Hey everyone, I'm Bryan Cantrill, co-founder and CTO of Oxide. Great to be with you this morning. Thanks for joining us. It's a fun space to be in because we're actually making real stuff, right? And as Steve outlined, we took a real, in this case literal, clean sheet of paper. This was not a napkin, but it was damn close. This is an actual sheet of paper. And when we started the company, we had been confined by the state of the industry, which had become commoditized, had become ossified, with different companies all delivering these different layers of the stack. And when you tried to get them to work together, you, the customer, were often the one who was performing that integration, and you've got fingers pointing at one another. I mean, and honestly, in many cases, you would have vendors blaming one another, and the actual uncomfortable truth is they're often both right.
They are often actually the problem is not crisply with one vendor or another. It's actually at that integration point, and where it's like, actually, your behavior is reasonable taken on its own, and you other vendor, your behavior is actually reasonable taken on its own. If you need me as your customer to tell you that. But when I combine them, I have a pathological, unusable system, and you are treating me like it's my fault. And you know, you're the only customer seeing this.
Yeah. And the number of times that, you know, if you've been a large customer, it's extremely frustrating to hear. And you will never hear that out of Oxide. You will hear that we have never seen this before, though. But we want to, because it feels like it's an excuse. You're the only customer seeing this. Feels like an excuse to not investigate your problem. We started with a clean sheet of paper, and we knew that with a clean sheet of paper we were going to be solving a bunch of problems. All of these problems had accrued over decades in this industry, going back literally to the dawn of the PC, because we were trying to run personal computers in our data centers. We, enterprise users. Hyperscalers weren't doing that. Hyperscalers had long since figured out that this doesn't work. They had designed their own machines. Their own machines, fit for a purpose.
They kind of, befitting each of them, they kind of acted with their own idiosyncrasies. Google built their own and then wrote a book explaining how awesome it was, the warehouse-scale computer. So you can read a book and appreciate Google's brilliance. You're like, 'Do I get to buy one?' He's like, 'No, no, no, no. But look, you can admire my great work here.' You're like, 'Okay, Google. Yes, it's great. It's amazing. Terrific.' Amazon, you could rent it, right? Amazon was very closed-lipped about what they'd done, but you could rent it. You actually could not buy it. Facebook, now Meta, interestingly enough, kind of did the most interesting motion where the technologists inside of Facebook said, you know, this is kind of not right that we developed this great innovation that the world doesn't actually have, and they started the Open Compute Project as an initiative to share those innovations. So you could actually go, as Steve and I did, to OCP Summit as we were contemplating Oxide. We went to OCP Summit and you look at these machines and you're like, this doesn't look anything like the machines that I have to run these commodity servers. This looks like, why don't I get to have these? And it's really basic. It starts, but certainly does not end with, the physicality of the system. The first thing you notice on these systems is they've got a DC bus bar-based design. So you've got a power shelf that rectifies from AC to DC, and you run direct current, DC, up and down the spine of the machine, and those sleds blind-mate into that direct current, which means there's no AC power supply per 1U, 2U. And once you see this, you can't unsee it.
Once you are aware of this and aware that literally every hyperscaler—Microsoft, Meta, Google, Amazon—all run with a DC bus bar-based design. This is not like new thinking. This comes out of the telco world. This happened for a long time. Anyone running at scale runs with a DC-based bus bar. Not enterprise computing. Enterprise computing has got AC power supplies per 1U, per 2U. It's cords everywhere. It's fans everywhere. It's heat everywhere. It's inefficiency. And that is the beginning. That is the most visceral thing we were seeking to disrupt. But we knew we were going to be—when you take a clean sheet of paper, when you take this thing, there are a whole bunch of things that we knew we would be able to go do. Yes, be able to do what the hyperscalers have done. We also knew we were going to be able to leapfrog the hyperscalers and develop yet new differentiators. Give you a very concrete example of that. We assumed—you looked at the Open Compute Project, you looked at like that Tioga Pass circa 2019, DC power in the back, cabling, networking cables out the front. We assumed we were going to do the same thing. You look at those pictures from, you know, Google when they're flexing on you, they're providing you those pictures of their DCs, and it's all like lovely pastel-colored cables, but it's cabling coming out the front. That's what we assumed we were going to go do. But we had the luxury of a clean sheet of paper. With that clean sheet of paper, we actually had more latitude. We were actually talking to a connectivity vendor, and they're like, 'Out of curiosity, why are you guys doing the cabling out the front?' We're like, 'Because everyone does it that way. I mean, the hyperscalers do it that way.' It's like, 'Yeah, the hyperscalers do it that way, but not if they were taking a clean sheet of paper today. They wouldn't. What would they do? They would blind-mate into a cabled backplane. Cable the backplane, blind-mate the network connector.' And we're like, 'Can we do that?' Like, 'Oh, absolutely you can do that.' 'Why don't they do it?' 'Well, they're afraid to do it.' You know, and that of course is like catnip. Like, 'Oh, they're afraid to do it. Okay, we've got to go do that then. That's great.' That cabled backplane was a bet-the-company moment. It has removed the cabling from the system. You take a sled, there is no cabling in that sled. The backplane is cabled, that sled goes in, locks in, no cabling. Because where there's cabling, there's miscabling.
