Turning Unused Home Power into a Controllable Resource with SPAN
Arch Rao is Founder and CEO of SPAN, the company behind the smart electrical panel that has expanded into utility grid management with SPAN Edge and distributed AI infrastructure with XFRA. What began as a smarter way to manage home electricity has evolved into a platform designed to digitize the electric grid, enable mass electrification, and unlock new AI compute capacity using existing residential infrastructure.
In this episode of Inevitable, Rao explains why SPAN's vision has grown beyond the home electrical panel, how utilities are using software-defined power management to avoid expensive infrastructure upgrades, and why distributed AI inference could become a new class of digital utility. He also discusses the intersection of electrification, grid modernization, and AI demand, arguing that the next major energy transition is the digitization of the electric grid itself.
Note: SPAN is an MCJ portfolio company
Episode recorded on July 8, 2026 (Published on August 4, 2026).
In this episode, we cover:
(0:00) An overview of SPAN
(1:53) How SPAN evolved from smart panels to grid infrastructure
(3:41) The inspiration behind XFRA and distributed AI compute
(5:21) What an XFRA node looks like and how it works
(6:04) Why homeowners would host AI compute infrastructure
(8:06) The homeowner business model: free panels, batteries, and lower utility bills
(10:48) Why SPAN is focused on new home developments
(11:56) How SPAN Edge helps utilities avoid costly service upgrades
(16:35) SPAN's go-to-market evolution
(17:16) Lessons from customer behavior and software adoption
(20:53) How SPAN Edge creates value for utilities
(26:16) Securing distributed AI infrastructure in residential neighborhoods
(27:48) Who is buying XFRA compute today
(29:45) Why distributed infrastructure offers flexibility over traditional data centers
(31:04) XFRA as a new form of digital infrastructure
(32:50) How SPAN differs from virtual power plants (VPPs)
(34:56) Why the real energy transition is analog-to-digital
(36:43) Why talent is SPAN's biggest hiring priority
-
[Cody Simms] (0:00 - 1:39)
Today on Inevitable, our guest is Arch Rao, founder and CEO of SPAN. SPAN makes a smart electrical panel that sits at the center of a home's power system, and has extended that platform in two directions. A utility-facing product called Edge that manages home load at the grid edge, and a new product called XFRA that turns unused home electrical capacity into a distributed AI compute network.
The company has raised nearly $500 million, and earlier this year launched XFRA in partnership with NVIDIA, with nodes already running at over 95% utilization in early deployments. XFRA arrives as AI compute strains the wires and transformers that move power around, and SPAN's bet is that the fastest path to increased capacity is the unused headroom already sitting inside millions of homes, with little mini data centers running in our backyards. MCJ is an investor in SPAN via our venture capital funds.
From MCJ, I'm Cody Simms, and this is Inevitable. Climate change is inevitable. It's already here, but so are the solutions shaping our future.
Join us every week to learn from experts and entrepreneurs about the transition of energy and industry. Arch, welcome to the show.
[Arch Rao] (1:40 - 1:41)
Cody, thanks for having me.
[Cody Simms] (1:42 - 1:52)
I think maybe we'll start here. For someone who last knew SPAN as a company that reinvented the panel at a home, what is SPAN today?
[Arch Rao] (1:53 - 3:25)
We've been on an incredible journey the last seven plus years now, initially starting with creating a product category for smart panels that didn't quite exist at the time, to becoming a product that can deliver a better customer experience for somebody buying a home battery backup system, and then scaling to being the first power control systems that can help you save money if you're electrifying your home, to becoming a staple for a number of new home construction partners, to last year introducing a utility version of the product called SPAN Edge, which has been now increasingly adopted by utilities as an alternative to building more poles, wires, and transformers. For what it's worth, one of the very first pitch decks that I had back in 2018-2019 had this progression as our ambitious growth to become a key part of how we think about decarbonizing the grid and how we think about electrifying homes.
Despite some headwinds that we faced along the way, I'd say we've done a pretty good job being able to scale and be, I think, at the top of the leaderboard still in terms of being the solution, the go-to solution for home electrification. Where we stand today is the core business is performing really well. We've brought on board some extraordinary customers and partners to scale, and we're thinking about the next logical evolution for us.
How do we solve for not just growth for SPAN, but how do we solve what is increasingly becoming the dominant topic of conversation, enabling load growth for compute, which hopefully has a collateral benefit to every consumer, every end user to reduce energy costs and electrification. That's XFRA. I'm happy to talk more about that today.
