theCUBE + NYSE Wired: AI Factories - Data Centers of the Future | Alan Butler, SF Compute

Clips
More from theCUBE + NYSE Wired: AI Factories - Data Centers of the Future

Alan Butler

Chief Business Officer

SF Compute

play_circle_outline Hybrid infrastructure resurgence: neo-clouds, hybrid on-prem and cloud

play_circle_outline SF Compute: SLA-backed GPU marketplace with resale, Genesis idle-cluster rentals, high-volume low-cost model, $40M Series A, 18x growth

play_circle_outline InfiniBand Bottlenecks, Multi-Tenant AI, and Agentic Workflows Straining Security and Isolation

play_circle_outline Customer segments: well-funded AI startups and underserved NVIDIA Inception members

play_circle_outline Back-to-basics analogy to Sun, utility computing, grid computing history

Info
Transcript

Alan Butler, SF Compute

Alan Butler

Chief Business Officer SF Compute

In this interview from theCUBE + NYSE Wired: Robotics & AI Infra Leaders 2026, Alan Butler, chief business officer of San Francisco Compute Co., joins theCUBE's John Furrier and Howie Xu to discuss how flexible, short-term GPU marketplaces are reshaping AI compute economics. Butler traces parallels between today's AI buildout and the early utility-computing era at Sun Microsystems, framing the current moment as "back to the future." He explains how SF Compute began by reselling excess GPU capacity and has grown into a marketplace built on data-center-backed S... Read more

explore Keep Exploring

How would you explain the current market transformation in infrastructure—considering the simultaneous changes in compute, networking, storage, user interfaces, data layers, and server/rack architectures? add

How did San Francisco Compute Company get started, what is its business model and growth trajectory, how does it position itself competitively (including SLAs and the ability to resell committed capacity), and how long do you expect current demand to last? add

Is multi-tenancy critical for AI workloads, and if so, what are the implications for infrastructure requirements (e.g., bare-metal provisioning, InfiniBand speed), SLAs, security/compliance, noisy-neighbor isolation, and the handling of agentic (multi-tenant agent) workflows? add

Why are AI startups (including NVIDIA Inception members) asking your company to provide compute resources? add

Is today's cloud computing fundamentally new, or is it essentially a recurrence of earlier ideas like 1990s utility/grid computing and Sun's Project Blackbox? add

bolt Powered by CUBE AI

Alan Butler, SF Compute

search

(INTRO)

John Furrier

>> Welcome back to theCUBE, I'm John Furrier, host here in our Palo Alto studio. This is our third annual AI Leaders Infrastructure Summit. Of course, we have a party, 180 people getting together tonight for our pool party reception, where people are going to interact, share their knowledge. As the AI infrastructure continues to be built out, I want to introduce my co-host filling in for Dave Vellante. Our next guest is Alan Butler, Chief Business Officer at SF Compute. Also the youngest ever employee at Sun Microsystems back in the day, we were just having a great retro throwback. Alan, great to see you on theCUBE. Thanks for coming on.

Alan Butler

>> Pleased to meet you guys.

John Furrier

>> You got your shirt, you got the swag there. Good to have you here.

Alan Butler

>> Good to have you advertising, San Francisco Compute Company. Yeah, we've got no outbound sales and marketing. It's all inbound. And so this is the marketing.

John Furrier

>> Yeah, let's get into it. Obviously, the little plug about Sun, I brought that up intentionally because the computer industry was really built on the backs of a few companies. Sun was one of them. A lot of Berkeley DNA, a lot of Stanford DNA, but that built that out. But we're seeing a revolution again happening now that feels the same but bigger, highly accelerated, a lot of new infrastructure enabling new kinds of things. That's what we're in the middle of. What are you up to? What's your take on this, given your history? You're a historian and an executive. What's your take?

Alan Butler

>> I've been called worse.It's a bit of back to the future, right? It's definitely a bit of back to the future. We should all go back and watch that movie, I guess, from years ago. Gigawatts. That's all I need is to power that car.

