SemiAnalysis on the AI power crisis: gas turbines, Meta's tent datacenters, and why training is still growing as fast as inference

Good to see you again. Second time on the show. So excited to have you here. Happy New Year.

Happy New Year.

Happy New Year. How are you? Are you uh are you shocked to see that uh the the G7 that's supposed to happen in France? Aren't you in France right now?

I'm in France right now. Yeah. Well, you're gonna have to wait for the summit because UFC takes priority in America apparently. I don't know if you saw that.

Yeah.

Anyway, um congratulations on the new article. Uh I would love for you to uh set the table for us and explain sort of what were the questions that you were trying to answer. What was the overall thesis that you came into this particular article? How AI labs are solving the power crisis. Uh and then I have a whole bunch of questions that I want to dig into.

Yeah. Look, the question that we keep receiving every day, every hour it seems, is how are we going to power the AI race? Yeah,

you know, uh is the grid able to handle all of that? And look, last time I came in, I think I said, hey, there's like over half a terowatt of data requests, uh in all the US grid, an insane amount of requests. And we we talk about sort of prisoners dilemma where because everyone is trying to find power then it creates sort of a vicious cycle of everyone starts being more speculative and putting requests everywhere and so basically the grid is overwhelmed.

Yeah.

People cannot find energy. Um and so that's why you're seeing the rise of on-site gas. Uh which is something that a lot of people have been talking about. But from our research, we just haven't found sort of any good any good way to understand what are people actually doing. What are the challenges? What are the systems that people are actually deploying? Uh what's the benefits? throughout the trade-offs and how to understand how to make sense of all of these new entrance. Yeah,

because one of the key highlights is okay, everyone talks about Ger Nova, everyone talks about Simmons energy, but we count actually 12 manufacturers that have secured orders of over 400 megawatts for US data centers on S gas power. There's way more people in the pipeline, but essentially we wanted we wanted to to show people how are the labs solving the power crisis. Uh talk about to some extent XAI was a bit of a pioneer because obviously it did that before everyone. Yeah.

And how are the other players following suit and uh actually solving this issue?

Yeah. Let's let's stay with that point about power requests from AI companies. Uh that's expanded significantly. Correct. Isn't it? It's over 10 terowatts now or something like that. Everyone's spamming with these requests, isn't it? It's an insane figure. Uh is that roughly correct or

I don't know about 10 terowatts.

Not it's over one.

Over a ter roughly a terowatt. roughly a terab.

Uh it's always complicated to know exactly. Yeah. Um but roughly a terowatt. Uh

and what are the mechanics of a of a of a request for power? Is that going to uh to uh governmentrun organizations? Is that a permitting process? Like what is the anatomy of actually a making a request for power if you're an AI company?

Sure. So typically you send a request to a trans transmission provider. Okay. Uh, so say American Electric Power, the largest in the US. You send them a request. I want Power in Ohio. Yeah. Um, you have to fill whatever some kind of form. You tell them what you want by when you want it. Um, and then if sort of that first phase moves through, um, you have to go through a system level study.

Yeah.

Okay. And the reason we have to do that and that process, the reason it takes time is because the way the grid works is demand and supply have to be always perfectly synchronized. And if they fall out of sick uh because there's too much supply or too much demand, then there's basically a blackout for everyone. That's the worst case scenario to be clear. Uh but it's possible. It has happened. It happened in Spain about a year ago

um because of an issue on the supply side. And so the implication is that if you want to interconnect a 1 gawatt data center, you're going to have a plenty of system level studies that are going to slow down the process.

Um and this is where you get into that sort of vicious cycle because everyone is sort of putting requests because they know it's going to take a while.

Yeah. Also,

isn't there isn't there some people putting in requests just at from a speculation standpoint? They're just like, "Hey, I know that if I get access to the power, I can resell it to somebody else, and if I just kind of lock into uh you know, you know, uh one of these uh

deals,

grids, basically giving me a contract, then I can go and flip it." Is that is that happening?

Everyone wants me these days or gigawatts.

Yeah.

