Subscribe
Share
Share
Embed
Hosted by Beth Rattner of Endeavour, Beyond the Grid is a podcast dedicated to reimagining the critical infrastructure that powers our rapidly changing digital world. As exponential AI growth pushes conventional power grids and finite resources to their limits, we sit down with the bold thinkers, engineers, and leaders building the solutions of tomorrow. From multi-gigawatt data centers and waterless cooling to decentralized energy hubs and digital twins, each episode cuts through the noise to explore how we can build resilient, community-first ecosystems. If you are ready to ask what’s next, what’s possible, and how we actually build it, you are in the right place.
00:00:00.000 — 00:00:06.520 · Beth
Hi everyone. Welcome to Beyond the Grid. We are here live in in person with James Feasey
today.
00:00:15.840 — 00:00:35.840 · Beth
Every company needs a suave British guy who can rescue people with, you know, a ballpoint
pen. That's our James Feasey. So we're very excited to have you on the show today, James.
And while you may not be rescuing people with ballpoint pens here at Endeavour, I am so glad
that you agreed to do this podcast today.
00:00:35.880 — 00:00:39.320 · James
I used to prefer underwater knife fighting, but is that.
00:00:39.360 — 00:00:40.920 · Beth
Is that really what it was?
00:00:41.280 — 00:00:44.720 · James
No, but it gives a good, you know, good visual.
00:00:44.760 — 00:00:54.400 · Beth
Exactly. Now I have a whole different image of you, so we'll go with that. But when you're not
doing that underwater knife fighting, let's say, um, what do you do for Endeavour?
00:00:54.520 — 00:01:01.050 · James
So here at Endeavour, I look after strategic sales and our go to market strategy for AI
infrastructure.
00:01:01.090 — 00:01:14.010 · Beth
Okay. And for people who aren't as familiar with what that might mean, how would you
describe as your main? Are your clients, cloud players and specific, uh, engineering buyers
within those cloud players?
00:01:14.050 — 00:01:57.290 · James
Yeah. So the focus is sort of twofold. It's one in terms of like an outbound articulating the
value of the technologies we've developed and the solutions that we've packaged together.
Um, to customers or would be customers coming from the technology space, whether that's
GPU as a service or traditional hyper scale players.
Um, it's also ensuring that input and feedback they have is fed back into our product teams.
And then the those products and services that we work to develop are effectively articulated,
um, at a marketing level. And, you know, throughout that customer journey.
00:01:57.450 — 00:02:05.860 · Beth
I love that. It's otherwise we're just in an echo chamber, creating things that we think
somebody might want one day without asking.
00:02:05.900 — 00:02:23.620 · James
Yes. Just making it easier for the customer to understand the value in terms that they care
about versus talking about a piece of equipment, products and, you know, products. Feature
set. Um, we try and turn that into what the value it is that it unlocks for the customers.
00:02:23.620 — 00:02:29.700 · Beth
And how did you how did you get here, maybe to infrastructure as a profession in general?
00:02:30.860 — 00:02:43.140 · James
Yeah. So I was really fortunate. I came to the States about ten years ago, 11 years actually
coming up on now. Time goes fast. Um, plus I, we had those two years of Covid that I always
kind of forget.
00:02:43.180 — 00:02:43.620 · Beth
Exactly.
00:02:43.700 — 00:04:04.640 · James
Maybe we'll get them back at the end or something. Um, so I was fortunate enough to get
picked up by a company called Corning Optical Communications. The people who make all
their optical fiber. And they kind of gave me a crash course in technology. So this is that the
ozone stack level layers one and two and three.
So networking and layered networking. And from there I was working with a customer that
was in data centers. And I thought, wow, you know this is really fascinating. Co-location was
getting bigger and bigger at the time. And I had the opportunity to move across there. And
then I got to this data center company and they had a cloud computing arm, and I thought,
wow, you know, that looks really cool.
So I, I went over there and started working very closely with engineering. And while I was
there, I was fortunate enough to work with the former director of research computing at
Novartis. So back in 2016, 2017, we were building, um, scale up AI systems for high
performance compute in healthcare life science.
So it was early into this. I remember him telling me I should be buying Nvidia stock when they
were mostly just known for gaming. Did you do it? Not enough.
00:04:06.040 — 00:04:08.000 · James
A very small amount. You know you.
00:04:08.000 — 00:04:08.880 · Beth
Wouldn't be here right now.
