Loading summary

Transcript91 lines

[00:01]
A
Welcome to Galaxy Brains. An infinite amount of cash. Cash. I'm your host, Alex Thorne. The US banking system is sound and resilient. Bitcoin made a new all time high. If you're not long.
[00:12]
B
If you're not long, you're short. Satoshi is going to come on there, laugh hysterically, go quiet. All bitcoin's gonna be erased. Bitcoin.
[00:20]
A
Bitcoin's the best crypto.
[00:23]
B
Bitcoin is going to zero.
[00:27]
A
Welcome back to Galaxy Brains. As always, I'm your host Alex Thorne, head of firmwide research at Galaxy. Bitcoin not Zero. We have a great episode for you this week. Simrat Dinsa, VP of Data Center Capital Markets at Galaxy is our guest and we'll talk with Sim at length about Galaxy's data center business Helios, phase one expansions, energy markets, generally the data center markets and of course we'll talk about AI generally. Some of the interesting controversies that everyone's been following and, and opportunities and growth and its impact on markets. Very good interview with Sim. I think you'll really like meeting Sim. He's worked here for a long time, so I'm really happy to have him on. Before we get to any of that, I need to remind you to please refer to the link to disclaimer in the show notes and note that none of the information in the show constitutes investment advice or an offer recommendation or solicitation by Galaxy or any of its affiliates to buy or sell any secure. Let's go now to our guest, Simrat Dinsa, VP of Data Center Capital Markets at Galaxy Sim. Welcome to Galaxy Brains.
[01:26]
B
Thanks for having me, Alex. Very excited.
[01:28]
A
I'm very excited too. You guys are very, very busy, so it's not often that we have someone from our data center business on this podcast, but we have had Austin Storms in the past. We've had Blake in the past. I think that's it. And actually I'm not sure if we've had either of them on since we fully transition from bitcoin mining to data center. So this is very exciting. I have a lot of questions for you, but what has it been like? You worked on our bitcoin mining portfolio before this giant transition that's now deeply underway into our data center business. What's that transition been like?
[02:01]
B
It's been super exciting. So have been at Galaxy now for four years along with Austin, along with Brian, along with Blake and a few others, and being able to kind of make, you know, our presence in, in West Texas grow out. That presence initially on the Bitcoin mining side and really then try and pivot the business to a higher value use case for that power capacity. Was, was challenging, but a very exciting ride for us. And you know, now we're at a point where we're really executing which is, which is phenomenal for the business.
[02:32]
A
Yeah. So I visited Helios I think maybe three years ago. Absolutely magnificent campus. Then the, the main data hall there, which of course at the time was bitcoin. Absolutely enormous. And the scale in person, you know, I know Galaxy releases sometimes the cool drone footage and stuff and it, it's really hard to actually recognize how big it is until you see it in person. But it's only getting bigger.
[02:59]
B
Yeah, I mean, I think when we, when we bought the site it was like 160 acres. We've grown that now into 2200 acres of just contiguous land. So massive, massive plots of land around, around there. But you know, it's a couple of football fields in length. You can see people working all over the place. I think we had like 1200 or something contractors during construction of just phase one.
[03:19]
A
Yeah.
[03:20]
B
And, and the scale of it is, is at another level compared to what we were used on the bitcoin.
[03:25]
A
So just for, I know a lot of our shareholders on On X are very tapped in to what these mean, but maybe not all of our audience. Phase one was the transition or retrofitting of the existing bitcoin mining data hall into a data center.
[03:43]
B
That's right. Yep, that's right. So when you look at our 800 megawatts that we have leased out to Core Weave in gross capacity, the first 200 megawatts is phase one, which at the end of Q2 we fully bought up, brought online, delivered to Coreweave, you know, rents, rents starting to come in. Very, very exciting milestone for us. And, and you know, it was call it 12 to 14 month construction timeline, which is a very, very fast schedule for building out full redundant systems. It is a tremendous feat for the business and super exciting for us because now when you look at kind of the broader AI data center landscape, these are still relatively new builds at this scale. The hyperscalers have been doing it for several years, but when you look at the traditional data center operators that are public, it's really a lot of the bitcoin mining guys that started to make their pivot into the AI side, started to sign these big deals. And with Galaxy being one of the first ones and now being one of the first ones to actually deliver on those deals which the execution capabilities of our team couldn't be more proud of the team and what we've done across the board from the finance side, construction side and now as we move into operations.
[04:54]
A
Because the most traditional data centers are like tens of megawatts, right? So you have many of them. You have them in and around like big metropolitan areas to serve those areas. Whereas like I guess, yeah, some of the other big hyperscalers have built at hyperscale. But bitcoin miners were among the first to really pioneer building just absolutely giant data center projects. I mean I'm thinking about all the big bitcoin miners. Some of them have absolutely massive. And we had, I mean Helios was massive for bitcoin mining. What can I ask like needed to happen to transition from a bitcoin mining data center to an AI training and inference data center?
[05:35]
B
I mean like what's different?
[05:37]
A
Like I remember when I visited like, you know, Helios was really cool for many reasons. One, because it was all liquid cooled asics. Like does, does that exist for this? Are these liquid cooled gp? I don't what else had to change.
[05:49]
B
It was a pretty dramatic change. I mean outside of kind of the main frame of of our data center. And so you just kept the four walls. Yeah, the four walls are still there. The substation is still there. The transformers. What made our acquisition of Helios so valuable was it came with 800 megawatts of power approved. Plus all the transformers basically installed are on po. So when we look at kind of phase one into phase two and phase three, we have a lot of the long lead infrastructure which enabled us to build on this timeline and deliver for coreweave. Just to kind of give a sense of the scale, the bitcoin mining side was probably sub a million per megawatt to build out a full a full bitcoin mining data center. Now we're looking at 10 to 15 and growing millions per megawatt as infrastructure as supply chains get tighter. So the scale of it has basically increased by a multiple of 10. The complexity of a lot of the tenant fit out. A lot of the internal infrastructure is far more complex than bitcoin mining. Although high level. It was kind of like liquid cool or immersion cooling that we were doing. Now we're doing direct liquid chip cooling. It is completely different hardware in the
[07:01]
A
actual sounds a little similar, but you're saying it's not really.
[07:04]
B
The only real similarity is from the high voltage electrical infrastructure down just basically the power capacity and usable power capacity and the infrastructure that comes with it that's kind of the same, but then when you look at the actual medium low voltage design, that's pretty much changed completely.
[07:21]
A
Totally different user of the electricity.
[07:23]
B
That's right.
[07:24]
A
With its own needs. And then phase two is what? Building more buildings, building out that remaining 800.
[07:31]
B
So phase two is a full greenfield build this time. You know, with phase one, we had the four walls as we were talking about, we had some electrical infrastructure build out. But this one is true. Like, you know, we're starting with the earthwork civil work site work for the, for the actual project, building out the new frame, building out kind of the next phase of substation capacity for it. So it is our first true greenfield build. But that being said, there isn't that much incremental work to add on top of what we had to do for phase one. It was still just basically a blank slate that we were starting with outside of a shell of a building.
[08:05]
A
I liked to think that it was not that different, Asics versus GPUs for AI. But I've been you saying that it basically, other than the electricity stuff, was basically just the four walls of the building.
[08:18]
B
I would say. I mean, the power density inside the data centers are similar. I mean, we're running the same 200 megawatts worth of, of total into the building. Yeah, that's just the, that's just the amount that the building can fit.
[08:29]
A
Right.
[08:29]
B
And, and when you layer on the cooling and everything. So, you know, there is that similarity. But when you look at the actual design, the redundancy components, there was no redundancy. With bitcoin mining, we ramp up, ramp down whenever we wanted to. Now we have to be online 100% of the time.
[08:44]
A
Right, right, right. So one of the other things that I know Galaxy announced, well, certainly since anyone from your team has been on the show, has been the approval by ERCOT for 800 more megawatts of electricity. What does that actually mean? Right, that's a regulatory approval. Right. It's not, it's. And then explain a little bit about ERCOT's process for, I don't know, deciding who gets these approvals or why, or why they're metering it and, you know, why we can't just stand up whatever we want. Like, I don't, I don't know a lot about that.