John Furrier

>> Yeah, yeah. Yeah, we need a DeLorean. Another Irish car. now we have gigawatt factories. we're in an AI factory revolution.

Alan Butler

>> Yeah, we are. And the pace of change is off the charts. I've been fortunate, actually, because I actually was involved in Lambda in 2019 and joined the company in 21 and moved on and helped sell a company to NVIDIA two years ago called Brev.dev, who's doing great for the developer community. But the core of the basics of it goes back to the early, early days. Sun and other companies like Sun that were very innovative and trying to drive solutions in a way that people could consume in a transparent, secure manner.

John Furrier

>> How would you explain this market? Because it's a lot of things kind of coming together at the same time. It's not just, compute, networking, storage. We've seen that before with cloud. Then you got unbundled. Now you have user interface change. That feels a lot like a revolution. You have the data layer completely changing, given the UI change. And then lower level at the semi, the density of the servers, rack scale systems are the new server. It's a bunch of servers, a lot of switches in that one scale-up system. That's just the racks. That's not including the scale out and scale across stuff going on. So a lot of new componentry, architecture changes. How do you explain that given the magnitude of this revolution?

Alan Butler

>> Well, the definition of a FRU has changed, Field Replaceable Unit. It feels like it's half the rack as opposed to a component, right? I think we're really back to basics. We're back to basics, and it's back to the world of hybrid, right? Jensen talks about it at NVIDIA, Michael Dell talks about it at Dell. It's a hybrid world that I don't believe ever went away, actually. It's just that the appetite and the risk profile for who's willing to jump in and lean in. We're in the Bay Area, right? So we've got the 49ers, and it feels a little bit like the gold rush on steroids, and the pace is just off the charts.

John Furrier

>> The demand, obviously, is there. How do you look at what you guys are doing? Explain your business that you guys are going after. This velocity, the architecture, the solutions, the supply chain challenges. The computing architectures are different but need to be high-performing. What are you guys up to?

Alan Butler

>> Yeah, we at the SF Compute Company, we started out by accident two and a half years ago. we had a cluster that we figured out we couldn't use all of and we had to sell some of. So we started partitioning, the team started partitioning off pieces of it to be able to rent out for when they weren't using it to help pay the bill. And that was the kind of part of the genesis of how the company came to be.

Alan Butler

>> How many cards? Thousands of them?

Alan Butler

>> No, it was small. It was a rack.

John Furrier

>> Even that?

Alan Butler

>> Oh yeah, yeah, yeah. Bare bones, right? Right. And one of my colleagues, Eric Park, who is the CTO of San Francisco Compute Company at the time, was the CEO of Voltage Park and involved in that. And SF Compute was dealing with Lambda and they were dealing with Voltage Park, which is kind of a small circle. Right. And all of a sudden they get creative on how to pay the bills, how to partition off pieces of compute when they weren't using it. And that was the genesis of having a compute layer where you can rent out on short -term provisioning that became then the basis of, in essence, a marketplace. And we don't like the word marketplace in all honesty. The New York Times called us the Airbnb of GPU compute from last December. It's not a bad analogy, but there's a lot of marketplace companies that are flipping GPUs over a fence with no SLAs behind it. And everything we do has an SLA standing behind it with a data center provider. So there's not a whack -a -mole when it comes to, I have an outage, go call someone. So we've started off with that. We closed Series A back in November when I joined the company. We went for 40 million on a 300 million valuation. We've grown 18x in revenue in the last 12 months. 100 % of it is inbound, so we've got no outbound in that context.

Howie Xu

>> So inbound means that demand is huge, and then how long do you think the party is going to last? Why?