[Cody Simms] (3:25 - 3:40)
Let's start there. And there's so much to unpack in everything you said, but maybe start with the most recent announcement, XFRA, because I think my guess is many people who are listening have heard of SPAN and kind of know the panel story at least somewhat, but are really curious about this announcement you all have made.
[Arch Rao] (3:41 - 5:21)
Sometime two years ago, almost to the date, in the summer of 2024, I came to this epiphany, if you will, that load growth that's coming from solar battery adoption, EV adoption, heat pump adoption is good, but continues to be at a lower velocity or pace than we would like. But adjacent to it, there's all this demand growth that's coming from AI and the build-out of data centers that's happening in almost in a parallel universe. It's happening in a way that is definitely begetting a lot of interest from utilities, definitely bringing in a lot of CapEx going into generation transmission, but not quite solving the distribution level problem that is the problem we need to solve for consumer level electrification.
In parallel, we had spent the last half decade leading up to that, really solving the digitalization of the electric grid with SPAN by saying, if you have sensing controls and intelligence at the very edge of grid, we should be able to enable a lot more electrification at a much lower cost, i.e. increase the utilization of our existing distribution network. So taking those, let's say input assumptions, I started to wonder if we could unlock the value of SPAN, i.e. get more existing distribution grid utilization to go up while delivering the speed to power problem that AI compute is facing. And it so happens that inference compute, which is where money is being made with AI, does not require the same profile of data centers that training does.
You don't need large contiguous racks of GPUs sitting inside a single large shell to generate value. You can deliver almost all of the inference value with distributed compute that can be approximate to customers, if you will. So that's what led to XFRA.
XFRA in its simplest form is a data center in a box.
[Cody Simms] (5:21 - 5:23)
What's it look like on a house? Describe it.
[Arch Rao] (5:23 - 5:50)
Think of it as like a larger air conditioning compressor unit outside your home. So it sits on like an eight by three, eight by four pad outside your home. It's around four feet tall, but what's contained in it is a heat pump plus all of the cool stuff that we've built around power quality, power management, resiliency, and direct liquid cooled rack with GPUs and CPUs in it.
So it feels very benign from the outside. It's intentionally designed to look pretty industrial and not particularly sexy, but it's a powerhouse of AI value.
[Cody Simms] (5:51 - 6:03)
Data centers today and consumers are, I would say, not in the happiest of relationships right now. If you judge consumer sentiment about data centers, why would homeowners want to put a little mini data center outside their house?
[Arch Rao] (6:04 - 6:31)
That's a great question. Quite frankly, that was one of the things I was trying to solve for. My somewhat altruistic view on this was if compute was going to be as fast a growth category as it's proven to be, it should have consumer value.
And even if the collateral benefit of the average customer gets a free SPAN panel and a free battery, and by the way, the ability to electrify at no additional cost to them because compute is paying for it, wouldn't that be a phenomenal win-win?
[Cody Simms] (6:32 - 6:40)
And maybe underscoring that, you're getting all of the SPAN goods essentially for free or a huge discount if you opt into XFRA that right?
[Arch Rao] (6:40 - 7:45)
That's exactly right. So from an average consumer perspective, maybe thinking about it through the lens of an individual homeowner, let's say. And for what it's worth, we are prioritizing new home construction simply because of the homogenous nature of the build-out.
We haven't spoken about this more broadly, but we're also solving for some small commercial that is proximate to urban and suburban areas where we can deploy this. But in the new home construction context, we are giving every home in a new home community a free SPAN panel and a battery. And then a subset of those homes, based on some of the power flow math that we've done, will be co-deployed with XFRA nodes because you don't want to saturate the entire community with XFRA nodes.
It's way too much of a power draw on the upstream network, so we're being judicious about that. But regardless, if every home now gets a free SPAN panel and a battery, you've got the ability to continue your own personal electrification journey while having resiliency, which is a large class of batteries that we can then use for other energy purposes. And all of that becomes a secondary value proposition because the subset of homes that have XFRA nodes are delivering an abundance of economic value with compute that we then share back to the customers.
[Cody Simms] (7:46 - 8:06)
And for those homeowners, what is their relationship with the XFRA node and the compute both in terms of power and in terms of internet bandwidth? Is it using their residential power bill to power this thing? And is it using their residential internet connection to send data up into the cloud for the hyperscalers to consume?
[Arch Rao] (8:06 - 9:23)
It's a little bit the other way, which is we are solving for a model where it's more like a landlord-tenant or third-party-owned infrastructure model. Today, you can have, let's say, a home solar system or a battery that gives you value that you consume, and then the vast majority of value could be through energy services or export value, if you will. That's kind of the idea.