So, you you try what you can to get some. Um, and if you if you're for example, say you're based in Ohio, you typically operate in Columbus, Ohio. Uh, then suddenly you tell your utility, hey, I I want another 500 megawatt. And they respond. Actually, I've had like 10 other people ask me this. So, you're going to be in the queue

and there's 10 other requests I have to evaluate before I get to yours.

Is there but this is why some of the Bitcoin miners have done actually very surprisingly well in just sort of like pivoting to AI because they had they had the the pre-existing power deals in place, right?

Yeah, it's not in the queue. It already exists because the queue is just an evaluation. But then once you have the once you have everything approved, you have to build a substation. You have to actually interconnect your data center to the grid and the crypto the Bitcoin miners, they already have the substation, they already have the transformers on site, they already have the energy flowing, the electrons are there. Um, so for them, it's just about retrofitting and just leveraging the megawatts that they have. Is there is there anything that uh I mean we were we were reviewing this uh this meta deal today and and they have uh someone on staff who's the director of global energy. Is there anything that a team or one of these hyperscalers can do to move through the queue quicker? I mean I imagine they can't just like overpay or bribe or it would otherwise just be an auction process. But is there is it sort of a battle of the for is that actually happening? Um I is there uh is it just a battle of a forums? Is it the regulatory team that you have on site? Is it is it different uh like scientists that are driving the work or is it just all the best legal team wins? Like what is the anatomy of actually winning allocation in power?

So the first thing we have to flag is there there are many different transmission providers uh in the US. So many different bodies that work differently.

Okay.

But there's a few ways you can do. So first of all, if you're a sophisticated end user, what you're going to do is analyze the grid um and try to understand where uh where does it make sense to have power from a I guess grid congestion perspective. Analyzing the network, you realize, okay, this area probably should have free power. So I'm going to talk to my utility and if you if you want to if you want to have a better chance of being ahead of others, then at least you have identified a specific spot that is likely to have available power.

Now, typically relationships obviously matter. uh if you've been in the business for 10 years or whatever and you're good friends with uh you know sea level executives, you can perhaps move up the grid just because just because of trust. It's it's it's not about necessarily bribing or something like that. It's more about you have a whole lot of new entrance. You have a whole lot of new entrance that don't have a lot of experience and then you have this other guy that maybe has been doing it for 10 years and has a lot of experience and success. So obviously for the utility it's easier to trust the guy that has the track record and has the existing relationship. So that's that's that that's one angle. Sure. Um yeah

uh I'm interested to know about your process to understand how much power each player is actually accumulating. Uh I I've seen you know the semi- analysis watermarkked screenshots of data centers. Is is satellite imagery actually that useful for understanding like the map of total data center capacity? Are you looking at regulatory filings? Are you looking at statements from the companies themselves? like it feels like you're at least one or two clicks ahead of what the executives at these companies are publicly saying and you're act the purpose of semi analysis is to you know uh provide analysis and data ahead of what's publicly available but what is the process for actually understanding Azure's compute capacity this year before anyone else does

yeah sure thing so there's um there's different um to some extent lead times Um for us the the the way we do the the call it the short-term forecast. Shortterm being roughly next quarters extensive satellite imagery. We love satellites. We pay a lot of money for satellites. Every single blog these days you you see satellite pictures. Um and it's pretty simple. One thing is using historical imagery figuring out when did construction of specific building start.

Two is what's a typical time to build for an operator. Uh so you have to you have to get a sense of like you know um having reviewed many of their buildings how much how long do they typically take or uh sort of understanding the design patterns how maybe they can accelerate the build out to what extent

um right and then if you know how how long it takes and when started you can know when it's going to be operational then it's about the size it's also about if it's a large project how fast can you get it up to speed um because one thing is to build the shell and the other one is to build all the electrical the mechanical and obviously deliver sequentially data hold by data hold typically. So there's a bunch of things you can review. Uh we use all of it. Satellites were big fans of course.

