Hello everyone. Welcome back to semi analysis  weekly. I'm here this week with the data center  

energy and industrials team. Uh that's Jeremy,  Ellie, and Rake. We're going to talk about  

a couple of articles they put out recently.  The first is stop saying half of 2026 US data  

center capacity is canceled. Nice title. We'll  clarify some stuff there. And the second one

pretty descriptive. Yeah. The second one is  about uh behind the meter data center uh power  

generation 40 gawatts by 2028. Guys, welcome  to the show. Thank you Jordan. Yeah, thanks  

for having us. We're covering three continents on  this one. We got a break in Singapore. The latency

we'll try. I never s analysis, bro. Normal day.  Yeah. Okay. So, let's let's dig in. Um, the the  

first article had a great title and obviously  the conclusion didn't bury the lead at all. Um,  

but I've been seeing this all over the media  where everybody keeps saying that half of data  

center capacity is canceled. You guys walk  through maybe reasons why people are saying  

this and then clarified some stuff as to like  what's actually happening. Um, can you give a  

lay of the land as to what's reality, what's fake?  Yeah. Um I can start with the why. Uh the there I  

think Bloomberg started this right with this big  headline of like half of US data center capacities  

delayed. Everyone else started piling in and  citing the same number and you know that most of  

these articles point to the same underlying source  right which uh is a report that's available out  

there. And you know when you look at the reports  the report says it was 12 gawatt of data center  

capacity scheduled to go online in the US in  2026 and only five is under construction right  

and I [laughter] think you know for us when we saw  this we're like why why is everyone talking about  

this like just it's just not possible uh and and  you know I think the funny thing is that like you  

don't even need to like do anything fancy you can  just disprove this data so easily like I I don't  

know man. Amazon announced publicly they built 4  gawatt in 2025. Is that going up or not in 2026?  

Obviously it's going up. Amazon alone is probably  going to add like you know uh 5 gawatt plus uh  

so that's basically all they think is going to be  going live like one company. Uh you know obviously  

got all the hypers skaters. Coreweave coreweave  you know is going to add a gigawatt in 2026. All  

of that is under construction. So you're telling  me core weave one gig under construction is  

like you know a fifth of the market like come on  there's a massive issue in the denominator you're  

just wrong just don't publish that. So yeah that's  a high level if you want to add more deeds but for  

me I think that's the that was the trigger is just  looking at the underlying source you're like oh  

guys you're just off it's just wrong. Yeah, we had  seen this getting like paraded around quite a bit  

and you know, I mean like everybody would go viral  just kind of reposting it. I was like, "My god,  

this is this is mind-numbing to see every time on  my Twitter timeline." Um, and obviously like our  

clients are pretty smart. Our clients are going to  take every like data source they can and compare  

them against each other. So, we also had to help  every client on that front. And it was like, well,  

why don't we just write a newsletter about this  and say like, "Yo, this this is just really fake.  

we don't have to really go into this as much  anymore. Please just read the newsletter. Um,  

a lot of the kind of baseline here is not worth  really talking about sometimes. Um, yeah. So,  

that's why we that's why we ended up writing it.  Yeah. And in addition to that, like um it's not  

even like the spirit of the article was fake in  a sense. Like you can imagine or you can forgive  

somebody getting the exact numbers wrong. But if  I look at this chart that you guys put out around  

how the outlook that you guys have forecasted from  roughly uh the the last year like April of last  

year to May of this year, it hasn't even changed  by 5%. I'll put this chart up on screen. So maybe  

you can talk through a little bit about like what  the reality of some of these forecasts are, which  

is that some stuff gets, you know, delayed, but  there's very few cancellations that are actually  

showing up in the market, right? Well, I would  disagree. I think there's a pretty substantial  

amount of cancellations. Uh just what we keep  saying is that these cancellations are are early  

stage projects, right? Um I'm not sure if you  have this one on the screen but like there's this  

map of the US where we show like the large load  requests right which is like slightly different  

but kind of goes to the same point as of today you  have over a terowatt of data center load that is  

requested by operators in the US alone right this  chart is as of six months ago it more than doubled  

now um so we have over a terowatt obviously you  know the whole US system right now peak load 750  

gawatt you're not you know double it uh right now  in sort of uh you know just a couple years uh it's  

just not possible it's not a reality so obviously  this is fake there's a lot of early stage projects  

and it makes sense essentially you know uh I think  the data center market started booming towards the  

end of 2023 uh that's when early folks started to  realize hey there's going to be a big constraint  

on this the first big deal started to get signed  at that moment um and you know since then there's  

been like this massive search for power and that  means that you know everyone was trying to find  

where is where is power available and that leads  to some behaviors where because everyone is doing  

it then if you want to be successful you also  yourself have to be aggressive plan multiple  

options um and this is that is actually nothing  new like hyperskers have always had multiple  

options when evalu evaluating projects it's even  more critical in times of constraint and so you  

would see sometimes for one final investment  decision see hyperskers maybe have 10 different  

options right and so that would be 10 projects  where they're sort of talked to the local counties  

um you know they they they talk to like uh uh the  utilities and so on and so forth. Uh and these are  