You're renting us, so to speak, a small amount of physical space in your yard that you otherwise would not use. And in exchange, we are providing you not just the free upfront value of the SPAN panel of the battery, but ongoing value by giving you a discounted energy and internet bill. So every home that we go into, we essentially will pay for all of the energy consumption of the XFRA node and most of the energy consumption of your own home.
We pay for high bandwidth fiber being landed right there into your home. And by extension, you get discounted internet service providers' value coming to you as a consumer as well. You're not in any means obligated to use the AI compute that's co-located with your home, because in theory, that compute is way too much for an individual customer anyway, because we're talking about clusters of GPUs in a single node, if you will, that are purpose-built for hundreds, if not thousands, of concurrent inference sessions.
So there's an enterprise user that's buying that compute capacity, but you get the economic benefit that we share with you on an ongoing basis.
[Cody Simms] (9:24 - 9:44)
So you're coming in, you're installing the infrastructure. The homeowner is, in theory, benefiting from a lower power bill and an upgraded or boosted fiber internet connection that has to share some load with this XFRA unit. But at the end of the day, the user is getting essentially free high power internet at home.
[Arch Rao] (9:45 - 10:48)
All the tens of thousands of homes that we've installed in across the US today, which by the way, I think when you guys invested, we were probably less than a thousand homes four or five years ago. We've seen some extraordinary growth in the last several years. And across all the 200-amp homes that we've installed in, empirically, what we see is that virtually all of them, north of 98% of those 200-amp homes have 80 amps, roughly 19 kilowatts of capacity, when I say capacity, the conductor capacity that is never utilized.
It's never accessed. As in your peak power, maybe 80 amps out of the 200 amps. You may very rarely go to over 100 amps, but above that, there's plenty of headroom all of the time.
Now, obviously, there's what's called a load factor map that happens upstream of it. Not every home that has 200 amps is linearly added up to the transformer upstream. That's why we don't put an XFRA node in every home in a community.
We'll put them in a third of the homes, if you will. But the economic benefit is across that entire community. That is existing distribution infrastructure that has the, quote unquote, interconnection required to build and aggregate the equivalent of a 100-megawatt data center.
[Cody Simms] (10:48 - 11:03)
So this is where you're working with the housing developer and you said you're focused on new construction, because then you can map out exactly how many nodes you need in a given multi-acre plot and how that maps to the utility's demand profile.
[Arch Rao] (11:04 - 11:35)
That's exactly right. And we're working in conjunction with the utilities. We're working in conjunction with the offtakers to ensure that and the fiber providers to ensure that the quality of service that we provide is comparable, if not, in some cases, better than a large-scale data center, because we have built-in resiliency and redundancy in the number of nodes that we have.
And it's not concentrated to a single site, if you will. We're working with the home developers to make sure that they're right-sizing their infrastructure as they build out and do land development all the way to building our homes. But the direct economic value there is they're now able to sell homes for lower cost.
[Cody Simms] (11:36 - 11:56)
And that's as opposed to, say, your product that you rolled out a year or two ago, which was the Edge product, which as I understand it is still home-specific, right? It's a demand response product that is specific to a given home's footprint. With XFRA, you're talking about managing the load profile of an entire micro-community or housing development.
[Arch Rao] (11:56 - 13:16)
Think of Edge as being a utility version of a SPAN panel, in the sense that instead of what was historically a rip-and-replace solution with SPAN Generation 2 products, if you will, or today a purpose-built version of a 40-circuit panel that's required in new home construction, there, the standard panels are going in by design. Where we found there to be an equally compelling opportunity is the vast majority of customers. Take, for example, PG&E here in California.
They get close to about 30,000 inbound requests from existing customers, homeowners specifically, every year that say that their contractor needs them to upgrade their service. So they go through often a very prolonged and expensive process to upgrade their service from 100 amps to 200 amps just to be able to add an EV charger or just to be able to add an induction cooktop. And in fact, now you can go on PG&E's website.
They've got a SPAN smart panel program, essentially. It's called Panel Boost. And they will essentially give you at no cost a SPAN Edge and install it, which sits at a meter.
And it has 8 to 16-circuit spaces to help you fully electrify. So in a retrofit case, we've taken the economic cost equation entirely off of the customer's purview and said, the utility should pay for it because there's a tremendous amount of value the customer gets for electrification now with the SPAN Edge panel. And the utility now gets this platform where they can see and control multiple assets.
So the demand response value is secondary, not primary.