00:04:08.920 — 00:04:15.160 · James
Yeah, exactly. Um, so, you know, that was really interesting. And then I just kind of kept
chasing,
00:04:16.200 — 00:04:25.360 · James
um, going up and up the stack, because that seemed to be at the time where more of the
action was happening. Um, infrastructure,
00:04:26.400 — 00:05:15.540 · James
you know, while it moved, it just wasn't moving at the rate of software development or the
things that you could be doing. So, you know, better code could overcome a multitude of sins
at the infrastructure layer, um, rack densities, you know, bring it back to a data center
standpoint. Yes, they were going up, but they, you know, weren't really going up by all that
much.
It was very, very incremental. And really the rate of pace of change that I've always found
appealing was happening. You know, at the higher orders of the stack at the software layer.
Um, so I then went from there. I had the opportunity to go client side and lead technology
operations at a healthcare company.
And that was, you know, really eye opening, being on the other side of the the vendor
landscape, seeing what made good vendors, what made bad vendors, how they could make
your life easier.
00:05:15.580 — 00:05:20.140 · Beth
I didn't know that about your that would really make a difference. What's an example of a
bad vendor?
00:05:20.860 — 00:05:23.180 · James
Um, transactional ones.
00:05:23.300 — 00:05:27.940 · James
With very short term mindsets? Um, people who.
00:05:28.220 — 00:05:29.060 · James
Also.
00:05:29.580 — 00:06:39.390 · James
Let the customer have too much choice. Now, don't get me wrong. You know, letting the
customer choose is really important, but particularly if you are the innovation company. So if
you're a cloud company and you are innovating and building these products, you should have
a defined point of view on what good looks like for that product, for that service, how you
should implement it, how you should use it, not sell a brand new service or technology to a
customer who is trying to very actively learn and catch up about what this new technology or
service is, and you don't offer them quite prescriptive guidance.
So I actually felt like what made very good vendors were people who had a point of view
about their products and services, particularly if they overlapped, which ones you should use
and why and how best to implement it. You know, that really helped from the customer side.
Um, ensure that you extract a maximum value from what you were spending money on.
So that was one of the sort of key lessons I learned there. And then from there, I ended up at
AWS, where I led go to market strategy for high performance and accelerated compute.
00:06:39.590 — 00:06:45.510 · Beth
Amazing. Is your job now very similar to what you had at AWS or is it pretty different?
00:06:45.550 — 00:06:46.190 · James
No.
00:06:46.550 — 00:07:14.400 · James
Um, it's it's it's different in a way, because, Um, we're a smaller, more agile company, and so
my ability to implement rapid change and effect change in a short order is far simpler here,
um, than at such a large company. And, you know, they did an exceptional job for a company
of their size. But, you know, at some point there's something really appealing.
00:07:14.480 — 00:07:27.040 · Beth
But you have to become more bureaucratic the larger you get. So do you actually have a say
over the product development team? Can you go in to the Turbo Sell team? Say, you guys, this
is really important. Yeah.
00:07:27.040 — 00:07:55.490 · James
So I'm really fortunate to work with some brilliant engineers who are far more intelligent than
I am. So what I try and do is just take pain points or feedback that we've received and ensure
that it's embedded into go forward design. And I work really closely with folks like Wes and
Chris Ellis, um, to ensure that, you know, that feedback is captured.
And as we're driving towards something we bring to market. It's being, you know, effectively
incorporated into the the solution we provide.
00:07:55.890 — 00:08:09.330 · Beth
Yeah. That's fantastic. That quick feedback loop. It's a very biomimicry thing, by the way.
Nature survives on quick information flows and quick feedback loop. So if we can do that here
all the better.
00:08:09.370 — 00:08:25.770 · James
Yeah. Yeah. So that's you know, one of the the really interesting pieces. Also I think if there's
one other difference. Yeah. Is the, you know, one of the things I saw was seeing was I spent
my whole career moving away from infrastructure going up stack.
00:08:25.810 — 00:08:28.130 · Beth
That's right. I was going to ask about that with.
00:08:28.650 — 00:09:00.620 · James
You know, what had happened or what was happening with AI and the impact that was
having. That value chain was actually collapsing back down to the hardware and the
infrastructure level. And now infrastructure was becoming critical and strategic and was
ultimately becoming the rate limited discovery.
So, you know, the other really interesting piece of this role is I think I'm right back, uh, dealing
with something that is at the very crux of the problem we're trying to solve if we're looking to
accelerate AI outcomes.
00:09:00.660 — 00:09:12.820 · Beth
I want to go back to my reference to you being the James Bond of Endeavour, of course, but
maybe talk a little bit about. So even before you were in tech, you were in Special Forces.