not all realistic and they're sort of just testing  the field and be like okay which county can enable  

me to build my large scale project where do I  have the workforce or do I have the supply chains  

uh you know where do I where can I build it? So  anyways, um there's an over supply of very early  

stage projects and you know our point and that was  sort of Rake's brilliant phrasing the cloudcoded  

projects is that [laughter] I just loved it. This  is amazing. Uh is that as AI is I guess not very  

good at filtering, right? So when you have humans  in the loop, yes, it's pretty obvious that some  

of these announcements are way too aggressive.  We gave a bunch of examples in the article of  

you know folks where you know they announce 10  gawatt project this they say the first trench  

of 500 meg is going to be available next year and  then you click on the website you see contact us  

nothing more and then you start digging into  whatever permits and whatnot you see nothing  

uh so at some point you're just like okay these  people probably have a lot of land in Texas or  

something like that but they don't yet have a real  project right uh any human judgment would sort of  

filter that but I guess for AI Okay, it's still  pretty hot these days. Uh so if you want to build  

a forecast of the industry, you have to base it on  realistic forecasts, you know, or semi analysis,  

extensive trade regulation has always been our  playbook. Not just data centers, but also chips  

because obviously this has downstream implications  on Nvidia upstream on the AI labs, their revenue,  

all of that stuff connects to each other. That's  you know the semi- analysis flywheel covering  

every single one of these industries and building  a a cohesive cohesive view. Um you know others  

don't do that and easily you know, struggle to  find the ability to quit the size outputs. Makes  

sense. So, in the article, you guys cover three  different uh types of data center delays. One,  

the aggressive announcement by a newer data center  developer. Two, an advanced project with overly  

optimistic construction timelines that might  just be a delay or pushed out or something. And  

then the third one is really uh something facing  permitting and local opposition issues. So Ellie,  

can you come in here and maybe explain one of  those examples where we do look in and find  

something in permits? Yeah. Yeah, sure. So, um  the the the project which um which um is facing  

quite a lot of push back locally um not yeah not  just New Mexico local opposition but also from  

uh from jurisdictional and reg um regulatory um  kind of uh characters. uh is a project Jupit or  

Oracle's project Jupiter um in New Mexico. So  yeah, essentially they're trying to construct a  

a uh pipeline to feed the uh the behind the meter  data center um uh in in in Da Da Anna uh province.  

Um and essentially um the the this pipeline hasn't  been built. Um and the route that they that they  

are planning to um that they want to build on is  is hasn't been approved. They haven't got a pro  

route. They have got a secondary approved route.  Um and essentially the they the whole all of the  

ways to kind of get gas to the site don't seem  very feasible. So looking at there's a pipeline  

issue which they're trying to um make make happen.  But um there seem every every FK filing that I see  

um that comes through um uh on the docket, they  haven't they haven't made much progress or any of  

progress at all because essentially it's defaulted  to a type of um uh uh regulatory process which  

essentially there is no precedent for it being  done sooner than essentially two years. And given  

the local opposition as well, uh that there's it's  very unlikely that it'll be sped up. Um there are  

other options to getting gas to a site like you  could ship um truck CNG and LNG to to a facility  

as as we point out. However, this at scale you  know getting to that's I think the only proven  

um like delivery of uh CNG at scale is kind of  around 200 megawatt or like max 200 megawatt.  

I think maybe even that's optimistic and there  aren't many there aren't many manu there aren't  

many producers um in the local area and there are  not not enough trucks to to facilitate that. So  

essentially yeah there's not only a pipeline which  doesn't exist and has no viable route of existing  

but getting gas to the site via CNG uh and LNG is  even harder um is is just not not viable. So yeah,  

no it's it just I think it br it like it it's  um it demonstrates kind of an interesting uh  

bottleneck which could occur with more behind the  meter facilities is like building gas pipeline  

um uh infrastructure uh especially in territories  um like New Mexico uh that's uh quite not not  

not extremely uh it's not extremely friendly  area towards that that kind of infrastructure.  

Yeah, I would um that that's how we we  basically look through the filings. We  

look through all of the filings. You know, we look  we we already had it on our radar because they had  

um issues with the initial turbines that  they were trying to use at the site. So,  

they ended up switching to Bloom fuel cells. Um  and through through kind of that that research,  

I I spotted this this timeline and I was just like  this is this doesn't make sense. Why are there so  

many local all the all these um all this push back  all these all these local comments and then yeah  