Um and like some of to us it's interesting because um we've been able to be very successful at predicting trends with hyperscalers by analyzing data because you have to think of it. I'm I'm saying you can just use satellite imagery but you have to do it for hundreds of data centers. Who has the time to do that way? We do. We're crazy.

That's awesome. But it worked out because if you think about it, Amazon, they accelerated growth, right, from 70% to 20% last quarter and a lot of people were surprised because people were saying no, they're losing AI. Azure has accelerated, Google has accelerated and Amazon hasn't. But actually, if you analyze data center construction, they've been accelerating like crazy on the construction side uh in the first quarter 2024. In the fourth quarter 2024, takes a year for them to build. Obviously, you see the acceleration fluid mechanically Q325, Q4 25. Uh that's one aspect. The other one which is a bit more complicated is tracking the the leases. Um and so sort of understanding where are the different hyperscalers leasing capacity from third party operators. Uh the big guys are QTS you know digital realy Equinex and there's all the crypto miners that are basically doing the same thing. And that's also something we track very closely. Uh and it's complicated because some of that information isn't public. But by triangulating many data points, you can actually get to the answer using only public data, which is what we do extensively. And using permits, digging into filings and all of that is a huge part of the process as well to get to the answer.

Yeah. As I understand,

how much how much uh is the are you guys benefiting from uh like actual AI tooling on the research side? Like are you guys running like you know 20 agents in parallel that are you know, basically like looking at a lot of filing, trying to find them, monitoring them, etc.

Are you feeling an unlock?

Yeah. Yeah. So, we're using AI quite a bit. Uh I think we can do more. Uh I wish we could do more to some extent. Obviously, it's bound by computer use. Um there's some complications. There's some portals for uh for per permit purposes, permit tracking purposes that we can automate. But the problem is that if you think about there are so many states, so many counties, so many different things. And some of them are fairly easy to use. And so today's agents can actually automate the process. Other ones are maybe more complicated, more archic, right? Some of them you have a clean PDF. Other ones it's like, you know, a scanner. It's written by hand. There's a lot of difference when you're going to the the weeds of the permitting process uh in the US. So we do automate quite a bit. Uh what what we're starting to do this is pretty interesting project that we have on the data center side is a vision model where you can actually uh based on satellite imagery uh u detect real time sort of what's the status of construction and understand sort of the inflections that you see um as well um on the in on the planet.

Yeah. Z uh zooming out it feels like uh I think you quoted it something like half a gigawatt of uh new data center capacity was being added for a period of year sort of linear growth uh and now we're seeing a break in the graph and we're seeing exponential growth and I'm I'm wondering how you're thinking about there's always this question of like when will AI show up in the GDP statistics it's obviously a big business and there's a lot of revenue uh But we're but we're not seeing the 10% GDP growth just yet. Maybe that's coming. Uh when are we

seeing productivity growth? I mean CNBC yesterday was trying to figure out why US economic productivity surged almost 5% highest in six years. Maybe uh but I mean my question with regard to like overall US power generation uh are you expecting to see a meaningful acceleration a break in the graph there this year next year? Uh or or is or is the data center overall energy picture still a small enough slice of the pie in terms of overall American energy production and consumption that uh we won't see it move the overall needle just yet? Oh, it's already moving the needle. It's just going to accelerate. There's just one direction at least for the next call it two years. Um, if you one thing you can do is a lot of the utilities are publicly traded. So, you can go one by one. Every single one of them, the only thing they're talking about these days is data centers is how much load growth I have and how how and then it's about capacity constraints. Can I deliver on that on that demand? But all of them are seeing tremendous load growth. And if you think about the leading indicators on the data center side, so for us, we use two things on the if you think about self build. So the data centers that are built and operated by hypers skaters construction starts are exploding. So they're all massively accelerating. They have accelerated tremendously versus 2022 and 23. So 24 was a big year. 25 insane. Um so this is a leading indicator with regards to what's going to happen in 26 for capex for revenue and obviously for load growth because these days it's all correlated. Um on the on the leasing side, same thing. If you think about uh how much data center commitments um are the hyperscalers doing these days, it's also just up and to the right. Uh 2025 has been an insane year uh for the leasing market. And so everything points to 26 27 just uh accelerating. Um and then if if you think about GDP growth because I think it's an interesting topic, there's uh two two things you can talk about. One is on the infrastructure side and the other one is on the productivity side. M