[09:14]
B
Yeah, absolutely. So I think earlier this year we got approval for an incremental 830 megawatts on top of the initial 800 megawatts that we had with the acquisition that's been leased out to core weave. So this incremental 830 megawatts has not been leased yet. We're having active discussions on that capacity, but it's a super exciting time because we were able to get that approval prior to ERCOT also starting this new batching process as a result of basically the queue growing to like 430 gigawatts of capacity. You have speculators, fake projects, this and
[09:46]
A
that, buying up land and then asking
[09:48]
B
for a queue, flipping land, basically throwing an interconnect study. So I think it is healthy in that it is real projects that are going to be sifted out of this. But the 1.6 gigawatts that we have approved came before this. So we're all set on that front. We can go out and have leasing discussions. We started to put in purchase orders for a lot of the long lead equipment, substation infrastructure transformers and that sort of stuff for this.
[10:13]
A
I remember seeing one of those transformers that had already was already there waiting to be built out like years ago on a cement pallet. Thing is absolutely massive.
[10:21]
B
It's huge.
[10:22]
A
It's like, I don't know, like four or five people tall and wide, like a cube. It's huge.
[10:26]
B
Yeah. No, and, and they take a long
[10:28]
A
time to build, right? Like you have to get a couple of years or something.
[10:31]
B
Yeah, I'd say it's, it's more so the, the delivery timelines now just because, you know, everyone, everyone wants to interconnect. You need these 345kv high voltage transformers to actually be able to pull from the grid. And if you have your private substation that, that's going to step down from high voltage to medium voltage down to the data center. This equipment is probably the most supply constrained across the board. And for us, I think we've had the foresight to lock in a lot of these pos very early on, put the galaxy's balance sheet capital to work on basically ensuring that we have the timelines locked down from the minute that we get approval. So we're not then waiting on a long poll of like, okay, we got approval, but then a year to get the device, then we need transform, you know, so we're trying to get ahead of the game on that front, ahead of the supply chains because it's just going to continue to get constrained just with how much data center growth there continues to be.
[11:25]
A
So. And then I want to ask about West Texas too, because I remember when I was driving to Helios, not only did I from Lubbock, not only did I drive, drive past maybe like 10,000 wind turbines, good wind there as well, apparently, which is also awesome. But also like there is the most giant like transmission lines I think I've ever seen in my life. Like, you know when you're driving down the highway and you see like the really big transmission lines that are moving electricity like from city to city, as opposed to the little ones on your street. These ones were like twice as big as the, the big ones that I, you normally see like in normal America. Right. Sort of next to the site. What, what is that? What is that?
[12:05]
B
Yeah, so I mean these are, these are the high voltage transmission lines. And, and I think it's why we really like the heliosite as well. Like outside of our private substation, we are 100 yards from the Cottonwood substation, which is probably one of the most liquid points in the entire of electricity. Yeah, exactly. So there's a lot of power that's flowing in from kind of the renewables rich area in West Texas down to kind of the main load centers in Dallas and Austin. And we're at a point where we're directly adjacent to one of the most critical pieces of infrastructure in ercot.
[12:38]
A
Right.
[12:39]
B
So from a reliability perspective, it's something that, when we have conversations, you know, that's a key point for a lot of our tenants is ensuring that we can get reliable power.
[12:48]
A
Because you're not like down at the fingertips of the grid. You're like right there, like a major artery.
[12:54]
B
Exactly.
[12:54]
A
Yeah, exactly. It's very cool.
[12:56]
B
So, you know, West Texas I think, and across the board, I mean we've started to see the hyperscalers move to kind of more of these remote regions. I think Google has a few, few data centers out there. You know, there's the Abilene projects with Stargate and now Meta and then, sorry, Stargate and then Microsoft and Meta has a few data centers out there. So people realize now the value that West Texas provide, provides, which is what has been a superfluous amount of capacity and it's easier to, you know, bring out the fiber, bring out the labor, find solutions for water, rather than trying to build out power infrastructure.
[13:30]
A
Right, right.
[13:31]
B
In some of these more metropolitan areas.
[13:32]
A
So super interesting. And then I, like I said, I'd be remiss if I didn't mention that Galaxy announced, I think just last week that it is now the stadium and a title sponsor for the Texas Tech Red Raiders football, which is in Lubbock, a great football team that has a long history of Playing well, but also like, is this. I was just like, the roots are growing deep here. You got all this land, you've now got tenant. You've, you're, you've got approvals. Like, we're big in West Texas now. Like, what was that about in your mind?
[14:05]
B
I think the biggest thing for us is, you know, as we continue to grow out in the area, it is important for us to kind of maintain and be strong stewards within our, within our community. And for us, West Texas has treated us immensely well. And I think for us in being able to bring jobs, high paying jobs, work with the local community on internships and all that, develop our labor force. I think it's just another step in the direction that this is the Kickstarter for us in our data center growth trajectory. And we want to be an important part of the community. We want to give back to the community. So.
[14:43]
A
So it's very cool. It is a cool community. I really enjoy Lubbock when I was there and I got to go back, I think maybe, maybe for a Red Raiders football game. Yeah, I think it's definitely on the top.
[14:51]
B
Champions last year.
[14:52]
A
I know, I know it's going to be exciting. I actually did go, we did go to a game when I was there three years ago and it was awesome. Huge stadium too. Like just great, big fan base. So that'll be fun. Let's pivot a little bit. Let's talk about sort of AI more generally because like I wrote a piece called the Last Model Problem. Pretty like substantially criticizing the Commerce Department for their phone call which they gave anthropic to ban Mythos or fable specifically from non Americans and sort of saying like that surely isn't meant to be the process by which this occurs. Like there was no letter even published. Like they just like called on My guess is my understanding, like, where do you see the, you know, the rise of these models being so powerful and effective, like clashing with or overlapping with government as one first part of the conversation, obviously that got lifted and Fable 5 is now available to people, but.
[15:48]
B
No, absolutely. And this is, this is a topic that I feel like has come really to the forefront as a lot of the kind of the Chinese open source models have started to get better and better. And look, there is an AI arms race happening right now. It's the, you know, the US and Frontier Labs versus basically China and open source. I think it's extremely important for regulators to get this right because if there is regulation that basically hampers the ability to actually accelerate Put out new models. That's something that China is never going to abide by. And it's something that as we kind of compete on the next frontier models, it's important that there's continued innovation and there isn't something that's hampering that growth. Because the minute that we lose the lead, then you have the rest of the world that's utilizing basically infrastructure that's not American infrastructure. So I think it has compounding effects not only on the AI labs, but when we look at the chip side of things where the data centers are being built out from the top down, it is important for us to have, obviously there has to be the right guardrails in place. But if it's anything that's slowing down the cadence of model development, I think that's gonna be really, really tough for the US to compete with China. And I think one last point. When you look at how the industry has evolved, a few months back people were saying open source is kind of 12 months behind, six months behind. Now
[17:21]
A
the new Kimi model is amongst
[17:23]
B
the bleeding edge and there's still concerns or questions on whether they distill the models, whatever. Basically they have a model that is competitive with some of our frontier models. And I think that that's just going to increase the cadence of new releases, ensuring that, you know, the American labs can maintain their moat basically and continue to utilize better and better models for their own internal development to kind of have this flywheel. So, you know, I think, I think it's, it's, it's certainly a very important topic to ensure that, you know, the right safety guardrails are there. But if it's going to lead to six 12 month review periods, we're just going to fall behind on that.
[18:01]
A