Alan Butler

>> Well, I grew up at Sun and I lived through the dot -com, as we all did, right? And the 401K is with us and everything else. I don't know, but I think you've got to be in the game and you got to plan for the downside, not hope it doesn't happen or hope you get out. Because, whether that's a 5 % or a 35 % drop, I think you just plan for it. I actually really like our position because when you look at a neocloud world and if you're someone like NVIDIA and you're looking at someone like CoreWeave or Lambda, they're at the top of the pyramid trying to sell, in essence, 10 deals for 10 billion each for five -year contracts, hypothetically, but it's a low -volume, high -dollar model. We're at the bottom of that pyramid working up, doing a high -volume, low -cost model, right, for shorter -term capacities. You can compute with us, subject to availability, one hour for one node, right?

Howie Xu

>> So you're more diversified.

Alan Butler

>> Far more diversified.

Howie Xu

>> Low volume serving just a small number of the labs, but for you, serve a lot of, which model do you think is going to be for the next three, five years? If you had the choice.

Alan Butler

>> I like our model from one hour to three years, and I think the higher risk is years four and five. We're building this with the right SLAs and with the right financial profile. We're the only ones in the industry that allow you to resell out of your commitment. So if you come to us hypothetically for six months for 100K worth of H100s, and you decide, three months I needed three months worth, or I only need 50K worth, we'll help you take that back. Now, you're not getting out of your obligation. We'll remarket it back to our marketplace. And now we've done a small percentage on the original resale, and we've done another small percentage on the new resale. So now it's a gift that keeps on giving, and that's very palatable and very powerful. You get those costs back.

John Furrier

>> Talk about the SLA piece because there's two different approaches. There's I'm going to stand up a bunch of stuff and sell it to the hyperscalers. And there's I'm going to stand up a bunch of stuff and actually make a business out of it and go to either enterprises or Neo clouds for capacity offload. And those are workflows, those are engineered systems. It's not like, hey, bulk raw GPUs, go configure it. It's going to have to have the software control plane. they don't want to run ages on some workload, deterministic workloads. Is that how you see you guys doing it?

Alan Butler

>> Yeah, we're extremely opinionated on the cluster we will onboard. And so when we go to a, let's call it new supplier, which we've got over a dozen so-called suppliers, which is data centers with compute capacity today in various stages, we've got a very strong technical opinion on what is minimum acceptable, including what's with InfiniBand. We need the bare metal provision and we need certain access that not everyone else would require. But that's because we tie it directly back to the SLA, right? And we want to stand behind our two nines and three nines capabilities that we live with. And so it's very important that we're pretty opinionated upon it.

John Furrier

>> You know, I would have many conversations and one of my revelations from my conversations today was the notion of single tenant, multi-tenant world. And we've been kind of talking about it, but I kind of locked in on this now because it's clear, I'd love to see if you agree or not, but to me it's clear that multi-tenancy is critical on AI workloads because that's what cloud does. You can do virtual private networks, you can do all those things like microsegmentation to get them in there, because it's faster.What's your reaction to that? Do you agree?

Alan Butler

>> Yeah, multi-tenancy, absolutely. With the right security provisions in place, all of which, frankly, are table stakes, they're not negotiable, right? And we're helping shape some of that. We have to put a volume in place for scale, right? And to be relevant, not as a one-trick pony or a snowflake. And multi -tenancy is absolutely a requirement.

John Furrier

>> On agents, I was riffing on, this is just an early thought, but I was riffing on the idea that, well, okay, multi -tenancy means company A, company B can be multi -tenant, but inside their own company, they have multi -tenancy of agents, because agents act with agency. And so you're starting to see now multi -tenancy move to agentic workflows. Thoughts on that, how that comes in from an enablement standpoint? Does that change the nature of SF Compute's, thinking around how you provision and scale up?

Alan Butler

>> Yeah, I'll have to have Eric on with me on that one, our CTO, but what I will say to you is noisy neighbors and security and division of everything and separation is really important. And having the compliance security around it is a no -brainer, right? And it's stuff that people are not paying enough attention to, actually.

John Furrier

>> Yeah, and security's huge.