[Cody Simms] (13:17 - 13:33)
It's interesting because in each of these cases, it feels like what you've unlocked is a shift in who ultimately controls the panel. And it sounds like to some extent, you're shifting control away from the homeowner in exchange for economic value to the homeowner.
[Arch Rao] (13:33 - 13:35)
If the customer chooses to do so.
[Cody Simms] (13:35 - 13:36)
Okay.
[Arch Rao] (13:36 - 15:06)
You can say, I don't want to do that.
I either want to just pay for my own panel or I just want a discounted Edge. And I don't want to enroll in a utility program because we don't think of it as demand response value. We think of it as dynamic service rating value.
And I'll explain that in a second. In a very simplistic sense, if you're an existing home and you're looking to electrify two out of three homes in the US will need to upgrade from 100 amps upwards. When we had a SPAN panel that you could buy yourself, the customers could afford to spend, let's say a few thousand dollars, $3,000 to $5,000 to install a SPAN panel.
Did it on their own accord. They avoided a $30,000 service upgrade and they got that benefit. If you did the same thing from a new home construction perspective, the home builder accrued that value and passed it down to you as a new home buyer, if you will.
But there's a vast long tail of customers that are either unaware of these solutions or don't have the economic means to even pay $5,000 for a panel upgrade. But that's the bottleneck for them to add an even electric water heater, let alone an EV charger or a solar battery system, if you will. That's what SPAN Edge is designed to solve.
Now, finding a pathway to now where the utility is the channel, as opposed to the contractor, the home builder, is the main unlock with Edge. And it needed to be a product that the utility is conceptually aligned with being able to own and operate, like distribution infrastructure. That's what we've materially unlocked.
So the principal value is economic benefit to the customer on day one. Once you get a SPAN Edge from, let's say, a utility, a PG&E-like utility, you no longer have to pay for a $30,000 system upgrade or for that matter, a subpanel that you typically have to add. The secondary value then becomes the utility being able to call on you if there is a grid service need.
[Cody Simms] (15:06 - 15:19)
Do the consumers who have the utility-focused product, SPAN Edge in this case, and now soon to be XFRA, do they still also have the ability to do their own granular controls if they want to?
[Arch Rao] (15:19 - 16:04)
100%. So one of the things that we have been incredibly focused on from the very beginning is the consumer needs to have agency. They need to be able to determine what's important to them.
And this was true when we did the nice-to-have versus must-have backup software-based control, if you have. Now, the same thing is extrapolated into our power controls, where you have a stack rank of loads. And the panel is autonomously determining how to shed load if there's a short duration of time where your loads are turning on concurrently.
And by the way, it very rarely happens, right? Less than 10 hours out of 8,760 hours in a year that you need to control anything. Even in the Edge version of the product, the software application and the controls are all exactly the same to the customer.
[Cody Simms] (16:04 - 16:34)
You started very clearly with the homeowner as the primary customer. And by customer, I mean buyer. Now you've moved to a world where you still have a homeowner customer.
Clearly, every SPAN instance is at a home that has a homeowner using it. But your sales and go-to-market has shifted more toward the utility buyer. What have you discovered in terms of what the homeowner actually thinks they want in terms of control and what they actually use in terms of control?
[Arch Rao] (16:35 - 17:07)
Well, first on the channel in go-to-market, it's intentional. We had to go DTC because we wanted to understand what the pain point of the customer was, what the installer need was. And it gave us a lot of data and input that we needed to build the right product.
Then we went through solar storage partners, the likes of Sunrun that helped us scale. Then we went to contractors. We have over 5,000 SPAN-authorized contractors around the country now.
And the product is sold through over 500 distribution endpoints. And then we moved to a model where new home construction was buying it. And then we got to utility.
At each stage, we demonstrated value and earned the right to be at the table with the next large channel, so to speak.
[Cody Simms] (17:07 - 17:16)
I remember when I first saw somebody wearing a "ask me about installing SPAN" t-shirt. It was an electric contractor. And I was like, hey, I know that company.
It was a pretty fun moment.
[Arch Rao] (17:16 - 19:06)
Yeah, the contractors are some of our best proponents. I think the data from the team was every first time SPAN electrical contractor, I'll digress for a second, but bear with me. The first time they install a SPAN panel, the very first panel, it typically immediately leads to three additional panels that they've sold for us.
Just because they're like, hey, this was a phenomenal experience. The customer loved it. It was easy for me to sell them on the product.
So we'd like to sell more. That's the kind of propagation effect that we'd like to see. Coming back to your question around what have we learned in terms of the data?