What was your job?
00:09:13.060 — 00:10:23.670 · James
So I was a British commando officer, and I was really fortunate to be part of a specialist
maritime counter-terrorism team. So we focused on international maritime counter-terrorism
and counter piracy. So it was a very dynamic role, as you can imagine. Very enjoyable role.
Yeah. Um, but, you know, one of the things I think that it teaches you, um, being in the
commandos or elite infantry anywhere is the ability to very rapidly synthesize information
and actually determine what is critical in that information.
And so what I like about moving into very fast paced environments, it's number one, you're
trained to have very good frameworks to to learn things very quickly. So being able to learn
things quickly is important when things are changing very quickly. And number two, in a
world where there is a lot of noise actually cutting through that to determine, you know,
which pieces of those of that noise is valuable, you know, and how do you message that and
ensure that, you know, you're really driving towards the critical piece of information versus
just not.
00:10:23.910 — 00:10:25.310 · Beth
Because you can't possibly respond to.
00:10:25.470 — 00:10:28.990 · James
It. You can't possibly respond to all the things you know. It's so multi-threaded.
00:10:29.030 — 00:10:32.870 · Beth
I hadn't thought about that. It's a great analog for what you now.
00:10:33.270 — 00:10:33.830 · James
You know,
00:10:35.070 — 00:10:51.400 · James
of the 20 possible pieces we could look at to, you know, move the needle and normally all of
those 20 things can have impact. Yeah. You Which one of those things is the thing that we can
move the needle on fastest and soonest?
00:10:51.400 — 00:11:05.880 · Beth
So do you think? Well, I often quote this. It's a question of sequence that they will ultimately
all be important. But the sequence matters that you go after and do a do you agree and
maybe.
00:11:06.360 — 00:11:40.520 · James
Yeah yeah I agree. I mean, there are so many things we can do to chase performance and
efficiency improvements in AI infrastructure. Some of the things are more easy to do than
others. Some of them have more mature technology ecosystems existing them around them
today. So, you know, normally that's a good place to start.
Some of them will have bigger impact, um, than others. So all things will have an impact. But,
you know, it's moving the big rocks first and then and then dialing in from there.
00:11:40.560 — 00:12:37.170 · Beth
Do you think it's easier to solve multiple problems at once, though? Sometimes because it
could be that a different paradigm is required. I'll give you an example. Um, in the in the
world of biomimicry, fans have sort of been the same design for a long, long time. And it's it's
all about pushing air at greater and greater levels of efficiency.
What's changed is there was an innovator who realized that nature often pulls through
vortices rather than pushing. And so he created something that looks a lot more like a
corkscrew that pulls. And the fan efficiency went massively up, the noise factor went way
down. And you had but you had a radically different design.
And I think it's because he was trying he was definitely trying to emulate nature, but he was
also trying to sort of solve for multiple problems at once efficiency, cost and ultimately and
performance, but also sound. Do you think that sometimes it does help to look at multiple
factors at once? Oh, it.
00:12:37.210 — 00:14:12.720 · James
It definitely does. I mean, I'm just talking about if you want to if you're looking at a particular
problem, There's normally one lever you can pull to get the impact soonest, but definitely in
the environment we're in co-design and co innovation is the name of the game. I mean if we
look to how we got here, particularly in the data center industry, you had people developing
hardware, the best AI hardware they could possibly develop, and they developed that not
entirely in a silo, but they developed that at a rate that it basically outstripped the the rate of
change in infrastructure.
So now you have a, um, company that manufactures chip IP in Nvidia. Then being very, very
proactive in terms of encouraging the entire ecosystem to collaborate and design and
innovate to solve the problem of getting their hardware into customers data centers. So I
think in the age of AI, We've realized that you can't look at these things in silos.
You've got to have vertical integration. You've got to minimize hand offs. You've got to
understand the impact across a grid to chip ecosystem in order to optimize for both
performance and efficiency. So we we definitely do need to look at a lot of things in tandem.
Because if you make one decision in a silo, it can have massive ramifications, you know, across
that entire continuum.
00:14:12.760 — 00:14:29.520 · Beth
That makes perfect sense. So maybe with that in mind, what's you know, what's coming.
We've heard these different projections, um, around 945 terawatts of power being used by
2030. What are your thoughts?
00:14:30.640 — 00:14:38.040 · James
Yeah. So my thoughts are that that is absolutely going to happen. So you know what.
00:14:38.040 — 00:14:39.880 · Beth
Is be more than double where we are today. Yeah.