That was again, it was the beginning of another bet-the-company moment. There is no BIOS in the system. BIOS, basic input output system. This is an idea, all caps, must be screamed. This comes from CP/M. CP/M predates DOS. DOS is from my childhood. As my children are quick to remind me, I'm old. This has been around for a really, really long time. The BIOS is this lowest layer of platform enablement that initializes the computer before the actual operating system runs. Everybody on x86, and sadly many people in ARM, run with a BIOS, with a UEFI BIOS. We knew that that thing was a huge problem. Wanted to rip it out. Wanted to do our own low-level platform enablement. This was something that AMD thought to be impossible. Why? Because Google had tried and failed. Google has tried and failed. It must be tautologically impossible. And it was really, really hard, and it was another bet-the-company moment. We got it working. We don't have a BIOS in our system. We boot holistically. First instruction is our instruction. We do all that low-level platform enablement and initialization and bring up the entire stack in front of it, on top of it, including all the distributed system, the control plane on top of it.
There are a lot of these. When we had a clean sheet of paper, we had a lot of opportunity to do things that people hadn't done. For each of the things that we did, it was not hard to find technologists to say, 'Yeah, of course, this is the way you would do it if you got a clean sheet of paper, but you never get a clean sheet of paper.' We got a clean sheet of paper. We didn't want to spoil it. What this meant is it's a little hard to be crisp about the differentiator of Oxide. What is the differentiator of Oxide? It's like, well, it's complicated. And we had a board member, the great Pierre Lamond. You don't know Pierre. Pierre is a truly singular individual. Pierre was at Fairchild back in the day. Pierre hired Andy Grove at Fairchild. So, yes, Pierre just celebrated his 95th birthday. We are the last company that Pierre helped found. Pierre was on our founding board. I would like to say that we are the last of the Fairchildren. If you know this idea in the '60s and '70s of companies that were spun out of Fairchild, Oxide is the last of the Fairchildren. Pierre's great. Pierre is infamously tough. And Pierre says to me—we had many direct conversations—Pierre says to me, 'What is the differentiator for Oxide?' Like, well, it's complicated. No, that's actually what I wanted to say. In an earlier and maybe the first or second board meeting, Pierre had banned my use of the word complicated. Didn't want to hear it's complicated. So, I had to use all these euphemisms for like, well, it's nuanced, it's multifaceted. I'm like, well, here, I know this is not what you want to hear, but there's not one differentiator. There are a bunch of differentiators, and it depends on what your biggest problem is. We have absolutely differentiated on power and efficiency. We've differentiated on reliability, on operability for sure. We've differentiated, just as Steve was talking about, about the time to deployment. That was a lodestar for us. We want that rack to arrive. We want it to come out of the crate—that importantly, sleds ship in the crate. We want that thing to be spun up and have developers on that thing in hours, not days, weeks, months. So there are many different differentiators, and that's what you get when you get a clean sheet of paper. That's the good news, right? Well, okay. Well, what vertical are you targeting? It's complicated. It's a general-purpose machine. There are a bunch of different verticals for whom this is really valuable. So part of what we honestly did not know when we started the company—we knew we were building it the way we wanted it to be built. We had experience deploying at scale at Joyent and then very much at Samsung. We knew these were pain points that existed, but you don't necessarily know the specifics of like who's going to grab whom. And when you—it took us four years to build this thing, three and a half. It took us—so I really should stick to three and a half and not four because—so we took our first capital in 2019. We shipped in June or July. June is when we shipped. July is when it arrived, of 2023, excuse me. So it took about three and a half years. The reason the three and a half years is important is because when we first started the company, we said it was going to take two to three years to build. And Pierre, quite famously, when we were originally pitching him, said, 'Which is it? Two or three?' And I'm like, eyeball to eyeball with him, and I'm like, 'Two to three.' It's like, 'Which is it?' He's like—okay, this is like—anyway, I love Pierre. It's great. Pierre asks very direct questions and gets ambiguous answers from me and manages to not—you know—manages to still stay in the room, which is great. So it took us three and a half years to build. When it takes you something three and a half years to build, the world can change a lot from underneath you. And we knew that on-prem compute was not going to go away. We felt very confident that the demand for running elastic infrastructure on-prem—we knew that that was not going to go away. We knew that was going to grow over time, and that has grown over time. It's a little bit amusing to me when venture capitalists that passed on us send me these like public reports showing the growth of on-prem usage. Like, why are you sending this to me? I should be sending this to you. Like, you passed on us because you didn't think this market existed. We knew the market existed. Absolutely knew the market existed and absolutely knew that there would be increase. You know, there was a question earlier in Steve's session about sovereignty, data sovereignty. We knew that that trend was going to continue. We knew that there were—and especially knew that there were economic drivers, and the economic driver is still like the driver, I think. I mean, if you were to really—if Pierre were to really get me down, like, what is the differentiator? Ultimately it has to be around doing more with less. But there's a lot of different ways in which that happens. So we didn't know what was going to happen necessarily. We knew