Not surprisingly, most customers want to set and forget. There are super users that like to tinker a lot. And the super users are in the app very often.
They like to go and see their energy consumption, see their priorities. Some of them are super, super users that want direct API access to their panel to build their own dashboards and widgets. We recently enabled that as well, where you can just have local API.
You can have your own authorization to pull the data because we generate so much high-quality data from the device that you can enable your own sensing and control scheme, if you will, with the data we provide. Overall, I'd say the software experience has been a critical part of how we've differentiated the product. The core hardware with its R controls listing is the functional safety piece of it.
Without that, there's no value in adding the product. With that, it's a very, very simple economic argument. But the software piece, where you have the consumer app, that's very compelling, but also by extension, all the downstream integrations we've done, the integrations with EV chargers, where we can throttle EV controls, integrations with heat pumps, water heaters, where you can do set point and state control, has dramatically improved the customer experience.
You're no longer sacrificing for an on-off type control, which is what you typically would have had to do in the past with a single load controller. Now you're able to say, here are my preferences, and if I only have to throttle your EV charger by 30%, so be it. If I have to throttle your compressor by 10%, so be it.
[Cody Simms] (19:06 - 19:13)
And the value to the homeowner, too, is you can use the SPAN software system to do all that versus all the clunky OEM softwares that are out there.
[Arch Rao] (19:14 - 19:23)
That's exactly right. There's one unified view, and it's like a persistent view because a panel, once installed, is there for the next 20, 30 years, regardless of what appliance you have. That's the idea.
[Cody Simms] (19:23 - 19:36)
You guys had to make a concerted decision as you rolled out, for example, XFRA, that you're going to give away SPAN, the panel for free, as part of doing that. Is there a risk in cannibalizing your core business as you go?
[Arch Rao] (19:37 - 20:43)
I think of it more as a flywheel effect. We are actually creating a massive amount of pull-through for SPAN panels by virtue of the fact that XFRA can accommodate the economics required for me to be able to give away a SPAN panel and a battery, if you will, the likes of a Powerwall, which is also another product I had the pleasure of working on a few years ago. It's all sort of coming together over the years, if you will.
I'm not worried about risk of cannibalization. I think the extent to which SPAN delivers economic value to existing homes is so significant that that market continues to grow every single year for us. XFRA in the residential setting is initially focused on new homes only, and that's usually a price-sensitive market anyway.
So giving a free SPAN panel actually drives up adoption there across the board. And again, we haven't spoken much about this publicly, but there's a whole swath of customers outside the residential setting. So for what it's worth, the third-generation SPAN panel can support small commercial 208 volts as well, not just residential voltage levels.
So we are looking at opportunities to deploy a vast number of SPAN panels and XFRA nodes in non-residential settings across the U.S. as we speak.
[Cody Simms] (20:44 - 20:53)
From a business model perspective, you obviously have been monetizing panel sales for years now. How does SPAN participate in these other models, both Edge and XFRA?
[Arch Rao] (20:53 - 22:55)
So Edge comes by design with a software-as-a-service agreement with utilities. And this was the DSR, or dynamic service rating, I was going to allude to earlier. So let's think about it as value at a home level, and then value at a community level, and then value at a network level.
One of the biggest challenges utilities face today is having to continue to deliver energy at really high reliability needs, while at the same time delivering it affordably. Affordability is front and center in every single PUC conversation across the country right now. What SPAN Edge, or for that matter, all SPAN panels combined allows the utility to do is to manage that risk of reliability and affordability by saying, hey, SPAN, can you look at all the SPAN homes downstream of this transformer and keep it below 90% operating threshold of this transformer?
Because it's a really hot day, lots of loads are coming online, and I don't want to break my system. I don't want to be at the risk of breaking my system. Or even simply put, I don't want to reduce the operating life of my system because I'm overrunning it above capacity for too many hours of the day.
And that's all they send us. They send us a set point. And what we do very elegantly is translate that into a new dynamic set point for each of the homes that are connected with SPAN.
So for the next three hours, Cody, your home is not going to be at 100 amp limit, it's going to be at a 73 amp limit, and my home might be at a 64 amp limit or whatever the case might be. And it still takes into account all of your priorities to ensure that the load control that happens, happens in a way that honors your preferences and does not cause you any on-off control. Now this has two benefits.
One, the utility is not doing individual load programs. They're not doing an EV charge control program or a thermostat charge control program, because what they really care about is the ability for them to protect their overall system and increase the utilization of the distribution network. And the customer has the benefit now of saying, I can either enroll or not enroll in this.