00:14:39.920 — 00:15:03.930 · James
More than double where we are today. Um, and there's, you know, a strong reason for that.
And that is AI scaling laws. So the AI scaling laws really underpin what we're seeing in terms
of this unprecedented rate of investment and build in the infrastructure world. And that is
that if you scale a number of things, namely model parameters,
00:15:05.090 — 00:16:26.820 · James
data set size and compute, your AI models get better in a predictable, consistent fashion. So it
is that predictable and consistent fashion that is encouraging these tech companies to spend
hundreds of billions of dollars on building bigger and bigger compute clusters, because there
is return in terms of value and performance of their model.
And you know, that improved performance of the model that makes a difference between
being a market leader and not being a market leader. These companies haven't historically
spent from a CapEx basis anywhere near the amount of money they're spending today, and is
really those AI scaling laws that are underpinning this investment and driving them to
continue to do this.
And so that's not going to slow down, because actually, we haven't got to the point in the in
AI scaling laws where we see diminishing returns. So at some point when and the law does
state at some point those returns do diminish. But we are so early in this process that we're
certainly not seeing that. And so when those returns start diminishing, we'll see a slowdown.
But you know, what that means is in the near future, we are not going to see a slowdown, in
my opinion, in terms of the rate of development and building that we're seeing.
00:16:26.940 — 00:16:29.300 · Beth
That could be higher than 945 even.
00:16:29.500 — 00:17:11.230 · James
It absolutely can be high. Yeah, I think we're doing things, you know, and some of the reasons
it can be higher is the assumptions made in that model. You know, traditionally it took three,
three and a half years to build a data center. Yeah, we can now build data centers. And, you
know, actually the longest pieces the the planning, the permitting.
Yeah, the actual building can be done in months, not years. And so, you know, when that
starts happening more and more and we continue to innovate, to drive speed, to market both
at the infrastructure level, at the hardware level. Um, I think those predictions will end up
looking conservative.
00:17:11.790 — 00:17:56.400 · Beth
So the thing that I, you know, I left 20 years of sustainability to come in to work in a data
center and an infrastructure company, a solution company, really. And one of the things I get
excited about is could that rapid demand actually drive alternative renewable energy
sources? Because at the end of the day, it doesn't coal has, uh, doesn't pan out or doesn't
pencil anyway.
And so turning back on all of our old coal plants probably wouldn't even meet the need.
Regardless, at the cost we need. And everyone sort of recognizes the environmental cost. But
do you think that do you think that heightened demand will actually be a beneficial driver for
new energy sources?
00:17:56.960 — 00:18:10.880 · James
Yeah, I think it is going to be a beneficial demand for new energy sources because energy is
becoming a very, very finite resource. Yeah. We are also going to see a big move towards
distributed versus centralized energy.
00:18:10.960 — 00:18:15.600 · Beth
Say more about that. Yeah. They are basically regional energy hubs.
00:18:16.040 — 00:18:20.800 · James
Local on site like at point of data center. Yeah, yeah. Energy.
00:18:21.000 — 00:18:23.320 · Beth
Um, you will be your own substation. You will?
00:18:23.360 — 00:19:42.380 · James
Yeah. Correct. Transmission is extremely difficult to build a transmission network from a
central generation generation plant out to a data center much faster rather than, you know,
waiting for that transmission system to be built. Also, there's loss in that transmission system.
If you just bring the power direct to where the data center is.
It, you know, drastically reduces your time to market. And then really what it's going to come
down to in terms of that is the most efficient way to deliver or generate energy. So we will see
a much broader mix that is going to take into account that drive towards efficiency, energy
efficient energy generation.
So, you know, in some of the places we're looking at, solar's going to play a big component in
that because you're going to have periods of time in the right conditions where solar is going
to be unmatched in terms of a cost basis for energy generation. You're still going to need
baseload power. And we have a number of technologies that are going to help with that.
But it is going to be definitely a catalyst for introducing more and different methods of energy
generation than our traditional centralized models of generation.
00:19:42.420 — 00:20:01.980 · Beth
That's right. Going back for a second, as you think about these increasing demand and
increasing scale, how is that landing with communities? You recently had a town hall. You
attended a town hall meeting where our one of our data centers was being approved. What
was the temperature in the room? What did you hear?
How did and how did you respond?
00:20:02.020 — 00:21:27.030 · James
Yeah. So I think the temperature in the room was was mixed. You have people who are
extremely excited. I think people can see the the impact that bringing these things has to a
community in terms of investment, jobs, just the sheer scale, the length of periods of time
that these things will now be being built over.