Yeah, I think it's very difficult, it's very tricky and a lot of people have hypothesized different outcomes. I love those essays. AI 2027 and AI 2040, which have different game theoretical paths that could occur depending on choices. And they, you know, I would say it seems pretty likely that the model development maybe in China as well one day, but definitely not now, really will become subject to substantial government oversight. Not just because they get so smart, but also because they're so critical to economies. I mean I was, somebody was asking on X, I saw a tweet that was like, are there developers that aren't using coding tools? AI coding tools, right. Do we, do you think there's even one at this point? Like, why wouldn't you be? Yeah, it's crazy. So and then do you think the open source, I mean that doesn't affect like, I mean like I'm using open source models at home, but they're, you know, quantized. Right. As I was called. They're much smaller, smaller models. The big ones still need like giant data center infrastructure. Right. So it's not about it affecting the
[19:12]
B
AI capex, it's about, if anything, I still think it continues to accelerate the, the CapEx. Because where these open source models are coming into play is actually for broader enterprise adoption. You can either Pay, I don't know, 50, $60 per million tokens or whatever for anthropic and OpenAI or you pay 50 cents for your own instance of kind of the bleeding edge.
[19:38]
A
So then you, rather than using Anthropic or OpenAI's data center, you end up trying to get your own data center capacity directly.
[19:45]
B
It completely diversifies the kind of model layer which for the labs there might be certain pricing pressure that comes with that where they may have to focus more on kind of the application layer. But for the NEO clouds, for the data centers, all of that demand still has to be serviced. And if you have cheaper and cheaper models coming out, they still have power footprints and the hyperscalers want to service this power on their clouds. Hyperscalers need data centers to service this, this demand. So it's all, it's all going back to Jevons paradox basically where intelligence is just getting cheaper and cheaper and that
[20:22]
A
induces more and more demand.
[20:23]
B
Exactly.
[20:24]
A
Because like. Right. You're making it possible for many more enterprises to use AI because they pay less overall.
[20:31]
B
Exactly. That's exactly.
[20:33]
A
I mean these are, I mean Fable 5 is a token burner. I used it fair amount. And you hit those limits.
[20:41]
B
Token maxing project.
[20:42]
A
Yes it is. Yes, it is good. It's quite good. But it is, I'm hitting that. I have the at home the max plan which is like the 5x of the regular pro plan and they're giving us all 50% of our weekly usage can be Fable 5 and I hit that resets on Wednesdays. I usually hit it by like Friday morning.
[20:59]
B
Yeah, yeah.
[21:00]
A
I mean back to opus.
[21:02]
B
Yeah. I feel like Alex, you, you've got some pretty unique projects you're probably working on.
[21:05]
A
Yeah, we'll, we'll talk more about that I think in the future. But what do you last sort of question here? I mean so much has changed. I feel like this, I want to ask you what you think the world, the data center world, the AI world, the Economy what it looks like in five years, it can be two or four, you know what I mean? Like, let's look ahead a little bit into Sims Crystal ball.
[21:25]
B
Yeah.
[21:26]
A
And it would love to hear any interesting things you think to expect. And before you answer, I just preface to remind the audience that basically no one was using AI coding tools in December of just last year. Everyone started using them like January, February, March.
[21:44]
B
Right.
[21:45]
A
So like that's only what, four or five months ago. And it has exploded in usage and quality just since like, you know, the difference of like the 4O models to the 5.5 GPT models and the, you know, the, I don't Even, you know, Opus 4 is to like 4.8 is dramatic. And all that really just happened in like the last seven months. So knowing that, where do you think we go in a couple years?
[22:13]
B
Yeah, yeah. Well, actually to that last point that you made on just kind of what has felt like almost a step function in intelligence over the past really few weeks, where it's actually interesting because it kind of goes down all the way down the infrastructure stack where these models that are coming out are really kind of the first Blackwell trained models at scale and larger and larger scale. The newer chips, the newer Nvidia chips.
[22:39]
A
Yeah.
[22:39]
B
And it's going to be really exciting to see as we kind of continue down, whether it be Nvidia or kind of across the chip stack, as well as newer chips are being deployed in larger and larger scale clusters, especially at Helios, how some of this intelligent step functions continues to increase. We're, you know, later this year. I think the Vera Rubins are kind of that generation of machines are going to start being deployed probably more so kind of into next year. But that's where things get really interesting and for the broader consumer, one, you have models that are getting cheaper. Two, you have a chip stack that is also getting way more efficient. How does this kind of diffuse into the broader enterprise ecosystem is something that I feel like we've gotten glimpses of, but it hasn't really been that pervasive that people work through kind of their internal legal frameworks and so on and so forth. And anecdotally, I'm not sure if you saw the CEO of Palantir, he had a CNBC interview where there is, I think what's going to be a very interesting trend in enterprises wanting to own kind of their own whole stack as well. I think Satya Nadella has basically called it your token capital, which is a new asset, effectively that's on your Balance sheet your internal intelligence at your company and how you can actually turn that into outputs with these new models, make it more efficient in helping with automating tasks or streamlining different projects. I think that is something that's going to be really cool to watch, is these enterprises coming in more and more into the participation in the AI game rather than kind of offloading it to, you know, kind of the labs to deal with, you know, people owning their own destiny and trying to really vertically integrate across that.
[24:30]
A
Yeah. One of the things that I've found so useful, a very useful use of AI tools is to have your own massive data lake and proprietary data and then ask the AI to evaluate it when you need it evaluated.
[24:46]
B
Right.
[24:46]
A
So like for example, I put out this report about whether or not Bitcoin had bottomed and whether the, the pricing indicators that had reliably, you know, on chain data into mvrv, all that stuff. If the ones that had reliably bottomed at prior price bottoms in other bear markets were those present today, I, I can do that, I can do that on glass note, I can look at them all and it can be like, okay, that one hit here and that one hit here, it would take me so long. But we have all those metrics, we've built all those metrics ourselves. So I can just be like, just go look, look across our database. And it's so good at that and so fast. And so to sort of buttress your point here, you know, companies have a lot of their own data and asking the AI to help them sift through it, synthesize it, present it is an incredibly powerful use case today. And I can imagine, rather than just like blindly asking Claude, hey, can you go find out if there are bottoms? Like, I don't know what data cloud's relying on. I want to rely on my data, but I may also not want to give all my data to Claude. So I can really see that becoming a. Companies are loath to give up their proprietary information. Right. So instead run the date, run it local basically.
[25:56]
B
Exactly. I think it's going to be existential for, you know, basically across industries to build your AI moat within that industry. And part of that is building your token capital base and being able to kind of create your stack that helps you do what you do best and kind of proliferate AI across kind of the enterprise.
[26:17]
A
So really crazy.
[26:19]
B
It's going to be interesting. It's going to be a wild ride.
[26:20]
A
We've come so far just in 2026, like I wanted to build this thing I'm sort of vaguely describing with this Bitcoin data. And I was going to, we're going to maybe pay to build it. And then we came back from like Christmas vacation basically in January, and I was asking my developer here, I was like, I was like, wait a sec, can we like, vibe code this? We're still calling it vibe coding now. It's legitimately just coding. He was like, way ahead of you, boss. Like, already did a lot of it. And that was six months ago. So, I mean, I just can't imagine where this is going. And I didn't even thought your point about the underlying stack, that the chip manufacturers getting better, the machine manufacturers getting better, that's allowing the. Even obviously from the top, the software model, the model design is also improving.
[27:07]
B
That's exactly.
[27:07]
A
But when they can train it so much better because they have so much more compute. I mean, I'm not even sure what, you know that we're going to be talking to a hologram the next year when you're here.
[27:18]
B
Maybe that's next time I'm on this podcast.
[27:19]
A
You can appear on 10 podcasts at once. Well, Sim Densa, thank you so much for coming on Galaxy Brains.
[27:26]
B
Thank you for having me.
[27:32]
A
Thank you for listening to Galaxy Brains, the weekly podcast from Galaxy Research. I'm Alex Thorne, head of firmwide research at Galaxy. Follow me on X at Intangible Coins, follow Galaxy Research on X at GLXY Research. Read our written reports at galaxy. Com Research. And don't forget, if you like Galaxy Brains to like and subscribe on your favorite podcast platforms like YouTube, Spotify, Apple Podcasts and more. We'll see you next time.