[ __ ] just um FK and the other regulators are  still thinking about it. So is this an example  

of a clawcoded data center project? So I I I think  what this speaks to is rather that there's like we  

have to re rethink the way we build data centers  in order to meet this demand. But it's it's not  

like uh other people had built one gigawatt sites  before this, right? Like we're still in the early  

in the evenings of bringing the first gigawatt  scale data centers to market. Um so we had we had  

to try a whole lot of different things. No one has  experience doing this. Uh maybe today yes but like  

as of you know one year ago, two years ago, no one  had experience. So obviously you know the you try  

all these new thing there's going to be errors. Um  so that that's why you know the way we analyze big  

projects like this we always try to think of what  is a reasonable timeline. Uh I think it's pretty  

um you know it's pretty common in the industry  that you have folks that um pitch more aggressive  

timelines but I think it happens and and that's  valid for you know every single industry not  

data centers contractors I think anyone that  has a construction we tell you they tend to be  

always over optimistic suppliers are always over  optimistic it's not like this is anything new  

uh and especially when you try something much  bigger at that scale with new designs and so on  

and so forth in new new locations new counties you  know new municipalities handling it then obviously  

you're going to face a few challenges. Uh so  that's a great fit for a category of projects that  

are already well advanced because they already  got financing. They got a release from Oracle  

uh OpenAI booked sorry Oracle charted to  OpenAI. So there's already a massive contract  

on this but facing execution challenges now.  Yeah, makes sense. Maybe we could talk about  

uh behind the meter now. So you guys in the the  second article published some some numbers on this  

uh first y-axis on one of your charts in a little  while there Jeremy uh which is pretty cool. So uh  

when we talk about the size and scale of some of  these projects I guess the behind the meter net  

additions when compared to uh like the available  grid capacity is a significant difference. Um and  

you specifically forecasted over 40 gawatts of  net additions of data center capacity behind the  

meter by 2028. Um that number is effectively like  it rails to zero right now, right? Like there's  

a few behind the meter projects, but people are  just getting going ordering turbines. The supply  

chain's ramping up. What does it actually  take to get there? Do you do you expect that  

behind the meter projects are more at risk of  cancellation based on some of the reasons Ellie  

just described or less at risk of cancellation  because they don't depend on the grid?

Excellent question man excellent question. Um so  I guess the first thing is okay this this uh this  

forecast sort of depicts uh our analysis mostly  of the US grid right and looking at looking at the  

gap uh context we're not adding enough generation  on the grid to meet that demand that is going to  

you know be in the tens of gawatts per year and  just keeps increasing every single year based on  

all of the signals that we keep seeing right um  so now is it realistic to assume there's going to  

be 40 gawatts of new behind the meter data centers  added by 2028. You know, you have to look at what  

is being planned right now. Um, you know, I can  tell you in the last two weeks, 3 gawatts of data  

center deals were signed [laughter] for behind the  meter purposes, right? So, deals are happening.  

Uh, site selection, I think, happened beforehand.  I think you saw a massive move in part of 24,  

mostly 25 where folks really starting started  to have like access to a gas pipeline as one of  

the main site selection criteria. uh I think  you started seeing sort of diversions where  

where maybe some of the traditional data center  operators were more focused on let's be in a tier  

one market like Northern Virginia uh you know  let's find grid connected sites because we need  

the five nights and you saw sort of more uh newer  operators maybe more you know AI fields that sort  

of uh tried to foresee this trend I think Cruso  is a great example uh you know they've been quite  

ahead of the curve on this and they've been able  to you know sign massive deals uh in you know  

like Abene 600 172 gig megawatt with Microsoft  announced publicly like Crusoe in Q1 that you  

know that's behind a meter for example. Um so  anyways I I think we're getting there in terms  

of is the supply ready for it because there's like  many many developers that have secured sites. Um  

in terms of manufacturers I think I can probably  do an extremely detailed rundown. Um in terms of  

probability of delays, I I I think the probability  of delays is lower with behind a meter because um  

of the the way these grid constraints play out is  that as a developer, you know, you talk to utility  

um they have no binding obligation to abide by  the schedule that they provided you. And what  

keeps happening time and time again is uh you know  with the anecdotes we hear from developers. Hey,  

I thought I was going to get 500 megawws by you  know 2027. I had a handshake with the utility  

or whatever. And then you know a couple months  later they tell me actually sorry bro I can't  

do it. Uh it's going to be you know 100 by 2028.  Uh you're 500 by 2032 because I got to do bigger  

network upgrades because hey actually I didn't  consider my analysis. Other guy also wants it.  

uh you don't have enough generation coming  and so on and so forth. Um and so yeah,  

I think utilities are sort of realizing they'll  face with more delays than they thought. Um and  

so in many cases if your strategy is only grid,  you're extremely likely to be disappointed. Um if  

anything that's more so on the on the power side  behind the meter, I guess adds a new risk which  

is more the ex execution layer like can you build  a power plant on time? Can you get the permitting  

that adds a new set of complexities? We're going  to see some high-profile delays. Uh obviously  

the New Mexico one based on our analysis is the  highest of profiles, right? Basically a hundred  

billion dollar deal for OpenAI with Oracle. Uh  so the highest type of uh you know delays that  

you can have. Um hope that hope they're going to  solve it on time. You know, wish them the best.  