um on on the infrastructure side, we you're already seeing it, but I think it's pretty easy to look at statistics. You can look at data center contribution to GDP. There's a bunch of stuff um from some of the agencies. You can look at um computer investments. There's a specific GDP road that tells you all of that is is exploding. Um roughly speaking, a guesstimate is over 50% of GDP growth today's AI infrastructure. Yeah.

Which is kind of insane if you think about it, right? And and you can do some monkey math as well, right? You see Nvidia's revenue. Uh what is it these days like 250 billion annualized or or more. Um if you do the math, right, a lot of that is going to the US. Now there's obviously it's complicated because there's imports and things like that you have to deduce, but you know, US GDP is over 40 trillion.

Uh right, 1% is 300 billion. Easily year-over-year additions, AI for investment is over 300 billion. Sure. Right. Power plants, data centers, chips and all that.

Yeah. And then that should drive significant GDP growth. That's

and then the second phase of this is uh on the productivity side. Obviously you have to see to some extent you you you would expect to see the people that are providing that productivity to to be beneficiaries. Um and so this is where it's important to track what are the AI labs doing, what are the startups doing, uh which is something we've been tracking pretty closely recently at analysis. And so if you look at the AI labs, yes, they're all accelerating pretty pretty fast as well on the revenue side. Like you see, you saw Open AI, tripling revenue, Antropic 10X and so on and so forth.

Can you uh take me through the latest bottleneck in gas turbines? We had Blake Schaw from Boom on the show and it feels like he's expanding his business to build turbines. And I'm interested in that specifically because uh we hear about nuclear power plants coming online. That's obviously a very heavy regulatory burden. Also, supersonic flight feels like incredible regulatory burden, but how difficult is it to just manufacture a new turbine or more of the same design, bring them online, actually ramp up capacity of natural gas turbines.

So, so typically to develop a new turbine, you're talking about seven to 10 years R&D process.

Um, so it's pretty fascinating because you have a few a few new entrance that are hitting the market today. Yeah. Uh a big one is Dan. Uh Dan is providing roughly 2 megawatt of turbines to XAI.

Uh Dan, Korean giants on the nuclear side. Um they've been developing their turbine for over 10 years. Um and it's like, you know, perfect timing. They have a

Yeah, what amazing timing. That's great.

That's pretty lucky. Pro Energy is another interesting example also seven year R&D um R&D program and late 23 finally they got all the approvals. uh it's just very a very uh complex technology a lot of very high precision materials so ramping up the manufacturing side is complicated another thing to to think about and uh that's the difference we have with folks like bloom is um what is the like how long can you make your investment when you build a new factory for gas turbines or some of these systems and this is an interesting thing to analyze because some for some folks it's actually easier to build new capacity because their payback period is much higher and because their revenue per megawatt is higher and because the their cost is also higher just because the prices are going up because AI labs are willing to pay more for the turbines.

So, so yes, but I'm more thinking about comparing different technologies. Like if you think about Bloom Energy, the cost to buy fuel cells is is very high is much higher than buying a turbine.

Sure.

But the flip side of this is that the payback period because the revenue is so high per megawatt, the payback period of building a factory is actually much shorter. So for Bloom's perspective, they take less financial risk if they expand capacity and they manufacture more. Yeah,

pretty good position.

It makes a lot of sense.

How did you react to the meta Oaklo news from the this morning?

Well, I mean it's um I think all of the all of the all of these folks are um are looking for energy that is uh you know cheap, stable, are looking for ways to ensure their the continuity of their supply.

Um and it just makes sense for all of the hypers skaters to to to work with the nuclear nuclear companies. It's not a surprise because everyone has done it already, right? You saw Google as well with Chyros. You saw a bunch of these deals already switched with Oak as well about a year ago. So, not too surprised about this one.