Summary6 min read

Galaxy Brains – "Datacenter Expansion and Strategy with Simrit Dhinsa"

Podcast: Galaxy Brains
Host: Alex Thorne (Head of Research, Galaxy)
Guest: Simrit Dhinsa (VP of Data Center Capital Markets, Galaxy)
Date: July 30, 2026

Episode Overview

This week’s episode delves into Galaxy’s transformational pivot from bitcoin mining to data center operations, focusing on the Helios project in West Texas. Alex Thorn and guest Simrit Dhinsa explore the technical and strategic evolution, challenges of retrofitting infrastructure for AI workloads, expansion plans, the intersection with energy markets, regional economic impact, and broader trends surrounding AI, regulation, and enterprise adoption.

Key Discussion Points and Insights

1. Transition from Bitcoin Mining to Data Centers

Background: Galaxy Digital previously operated large-scale bitcoin mining facilities, with Helios as a flagship site.
Pivot: The company has aggressively transitioned to focus on high-value use cases—namely, AI training and inference data centers.
Scale: The Helios campus evolved from 160 to over 2,200 contiguous acres, supporting massive expansions in electrical capacity and physical infrastructure.
Team Experience: Many team members, including Simrit, have been with Galaxy through this transformation, contributing to both bitcoin mining and data center arms.
- Quote:
  “Being able to pivot the business to a higher value use case for that power capacity was challenging, but a very exciting ride for us.”
  — Simrit Dhinsa [02:01]

2. The Helios Project: Technical and Strategic Expansion

Physical Scale: The site now houses massive data halls—“a couple of football fields in length” with 1,200 contractors involved during phase one. [02:58]
Phase One: Involved retrofitting existing bitcoin mining infrastructure for AI, specifically for CoreWeave. Achieved 200 MW of capacity delivered in “12 to 14 months,” an unusually rapid schedule.
- Quote:
  “It is a tremendous feat for the business and super exciting for us because now when you look at the broader AI data center landscape, these are still relatively new builds at this scale.”
  — Simrit Dhinsa [04:05]
Phase Two: Marks Galaxy’s first true greenfield build—entirely new construction, leveraging prior experience and pre-installed infrastructure for speed.
Cost Differential: Retrofit and expansion toward AI data centers cost “10 to 15 and growing millions per megawatt,” up from sub-$1 million per MW for mining. Complexity and redundancy demands have risen dramatically.
- Quote:
  “The complexity of a lot of the tenant fit out. A lot of the internal infrastructure is far more complex than bitcoin mining.”
  — Simrit Dhinsa [06:00]
Technical Evolution: Retained only “the four walls and the substation”—all internal electrical and cooling systems substantially redesigned for AI GPU clusters and 24/7 reliability needs.

3. Power and Regulatory Approvals in West Texas

ERCOT Approval: Galaxy secured approval for an additional 830 MW (totaling 1.6 GW) of capacity pre-batching queue, giving them a regulatory and supply chain advantage.
Background: Texas’ grid operator ERCOT has imposed new batching due to “430 GW” of speculative queue requests.
Supply Chain: Transformers and other HV equipment are major bottlenecks; Galaxy preemptively secured them with early POs, aided by its capital strength.
- Quote:
  “This equipment is probably the most supply constrained across the board. And for us, I think we've had the foresight to lock in a lot of these POs very early on...”
  — Simrit Dhinsa [10:31]

4. Geographical and Grid Advantages

West Texas Location: The Helios campus sits adjacent to one of ERCOT’s most liquid substations, surrounded by dense transmission infrastructure and renewables (notably wind).
- Quote:
  “We're directly adjacent to one of the most critical pieces of infrastructure in ERCOT.”
  — Simrit Dhinsa [12:05–12:38]
Trend: Hyperscalers and major tech firms (Google, Microsoft, Meta) are increasingly siting data centers in remote, power-rich regions outside congested metros.
Community Engagement: Galaxy sponsors local initiatives such as Texas Tech Red Raiders football, emphasizes jobs, internships, and integrating with the regional community.
- Quote:
  “It is important for us to be strong stewards within our community... this is the Kickstarter for us in our data center growth trajectory.”
  — Simrit Dhinsa [14:05]