You know, they're much more long term in terms of immediate impact from a, um, from a
jobs job creation standpoint. Um, I think also the fact that these facilities attract high tech
companies that are willing to make investments in the community is also something very
beneficial. But, you know, right alongside that, there's a lot of bad press about data centers.
Um, and people are understandably nervous about having data centers come into their
communities. I think part of that is actually because now things like power are a finite
resource. It's driving a very different lens in terms of power efficiency and power utilization
than there traditionally had been in data centers where power was plentiful and actually
chasing those incremental efficiency gains, it was just cheaper to get more power and being
efficient.
You know, we've left that world behind. But from a, you know, from a power standpoint, from
a cooling standpoint.
00:21:27.070 — 00:21:36.810 · Beth
And so that's really interesting. I just have to pause it. That's really interesting. That is a that's
such a different reality than we've ever been in. You can't there's not just more to get.
00:21:36.850 — 00:23:21.980 · James
That's not small to get. And when there isn't more to get, you know, you start focusing on
how to be as efficient with utilizing that finite resource as you can possibly be. Yeah. Um, and
you, you hear, talk about that, um, you know, in terms of optimizing the number of tokens
you generate per watt of power, you know, how do you do that?
That's actually the mark of a for an efficient facility. And so we're starting to see that more
industry wide. Of course here that in devel we've been doing that for much much longer than
that. It's always been a hallmark of what, you know, Jake set out to do even before he was
here at Endeavour. So maybe rightly so.
Data centers have got a bit of a bad rep, and there's still actors out there who are doing things
the most expedient way. Um, you know, example of that is the most, the cheapest and most
efficient way. Traditionally, to cool a data center has been to directly evaporate water, use
some sort of direct evap technology that consumes millions and millions of gallons of water a
year.
Yeah. You know, now at the scale we're building, at the size we're building, that is utterly
unsustainable. Yeah. Um, and now people aren't just asking for a little bit of water. They're
asking for, you know, monumental amounts of water. We aren't building 30 40 megawatt data
centers. We're building multi gigawatt data centers.
So, you know, it's fundamentally different in terms of those resource demands. And so water
was actually probably the biggest thing that, you know, stakeholders in that town were
concerned about in terms of a data center development coming in. So having, you know, ultra
efficient, uh, waterless cooling technology was an extremely beneficial way to placate.
00:23:22.180 — 00:23:22.780 · Beth
They didn't know that.
00:23:22.820 — 00:23:23.700 · James
They didn't they didn't know that.
00:23:23.700 — 00:23:25.420 · Beth
They didn't know that we had waterless cooling.
00:23:25.420 — 00:23:27.220 · James
They didn't know that. No. Because
00:23:28.470 — 00:24:44.880 · James
there aren't really articles, sensationalist articles posted about the evils of data center that
actually say, oh, not all of them do, you know? They kind of focus on, again, those bad actors.
And, you know, we are very early in that space. Yeah, we we spent a lot of time convincing
people that actually waterless is the way to go.
And you can drive the same efficiencies you can with, you know, even the most efficient direct
evap systems. Um, but again, when we get into these scenarios where speed to market is
important and efficiency of resource utilization is important, it's kind of reframed that
argument, you know, with waterless cooling, we don't need to go out and get those permits.
We don't need to go out and secure that water source. It's a dependency we've now taken off
the table. It's huge. It's a yeah, you know, permitting piece that we don't have to jump over a
hurdle. We don't have to jump over anymore. Um, and we don't take any, you know,
performance penalty because of doing it.
And so it was really nice to actually be able to get rid of 90% of the people's concerns about a
data center coming into town with a, you know, technology that we've pioneered, developed
and brought to market and now are using it ourselves.
00:24:44.920 — 00:25:06.640 · Beth
That makes me happy. And I'm also excited about the idea of us capturing water, even
certainly in drought ridden regions like Texas or the southwest in general, or even the
American West. But also just why not bring water where we can? You know, do you think that
we can sort of counterbalance the the bad press that's out there?
00:25:07.040 — 00:25:24.969 · James
I think we can, you know, whether everybody does. I think we're going to reach a point where
you're going to have to because even just talking about scale, accepting something, one data
center that's a 20 megawatt data center in your town is a very A different
00:25:26.290 — 00:25:36.570 · James
kettle of fish to accepting a five gigawatt data center in your backyard. So there is going to be
a lot more impetus.
00:25:36.570 — 00:26:02.290 · Beth
I think we're the thing I get most excited about is that we can move to sort of a positive.