But if I do enroll in this, I'm getting very clear economic benefit without having to worry that the utility might turn off my air conditioning system.
[Cody Simms] (22:55 - 23:03)
You know, if you're going from 100 amps down to 73 amps, how do you as a individual homeowner control what is a priority and what is not?
[Arch Rao] (23:03 - 23:44)
You go into SPAN app and you can say, during a power up event, here's a stack rank of loads that I want you to control. And it sends you a push notification saying, hey, Cody, you're going to have a power up event tomorrow between 3 p.m. and 6 p.m. If you'd like to change your priorities, you can, or if you'd like to opt out, you can. And more often than not, like nine times out of 10, what happens is we are partially throttling EV chargers.
So even then there's a very low perceptible impact to you. And we are partially throttling two biggest thermal loads in your home. So your air conditioning system and water heater, and that actually delivers a massive amount of value.
But now we've abstracted that layer of load control from being a single load demand response program to a utility level infrastructure value program.
[Cody Simms] (23:45 - 23:50)
Maybe walk us through the same with XFRA in terms of how SPAN participates from a business model perspective.
[Arch Rao] (23:51 - 24:45)
So XFRA is obviously one of the largest loads you're going to see at the distribution grid if you deploy it at scale. And so we have our SPAN fleet orchestration layer. We've got the XFRA secure orchestration layer on the compute side.
Each home, as I described, has a SPAN panel that can control loads in the unlikely event that you approach close to whatever 200 amps. We also have an on-site battery that can inject power into a bus bar that can actually operate as an on-grid microgrid, even if it's greater than 200 amps. So your import from the grid is restricted, but the system can operate at higher levels.
And we have the ability to throttle compute. We are able to throttle down the GPU power level, and we're also able to move compute workloads around. So we can say this portion of the grid, let's say in California, is being overutilized right now.
There's too much loading over here. Let's move the inference workloads over to Texas because we have a bunch of XFRA nodes over there. And that's the kind of control that just doesn't exist in the market today.
[Cody Simms] (24:46 - 24:59)
So basically, it's taking the same model you've built with Edge, but you're now owning some of the load yourself, and it's allowing you to actually control that throttle of your own load.
[Arch Rao] (25:00 - 25:26)
That's exactly right. So SPAN panels or SPAN Edge as a utility product have the same core functionality. Initially, power controls and an ongoing value of energy management.
And then compute now becomes another very, very controllable load for us. In theory, we very rarely have to regulate that. So I think one of the things that we've discovered by operating a few nodes around California now is that we don't actually have to regulate compute quite as much because we're able to move workloads around.
[Cody Simms] (25:27 - 25:41)
I would think the compute load becomes the first thing you can move around, and it allows you to not have to mess with people's air conditioning and water heater quite as much once you are fully deployed, right? Once you have a large scale deployment of XFRA out there.
[Arch Rao] (25:41 - 26:15)
That's exactly right. I think a loose analogy here is thinking about bandwidth management on like a cellular network. If you have a lot of people that are all connected to the same tower, the infrastructure on a cell tower operator has the technology to move some of the connectivity, some of those connected users to an adjacent tower, if you will, just to manage bandwidth across the fleet of users.
And because inference AI workloads are by design smaller than large contiguous training workloads, we can finish a batch of work and we can sort of shift the other batch of work over to another XFRA node if you wanted to.
[Cody Simms] (26:16 - 26:25)
How are you handling the security of these boxes? I mean, these GPUs are highly valuable. Are we going to have a porch pirate phenomenon with people cracking into distributed home inference data centers?
[Arch Rao] (26:26 - 27:37)
I'll tell you what all the things we're doing from a security perspective that would not surprise you whatsoever, but also tell you anecdotally. You know, when I was at Tesla, for example, the company deployed billions of dollars worth of infrastructure for charging, for example, for batteries, for example. And the hot commodity at the time may not have been GPUs, but it was definitely power electronics and copper and this and that.
We didn't have a high incidence rate of porch pirates. And these were units that were not in the backyard of a home. These are units in the Target parking lot, much more accessible to people.
That concern is a little bit overblown, in my opinion. That said, we take security seriously. And so each device that we're building is in a fully self-contained cage.
There's no access to it. There's no physical access to it. You can't just open a door and access the GPUs.
They are anchored down with earth poles or tankers into a concrete pad. They weigh over a thousand pounds. You couldn't just lift it out of there.
We have sort of digital security. We have ways of sensing if there is unauthorized entry and we shut down all of the compute and basically break the server. So the value of the thing that you get access to is zero, is marginal.