Howie Xu

>> So you mentioned that you are the Airbnb for GPUs, right?

Howie Xu

>> Well, the New York Times called us that, but yeah.

Howie Xu

>> Okay, but what about the rest? do you see that a GPU may become more and more of sort of a less scarce resource, and the memory, networking, storage. What do you see? I'm pretty sure all of them are still scarce resources, but how do you see the evolution at this point?

Alan Butler

>> you're dealing with what's the long pole in the tent. I had two conversations with a tier one OEM today, and you're dealing with what am I doing with networking, and we've actually been strategically buying ahead some networking that gets us out of some of that long pole in the tent element. So network is definitely one of the long poles. Yeah, and I think you have to plan for it to be available, right? Today it's networking, tomorrow it's something else. So you just have to deal with that, and in some ways, back to 1849, it feels like a bit of a combination of Gold Rush days and Las Vegas gambling.

John Furrier

>> It's some picks and shovels versus blackjack.

Alan Butler

>> That's right.

John Furrier

>> On your customers, take us through the motion. Who are you targeting? what's the ideal customer for you? Because you're seeing a lot of different diversity of where enterprises are leaning in. It's usually different by company, seeing physical AI is coming in for robotics. But a lot of on -premises, but also distributed computing is in play. You said hybrid.

John Furrier

>> Yeah.

John Furrier

>> Hybrid is essentially now the standard. Essentially everyone has their own cloud, so why not connect through a service? What's the, who's the customer right now and as you go to the one year to three year contracts, which I think makes sense, no one really knows what they're going to do in year four and five.

Alan Butler

>> Well, maybe there's two answers to that I'll give you. One is we have a lot of well -funded AI startups. A lot of them are headquartered in San Francisco, but that's where the genesis of it, hence the SF Compute company. But they're asking us for compute. We're trusted. We've got a team that's well -recognized technically and so forth. But there's a second bucket that's meaningful for NVIDIA. NVIDIA's got 40 ,000 Inception members today. That's everyone from two -person startups globally to series C, D, and E. I'm going to suggest that 90 -odd percent of them are underserved or unserved today. When they go to get compute, they've got to go off and say, where can I get it? And they don't have the ability to buy five -year contracts. And there's quite a bit of collaboration going on between us and NVIDIA on how to help fulfill some of that.

John Furrier

>> That sounds a lot like AWS around 2012, 2013, a lot of the underserved startups buy a Supermicro box, get a data center, put it in a shared rack, good luck with that, that's a pain in the ass, put my credit card down. OpenStack, back end. Go to AWS, much easier for a startup, hey, I don't even know what I'm building, but I got some traction, I don't want to stop. Yeah. There's a lot of the startups who are like, in that version of AI, it's like, oh, I got 10 ,000 customers, what do I do?

Alan Butler

>> Right, and whether it's on the supply side or on the demand side for us, because we're looking at both, right? We find it really fascinating. We have got to talk to a lot of people that don't know what they're talking about on both sides of the fence to get to the people who do. And we end up having a lot of almost on-the-job training and education on configuration, compute, elasticity, availability, and oh, by the way, the pricing demands and the pricing models. They want it tomorrow. They want it today.

John Furrier

>> Yeah, yesterday.Yeah, and it is truly crazy. And what's your strategy now as you guys look at this? Because obviously the demand curve's there, you're balancing both sides, supply chain is constrained, so you have to engineer around those. What's the focus, what are you optimizing for?

Alan Butler

>> Yeah, I think there's a couple of things. One is, number one, it's all about the customer, right? it needs to be about the customer, and to do that, it means you're delivering a high quality service that can be counted on, right? And you've got SLAs and some reasonable penalties to back up what you're talking about. And B, it's a very heavily underserved community globally, right? We were just at the RAISE Summit event in Paris a couple of weeks ago. We've done Dell World a few weeks or a month or two before that. AWS re:Invent's on the list. GTC, we had a booth at the Inception area at GTC. Open Compute's coming up. Supercomputing is coming up.