Yeah.

How much does the nuclear uh fision or even the fusion stuff got Google's uh partnered with Commonwealth Fusion Systems on a few things. Um how much does that play into uh the analysis that you do when you're looking out six quarters? Because when we saw the date 2029, 2032, 2035, our eyes kind of roll back in the in the bottom in the back of our head and say, well, that feels post singularity, so what's the point? Probably important at some point, but uh certainly less of a less of a critical decision in the horse race of, you know, which lab will get power next month to train the next model that we're all going to be focused on.

Yeah, exactly. That's why if you read the report that we we wrote, it's all about gas because if you think about the next few quarters, it's gas. There's just no other solution. Nuclear is going to take a few years. Solar and batter is not ready yet. All of these other alternatives, I think they all have good potential and I'm sure we're going to see a lot of different solutions in 5 10 years, but today it's just gas.

So, so you're modeling energy. It's mostly gas. Um uh any plans to mo uh to to model or uh or analyze or model water? Is that important at all?

Oh, interesting. Yeah. Big topic these days, right? Yeah.

Um uh our view generally speaking is that water is not that big of a problem because in the data center space you have this trade-off between energy and water. Yeah. Um and so you can actually enclose loop systems that consume pretty much no water in some cases zero water. There is some water required uh just to to build the initial tank and the initial loop. Yeah.

But it's closed loop so you don't need water. So what's the issue, right?

Yeah. Exactly. Yeah. It is it is a funny retort to anyone who's worried about the water uh the the AI water usage. Just if it's important, why doesn't semi analysis talk about it ever? Why are people not trading the water stocks if it's if it's important? Um I am interested in the

we might talk about it soon.

Yeah, I mean I'm sure there's some sort of angle, but uh on on the on the question of water usage, it does seem like uh Meta is moving from an air cooled system to a water cooled system. I think I have that right. They moved away from the H design of their data center. Um can you tell me more about why the Hshaped data center was not suitable for water cooling? It felt like a very modern building. Uh why was it impossible to retrofit that? Why did they have to go with an entirely new design?

Yeah. So the the the thing about the edge design above everything is it was really designed by Meta for uh leading cost efficiency.

And so typically the the ratio people use the PE which tells you what's the energy efficiency of a given facility. Meta had the world's best uh the world's most efficient facilities. um the the energy required to cool the data center was extremely low

and that's because it had a fairly complex structure three stories um

okay

the drawback of that is is that the time to build a facility was about two years so that doesn't work in the AI era we're talking about month right it's 122 days for XAI so you need to go faster so that's one of the main issues now the other one is regards to cooling um the the way that we're cooling this specific data ser is like it's I call it an air-to-air system you could you could simplify it and say They open the window. That's basically how they cultivate this. They open the window. Obviously, they have a bunch of

We tried opening the window.

You know, I I actually I actually toured George Hans's uh uh he has a a miniature data center, just a couple racks of GPUs that he trains for uh autonomous driving for self-driving cars that he builds. And uh and his cooling, it really is just like a window unit that just flows air through the this particular room in his office. And and he he's using air. a car with no AC once.

Open the windows.

Just open the window.

It's a time honored tradition, but sorry I cut you off.

It works. It works. It doesn't work that well if your your hot is liquid cooled. Yeah. Obviously, if you have a cold plate that goes through the chip, then the question is how do you cool uh the fluid that you put into the cold plate to remove the heat from the chip? Yeah,

opening the window doesn't work that well. You can do some kind of retrofit with liquid to RCDUs. It's expensive. It's not very efficient. Uh so the best way to do is to have a dedicated fluid cooling system which involves building a whole dedicated piping infrastructure and all of that which is what meta did their new data center can handle seamlessly uh liquid cool chips whereas the old one uh it was much more complicated.

Do you have a view on um you said uh Meta's uh Meta's previous data center I think it was 150 megawws um two years to build it. Uh how fast are they now? Are they at 6 months a year? Do you have any idea of where they will be on the on the speed frontier since that seems so critical?