Um I think you're going to see more of that, but I  think in terms of raw volumes because there's like  

so many options now for developers, you're going  to see a lot of success stories as well. And you  

know, Colossus one, sorry, Colossus one and two  are proof that you can do this at scale. Well,  

I don't know if they're the best examples  because of the issues they have with regards  

to permitting, but they're demonstrating like it  can be done and there's other doing it. Crusoe  

is another example. I was just going to add to  that like we're seeing the OEMs really reacting  

to the behind the meter story as well. So, you  know, we're seeing obviously the ma, you know,  

the major OEMs, the established players like,  you know, your Seammens, G, you know, all all  

those guys um expanding manufacturing capacity.  But then throughout the stack, you're seeing new  

new players, new entrance, um new new types  of technology, and then recycled all types of  

technology uh which we've written about in several  different um notes. But um yeah, I think that  

that's that's a really like increasing that that's  a good tell for like you know how this market and  

is is developing at the speed and and also I  think on on the permitting thing that's that's  

something which could delay um since we've seen  you know companies switch to bloom fuel cells in  

order to kind of get speed to well possibly speed  to power because of manufacturing footprint of of  

bloom fuel cells but also it's lower um NOx and uh  socks and [laughter] emissions. So yeah, so they  

um yeah and I think that was the Nubius New Jersey  facility for example is um they switched to Bloom  

uh um I think it was Bergen uh Bergen turbines  prior. So yeah um in order to yeah to combat the  

uh permitting issues there. So we might see, you  know, shifting around of technology. Um maybe  

relocating to more to other states like to Texas  rather than in on the East Coast. Um but yeah,  

know that's that would be kind of the the story  going on in the BTM market right now. Full full  

broke. Makes sense. Yeah. Um Jeremy, maybe we  can go back to one thing you said related to  

uh the like size of some of this. So uh at this  point when you say that you signed you you know  

about 3 gawatt signed in the last week um and  we're talking about in the in the context of  

the article 40 gawatt of behind a meter power  generation by 2028. Can you just high level for  

the general audience that listens to this podcast,  describe a rough timeline for when something gets  

signed versus when construction starts versus when  you can power the first trunch of GPUs on versus  

when a large site like a g gawatt scale site might  actually be completed? Is 3 gawatt signed in mid  

2026 actually going to have anything by 2028? Is  it a 2029 2030 story? Like is this a year and a  

half, 2 and 1/2 years, three, four, five? Yeah,  I would say it varies a lot depending on what  

what signed actually what actually is signed.  Um there there's a bunch of deals out there.  

Um the bulk of the volumes these days would  be turnkey leases where let's say a company  

like I don't know digital realy QTS builds a data  center for uh Microsoft you know they sign up for  

lease uh turnkey lease where sort of QTS takes on  everything and Microsoft just rents you have power  

shells which are a bit different which is sort of  a lower um lower bound on the developer and sort  

of more capex from the tenant. you have just power  deals, right? So it's basically a PPA uh let's say  

Oracle with Volta grid and in the Czech 4 county  Texas site essentially a PPA there's no data  

center involved right so depending on the deals  the timelines can vary uh rule of thumb would be  

say deal sign 18 month you have capacity but you  have the first ranch um and then depending on how  

fast you sort of can build it uh but I think it's  pretty clear that these days the expectation is  

that the ramp from sort of first capacity first  phase to full ramp is expected to be you know  

faster and faster Um, and the point we make in  that article is that from a buyer's perspective,  

behind the meter is now becoming much more  attractive than grid because you're in control of  

your destiny, right? Like, hey, you know, you're  going to have gigawatts by x amount of time. Now,  

obviously, it has to be permitted and so on and so  forth, but at least from a power standpoint, you  

know exactly uh what you're going to have on site  and when, provided that there's, you know, delays  

from the suppliers and so on and so forth. Um, so  the idea is, hey, if you have all of this power,  

then you also need to build a data center because  otherwise useless, right? Uh so the expectation  

is that is that that that the ramping up this 3  gawatt is not going to take 5 years. is going to  

be much faster than this, right? And generally  what you observe on the marketplace is that when  

there's a gigawatt scale deal being signed, let's  say uh you know Oracle with uh stack in New Mexico  

with Vantage in Texas, you see that financing  for the whole project takes place shortly after,  

right? We saw like, you know, um Digital Bridge  money for Vantage uh just maybe a month after  

um after the lease was signed. We're talking  about, you know, $25 billion of financing. So the  

whole the money is already secured, right? Once  you have that deal, you secure all the money and  

then you just try to build as fast as possible.  Um it's not going to take five years. It's going  

to be much trouble than that. Yeah. So not one,  not five, somewhere in between. Makes sense. Buy  

the model. Buy the model. Indeed. Yeah. Everybody  listening, buy the model. [laughter] Okay. Okay,  