5. AI Models, Regulation & Global Competition

US–China AI "Arms Race": The pace of frontier model development is entangled with regulatory, national security, and open source debates.
- Concerns about the US losing its lead if regulation “hampers innovation” while competitors progress unimpeded.
- Quote:
  “If there is regulation that basically hampers the ability to actually accelerate, put out new models... China is never going to abide by [that].”
  — Simrit Dhinsa [15:48]
Open Source Acceleration: Open source models (notably some from China) have closed the quality gap with US frontier labs—accelerating the cadence of releases and increasing pricing pressure on major labs.
Enterprise CapEx: Proliferation of open source models is driving more enterprises and “neo-clouds” to pursue their own infrastructure, intensifying demand for data center capacity (Jevons Paradox applied to AI).
- Quote: “It's all going back to Jevons paradox basically where intelligence is just getting cheaper and cheaper and that induces more and more demand.”
  — Simrit Dhinsa [20:21]

6. Future Trends: Crystal Ball

Rapid Model Evolution: Dramatic, stepwise leaps in AI capability as new NVIDIA (Blackwell, Vera Rubin) and similar chips come online, with scale and efficiency improvements feeding faster, better training.
- Quote:
  “We're, you know, later this year... larger and larger scale clusters, especially at Helios, how some of this intelligent step functions continues to increase.”
  — Simrit Dhinsa [22:39]
Enterprise Stack Ownership: Shift toward enterprises building proprietary “token capital”—their own AI/data stack, leveraging internal data lakes for custom use cases without surrendering proprietary info to third parties.
- Quote:
  “I think that is something that's going to be really cool to watch, is these enterprises coming in more and more into the participation in the AI game rather than offloading it...”
  — Simrit Dhinsa [24:29]
Explosion in Usage: AI tools, especially for coding and data synthesis, have become nearly ubiquitous in the developer world—with recent adoption and improvement compressed into just a few months.

Notable Quotes & Memorable Moments

“I mean, I think when we bought the site it was like 160 acres. We've grown that now into 2200 acres of just contiguous land.” — Simrit Dhinsa [02:58]
“Retrofit and expansion toward AI data centers cost ‘10 to 15 and growing millions per megawatt,’ up from sub-$1 million per MW for mining.” [07:00]
“We're directly adjacent to one of the most critical pieces of infrastructure in ERCOT.” — Simrit Dhinsa [12:05]
“If there is regulation that basically hampers the ability to actually accelerate... that’s something China is never going to abide by.” — Simrit Dhinsa [15:48]
“It's all going back to Jevons paradox basically where intelligence is just getting cheaper and cheaper and that induces more and more demand.” — Simrit Dhinsa [20:21]
“I think that is going to be existential for basically across industries to build your AI moat within that industry... build your token capital base and... vertically integrate.” — Simrit Dhinsa [25:55]

Timestamps for Key Segments

00:27 — Introduction to Galaxy’s data center pivot and Simrit’s background
02:01 — Transitioning the Helios facility for AI (scope, team, challenges)
03:42 — Retrofitting vs. new build costs, technical challenges
08:43 — ERCOT approvals and Texas energy markets
12:05 — West Texas grid advantages, regional strategy
14:05 — Galaxy's community engagement with Texas Tech
15:48 — The US–China AI arms race and regulatory risks
19:11 — Open source models, enterprise CapEx, and data center demand surge
22:12 — Rapid progress in AI, implications for infrastructure and enterprise
24:29 — The future of enterprise stack ownership and data utilization

Summary

This episode offers a deep dive into how Galaxy is leveraging its bitcoin mining heritage and massive West Texas location to ride the enormous wave of AI compute needs. With careful regulatory strategy, robust supply chains, and an eye for future enterprise trends, Galaxy is positioning itself at the cutting edge of the data center and AI revolution. The conversation underscores not just technical and business transformations, but also the broader significance of energy markets, US–China rivalry, regulatory policy, and the explosive, unpredictable pace of AI adoption.

