- 3 minutes ago
As AI demand surges, LT350 is reimagining parking lots as data centers. Founder and CEO Jeff Thramann explains this innovative approach to AI infrastructure.
Category
📺
TVTranscript
00:00As AI demand continues to grow, innovation is taking over, changing how and where data centers can be built, including
00:07the unexpected idea of using existing car parking lots.
00:11Joining us now is the CEO of one company doing just that, Jeff Trayman, founder and CEO at LT350. Thank
00:18you so much for being here, Jeff.
00:20Yeah, thank you. Thank you for having me.
00:22Of course, this is such an interesting idea and we're so excited to learn more about it.
00:27So tell us a little bit more about what inspired the idea of taking a whole data center and putting
00:32it into a car parking lot.
00:35Yeah, like like a lot of ideas, it it evolved over time.
00:40My background is that of a serial entrepreneur and an inventor.
00:44And I'm the named inventor on somewhere north of 130 patents.
00:48So I was out there looking at the clean energy space at first was a really early adopter into Tesla.
00:56I bought car five hundred and forty one off the lot and wanted to get into the space really to
01:02clean the air.
01:03So the original focus was about battery storage, clean energy generation, that sort of thing.
01:10But the economics just weren't there. So we waited to see what other things we could do.
01:15And as we built these canopies or came up with the idea to build these canopies, put solar on the
01:21top to generate energy,
01:23put batteries in the top of the ceilings of the canopies to store the the energy that we look for
01:29other revenue sources.
01:31Like what else could we do up there? And EV charging became one.
01:36And when you started looking at the data center space coming out, we started looking at what we could put
01:42GPUs up there,
01:43power them with clean energy and the backup storage.
01:47Yeah, I mean, it really sounds like all of that experience you have with innovation, all of these patents,
01:52all of that has kind of led you to where you are today.
01:54Now, tell us a little bit more about how the LT 350 model differs from traditional data centers that we
02:01might be more familiar with.
02:03Well, it's a lot of it's the same problems we started seeing when clean energy started coming onto the grid
02:09because it's intermittent.
02:11So it would come on and then it would be there and then all of a sudden it wouldn't.
02:15And if you're cutting down your base load and that goes away, then it becomes it becomes somewhat problematic.
02:21So you needed batteries to to back that up, to to store it.
02:26So we, you know, we're we we started looking at what were people doing out there.
02:33And we thought just with battery storage, they were building these big centers way out in the middle of nowhere
02:38where generation was.
02:40And then you you had issues with creating big solar farms, putting the batteries there.
02:45How do you transmit it back over?
02:46And we looked at the parking lots and we were bidding into some RFPs with some of the utilities at
02:52the time, Southern California, Edison in particular.
02:54And they were putting out battery storage requirements and we saw the parking lots and we're like, these are everywhere.
03:01We don't see any place else where you can site this.
03:04So if you break it up into smaller pieces and make it and make it an asset to the grid,
03:11so you're not hurting the grid, meaning that you can charge when the grid has excess energy, that would be
03:16off peak hours or in some places like California, where they're driving so much solar during the day, they're just
03:22curtailing it.
03:23They're dumping it into the ground. And we can capture that.
03:26We can capture that with batteries. So being right there to be able to capture that that energy that is
03:33getting curtailed today, I think, is a big advantage over these big data centers.
03:37When you go into someplace and you have all of these upgrades that are required from from power and transmission
03:45that just take that just take forever.
03:47So this is a lighter footprint that helps the grid.
03:50Yeah, I mean, it's something that, to be honest, I probably would not have thought of.
03:53I think so many people would have just kind of, you know, never really even gone there.
03:57So to actually hear you explain it like that, it makes a lot of sense.
04:01And it seems like it is something that could really be helpful as we move forward with so many increased
04:06needs for power and energy with all of this AI that's heading our way.
04:10Tell us a little bit more about how you maintain the parking functionality, but also while servicing the data needs
04:16and having these data center capabilities be running at full speed.
04:21Yeah, so the parking is pretty easy because we're really canopies, right?
04:26So we're above the parking lots.
04:29It's the airspace of the parking lots that we're really leveraging as the asset that we're turning into a data
04:37center.
04:37Every parking lot has these and, of course, you can't come in with a container and just take up parking
04:42spots because you have zoning requirements.
04:44Those parking spots are typically there for a reason, especially if they're attached next to a building.
04:49Nobody really wants to have parking spots, but if you're putting up a building, you've got to have a certain
04:53amount of parking.
04:54So we come in to take that wasted asset, essentially that airspace, and turn it into the parking lot.
05:00So we don't disturb the parking at all.
05:02You might have seen these solar canopies that are out there today.
05:04Ours would look almost the exact same.
05:07The only difference is between the girders up on the top where you have these big steel beams.
05:12These steel beams are a foot in depth.
05:16We would just have cartridges that go up there to carry the GPUs.
05:22Wow.