So I think there's a lot of thought that's been put into security. And we've had incredible partners in companies like NVIDIA that have helped us think through this stuff. So we feel pretty good about that.
[Cody Simms] (27:37 - 27:48)
And who is actually buying the compute from you today? Or are you at the point of, we're going to build this out and once we have amassed enough capacity, we'll go into commercial agreements on it?
[Arch Rao] (27:48 - 28:17)
That's a really good question because at the end of the day, you want to make sure that you're building a product that has real offtake. So we will own and operate the nodes at scale. So this is the XFRA, the hardware that's being capitalized and deployed.
We are selling compute capacity to the market. There's very broadly speaking, sort of three tiers in the market. Purely the merchant market, which is where we are today in the handful of nodes that we have.
And that's just you and me as users or academic institutions or startups being able to go and rent GPU capacity for weeks at a time, months at a time.
[Cody Simms] (28:18 - 28:20)
So that's your own neocloud product then that you'll be offering?
[Arch Rao] (28:20 - 29:27)
We are offering it through existing marketplaces like the vast AIs of the world, if you will, right? And successfully so. Every unit we have out there is being rented at 95% plus utilization today.
That means we're getting paid handsomely and the units are well on their path to being cash flow positive. The second category is you have aggregators. You have inference providers and aggregators that are solving for a combination of open source models and proprietary models that are running to serve enterprise use cases.
That market is probably in the hundreds of megawatts of compute capacity a year. The first category would have been like tens of megawatts of compute capacity, let's say. And we're building those out with like two-year, three-year type of contracts, if you will.
As we are getting to scale, the interest we have from the very large operators, the frontier labs, the hyperscalers, et cetera, is, hey, can you deploy gigawatts of compute capacity for us over the course of the next 12, 18 months? Because an equal amount of compute, if you went the traditional path, would take us four to five years to build. And today, all the projects that are in the pipe are taking longer and are costing more.
It's like a home remodel project almost. So this becomes an easy way for them to access compute capacity. And there we're signing up to multi-year, five-year, seven-year type contracts.
[Cody Simms] (29:27 - 29:37)
How do you think you are impacted with this distributed model in a world where the AI infrastructure build-out slows down?
[Arch Rao] (29:37 - 29:40)
You mean like the demand slows down or the data center build-out slows down?
[Cody Simms] (29:40 - 29:45)
If the data center build-out slows down, if CapEx spend starts to slow down in particular.
[Arch Rao] (29:45 - 31:03)
Yeah, interestingly, I was having this exact same conversation with an investor earlier today, right? Unlike a single large 100 megawatt data center, that 100 megawatts and let's say $50 million a megawatt is a roughly $5 billion CapEx investment to get it up and running. Is construction capital, supply chain commitments, land development, building out a site, bringing in compute, et cetera, with the hope that you can deliver a project in four to five years.
Whereas we are able to bring online single digit 10 megawatts every single week, starting this week, next week, or starting a few months ago. So we have the unique ability to throttle down the amount of compute we deploy. If for whatever reason the market is saturated or the infrastructure capital is running low, in a way that no other data product can today.
So I think that's one unique advantage. But the objective reality, at least from folks that are much closer to the AI demand or the actual use of inference compute, there doesn't seem to be a forecastable moment in time in the next few years where the supply outpaces demand. I don't think that's a realistic problem.
And longitudinally how we're thinking about this is at the end of the day, what SPAN is building with XFRA is digital infrastructure for power and compute across the country. It could be for inference today, it could be for physical AI tomorrow, we're agnostic to it.
[Cody Simms] (31:04 - 31:29)
It seems to me it's almost a new category of utility. It's like a compute utility with XFRA in particular, in that you have hardware on residential property that is not owned by the property owner and that you're going to need, I assume, easements to go service and upgrade and manage and this, that, and the other, much like a utility does today with poles and wires, right?
[Arch Rao] (31:30 - 32:17)
Like electricity, water, and sewage. Yeah. Look, I think there are similarities and meaningful distinctions here.
Electricity, water, sewage, obviously, are necessities for every consumer. I don't think we're quite there yet with AI compute phenomenon where everybody's using it yet. But I do think increasingly, if you think about, again, the proliferation of robots, any kind of machine that does work for you, that's going to increasingly be pervasive over time.
Let's even say in a five-year to 10-year horizon. What we're really building is that underlying infrastructure there. You're right.
It brings together power, internet, in a very important way. We are not using water the same way large data centers do because we're a self-contained system and it's got its own coolant loop and we're much less dependent on that. But you're right.