Alan Butler

>> We'll be at Supercomputing.

Alan Butler

>> Exactly right, right. And so it's a crazy appetite for what's going on. But it still goes back to basics. And it goes back to basics on the technical components, the technical team, the pricing ability, and then, frankly, from our point of view, the unique resale capability we have that others don't have is very, very appealing. And frankly, our OEM strategies are aligned very well with some of the key OEM providers.

Howie Xu

>> What is the primary pain point that they have with the other solutions, like a hyperscaler or Fireworks of the world? Is that a cost or the privacy? What is the thing that drives them to go to you?

Alan Butler

>> I think it's specialty, more of a focus, if you will, right? Not trying to be all things to everyone. And I'm sure there'll be plenty of opportunity when someone will pivot from one orchestration to the other for multiple reasons. But I do think the landscape is shifting. And I think, and you have to plan for, I mean, AWS is a great company, right? And GCP and the list goes on. Great companies that will absolutely be around. I think they're learning some things from the likes of us as well.

John Furrier

>> Yeah, and you've got to build that durable business. Amazon started out as basic building blocks, just the ease of use for developers, underserved market, they build high level services, build out scale. Similar thing going on with the Neo Clouds, and it's interesting, we're just riffing on the whole Sun. Back in the day, you had to design the systems, manufacture them, put them together, sell them, support them, run a business. That was a real undertaking. That was ARR, we got some service revenue in there. But that was the business, and in a lot of monetization conversations, we're seeing the same pattern, albeit at a little bit different scale, but a NeoCloud's got to stand up and they got to be in business. And this is now fundamentally, you mentioned going back to the basics, this isn't as easy as putting a SaaS app in an app store.

Alan Butler

>> No, it's not.

John Furrier

>> This is real physical gear, infrastructure. Not the same as the old school days, but it's the same concept. You spend a lot of cash ahead of bookings.

Alan Butler

>> Yeah.

John Furrier

>> So get down to basics. What is that basic playbook? I see customer, but how do you guys think about this? Because we're kind of in a similar era.

Alan Butler

>> Well, I think it's not just NeoClouds, by the way. It's, for instance, NVIDIA's NCP partners, if you will. That's a big population that have compute already. Think about this. At Sun, we were doing this in the late 90s, utility computing.

John Furrier

>> Yeah, I remember.Right? Grid computing, utility computing.

Alan Butler

>> Uh-huh, right.Cloud computing in the 90s. And Project Blackbox with the containers for the data centers, talking about back to the future. Here we are again. and I think that there is a lot of rinse and repeat that's highly appropriate to your sentiment, even if there's variances and differences logically, right? But the basics are the basics, and at the end of the day, if you're delivering a high quality solution that's reasonable value, then you will get stickiness more than your IP.

John Furrier

>> Great. Alan, thanks for coming on theCUBE. Put a plug in for What you're working on. What are you looking to do, hire customers, why should they go to you guys, put a plug in.

Alan Butler

>> Appreciate it. Two things, if you're looking for H100s and above, now or over the next six months and beyond, give us a call, sfcompute.com, right? It's a website. If you've got capacity that is, in your opinion, underutilized, H100s and above, come to us and we'll help you improve your utilization in a very dynamic way that you cannot do on your own today.

John Furrier

>> Great, thanks for coming on, I appreciate it. And we should definitely do a session, get Scott McNealy on, we'll do a little Sun throwback. Lot of parallels between the system revolution back at the Berkeley, Stanford days, East Coast, that early computer industry we're seeing kind of a redo, but in a whole modern way.

John Furrier

>> It is.

John Furrier

>> Thanks for coming on, I appreciate it.

Alan Butler

>> Thank you very much.

John Furrier

>> I'm John Furrier, Howie Xu, we're here. Our third annual AI Summit of AI Leaders here in the studio. I'm John Furrier, host of theCUBE, thanks for watching. Thank you.