05:22I mean, in thinking about building all of this, it seems like you don't need as much space in terms
05:27of building a warehouse or something like that,
05:29but you still need a lot that goes into it to actually get these centers up and running.
05:33So what have been the biggest challenges for you as you've been on this journey?
05:38I think really the biggest challenge is finding the need that drives the margins to justify the spend on the
05:46design and engineering work to really create these modular cartridges.
05:49And that's what you're seeing today, and that's what's so exciting about the transition over to using this, not just
05:55for batteries and for energy, but also marrying it up to the GPUs.
06:02The opportunity there is incredible.
06:05The economics work really well.
06:08So therefore, you can justify the spend to design these cartridges because you're not just going to take batteries off
06:15the shelf.
06:15You're not going to take GPUs off the shelf.
06:17You have to customize those to go into a cartridge.
06:20And we're now starting to see people recognize that.
06:22You saw the NVIDIA SPAN EXPRA announcement where they started saying, OK, look, we're having all this pushback on the
06:29big data centers, so we're going to go to this distributed model.
06:33So they're putting like two GPUs in there, so they're kind of validating that you can do this.
06:38Their idea is to put them at homes.
06:40We think that might be a little bit dangerous just because of the cost of these things and the potential
06:45theft in a commercial parking lot we think is better.
06:48You get scales a little bit better.
06:50It's more secure.
06:51And you can do a little bit more with it.
06:53Yeah.
06:54Tell us a little bit about where these have rolled out so far because, again, such a fascinating idea.
06:59If people are out there, will they see these in different locations already or where are you in that process?
07:05Yeah, we're in the early stages.
07:07So we think we're about 18 months out from the first pilot being up and operational.
07:13And what we're doing in the meantime is we are identifying the location.
07:17We have that already specified for the first canopy that we would go up.
07:22We're trying to lock in the contracts to get that done because you've got to bring these utilities in early.
07:27Even on any type of thing, we're going to bring extra power and stuff into the area.
07:31So we just initiated our big design engineering effort.
07:36That's with a group called Fresh Consulting, where they're starting to really put together the pieces of this, the canopy.
07:43How do you put the cartridges?
07:44How do you connect those two?
07:45Who's going to be the supplier of the GPUs, the CPUs?
07:49Who are we going to use for the batteries?
07:51All that stuff is being worked out right now.
07:53And we expect to have the first operational canopy up 18 months from now.
07:57Well, you have a lot of work ahead of you, but that's so exciting to thinking about all of those
08:01different pieces of the puzzle coming together and really getting operational there.
08:05How do you see the future of distributed AI infrastructure really evolving over the next five to 10 years?
08:11I mean, if you guys have your first center operational in about 18 months, what do you think comes next
08:16after that?
08:18Well, I think it's just scaling and rolling it out.
08:20You know, the one thing we haven't talked about so far is really where is AI going?
08:26And when you talk about these big data centers that all the folks are out there building, most of that
08:32is designed for training the models.
08:34And when you train the models, it doesn't really matter when you do.
08:37It's not like you're in any immediate hurry to train a model, right?
08:40The data can get there on its own time.
08:42You can run the training.
08:43You can do whatever you need to do.
08:45You can get the model working.
08:46But once the model is working and enterprises are starting to use it and they're running this, what they call
08:52inference, as opposed to the training, that needs to be fast.
08:55And the only way you can be fast is you have to be close.
08:58So it takes about five microseconds per kilometer, you know, to transmit something through fiber optic.
09:03And if you're far away, it's just not going to meet the latency requirements of things like controlling a fleet
09:09of robots.
09:10That needs to be under 10 milliseconds.
09:11So how do you get that close?
09:14The best answer is put it right adjacent to the enterprise.
09:17Put it in the parking lot.
09:18And the same thing applies to security of the data, right?
09:21As you, for certain industries like healthcare, where you have to have all your data has to be HIPAA compliant,
09:26you want to minimize the runs back up to the cloud and keep it as secure as possible.
09:32If you can run your inference locally behind your own firewall through a cloud model in your parking lot, I
09:38mean, that's ideal.
09:39Nobody else can do that with this type of capability.
09:42So we think as inference grows, and I think the CAGR there is expected to be like 35% per
09:48year as we move forward because more and more people will use it.
09:52But as that grows, you're going to need to be close.
09:55And parking lots are just an ideal location.
09:58They have the power to them already.
10:00They have water available.
10:01Now, our particular cartridges are going to run on a closed-loop water cooling, so we don't use a lot
10:06of water.
10:07But it's nice to have that infrastructure there, you know, right away.
10:11And we have the land.
10:12It's not like you're fighting to find where you're going to site these monstrosities.
10:17Yeah, it really is a very unique solution that you've come up to solve this problem of needing all of
10:23this energy.
10:24And it's very exciting to see where it goes next.
10:26So thank you so much for breaking all of this down.
10:28We really appreciate it.
10:29Jeff Trayman, founder and CEO at LT350.
10:32We can't wait to see what's next.
10:35Yeah, thank you.
Comments