It's building, let's call it, the cell phone tower model for AI.
[Cody Simms] (32:18 - 32:49)
In the last few weeks, there have been a number of virtual power plant or VPP announcements. Tesla, Sunrun, Renew Home announced a big one. Voltus and Google announced sort of a large one.
How are you thinking about what SPAN is doing as maybe describe what those are and what SPAN is doing is similar or different to those other large announcements in the space, which people could broadly lump into the category of a bunch of distributed stuff that provides power to data centers.
[Arch Rao] (32:50 - 34:38)
There's been a lot of acronyms to describe, in my point of view, very similar constructs, be it Demand Response or DERMS or VPP as the case may be. In the class of VPPs, I would say more so than load control, the most valuable asset is probably energy storage because it has the ability to actually inject power back into the grid. I would say the programs that either use Tesla batteries or Sunrun's batteries, et cetera, can be and will be meaningful contributors to how you think about virtual power plants.
But that still isn't the class of how do you at a non-co-located basis create room on the grid for say, continue to run compute. So as a compute AI provider, can I then compensate somebody else for injecting power from a battery or reducing the load that has some economic value, if you will. And there's going to be value for that.
There's going to be value for that from the point of view of grid services. But for us, that's again, almost a secondary value proposition. Every node that we're building in and of itself is being built because we have SPAN technology that allows us to use the existing headroom in the physical distribution network.
We are capturing that underutilized capacity right there, as opposed to some non-co-located solution that might be able to throttle down for me to be able to operate this node. So that's one. Two, by having the layers of control that I described earlier, the load controls, the battery-based controls and the cloud compute controls, we are effectively able to provide that value in real time in a primary sense, not a secondary sense.
In some ways, one of the things that has come to mind to me is when you think about the entirety of a base power business model, that's our secondary value proposition. Because the primary value proposition is the actual load that we can control, i.e. compute.
[Cody Simms] (34:39 - 34:55)
What do you think the world looks like in five to 10 years? Fast forward out, the business model that you're building today and not just from a SPAN perspective, but both from a power management perspective, as well as from an AI distributed compute perspective, what are we looking at?
[Arch Rao] (34:56 - 35:58)
I've been on a soapbox about this for quite some time. So I think this is an easy one for me to answer. In a lot of ways, I am of high conviction that the energy and electrical infrastructure is having its time and day where it's analogous to what the telecom industry went through 20 years ago.
We are, you know, the early days of the internet, followed by the early days of everybody having a device that was connected to the internet, now having essentially multiple streaming devices. We are in that moment in time where there is an energy transition afoot. And for me, the energy transition is not about gas to electric.
That's the effect, not the cause. The cause or the real energy transition is analog to digital or electrification. So the electric grid becoming digital.
And I'd like to think humbly that SPAN and XFRA and all the exceptional folks of the team over here are working on is at the forefront of that grid digitization that comes with better visibility, better controls, better fleet level optimization, better understanding of the system level problem as opposed to a nodal level problem, if you will. And I'm really excited for that.
[Cody Simms] (35:59 - 36:11)
I think a unique way to frame the energy transition, which is analog to digital transition. And with that becomes the necessary control layers to navigate this digital world.
[Arch Rao] (36:11 - 36:41)
That's right. Compute is just one more flavor of an accelerant. Even if there wasn't compute, I'd like to think that what we were all materially solving for is a time and place where there are more electric machines than gas-powered machines.
Even just that transition would have required the grid digitization. That was the genesis of SPAN. And where we are now is we are dramatically 10x-ing, 100x-ing that load growth, which is forcing that digitization to happen much, much sooner.
And we are enabling that.
[Cody Simms] (36:41 - 36:42)
Arch, where do you need help right now?
[Arch Rao] (36:43 - 36:54)
Talent.
We are solving hard problems. We need smart people to come and solve it with us. So MCJ followers, come and join us.
We're looking for people across all fronts, technical, non-technical folks to come and join us.
[Cody Simms] (36:55 - 36:59)
Arch, this has been great. Thanks for jumping on here and it was amazing to get the update.
[Arch Rao] (36:59 - 37:00)
Thanks for having me on.
[Cody Simms] (37:01 - 37:27)
Inevitable is an MCJ podcast. At MCJ, we back founders driving the transition of energy and industry and solving the inevitable impacts of climate change. If you'd like to learn more about MCJ, visit us at mcj.vc and subscribe to our weekly newsletter at newsletter.mcj.vc. Thanks and see you next episode.