Maybe it becomes energy positive in a region, water positive in a region. But just because we
showed up, right, just because we chose not just to be waterless in terms of our cooling, but
maybe also capture rooftop water, maybe also do desiccant meals.
We can be a resource back to that, that community. Do you think that that's something that
we can convince others of?
00:26:02.290 — 00:27:08.700 · James
I think we can definitely continue to lead the market in that. I think others are going to have
to be convinced. It's quite easy to hide a 15 20 megawatt data center in a community. I'm sure
people all over the place already live pretty close to data centers and have no idea that
they're living there. But when you start building at multi gigawatt scale.
You can't hide that. You cannot you know. So it's going to be very obvious and it's going to
have a lot more optics on it. In terms of the local community. And so anybody who wants to
build at that scale and wants to build with a heavy impetus on speed to market, is really going
to think about how you overcome these community concerns.
We're going to have to work much, much closer with communities in order to keep on driving
the AR, AI outcomes we're driving towards. So that is definitely going to lead to data centers
in general across the board becoming better custodians and looking to actually bring much
more lasting value to the communities that they are building in.
00:27:09.460 — 00:27:26.390 · Beth
So speaking of that, what's the most enticing use of waste heat that you can think of? I mean,
we heard about like at the Olympics in 2024, there were heating pools with district heating.
What? What do you like for waste heat?
00:27:26.870 — 00:28:55.120 · James
Yeah, I think rather than wasting it, you know, reusing any waste product or any waste stream
from a data center is really important. That did a great job of, you know, raising awareness for
how to do it. I think the next iteration of that is going back to how you were talking about not
wanting to do things in, in silos or there was, you know, actually, you need to look at
something as a whole rather than looking at a data center being the the entire scope of a
project, you actually start day one master planning these campus environments to be
multiple use campuses.
So it's not just a data center out in the middle of nowhere. Because the thing about having,
you know, capturing waste heat and then having the ability to use it is you need something to
use it on to make that make sense. And so it's got to be a mind shift change away from just
building data centers to looking at, you know, building more holistic communities where we
can take that waste heat and we can use it for district heating, um, or, you know, other
beneficial uses.
There's ones with greenhouses as well that we could look at, but I think those things have to
be looked at day one rather than, oh, build the data center first and then we'll worry about all
that stuff afterwards. Because inevitably if you leave it to afterwards, it becomes infinitely
more complex to do it.
Then if you just masterplan it and do it up front.
00:28:55.160 — 00:29:28.250 · Beth
It's yeah, we'll data centers continue to get bigger and right centralized. So energy systems
may be decentralized around these data centers, but will the data centers themselves, you
know, edged as a term came up with this idea of smaller data centers regionally located near
urban environments. But now we're seeing all of the data center anywhere in the middle of
nowhere.
If I can make it a two gigawatt campus or more. What do you think of the future? Do you
think the future ever comes back around?
00:29:28.370 — 00:30:14.170 · James
It definitely does. We've seen this in computing time and again, you know, from
decentralized, centralized, decentralized, centralized. Yeah. Um, I think in the near term, we're
going to continue to see very large centralized campuses, just because that's where you get
maximal benefit in terms of AI cluster usage.
But, you know, talking to customers like meta, they're looking at now rather than doing, you
know, if we just rewind the clock two years ago, a cluster was a subset of a data center. Yeah,
three years ago now, you know, it was a subset of a data hall. Then it became a data hall.
Then that cluster became a data center.
Yeah. Now that cluster is a campus. You know, the next thing they want to look at is how do
we then make, you know, multi-campus clusters.
00:30:14.450 — 00:30:57.270 · James
City. And so that starts Realizing that as well. Also, as we move away from a time that is
dominated by AI training. So, you know, AI clusters, you know, pumping out better and
improved AI models to the uptick in use case and integration of those, um, of AI. We all see
this sort of division between where we go and do training, you know, probably still
predominantly at centralized large data centers and where the inference happens, you know,
closer to the edge, closer to the people using it, you know, closer to those use cases.
00:30:57.790 — 00:31:54.000 · Beth
Interesting. You you just made me think of. So in nature, they're called ecotone. So when the
forest meets the the river, that's the richest in terms of biodiversity. Wherever two
ecosystems come up against each other, that ecotone is fraught with life. And I'm wondering
if the future is if we get good at using waste streams.
Not even just our own, but also food waste, plastic waste, fashion waste. And those things
start to subsidize the bill for a data center, even. And we create new materials out of it with
our waste heat, whether that's from greenhouses or algae ponds, new, you know, new
practices, and we're water positive.