Yeah. So two two two and a half years roughly speaking. Uh then they built some sort of rectangular design which is 12 to 15 months. Okay.

Uh and then they realized actually we need to go faster and that's when you started seeing the tents.

Um and these tents uh yeah the goal is to be be able to put out a GPU cluster in six months. One of these tents. Wow, that's fast.

And it's interesting because they go back to an air to air cooling system. So what I told you earlier actually is wrong because despite hardware being increasingly liquid cooled, now they're doing air to air again opening the window. Sure. But they have those side cars which are liquid to air which are expensive again. Okay. It's one way to go faster. That way you don't have to build a whole like piping in front and all that.

Yeah. Yeah, that makes a lot of sense. Um what about um how the other hyperscalers are are are matching up in terms of in terms of speed, but also uh a lot of the other hyperscalers, a lot of the big tech companies had made commitments to maybe move away from natural gas to maybe go more uh net zero, more energy efficient, more carbon neutral. uh that feels like the water debate is maybe a moot point but there will be some sort of uh climate discussion in the future as more and more natural gas gets brought online. It is a fossil fuel after all or it is not a renewable energy source. So, um, how how are are there any big tech companies that are grappling with that or struggling to get through previous commitments that they made um to, uh, to be more environmental or more net zero and now they sort of have to retool their business and messaging. Yeah, you've already starting saying this in 2025, but in 2024 where they're all they all said uh for the time being we keep our commitment for whatever 2030 or something uh net zero, but uh short term

we're going to have we're not going to be able to meet our goals and to uh I guess um clean our fleet as as as fast as we expected. But yeah, there there's just no way. Um so there are a few interesting things. So the the first thing I would say is you're seeing some projects that are natural gas based but they have uh they have ways to become more sustainable. For example, there's a it could be a site where you have great geology to do carbon capture.

So for Cruso has a one like that in in Wyoming. So that's those types of projects obviously have a long-term potential as well because then you can meet your commitment uh if you're still committed to that. It's unclear they're already that committed. But other topic the the other thing is so Google say generally from our analysis Google seems to be still the most committed hyperskater. Um and they're doing some pretty interesting stuff in in Texas right you saw the acquisition of intersect power. They're building some of these some massive campuses in Texas where they actually have on-site solar energy and battery. But to be clear it's not behind the meter. The reason we talk don't talk about it in the um in the optical is because it's not fully off-grid. Uh there's still a grid connection.

Okay. So it's not comparable to the off-grid deployments of folks with turbines and

interesting. Yeah. On on the Google topic, uh is there any sort of durable advantage in multi- data center training that you're seeing from Google? Are are are you seeing evidence that they're leaning into that more? Are they building more smaller data centers or are they also I I I don't know off the top of my head. Are they competing with like the Colossus 2, these, you know, Mark Zuckerberg comes out with the picture of the cube in Manhattan? And it's very clear that uh that Meta is in the one big massive data center race. Uh at least they're trying to visualize it that way. I haven't seen that from Google. Is that intentional? Is there anything we can read into that about their actual training and deployment strategy?

No, no, you're right. That's a good point. It is it is interesting to to analyze the different I guess frontier AI training architectures from different players.

Yeah.

If you if you look at the what Meta is trying to build in in Louisiana,

uh it's 2.1 gawatt campus for the first phase.

Um and they have individual buildings that are 400 megawatts each. Okay.

Per building.

It's pretty insane. Microsoft open.

But it's still split up. It's still split up. Uh but then putting it all on the same campus, there's some sort of economy of scale around power generation or is it just or is the latency with the fiber connections actually relevant? Uh and you wouldn't want to have it across town so you put it on the same campus like when would when would a a big tech player choose to split across state lines, across the country, across the world versus centralize everything in one campus? So I would say for now Google is really the only one that has been that that had adopted this strategy and is very unique about it to some extent. You could argue it's because of uh first of all they're the most sophisticated on the networking side for for a while. Like that's one of the advantages with Google is for the last 15 years or whatever they've been the best at infrastructure on every single part of infrastructure and so networking they've been building their own fi fiber networks for a while and so they have much higher bandwidth inside of metro and between metros and other other hyperscalers because they've been doing that longer. They've been planning ahead for for all of that.