I want to show you guys one chart and and get your  reaction here because uh I thought this one was  

fascinating in terms of understanding the market  share. Uh I think a lot of the people who buy the  

model are using the data that you guys provide  in order to make decisions for companies on the  

legend here. AWS and orange, Google and blue,  Meta in this fuchsia pink color let's call it,  

and then Microsoft and green. So these these are  the hyperscalers. They make up a large percentage  

of the market in terms of how many megawatts  they have under construction. Um the market  

share that these guys represent going forward  is I don't know just really interesting as we  

see them kind of compete with each other. So,  I don't know, Rake, what's your take on the  

relationship between the hyperscalers competing  each with each other and then maybe hyperscalers  

versus everyone else in the market in terms  of just how big everybody's going right now

in terms of competing with each other? Well, I  mean, I think when we take a look at kind of the  

capacity race, you're going to notice that it's  mostly driven by the a the respective AI labs. So,  

most of the time like an AWS capacity build out  will be kind of fueled by anthropic, right? Or  

meta will typically be for MSL. Um Microsoft  typically open AI as well. We know that there's  

some anthropic deals going on there as well, but  by the model. Um, in any case, these are going  

to be [laughter] in any case, these are going to  be yeah, like the the drivers of the data center  

buildout right now, right? Because they're the  ones who have the capital able to actually build  

out the data center capacity that's needed for  the AI labs. They're the ones who have the kind  

of investment grade back like investment grade  um financing able to start these projects and get  

them going forward. most projects struggle with,  you know, yeah, actually getting the financing.  

Like if you're a Neocloud, um, which if you're  listening to this later on, maybe you already  

have read the other article. Um, but if you're a  NeoCloud, it's a bit hard to get the financing to  

start and go ahead and like get a data center and  go ahead and buy the GPUs and expend the capital  

for all of this like build out, right? Because  it's going to come out to probably billions of  

dollars. Um so in essence you're mainly left with  hyperscalers driving the AI buildout or it's going  

to be very very wellunded or well capitalized uh  neoclouds like if we look at any of the kind of  

Blackstone backed guys any of the KKRbacked guys  um they've got plenty of money coming in from them  

to support their own buildout but in any case um  I forgot your second question actually so if you  

could say that again. No problem. Yeah. Yeah, it  is interesting. Maybe let me ask a different one  

because I'm thinking of it now listening to you  talk. So, at the very beginning of this article,  

we were uh commenting on the public research that  has driven all of these articles in the media  

about how half of [snorts] 2026 US data center  capacity is canceled. But that chart that I put  

on screen and roughly what you're saying about  the hyperscalers building for the AI labs says  

that you know even if you just take two of those  four hyperscalers who've got more than 5 gawatts  

under construction and then there's the whole rest  of the market the other two hyperscalers and then  

everybody else. So can you opine on uh where the  gap is in terms of people trying to do research on  

data center capacity under construction and just  missing entire gigawatts worth of capacity. That's  

a real thing for next year. Yeah. Yeah. Yeah.  Um, luckily Jeremy was able to explain the whole  

article in 3 minutes earlier. So I can kind of  expand on what he said. It's a bit too efficient  

with this communication. Um, basically when you  have a model like ours, um, we do everything kind  

of in a bottoms up way, right? Where we have like  each individual data center is like for example a  

row in an Excel file for us, right? And so you  can kind of get the forecast for each individual  

building, the tenant. If the end user is an  open AI or an anthropic um you already have  

that information for that building in specific,  then you can go ahead and find the timeline for  

the building. Right? A lot of people don't a  lot of people don't have the expertise to judge  

timelines on these things. Right? when we go ahead  and do our data center research. I like to joke  

that Jeremy and I have probably seen probably seen  over 10 to 20,000 satellite images, right? So,  

not a lot of people can say that they understand  like what that single gray pixel on a brown floor  

means. It's like, oh, that's actually the concrete  padding coming in. Um, that's going to be a good  

sign. Hopefully, we see vertical construction,  which is the darker gray pixel in the corner  

over there that most people wouldn't get. Um,  so the satellite images are super useful for us,  

but like that's just one that's really fun to talk  about. Um, the other stuff is a bit more boring,  

which is where we go through every kind of  permit portal for every county, city, state,  

uh, country in the world, right? Um, we like  to talk about how much we use Claude as you see  

in the article, 170K in a week. Um, the cloud  part, everybody loves the cloud part. Anyway,  

um so we got, you know, our lovely chart on  the kind of front of the page where it's, hey,  

$170,000 spent in a week, right? We're we're  token maxing, you know, we're token mogging,  

meta or whatever you want to call it. Uh more  tokens per employee than, you know, yeah, Meta is  

one of the But in any case, we tend to use Claude  and we've actually built out basically an entire  

uh harness and workflow for our agents. We love  our agents um that go ahead and scan and scrape  

every permit portal basically around the world  down to state, municipality, and city and county  

level. And with this, we get a ton of public  filing information that allows us to basically  

estimate timelines with that. And without going to  too much more detail, that's one part of basically  

three in our methodology. And for the other three,  you can also buy the model to figure that out.  

Um, [laughter] and so I put the uh put the cloud  code spend chart on screen here. So we've got  

uh to be clear to describe this, the  color coding is a per user breakdown,  

I believe. Yes. And uh people who are looking can  see the really big yellow block at the bottom.  

Yeah. Yeah. 90% sure. The culprit may be on this.  The culprit is on the phone right now. Yeah.  

That's Jeremy just abs just melting  anthropics GPUs at the bottom there.

Yeah. Yeah. Yeah. The ones that he's looking  at from the sky are also uh cooking up some of  

his requests there. Yeah. Yeah. Exactly.  They saw they saw Jeremy spend and they  

went and bought out the uh SpaceX capacity  right after. So, thanks to him. [laughter]

Good forecast.