Loading summary

Transcript91 lines

[00:01]
A
Welcome to Galaxy Brains. An infinite amount of cash. Cash. I'm your host, Alex Thorne. The US banking system is sound and resilient. Bitcoin made a new all time high. If you're not long.
[00:12]
B
If you're not long, you're short. Satoshi is going to come on there, laugh hysterically, go quiet. All bitcoin's gonna be erased. Bitcoin.
[00:20]
A
Bitcoin's the best crypto.
[00:23]
B
Bitcoin is going to zero.
[00:27]
A
Welcome back to Galaxy Brains. As always, I'm your host Alex Thorne, head of firmwide research at Galaxy. Bitcoin not Zero. We have a great episode for you this week. Simrat Dinsa, VP of Data Center Capital Markets at Galaxy is our guest and we'll talk with Sim at length about Galaxy's data center business Helios, phase one expansions, energy markets, generally the data center markets and of course we'll talk about AI generally. Some of the interesting controversies that everyone's been following and, and opportunities and growth and its impact on markets. Very good interview with Sim. I think you'll really like meeting Sim. He's worked here for a long time, so I'm really happy to have him on. Before we get to any of that, I need to remind you to please refer to the link to disclaimer in the show notes and note that none of the information in the show constitutes investment advice or an offer recommendation or solicitation by Galaxy or any of its affiliates to buy or sell any secure. Let's go now to our guest, Simrat Dinsa, VP of Data Center Capital Markets at Galaxy Sim. Welcome to Galaxy Brains.
[01:26]
B
Thanks for having me, Alex. Very excited.
[01:28]
A
I'm very excited too. You guys are very, very busy, so it's not often that we have someone from our data center business on this podcast, but we have had Austin Storms in the past. We've had Blake in the past. I think that's it. And actually I'm not sure if we've had either of them on since we fully transition from bitcoin mining to data center. So this is very exciting. I have a lot of questions for you, but what has it been like? You worked on our bitcoin mining portfolio before this giant transition that's now deeply underway into our data center business. What's that transition been like?
[02:01]
B
It's been super exciting. So have been at Galaxy now for four years along with Austin, along with Brian, along with Blake and a few others, and being able to kind of make, you know, our presence in, in West Texas grow out. That presence initially on the Bitcoin mining side and really then try and pivot the business to a higher value use case for that power capacity. Was, was challenging, but a very exciting ride for us. And you know, now we're at a point where we're really executing which is, which is phenomenal for the business.
[02:32]
A
Yeah. So I visited Helios I think maybe three years ago. Absolutely magnificent campus. Then the, the main data hall there, which of course at the time was bitcoin. Absolutely enormous. And the scale in person, you know, I know Galaxy releases sometimes the cool drone footage and stuff and it, it's really hard to actually recognize how big it is until you see it in person. But it's only getting bigger.
[02:59]
B
Yeah, I mean, I think when we, when we bought the site it was like 160 acres. We've grown that now into 2200 acres of just contiguous land. So massive, massive plots of land around, around there. But you know, it's a couple of football fields in length. You can see people working all over the place. I think we had like 1200 or something contractors during construction of just phase one.
[03:19]
A
Yeah.
[03:20]
B
And, and the scale of it is, is at another level compared to what we were used on the bitcoin.
[03:25]
A
So just for, I know a lot of our shareholders on On X are very tapped in to what these mean, but maybe not all of our audience. Phase one was the transition or retrofitting of the existing bitcoin mining data hall into a data center.
[03:43]
B
That's right. Yep, that's right. So when you look at our 800 megawatts that we have leased out to Core Weave in gross capacity, the first 200 megawatts is phase one, which at the end of Q2 we fully bought up, brought online, delivered to Coreweave, you know, rents, rents starting to come in. Very, very exciting milestone for us. And, and you know, it was call it 12 to 14 month construction timeline, which is a very, very fast schedule for building out full redundant systems. It is a tremendous feat for the business and super exciting for us because now when you look at kind of the broader AI data center landscape, these are still relatively new builds at this scale. The hyperscalers have been doing it for several years, but when you look at the traditional data center operators that are public, it's really a lot of the bitcoin mining guys that started to make their pivot into the AI side, started to sign these big deals. And with Galaxy being one of the first ones and now being one of the first ones to actually deliver on those deals which the execution capabilities of our team couldn't be more proud of the team and what we've done across the board from the finance side, construction side and now as we move into operations.
[04:54]
A
Because the most traditional data centers are like tens of megawatts, right? So you have many of them. You have them in and around like big metropolitan areas to serve those areas. Whereas like I guess, yeah, some of the other big hyperscalers have built at hyperscale. But bitcoin miners were among the first to really pioneer building just absolutely giant data center projects. I mean I'm thinking about all the big bitcoin miners. Some of them have absolutely massive. And we had, I mean Helios was massive for bitcoin mining. What can I ask like needed to happen to transition from a bitcoin mining data center to an AI training and inference data center?
[05:35]
B
I mean like what's different?
[05:37]
A
Like I remember when I visited like, you know, Helios was really cool for many reasons. One, because it was all liquid cooled asics. Like does, does that exist for this? Are these liquid cooled gp? I don't what else had to change.
[05:49]
B
It was a pretty dramatic change. I mean outside of kind of the main frame of of our data center. And so you just kept the four walls. Yeah, the four walls are still there. The substation is still there. The transformers. What made our acquisition of Helios so valuable was it came with 800 megawatts of power approved. Plus all the transformers basically installed are on po. So when we look at kind of phase one into phase two and phase three, we have a lot of the long lead infrastructure which enabled us to build on this timeline and deliver for coreweave. Just to kind of give a sense of the scale, the bitcoin mining side was probably sub a million per megawatt to build out a full a full bitcoin mining data center. Now we're looking at 10 to 15 and growing millions per megawatt as infrastructure as supply chains get tighter. So the scale of it has basically increased by a multiple of 10. The complexity of a lot of the tenant fit out. A lot of the internal infrastructure is far more complex than bitcoin mining. Although high level. It was kind of like liquid cool or immersion cooling that we were doing. Now we're doing direct liquid chip cooling. It is completely different hardware in the
[07:01]
A
actual sounds a little similar, but you're saying it's not really.
[07:04]
B
The only real similarity is from the high voltage electrical infrastructure down just basically the power capacity and usable power capacity and the infrastructure that comes with it that's kind of the same, but then when you look at the actual medium low voltage design, that's pretty much changed completely.
[07:21]
A
Totally different user of the electricity.
[07:23]
B
That's right.
[07:24]
A
With its own needs. And then phase two is what? Building more buildings, building out that remaining 800.
[07:31]
B
So phase two is a full greenfield build this time. You know, with phase one, we had the four walls as we were talking about, we had some electrical infrastructure build out. But this one is true. Like, you know, we're starting with the earthwork civil work site work for the, for the actual project, building out the new frame, building out kind of the next phase of substation capacity for it. So it is our first true greenfield build. But that being said, there isn't that much incremental work to add on top of what we had to do for phase one. It was still just basically a blank slate that we were starting with outside of a shell of a building.
[08:05]
A
I liked to think that it was not that different, Asics versus GPUs for AI. But I've been you saying that it basically, other than the electricity stuff, was basically just the four walls of the building.
[08:18]
B
I would say. I mean, the power density inside the data centers are similar. I mean, we're running the same 200 megawatts worth of, of total into the building. Yeah, that's just the, that's just the amount that the building can fit.
[08:29]
A
Right.
[08:29]
B
And, and when you layer on the cooling and everything. So, you know, there is that similarity. But when you look at the actual design, the redundancy components, there was no redundancy. With bitcoin mining, we ramp up, ramp down whenever we wanted to. Now we have to be online 100% of the time.
[08:44]
A
Right, right, right. So one of the other things that I know Galaxy announced, well, certainly since anyone from your team has been on the show, has been the approval by ERCOT for 800 more megawatts of electricity. What does that actually mean? Right, that's a regulatory approval. Right. It's not, it's. And then explain a little bit about ERCOT's process for, I don't know, deciding who gets these approvals or why, or why they're metering it and, you know, why we can't just stand up whatever we want. Like, I don't, I don't know a lot about that.