that there was going to be demand for this, but it's like you just don't know. You just don't know what's going to happen. When we first started the company, one of the earliest advisers of the company is like, 'You've got to run ESX on this thing.' It's like, bad news. That's not going to happen. We're not running ESX on this thing. Let us explain to you what we're doing. We are doing our own hypervisor, our own control plane, our own distributed system to provision the primitives of our system, to provision compute, storage, networking. So we are not going to run ESX on this thing. He's like, 'I think you're making a huge mistake.' Well, needless to say, a couple years down the road, he's like, 'You guys are geniuses.' And like, you were so—like, 'Did you know this was going to happen?' Like, well, no, not exactly. Didn't really predict the acquisition of VMware by Avago. I insist on calling them Avago because Avago has taken on the skin of Broadcom to give themselves like legitimacy. Sorry, this is private equity. The ticker is still Avago. I'm going to call them Avago. So, they're just a point of principle, I guess. But no, we did not predict that. We did not—but with that happening, the amount—the customer pain is definitely predictable. We've gotten a lot of people who are all of a sudden newly interested in Oxide because of that. But we didn't know exactly what was going to unfold. We also knew that there would probably be something—just like when you take a clean sheet of paper, you know that there's probably something that you're not considering, like that cabled backplane, an opportunity that you're going to have that no one else is going to have because you took a clean sheet of paper. And so we know like there's probably going to be a market that we can't really predict right now that's going to be a huge tailwind. If you look at successful companies, they have this. I mean, famously with Nutanix, this is VDI, right? I was at Sun back in the day. I was at Sun for kind of exactly the latter half of its life, was there from '96 and I went down with the ship in 2010. But if you look at Sun in the '80s, EDA, right? EDA was explosive for those workstation companies. So there was a market that they didn't necessarily predict, but it was a huge tailwind. So, we thought there might be one of those markets, but I'll be candid with you that I'm surprised a little bit about what that market is and the differentiator that they really care about. So, to give you a little bit of context for this, when we were first announcing the Series A, we were working, you know, with a PR firm, as one does, and they're talking to reporters, as one does. They're like, 'This really needs to be an AI company.' It's like, what do you mean this needs to be an AI company? I mean, come on. And like, well, what is your AI story? Like, we're not a—we're making this general-purpose computer, and AI is obviously important, and we saw that, but that's like being—it's the accelerators that are targeting that. Clearly, Nvidia and then a bunch of other folks. It's like, well, what's your accelerator story? We get this a lot. 'What are you guys doing about GPU?' Okay, what are we doing about GPU? Well, the GPU, as I tell people, there are two doors. One is labeled 'partner with Nvidia' and the other is labeled 'compete with Nvidia.' And what they have in common, those doorknobs are both hot. There's an inferno blazing on the other side of both those doors. Those are both very complicated doors to go through. And we've been like, you know what, we've got enough to go do on general-purpose compute. General-purpose compute, networking, storage. We absolutely are keeping an eye on what we wanted to do with an accelerator-based product, but it's a bit of a gory market and Nvidia is executing well and so on. So, it's like not really rushing to go partner with Nvidia or compete with Nvidia, but we're happy to kind of be off to the side. All right. Well, okay. They, you know, okay, they kind of throw up their hands like, well, they can't get a reporter saying it's not AI. All right, fine. But then some of the big AI companies came knocking. It's like, well, okay, what's this about? And as it turns out, and again this is like, you know, Steve was describing having the national labs having to kick our door down sometimes, you know, it's a little bit thick for—we're a little bit thick, I'm a little bit thick to understand. We try to explain to these, you know, large AI companies that like, no, no, we don't have an accelerator. 'Yeah, yeah, yeah, we know that. You think these models are all GPUs?' Well, I guess not. So if you go to your LLM of choice, ChatGPT, what have you, you see 'searching the web' and you type something in, you think a GPU is searching the web? No, that's a CPU. CPU is searching the web. You think that when they're training their next-gen model, someone's got to go out and crawl the web, pull that in. You think the GPU is out there crawling the web? No. You think a GPU is doing all the processing it needs to do kind of traditional MapReduce-like data pre-training on that? No. CPUs. You're like, 'Oh, right. CPUs. That's what we make.' They're like, 'Yes, yes, that's what you make.' And by the way, when we look at—we companies with a lot of compute, and the large model companies have a lot of compute. Like, I've got some bad choices right now. I'm in the public cloud and it's killing me. Like, this is economically ruinous to be in the public cloud. So I know I'm going to need to go on-prem. I'm going to need to own my own compute. Especially when you look in the even not-so-indefinite future, I need to own my own compute. Okay. So that was—I mean, again, I'm a little embarrassed to be like, 'Oh, right. Okay. Yeah, I guess that makes sense.' And you realize that like all of the trends—I mean, probably was true two years ago that there was much less of a CPU component. But if you look at all of the trends towards those large models, around test-time compute, around, you know, whenever you are using these models—and I love it when they write a program for you and you can have them show their work and they're like writing a program for you—that program is executing on a CPU. There's a lot of CPU involved.