[snorts] Yeah, we're we're moving markets now  like that. Go ahead, Ellie. What you thinking  

about over there? No, I was just gonna uh yeah,  make a joke about [laughter] Jere Jeremy's usage  

and making a profit good with Pope. Um that was  [laughter] um Yeah. No. Um yeah. So interest Yeah,  

it's going to be interesting to see um uh how  much of these um uh Yeah. how much is going to  

be built behind the meter from from the the main  AI labs as well. Um yeah see yeah saying and I  

think also the kind of vary you know behind the  meter is there's one that's one term there are  

kind of various different kind of definitions  um of kind of moving energy that moving kind  

of energy on site not taking all of it from the  grid um I mean you can colllocate your energy can  

um have a net metering kind of solution which  is um yeah more of a kind of interactive um  

uh interactive uh relationship with  the grid you can have fully islanded  

um off-grid uh um you know energy energy supply  for your data center. But yeah, I think that's a  

that's going to be the kind of thing that um that  that is going to be interesting to observe in the  

future as the more of them are built. Um seeing  what's preferred I mean which state fully islanded  

bitter grid connection maybe that might hedge some  you know interconnection in the future. Also then  

there's a you know what happens to these assets  these turbines um once you know once maybe they  

do connect to the grid they might end up being  part of the grid themselves. So like a lot of  

these turbines could be uses um uh peers gas peers  um uh and help solve um help with uh solving the  

actual like lack of power on the grid itself. I  mean, obviously there's transmission to sort out,  

but um yeah, there so I think it's a it's a it's  it's more symbiotic than um than than it's kind  

of portrayed by a lot of other outlets. I think  it's more of it's not I don't think it's one or  

the other. It will be Well, I don't think it'll be  one or the other forever. Makes sense. Yeah. Well,  

let me throw this chart on screen that I thought  was really awesome from the second article about  

behind the meter power generation, kind of  based on what you're talking about. Um this  

here's a breakdown of the supply ceiling by OEM  per year. So we can see that increase as some of  

these suppliers that would be responsible for  the power generation in behind the meter gas  

settings and how much capacity they're increasing  in terms of how much they can produce. And I mean  

two things jump to mind when I look at this.  First of all is just how much it's increasing.  

So it seems to be going up at roughly the rate at  which we're going to increase the power generation  

requirement. Uh that's a good thing. And then the  second thing is just how many different suppliers  

there are available. We look at stuff that's a  constraint in the market. Whether this is like  

GPUs or wafers or memory or um anything else rough  generally speaking when something's a constraint  

it depends on like one two three suppliers but  when you have this diversity where the top eight  

that are on screen and colorcoded uh make up just  more than half let's say but then your footnote at  

the bottom of this chart uh shows like 25 more  names that can possibly produce power. I mean,  

we've about this previously, but maybe you guys  can walk through like just how diverse the options  

are available for power generation is and and what  sort of shared supply chain there is that benefits  

um people who are looking for alternatives when  they need power and they're not sure where to get  

it for behind the meter. Yeah, sure. So um I mean  there as I I was saying like there would be kind  

of established players who have like a large um  uh supply chain where you know and they've been  

in the business for years and years and years.  Um they have manufacturing capacity globally.  

um they have quite uh they you know they're  they're quite they quite often say in earnings  

that you know the bottlenecks are kind of the  castings the bottlenecks are the the blades  

um for whatever kind of turbine or aero  derivative um or resip kind of whatever  

solution they're they're they're offering but um  it it seems it seems that like this is not slowing  

anything down because they're still announcing  capacity increases and they're still and they're  

still saying that these capacity increases are  driven by AI AI demand growth. So it it seems  

to be actually a little bit of a narrative  talking about the constraints to castings and  

um you know kind of upstream constraints to to  building building these turbines. Um is a bit of  

a narrative from last year. I think this year it's  now the the large OEMs are sort of like realizing  

that actually uh more many more players are coming  to the market uh and they're trying to keep their  

market share not and not lose it. Um, I think that  that literally that was in a seaman's um uh uh um  

pre-close call this morning uh or yesterday. So  they're very much kind of um aware of this growing  

market and that you know they can't kind of play  the scarcity game and try and get everyone to to  

freak out because people did not freak out but  people saw an opportunity and are converting you  

know old old parts of you know plane technology  into kind of um into engines and then and then  

you know um boiler boilers and obviously there's  um a kind of there's the autoderative kind of  

um options as well. So that's we're seeing all  of these different types come into the um uh into  

the market. Some with kind of more the that have  more of an advantage in terms of permitting and  

some some less. Um I think there's also something  which we I've I've been noticing is um more of the  

um yeah more of a car industry um kind of having a  larger play in in in larger playing a larger role  

in the data center industry in the sense that  like EVs are not um are not as I mean they're  

quite popular in California and the east coast  but like in lots of the US um there's kind of EV  

um growth has been quite slow and so a lot of that  manufacturing capacity I think is is going towards  

well we have the Tesla as a good good one and  then we saw and then Ford as well um uh providing  

best um so battery storage battery energy sort of  storage solutions to the data center industry as  

well so we're seeing we're seeing kind of these um  existing uh players uh switch to the data center  

industry um as well so yeah I hope that answers  your question but yeah it's cool yeah makes sense  