I wonder if there could be a richness of eco tones created because we've centralized all of
these resources, and then the cities and the urban use cases are right there. I wonder if
there's an analog analogy there.
00:31:54.040 — 00:33:04.530 · James
Yeah, I think with what you're talking about, it would be nice if there's still heavy
centralization, because ideally what we're trying to create is bring value to a place where
there probably wasn't value or as much value before versus in a city, you might have elements
of that value, um, already at place, you know, those broader ecosystems.
Urban planners have been thinking about that for a long time. That's true. Some of the
centralized places we're going, you know, maybe they haven't had any of that benefit. And we
can be the catalyst as those large centralized campuses to bring it. Um, but but I but I think
over time, you know, we will see there's a heavy drive to be more flexible, more agile.
Right now, we're limited by the technology sets that we have available us today from a
networking standpoint in terms of how we can do training. Um, if we scale that wall, you
know, we'll we'll increase decentralization again. But then we want to be careful that we've
built these large centralized campuses that they continue to have, you know, lasting impact to
those communities and ecosystems we've built around them that value.
00:33:04.570 — 00:33:23.180 · Beth
That's right. Are you. So again, back to James, the prognosticator here. Um, what are some of
the. I can go either way. Let's start with what's a leapfrog technology that you think right now
might be a little bit in the science fiction category, but will probably show up as real.
00:33:23.180 — 00:34:25.590 · James
So I think the leapfrog technology that is it isn't science fiction. It's more real than that, but
it's not next 2 or 3 years real. It's maybe a little bit longer than that is going to be new
networking technologies that allow us to break up, break out of this scale up copper domain
that we're trapped in. So right now, you know, a lot of what's driving these densities is the,
you know, limitations of an alternate to copper for scale up networking.
And so we've got to keep that power efficiency and that performance that you need for GPU
to GPU networking in copper. That's leading to bigger racks and denser racks. Yeah. Um, that's
been driving a lot of what we've been chasing for the last 2 or 3 years in the AI infrastructure
space, because we've got to catch up with, you know, these rack densities that have all of a
sudden gone from 15, 20kW of rack into the hundreds of kilowatts of rack, and now we're
talking megawatts of rack.
00:34:26.510 — 00:34:28.710 · Beth
By the way, is that true megawatts per rack.
00:34:28.710 — 00:35:23.880 · James
It's coming in, you know, the not too distant future on current trajectory. It's crazy where
what some of these new networking technologies will enable us to do is disaggregate racks
again. So go away from building bigger, like physically bigger and denser racks to new
architectures where we can now spread that scale up domain across a larger footprint in the
data center at lower power consumption.
And we'll probably find, you know, some happy medium where we get optimal price
performance. And I'm just imagining it's probably not going to be at, you know, having
multiple two megawatt racks. Yeah, I think from a price performance standpoint. We're going
to find its. It's below that. So I think we're going to see rack densities go up and up and up.
And then we're going to reach a point where we actually start to see that plateau and come
back down with some of these new networking technologies that are coming.
00:35:24.000 — 00:35:45.320 · Beth
Sorry, I'm like peppering you with so many biomimicry terms, but they're really just coming
up. Another sort of like truism in biomimicry is that nature optimizes it doesn't maximize. So I
feel like a lot of what humans try to do is we maximize an opportunity, but nature's actually
always looking for that kind of that appropriate balance or optimize optimization path.
And that's to be what you just described.
00:35:45.360 — 00:36:35.370 · James
Yeah. So right now we're just I think going back to what I was saying earlier about with the
technologies we have available, how can we have the most impact that normally comes first?
Right. So we're in that we're in that phase where we're just chasing performance. And then
when we've got that nailed down, we start looking at, well, how can we get that performance
with better efficiencies.
And so I think, you know, this is going to be one of those technologies that is going to
fundamentally shift a lot of what has been driving the data center industry over the last
couple of years, and it'll continue to drive it for the next couple of years. But it's going to be
really interesting when we get a technology that like that, that allows us to break out of these
scale up, these copper scale up domains.
Um, you know, that's going to be the next really interesting moment in AI infrastructure
ecosystem development.
00:36:35.410 — 00:36:53.330 · Beth
What do you think within within our sector or maybe just within this realm of sort of the rise
of AI in general or data centers in general, like what's a core issue or problem or challenge
coming for a society that maybe not everyone has their eye on? Yeah. Even if we don't know
how to solve it.