[09:14]
B
Yeah, absolutely. So I think earlier this year we got approval for an incremental 830 megawatts on top of the initial 800 megawatts that we had with the acquisition that's been leased out to core weave. So this incremental 830 megawatts has not been leased yet. We're having active discussions on that capacity, but it's a super exciting time because we were able to get that approval prior to ERCOT also starting this new batching process as a result of basically the queue growing to like 430 gigawatts of capacity. You have speculators, fake projects, this and
[09:46]
A
that, buying up land and then asking
[09:48]
B
for a queue, flipping land, basically throwing an interconnect study. So I think it is healthy in that it is real projects that are going to be sifted out of this. But the 1.6 gigawatts that we have approved came before this. So we're all set on that front. We can go out and have leasing discussions. We started to put in purchase orders for a lot of the long lead equipment, substation infrastructure transformers and that sort of stuff for this.
[10:13]
A
I remember seeing one of those transformers that had already was already there waiting to be built out like years ago on a cement pallet. Thing is absolutely massive.
[10:21]
B
It's huge.
[10:22]
A
It's like, I don't know, like four or five people tall and wide, like a cube. It's huge.
[10:26]
B
Yeah. No, and, and they take a long
[10:28]
A
time to build, right? Like you have to get a couple of years or something.
[10:31]
B
Yeah, I'd say it's, it's more so the, the delivery timelines now just because, you know, everyone, everyone wants to interconnect. You need these 345kv high voltage transformers to actually be able to pull from the grid. And if you have your private substation that, that's going to step down from high voltage to medium voltage down to the data center. This equipment is probably the most supply constrained across the board. And for us, I think we've had the foresight to lock in a lot of these pos very early on, put the galaxy's balance sheet capital to work on basically ensuring that we have the timelines locked down from the minute that we get approval. So we're not then waiting on a long poll of like, okay, we got approval, but then a year to get the device, then we need transform, you know, so we're trying to get ahead of the game on that front, ahead of the supply chains because it's just going to continue to get constrained just with how much data center growth there continues to be.
[11:25]
A
So. And then I want to ask about West Texas too, because I remember when I was driving to Helios, not only did I from Lubbock, not only did I drive, drive past maybe like 10,000 wind turbines, good wind there as well, apparently, which is also awesome. But also like there is the most giant like transmission lines I think I've ever seen in my life. Like, you know when you're driving down the highway and you see like the really big transmission lines that are moving electricity like from city to city, as opposed to the little ones on your street. These ones were like twice as big as the, the big ones that I, you normally see like in normal America. Right. Sort of next to the site. What, what is that? What is that?
[12:05]
B
Yeah, so I mean these are, these are the high voltage transmission lines. And, and I think it's why we really like the heliosite as well. Like outside of our private substation, we are 100 yards from the Cottonwood substation, which is probably one of the most liquid points in the entire of electricity. Yeah, exactly. So there's a lot of power that's flowing in from kind of the renewables rich area in West Texas down to kind of the main load centers in Dallas and Austin. And we're at a point where we're directly adjacent to one of the most critical pieces of infrastructure in ercot.
[12:38]
A
Right.
[12:39]
B
So from a reliability perspective, it's something that, when we have conversations, you know, that's a key point for a lot of our tenants is ensuring that we can get reliable power.
[12:48]
A
Because you're not like down at the fingertips of the grid. You're like right there, like a major artery.
[12:54]
B
Exactly.
[12:54]
A
Yeah, exactly. It's very cool.
[12:56]
B
So, you know, West Texas I think, and across the board, I mean we've started to see the hyperscalers move to kind of more of these remote regions. I think Google has a few, few data centers out there. You know, there's the Abilene projects with Stargate and now Meta and then, sorry, Stargate and then Microsoft and Meta has a few data centers out there. So people realize now the value that West Texas provide, provides, which is what has been a superfluous amount of capacity and it's easier to, you know, bring out the fiber, bring out the labor, find solutions for water, rather than trying to build out power infrastructure.
[13:30]
A
Right, right.
[13:31]
B
In some of these more metropolitan areas.
[13:32]
A
So super interesting. And then I, like I said, I'd be remiss if I didn't mention that Galaxy announced, I think just last week that it is now the stadium and a title sponsor for the Texas Tech Red Raiders football, which is in Lubbock, a great football team that has a long history of Playing well, but also like, is this. I was just like, the roots are growing deep here. You got all this land, you've now got tenant. You've, you're, you've got approvals. Like, we're big in West Texas now. Like, what was that about in your mind?
[14:05]
B
I think the biggest thing for us is, you know, as we continue to grow out in the area, it is important for us to kind of maintain and be strong stewards within our, within our community. And for us, West Texas has treated us immensely well. And I think for us in being able to bring jobs, high paying jobs, work with the local community on internships and all that, develop our labor force. I think it's just another step in the direction that this is the Kickstarter for us in our data center growth trajectory. And we want to be an important part of the community. We want to give back to the community. So.
[14:43]
A
So it's very cool. It is a cool community. I really enjoy Lubbock when I was there and I got to go back, I think maybe, maybe for a Red Raiders football game. Yeah, I think it's definitely on the top.
[14:51]
B
Champions last year.
[14:52]
A
I know, I know it's going to be exciting. I actually did go, we did go to a game when I was there three years ago and it was awesome. Huge stadium too. Like just great, big fan base. So that'll be fun. Let's pivot a little bit. Let's talk about sort of AI more generally because like I wrote a piece called the Last Model Problem. Pretty like substantially criticizing the Commerce Department for their phone call which they gave anthropic to ban Mythos or fable specifically from non Americans and sort of saying like that surely isn't meant to be the process by which this occurs. Like there was no letter even published. Like they just like called on My guess is my understanding, like, where do you see the, you know, the rise of these models being so powerful and effective, like clashing with or overlapping with government as one first part of the conversation, obviously that got lifted and Fable 5 is now available to people, but.
[15:48]
B
No, absolutely. And this is, this is a topic that I feel like has come really to the forefront as a lot of the kind of the Chinese open source models have started to get better and better. And look, there is an AI arms race happening right now. It's the, you know, the US and Frontier Labs versus basically China and open source. I think it's extremely important for regulators to get this right because if there is regulation that basically hampers the ability to actually accelerate Put out new models. That's something that China is never going to abide by. And it's something that as we kind of compete on the next frontier models, it's important that there's continued innovation and there isn't something that's hampering that growth. Because the minute that we lose the lead, then you have the rest of the world that's utilizing basically infrastructure that's not American infrastructure. So I think it has compounding effects not only on the AI labs, but when we look at the chip side of things where the data centers are being built out from the top down, it is important for us to have, obviously there has to be the right guardrails in place. But if it's anything that's slowing down the cadence of model development, I think that's gonna be really, really tough for the US to compete with China. And I think one last point. When you look at how the industry has evolved, a few months back people were saying open source is kind of 12 months behind, six months behind. Now
[17:21]
A
the new Kimi model is amongst
[17:23]
B
the bleeding edge and there's still concerns or questions on whether they distill the models, whatever. Basically they have a model that is competitive with some of our frontier models. And I think that that's just going to increase the cadence of new releases, ensuring that, you know, the American labs can maintain their moat basically and continue to utilize better and better models for their own internal development to kind of have this flywheel. So, you know, I think, I think it's, it's, it's certainly a very important topic to ensure that, you know, the right safety guardrails are there. But if it's going to lead to six 12 month review periods, we're just going to fall behind on that.
[18:01]
A