Jeremy what's uh jumping of mind here when you're  hearing us talk about all the different suppliers  

that you can use for biometer power generation.  Yeah. I mean like I think it was always to be  

expected. Um last year the narrative was behind  a mirror is not possible because there's this  

free manufacturers that are so slow and you know  everyone was sort of throwing that chart uh the  

famous chart of like you know how much gas orders  there were in like the early 2000s where you had  

this massive sort of you know wave and everyone  was saying GEV and Seammens are so scarred of this  

era because they invested massively in capacity  and so they're going to be very conservative  

right but you know the that that leaves a market  opportunity um and one one question would like to  

ask uh management teams at these power companies  is you know how AI pill are you essentially you  

know how much do you believe how much risk are you  willing to take um there's also a function of how  

easy is it for you to take risk in the sense that  you know what's your economics on building more  

capacity near factories and there's a bunch of  companies that you know score very well the one  

that we've been flagging for a while we especially  doubled down like at the end of 25 when we did our  

big deep dive is blue energy uh because you know  I think clearly the management team is very AIP uh  

I think they have economics that sort of enable  uh them to build capacity faster than others um  

and the the the behind the mirror conversation is  is really interesting because the like I I think  

it's like you know company by company they're  sort of adjusting their mindsets as they adapt  

to the new reality and also sort of solution by  solution you keep uh adjusting your expectation  

just based on constraint so what I'm saying is  that initially everyone was like okay this is  

going to be bridge power it's going to be matter  of like one two years I run off up offg grid and  

then I'm going to have my uh you know my grid  come in and maybe it's going to be backup and  

so you you would only consider like uh systems  that are good at backup um but that's the ultimate  

disadvantage for something like bloom is that  you know it's like not very good at backup you  

has to be run extremely hot it's 15,000 degrees  C uh takes like two days as far as I know to go  

from zero to 100 uh so for backup it's you know  really not the system. But hey, if you have no  

other alternative, then may you know, maybe it's  that maybe you just have to to go for it, right?  

Like if your if your option is your other option  is I'm not going to have power, then you know,  

you're screwed. You're not going to be competitive  in the marketplace. Um, and you know, other issues  

are like power costs. And I think when you look  at the the recent like SpaceX deals, I think it's  

pretty clear that sort of the revenue per megawatt  that folks are making on the cloud side also on  

the lab side on the model side like revenue per  me go up [laughter] at every layer I guess uh and  

as that increases it means that power costs are  increasingly irrelevant. Um and power costs for  

solutions like Bloom are not that expensive  anyways. Uh with regards to you know what we  

have in Europe for example with the grid. Uh well  certainly not as a as a percentage of the total  

construction costs of the project or when you  consider the GPUs and stuff. People can pay double  

for the turbines they want compared to I don't  know somebody else that's competing in the market  

for them and not really affect the total capex of  train. And the other thing is like also estimating  

the bit of materials I think has proven quite  complicated for many of these vendors especially  

as labor rates are going to the moon and you  thought you were going to have local labor but  

hey actually everyone in Texas is you know already  occupied building data center so you have to call  

people from Denver or from Ohio and obviously it's  much more expensive um and this means that this  

also favors solutions that are fast to install  first fast to deploy uh because your B is sort  

of more predictable your time are more predictable  and uh both on the labor side and on the sort of  

full deployment side. So uh you know I think  uh again like analyzing timelines and ease of  

deployment is key and I think Tom solutions score  extremely well and I also think we're going to  

keep seeing more and more new entrance uh because  you know everyone looks for capacity earlier and  

the standards are sort of dropping right the the  there's a few new companies that entered recently  

that are basically coming from the automotive  industry not themselves like their source engines  

from the automotive industry. And then when you  think of this you're like wow automotive like you  

know we're talking about like 100 gawatt plus per  year of production and that's an industry everyone  

knows runs at very low utilization rates you know  50% 40% these factories are not doing too well. Um  

so the capacity and the incentive that these guys  have to sell their engines through the data center  

market is also tremendously high. So uh you know  that unlocks gigantic capacity. Absolutely. At a  

minimum it's diversification for their business.  So the the one thing that we haven't talked about  

necessarily yet is alternatives to uh gas let's  say uh I'm sure a lot of people who are new to  

this space who are listening are thinking in  their mind about you know what about renewables  

what about solar wind hydro nuclear like you  know who's considering that what's going on so  

I'll throw two charts on the screen and then  hopefully you guys can comment just on the  

reality of being able to use um renewables plus  batteries or um you know any renewables let's say.

So yeah. Um yeah, go for it. So yeah, this  is uh grid connected. Um so the the point  

of this obviously like name plate additions are  overall growing and we're talking about you know  

60 gawatt 50 gawatt per year. So it's a lot but  as everyone knows you know the 100 me gawatt of  

solar is not a true gawatt for the grid because  you know it's intermittent and only turns on at  

certain hours and so on and so forth right  pretty simple. So you know depends on the  

areas you see an ELCC value. So adjusted for the  actual capacity value it brings to the grid. Uh  

depending on the area it can be like 10% of it,  20% of it, sometimes below 10, sometimes over 20.  