00:36:53.370 — 00:39:25.730 · James
No, I think we don't. I think it's articulated in a lot of ways that there's, you know, we've got a
transformational technology here. How do we use it to maximize societal impact in a way
that's positive for humans. I think if we look at, you know, what happened in the era of the
internet, the iPhone, there is, you know, undoubted benefit.
You know, a lot of societal benefit that's come out of that. But on the flip side of that, there
have been a lot of negative consequences. You know, particularly with social media, if you
look at things like depression rates and some of the concerns about people becoming very
insular, not being great at, you know, communicating there's been impact in terms of
polarization, because now algorithms will just feed you more of what you want to hear, you
know, and it keeps you going down that rabbit hole and further or further away from a, you
know, sort of centralized viewpoint.
You very rarely get exposed to other arguments or viewpoints. Yeah. Um, unless it's even
done. So just to carry on getting you excited about, you know, your particular end of the
spectrum you sit on. So I think there's been, you know, an undoubted amount of good that's
been done for society by the internet and widespread, um, very easily accessible media.
But there's also been a lot of negative consequences because I don't think we knew how to
deal with it. I don't think we knew how to gear up a generation of humans who were going to
experience that. I mean, I myself grew up before the age of there was internets and cell
phones, but, you know, internet was a very slow dial up connection you couldn't really do
much on.
It wasn't very exciting. So going from that to now, being in a society where it's at my children's
fingertips from the moment they're born, they're exposed to, you know, the entire internet,
how do you how do you deal with that? It's really challenging as a parent. it's it's even more
challenging from a government level.
Um, with new technologies coming online to ensure that, you know, when these technologies
are do come to bear at scale, they're actually benefiting humans and not, you know,
detracting from the human experience.
00:39:25.730 — 00:39:27.490 · Beth
Even if we don't know how to solve for that yet.
00:39:27.530 — 00:39:48.010 · James
Even if we don't know how to solve for it, we should be thinking about it. We should. We
should be thinking about no. Where does that? Where do we strike that balance between um
and who who gets to control that balance? You know, these are very challenging questions,
but it's something that we yeah, we should be thinking about.
00:39:48.050 — 00:40:14.860 · Beth
These are these are, uh, philosophical questions almost. Um, you mentioned your kids. And
this brings me to my last question, I think, which is what do you think their jobs are going to
be? Or what do you think the their relationship to technology. Overall, I make jokes to my
nieces all the time, like, you're going to have to help me reset my hologram machine, which
I'm sure I will have, and they will.
You will mock me for my inability to use it.
00:40:14.860 — 00:40:25.700 · James
So I think that technology is moving faster than robotics, so you're probably better off doing
something with your hands. Yeah.
00:40:25.740 — 00:40:29.340 · Beth
So I think learn how to carve, learn how to weld.
00:40:29.340 — 00:41:02.790 · James
I think something that is going to make you. If I'm looking cynically at this harder to replace by
technology. But on the flip side of it, I actually think in a world where we spend so much time
on the other side of screens, just consuming content and not doing. Yeah, just passively
consuming. Yeah. At least by having a job where you're physically doing something.
Yeah. With your hands I think will be extremely rewarding. Um, especially going Forward.
00:41:02.870 — 00:41:14.110 · Beth
There was a book written. I think it's called The Case for Working with Your Hands Why Office
Work is Bad for us. But they didn't even talk about the future of sort of like advanced
digitization. But yeah.
00:41:14.510 — 00:41:20.350 · James
Yeah, I mean, I had a previous job where I worked with my hands and, you know, that was
extremely rewarding in a way that.
00:41:20.390 — 00:41:25.070 · Beth
Our cognitive skills are are wired for using our hands.
00:41:25.110 — 00:41:31.230 · James
So it's hard to replicate that level of reward without doing something with your hands.
00:41:31.270 — 00:41:33.830 · Beth
So do you think you'll go back to it? You think you would go back to?
00:41:33.870 — 00:41:35.270 · James
No, no, I'm good here now.
00:41:36.510 — 00:41:38.070 · Beth
No more underwater knife fights for you.
00:41:38.110 — 00:41:39.310 · James
No, not for me.
00:41:39.470 — 00:41:46.510 · Beth
Oh, James, this is such a pleasure. Thank you for being on the show with us. And I'm really
grateful that you're part of Endeavour.
00:41:46.870 — 00:41:47.830 · James
Thank you. Thanks for having me.
00:41:47.830 — 00:41:51.470 · Beth
Yeah, it's a pleasure. Thanks, everybody. We'll see you next time.
______________________________________________________________________________