Yeah, I think it's very difficult, it's very tricky and a lot of people have hypothesized different outcomes. I love those essays. AI 2027 and AI 2040, which have different game theoretical paths that could occur depending on choices. And they, you know, I would say it seems pretty likely that the model development maybe in China as well one day, but definitely not now, really will become subject to substantial government oversight. Not just because they get so smart, but also because they're so critical to economies. I mean I was, somebody was asking on X, I saw a tweet that was like, are there developers that aren't using coding tools? AI coding tools, right. Do we, do you think there's even one at this point? Like, why wouldn't you be? Yeah, it's crazy. So and then do you think the open source, I mean that doesn't affect like, I mean like I'm using open source models at home, but they're, you know, quantized. Right. As I was called. They're much smaller, smaller models. The big ones still need like giant data center infrastructure. Right. So it's not about it affecting the
[19:12]
B
AI capex, it's about, if anything, I still think it continues to accelerate the, the CapEx. Because where these open source models are coming into play is actually for broader enterprise adoption. You can either Pay, I don't know, 50, $60 per million tokens or whatever for anthropic and OpenAI or you pay 50 cents for your own instance of kind of the bleeding edge.
[19:38]
A
So then you, rather than using Anthropic or OpenAI's data center, you end up trying to get your own data center capacity directly.
[19:45]
B
It completely diversifies the kind of model layer which for the labs there might be certain pricing pressure that comes with that where they may have to focus more on kind of the application layer. But for the NEO clouds, for the data centers, all of that demand still has to be serviced. And if you have cheaper and cheaper models coming out, they still have power footprints and the hyperscalers want to service this power on their clouds. Hyperscalers need data centers to service this, this demand. So it's all, it's all going back to Jevons paradox basically where intelligence is just getting cheaper and cheaper and that
[20:22]
A
induces more and more demand.
[20:23]
B
Exactly.
[20:24]
A
Because like. Right. You're making it possible for many more enterprises to use AI because they pay less overall.
[20:31]
B
Exactly. That's exactly.
[20:33]
A
I mean these are, I mean Fable 5 is a token burner. I used it fair amount. And you hit those limits.
[20:41]
B
Token maxing project.
[20:42]
A
Yes it is. Yes, it is good. It's quite good. But it is, I'm hitting that. I have the at home the max plan which is like the 5x of the regular pro plan and they're giving us all 50% of our weekly usage can be Fable 5 and I hit that resets on Wednesdays. I usually hit it by like Friday morning.
[20:59]
B
Yeah, yeah.
[21:00]
A
I mean back to opus.
[21:02]
B
Yeah. I feel like Alex, you, you've got some pretty unique projects you're probably working on.
[21:05]
A
Yeah, we'll, we'll talk more about that I think in the future. But what do you last sort of question here? I mean so much has changed. I feel like this, I want to ask you what you think the world, the data center world, the AI world, the Economy what it looks like in five years, it can be two or four, you know what I mean? Like, let's look ahead a little bit into Sims Crystal ball.
[21:25]
B
Yeah.
[21:26]
A
And it would love to hear any interesting things you think to expect. And before you answer, I just preface to remind the audience that basically no one was using AI coding tools in December of just last year. Everyone started using them like January, February, March.
[21:44]
B
Right.
[21:45]
A
So like that's only what, four or five months ago. And it has exploded in usage and quality just since like, you know, the difference of like the 4O models to the 5.5 GPT models and the, you know, the, I don't Even, you know, Opus 4 is to like 4.8 is dramatic. And all that really just happened in like the last seven months. So knowing that, where do you think we go in a couple years?
[22:13]
B
Yeah, yeah. Well, actually to that last point that you made on just kind of what has felt like almost a step function in intelligence over the past really few weeks, where it's actually interesting because it kind of goes down all the way down the infrastructure stack where these models that are coming out are really kind of the first Blackwell trained models at scale and larger and larger scale. The newer chips, the newer Nvidia chips.
[22:39]
A
Yeah.
[22:39]
B
And it's going to be really exciting to see as we kind of continue down, whether it be Nvidia or kind of across the chip stack, as well as newer chips are being deployed in larger and larger scale clusters, especially at Helios, how some of this intelligent step functions continues to increase. We're, you know, later this year. I think the Vera Rubins are kind of that generation of machines are going to start being deployed probably more so kind of into next year. But that's where things get really interesting and for the broader consumer, one, you have models that are getting cheaper. Two, you have a chip stack that is also getting way more efficient. How does this kind of diffuse into the broader enterprise ecosystem is something that I feel like we've gotten glimpses of, but it hasn't really been that pervasive that people work through kind of their internal legal frameworks and so on and so forth. And anecdotally, I'm not sure if you saw the CEO of Palantir, he had a CNBC interview where there is, I think what's going to be a very interesting trend in enterprises wanting to own kind of their own whole stack as well. I think Satya Nadella has basically called it your token capital, which is a new asset, effectively that's on your Balance sheet your internal intelligence at your company and how you can actually turn that into outputs with these new models, make it more efficient in helping with automating tasks or streamlining different projects. I think that is something that's going to be really cool to watch, is these enterprises coming in more and more into the participation in the AI game rather than kind of offloading it to, you know, kind of the labs to deal with, you know, people owning their own destiny and trying to really vertically integrate across that.
[24:30]
A
Yeah. One of the things that I've found so useful, a very useful use of AI tools is to have your own massive data lake and proprietary data and then ask the AI to evaluate it when you need it evaluated.
[24:46]
B
Right.
[24:46]
A
So like for example, I put out this report about whether or not Bitcoin had bottomed and whether the, the pricing indicators that had reliably, you know, on chain data into mvrv, all that stuff. If the ones that had reliably bottomed at prior price bottoms in other bear markets were those present today, I, I can do that, I can do that on glass note, I can look at them all and it can be like, okay, that one hit here and that one hit here, it would take me so long. But we have all those metrics, we've built all those metrics ourselves. So I can just be like, just go look, look across our database. And it's so good at that and so fast. And so to sort of buttress your point here, you know, companies have a lot of their own data and asking the AI to help them sift through it, synthesize it, present it is an incredibly powerful use case today. And I can imagine, rather than just like blindly asking Claude, hey, can you go find out if there are bottoms? Like, I don't know what data cloud's relying on. I want to rely on my data, but I may also not want to give all my data to Claude. So I can really see that becoming a. Companies are loath to give up their proprietary information. Right. So instead run the date, run it local basically.
[25:56]
B
Exactly. I think it's going to be existential for, you know, basically across industries to build your AI moat within that industry. And part of that is building your token capital base and being able to kind of create your stack that helps you do what you do best and kind of proliferate AI across kind of the enterprise.
[26:17]
A
So really crazy.
[26:19]
B
It's going to be interesting. It's going to be a wild ride.
[26:20]
A
We've come so far just in 2026, like I wanted to build this thing I'm sort of vaguely describing with this Bitcoin data. And I was going to, we're going to maybe pay to build it. And then we came back from like Christmas vacation basically in January, and I was asking my developer here, I was like, I was like, wait a sec, can we like, vibe code this? We're still calling it vibe coding now. It's legitimately just coding. He was like, way ahead of you, boss. Like, already did a lot of it. And that was six months ago. So, I mean, I just can't imagine where this is going. And I didn't even thought your point about the underlying stack, that the chip manufacturers getting better, the machine manufacturers getting better, that's allowing the. Even obviously from the top, the software model, the model design is also improving.
[27:07]
B
That's exactly.
[27:07]
A
But when they can train it so much better because they have so much more compute. I mean, I'm not even sure what, you know that we're going to be talking to a hologram the next year when you're here.
[27:18]
B
Maybe that's next time I'm on this podcast.
[27:19]
A
You can appear on 10 podcasts at once. Well, Sim Densa, thank you so much for coming on Galaxy Brains.
[27:26]
B
Thank you for having me.
[27:32]
A
Thank you for listening to Galaxy Brains, the weekly podcast from Galaxy Research. I'm Alex Thorne, head of firmwide research at Galaxy. Follow me on X at Intangible Coins, follow Galaxy Research on X at GLXY Research. Read our written reports at galaxy. Com Research. And don't forget, if you like Galaxy Brains to like and subscribe on your favorite podcast platforms like YouTube, Spotify, Apple Podcasts and more. We'll see you next time.