Depends on you know it's all sort of very complex  system level calculation. Uh the problem is again  

like the grid is also sort of facing all of the  transmission issues. It's intrinsically slow.  

There's giant interconnection cues. So I think  the question the real question for solar and  

batteries is can you do it behind the meter? Like  you know can people build these things on site?  

Um I I I think it's going to happen. Uh and we're  gonna have much more content on that published on  

our website. So stay tuned. Uh obviously the high  level challenge is that the the land required  

to build like massive amount of solar is just  tremendous. Uh so logistically these projects  

are fairly complex. You know if you're talking  about I need 20,000 acres just to throw solar  

panels at it. Uh it's going to be for a gigawatt  data center or something like that then you know  

buying all that land can get complicated. Um  sometimes what happens you know during the  

course of this uh sort of land buyout is that  the land holder sort of realize hey this guy  

wants to build you know hundred billion dollar  project maybe I'm going to sell my land 10x is  

more expensive than what he thought I was going  to get right so that sort of stuff happens as  

well so anyways the complexity of building these  massive projects fairly elevated which removes  

the time to power angle to some extent for solar  and batteries um but as folks sort of get bigger  

and bigger and start planning you know multiple  years ahead. I think there there's already a lot  

of big projects like this underway. Um and  the other issues obviously SMRs and nuclear  

uh you know no you know I'm not going to reinvent  the wheel. This is slow stuff right. Uh everyone  

knows building nukes takes time. Uh so again  not a time to power. So right now you have  

a lot of non-binding LOIs in the market  or non-binding deals that are contingent  

contingent on like milestones and execution and  getting the regulatory approvals and so on and  

so forth. I think in the 2030s we're going to see  a gigantic diversification of energy sources to  

power data centers. Uh I think for the 2020s  we're going to be very much in the gas world.

Makes sense. Okay. As we move to wrap here  guys, what's Yeah. Anything left unsaid? Ellie,  

what you got? I was just I was just going to  say that we might see um kind of as in moving on  

uh more kind of carbon I'm start to see more  kind of carbon capture um utilization storage  

um facilities being kind of colloccated with BTM  sites and with data centers as well. Um uh I think  

that's something interesting to observe um as well  going forward to know. Yeah. Rick, how about you?  

What's jumping a mind man? [sighs] Um, you know,  I don't know if I have too much here. Uh, I think,  

uh, I think we covered that the listener  needs to know. I'm not going to lie. No,  

I don't know. I think, yeah, I think the BTM  movement is going to be very, very big. But  

I think I think maybe Jeremy touched on it um  already, but it's interesting to see the like  

what we call kind of like peak turbine in 2026  where it's like there was a huge overby or like  

a huge overpurching of turbines in 26 where not  everybody knew what to do with them. Not everybody  

could figure out how to get the permits or not  everybody could figure out how to build the data  

center to actually use the turbines. And so you  ended up seeing a lot of these turbines go kind  

of underutilized or I mean later on start to hit  the secondary market for example. But I think this  

is going to read pretty bearish to people at first  and people might kind of freak out on it. But long  

term we're still very very probe behind the meter.  I just think it'll be interesting to watch how  

people interpret this though because it's going to  it'll probably keep picking up, right? Like you'll  

probably see your favorite projects turbines start  to hit the market at some point and you're going  

to freak out and then the question is is it  over, right? So, we don't think it's over. Um,  

we still quite like BTM, but I think that's going  to be a fun dynamic to watch play out. Okay,  

we got to end going around the horn here.  Rick, what what's your favorite project?  

project. My favorite project you Oh, actually I  might have one. I might have one. I I would go the  

NScale Norway project. I love this one. Actually  [laughter] love this one. This one. Jeremy,  

what's your favorite project or or top three?  You You don't have to choose. Colossus 2,  

bro. Colossus 2. Fast, efficient, scale,  uh, super high revenue. W [laughter] All,  

what's yours? Um I I would say Venevius  switch and new see um Si right why not  

quite Oracle New Mexico all right go ahead I  know I don't know Oracle New Mexico [laughter]

that's a lovely we spent the most time with that  one I think yeah yeah where are we going to get  

robots building data centers we need robots  we need robots building data centers They're  

coming. They're coming. Hey, we should uh you  guys should watch the other podcast. You should  

watch the other podcast. We talk about this.  [laughter] I don't even think we did. What year  

are you going to see a data center fully built  by robots? Fully built by robots. Wow. Um Oh,  

that's a lot of manhour. Electricians. Lot of  man hours. What is that like eight and a half  

million manh hours? Um [laughter] dude,  lot of man. Uh maybe 2029 at the earliest  

20 2030 2030 breaks going to back calculate  actuator constraints in the supply chain let  

me let me think about how many permanent  magnets are being manufactured right now

that's a tough one that's a tough  one man yeah the lord pull oh man  

all right thanks for joining good  All right. Cheers. Bye. Bye, Dylan.