Yeah. Um so Google is sort of ahead of the others and on the tech side um they've uh they've figured out multi data center training from a from a model perspective way ahead of everyone else but they've been very open about this right you had a podcast from Jeff Dean for example a few months ago where he said openly yes we're doing a multi data center it works pretty well

yeah no one else has really done it uh at that scale for now we have a bunch of startups they're not doing it at scale so yeah just Google is better at doing it and

what it enables them to do is And it's it's it enables them to to some extent have more options with regards to site selection.

Yeah.

Uh you they're not really limited to finding the one 2 gigawatt site or the one you know 1.5 gig sites. Uh they can just go around the metro maybe a 50 mile or even 100 mile area and find a bunch of sites that are each 200 300 400 500 connect them and bam they have a you know two giga campus. Yeah. So that's the the strategy they're pursuing.

That makes sense. Last question for me and we'll let you get back to your day. I know it's late there. Um uh have have uh AI workloads has a shift in AI workloads had any effect on power decision makingaking? Uh it feels like you you go back a couple of years ago and uh the vast majority of power that was being used by AI was for training. Now we're moving more towards inference. Uh does that change the landscape of power acquisition for AI companies broadly or is it sort of an irrelevant point? So first of all, I would actually disagree with the premise. Uh based on our analysis, uh training is still the majority and it's growing equally as fast as inference.

Inference is surging, but training is surging as well. And that's normal because you have an incentive to do it.

Uh everyone wants to, you know, invest in the model that is going to unlock revenue growth next year. We haven't seen the limit. So there's the incentives are aligned to invest today and that's what everyone is doing.

Interesting. Um anyways to answer your question yes inference does have uh different requirements. There's different types of inference as well. Um if you think about open AI for example they have two main businesses. Chad GPT is a vertical application. It's fully controlled by themselves. So for them it's it can larger campuses can be gigawatt scale can be a few hundred megawatt. It's going to be big campus anyways. Uh they can make use of smaller ones. If you think about it from an infrastructure perspective, uh what is their biggest pain point uh as a company is they want to maximize GPU utilization rate is the single largest expense by far is GPU. So they need to maximize it and that's easier to do when you have large campuses. So that's why you still see uh campuses are fairly large even for inference um they also in terms of latency that's a common sort of uh topic of discussion. Do you need to be very close to the end users? Well, if you think about it, what is actually consuming power for OpenAI? It's, you know, deep research. It's GPD 5.2 Pro. Uh it's this thinking models and they they take minutes to answer and so you don't really need to be near the metro. You can just uh be sort of far away and still find large pockets of power.

Yeah, makes sense. Uh well, thank you so much for coming on the show. Uh congratulations on the progress. Really appreciate the semi analysis is hiring, correct? Can you take us on a brief summary of of roles or how to apply? Oh yeah, we're hiring a lot of people. So we have a careers uh section in our website. So you can all check out uh we're looking for folks u in the AI space. So if if you're interested in digging into what we call tokconomics, which is analyzing the economics of AI, analyzing the latest trends in terms of LLMs, different types of model architectures, reach out. Uh if you're an engineer, you have experience with GPU clusters, uh reach out as well. We're hiring to uh to increase our technical team. uh inference max really cool project as well where we benchmark all of the different models. If you want to work on inference max and work on TPUs and tranium and GPUs and AMD and Nvidia and all of that, reach out as well. We have a lot of pretty cool offers. Uh so you should all check out the website and uh yeah, it's a good good adventure.

Amazing. Well, thank you so much for taking the time to come up for us. Great to see you and happy new year. We'll talk to you soon.

Cheers. Goodbye.

Turbo Puffer serverless vector and full text search built from first principles on object storage. fast, 10x cheaper, and extremely scalable.

Um, next we are entering our Lambda Lightning round with Andre Horow. It's