed to build them out. But that is now lengthened out to more closer to 30 days.

Brian:
Yeah. And on the OEM side. So we've got a bunch of stuff that stuff that we're waiting on deliveries for right now, we've been talking about it probably like high tens, maybe 15 to 20 million in CapEx. That's out there kind of being fabricated. And this is with OEMs and you know, when we were doing the deals, it was, hey, we're committing to do this within four weeks. And now it's like out to 12 weeks and you go through the conversation with them and you're like, all right, like I get it. Like, I know what's going on to you is taking me to like understand problem. All ability to plan in real time is like completely out the window.

Joe:
Huh? Wow. Yeah. That's the thing. I mean, that seems to be the theme of so many different industries, whether we're talking about shipping logistics, trucking, lumber and so forth. Planning ability has kind of gotten blown up.

Brian:
Yeah. Yeah. I mean, this shipping thing is crazy too, because like stuff that I would normally put on a boat and I'm like, put it on an airplane and the boat may sink, like may get stuck. I have no idea what's going to happen. Like get me my stuff.

Joe:
I just want to go back to one question. Cause I, I still feel, felt a little bit unsatisfied by the answer. There let's say cryptocurrency disappeared. I don't know how that would happen. Maybe the prices crash, maybe somehow there's a coordinated national effort to ban it. Or maybe financial institutions are banned from dealing with it. Somehow that industry, which is extremely big and growing and worth like $2 trillion in market cap. Let's just say somehow it gets obliterated.

Brian:
We can put a real scenario with this, let’s say that ETH devs actually shift proof of stake. Right. That's the biggest question that I get from investors is what about proof of stake?

Joe:
And so proof of stake for people who don't know is basically currently Ethereum operates with the same security model as Bitcoin mining more or less, with a sort of high chip and electrical spend in order to secure the network and acquire coins. They want to change the model to something where there's less mining.

Brian:
I'm not going to talk about the probability of that happening. Right. I mean, I think they've shown like, well, I'm going to give one comment on it. Um, you know, starting in 2016, proof of stake was shipping every three months. Yeah. And here we are in the middle of 2021. And like, do I think they're eventually going to get there? Yes I do. But like look at the market, look at how much money has been spent on mining hardware and building mining operations. Like the miners are not going down without a fight. And without a fight could be launching new coins. It could be contentious forks. It can be anything. But you know, this idea of GPU mining is going to die, it's been going on since the advent of GPU mining, right? Like when Bitcoin went from GPU, GPU mining was dead. And then like six months later, Litecoin showed up. Right.

So there's always going to be this, this demand for security driven by GPU compute. I think it goes back to that kind of like fair launches and the idea of getting coins out to the peak ‘to the people.’ Is that as soon as, as soon as the theory goes, proof of stake, like that's going to be rewarding the incumbents more so than new entrance, right. You're not going to have a hundred thousand people go plug in a GPU and learn what a theory is. I just think that the idea that when Ethereum goes proof of state that GPU mining dies is a little bit crazy. We don't necessarily know what it's going to be yet, but I'm pretty confident that there's going to be something. Now in a price crash environment, you know, this is really where we built ourwhole business is taking the contrarian bet in the binary outcome environment, which was really like January, like December of 2018, January of 2019. Right.

It was like, I don't know, but crypto is either going to zero or it's not like, I think that in, in the mining space, like where you make real money is when you're in a position to take that contrarian distressed asset bet. We look at that as a company is like, yeah, it would really suck if crypto went to zero again, but look how awesome it worked last time. And I don't think we're the only ones that think that way.

Joe:
Bottom line still, if I'm a gamer and crypto goes to 10. Crypto crashes 90%. Does it get easier for me to acquire GPU so that I could play a game at a higher quality,

Brian:
We'll get easier for you to acquire use GPUs to play games, but I don't necessarily know if you're going to want them.

Joe:
Why not? Why not the new ones?

Brian:
Because I mean, Nvidia is one of the most valuable and forward-thinking companies in the world for a reason, right. Then they understand market demand. And if they see crypto crash, like they know that a portion of their demand is from crypto, right. They're launching the specific lines. They're going to book less foundry space. 


Joe:
Why not just shift some of that foundry space over to gamers though?

Brian:
Well, I mean, before we already concluded that it's the same, it's the same chip sets. Right? And they're just being packaged differently. Right. So, and one of the things that people don't necessarily understand is that, you know, in a lot of the cards, whether it's like a 30, 60, or 3060 TEI or a 3070, that that may be the same chip set with just features disabled. Or maybe it’s been packed differently because it has like, the quality of the run was worse for some portion of it. Nvidia is not making all these decisions like, Hey, let's go build like 50 million, 3060s. Right. They've got their, their kind of foundry space and what that may be booked and they know what their overall is going to be. And then it's, they're just kind of optimizing what they bring to market. Right. It's just in a down crypto year, I think that what they did with the touring line where there the RTX 2000 series, I think that the production of that line was way lower than the GTX 1000 series. Right. And that was kind of coming out of crypto winter/during crypto winter. And I don't think those GPUs were hard to find got it. But you couldn't, but you also couldn't buy 50,000. 

Joe:
Got it. Well, this was a fascinating discussion. Like I said, we've talked a lot about chips in the past, but we really haven't gotten into the actual fight to acquire them. Plenty more to dive into on this topic alone, I feel like, but Brian, really appreciate you coming on Odd Lots!

Brian :
Thanks for having me. I'd love to do it again sometime.

Joe:
Yeah, for sure. Take care Brian.

Joe:
Well, I would love to be able to talk about that episode with Tracy. Unfortunately, she's not here, but I thought that was completely fascinating. You know, this idea, so many parallels between everything else we're seeing, like the inability to plan the requirement to buy further out into the future, this idea that sort of just in time isn't working. And it really does seem like the rise of these new coins, FileCoin, Chia.

Like every sort of like commoditized, technological capacity that you can think of somebody, it feels like is making a coin that can mine it. And it's fascinating. And I thought it was fascinating that core weaves owned business model is kind of enabled, like on one hand, they're competing with all these miners for chips, but on the other hand, their ability to use idle chips in order to mind and get revenue when they're not being used for the normal, uh, computing by their clients is super interesting.

©2021 Bloomberg L.P.