WEBVTT
1
00:00:00.080 --> 00:00:07.860
What's up, y'all? Welcome back to Blockspace Live, presented by CleanSpark for another beautiful week of compute, powered land, and energy.
2
00:00:07.900 --> 00:00:16.950
Terawatt just received approval for 482 megawatts of energy load at its Justified data center in Kentucky.
3
00:00:17.480 --> 00:00:36.300
And Charlie, the deal shines a very useful light onto exactly what goes on under the hood when one of these data centers comes in, starts soaking up power from the grid, specifically their taxes, who's really paying for the transmission, in this case, TerraWolf, and also the design of the data center.
4
00:00:36.400 --> 00:00:43.620
A lot of good ammo here for refuting some of the most basic FUD against data centers. So we'll be covering that for our lead story.
5
00:00:44.000 --> 00:00:56.920
Following that, we've got a research note from Morgan Stanley where they expect a flurry of powered shell deals to hit throughout the rest of the year, and they've got some incredibly bullish price targets on some of our favorite stocks.
6
00:00:57.000 --> 00:01:00.320
Now, pretty sure Morgan Stanley's on the sell side for all of these.
7
00:01:00.740 --> 00:01:17.080
Let's take it with a grain of salt, but still some pretty good analysis here, especially with regards to evaluating these companies and also taking a look at what lease terms they might expect as we see these powered shell leases continue to move it up and to the right in terms of dollars per megawatt.
8
00:01:18.000 --> 00:01:32.410
After that, we have two interviews. First, with Luxur COO Ethan Vera on their new BentHouse announcement, plus flexible AI load. You didn't think it was possible, but Luxur and Ethan are gonna tell us how it is.
9
00:01:32.720 --> 00:01:45.420
Then we will wrap up the show with Craig Taveras, the lead of Hive's Buzz HPC segment, to talk about their Q2 results, specifically as it relates to timelines on their latest AI builds and their strategy going forward.
10
00:01:46.620 --> 00:01:59.440
Blockspace goes live every single weekday at 1 p.m. Eastern. We are Compute's daily live show, AI, data centers, powered land, markets, and emerging technology.
11
00:01:59.560 --> 00:02:10.039
If you like the live stream but miss it, it's on podcast form. Anywhere podcasts are found, you can search Blockspace in the search bar of whatever app you listen to podcasts on.
12
00:02:10.759 --> 00:02:22.880
Leave us a review and hit our website, blockspace.media, for expanded written articles, a newsletter that hits your inbox every day, and everything Blockspace content-wise.
13
00:02:23.320 --> 00:02:33.160
This show is brought to you by CleanSpark, Nasdaq-listed ticker CLSK. More on CleanSpark later on in the show. Colin did a deep dive on TerraWolf.
14
00:02:33.320 --> 00:02:46.740
Big announcement this morning, and it's bullish TerraWolf and great ammo for beating that anti-data center FUD. Colin, what are we looking at? Yeah. So this hit the wire this morning.
15
00:02:47.480 --> 00:02:57.620
TerraWolf getting approved for 482 megawatts for a power agreement to power its Justified data center in Kentucky. Reading here from Blockspace,
16
00:02:58.710 --> 00:03:06.880
"The Kentucky Public Services Commission approved electric services for up to 482 megawatts for TerraWolf's Justified data campus in Hancock County.
17
00:03:07.320 --> 00:03:19.040
The August 21st order authorizes Big Rivers and K Energy to implement the agreement." Big Rivers is the utility, and K Energy is one of the owners of that utility.
18
00:03:19.090 --> 00:03:31.180
[lip smack] Um, I believe the exact term is an owner member of Big Rivers. [lip smack] "The services agreement runs for an initial 15-year term and then extends in rolling five-year periods.
19
00:03:31.190 --> 00:03:40.030
And TerraWolf, importantly, will bear the market transmission, delivery, and infrastructure costs specific to, um, the data center for Anthropic."
20
00:03:40.560 --> 00:03:53.020
So just to back up, this is a data center for Anthropic on a 20-year deal. [lip smack] I believe we talked about this earlier in the summer. I'm not-- I don't remember quite exactly when the deal hit.
21
00:03:53.600 --> 00:04:02.780
But that last note about transmission and all of that, Charlie, I think is useful to point out with regards to we're st- we're seeing this kind of across the board at this point.
22
00:04:03.300 --> 00:04:17.320
ERCOT in some sense is leading the way here with the PUC, um, and ERCOT changing rules on data center operators and developers paying for transmission to their sites rather than the utility doing that, as has been kind of standard in the past.
23
00:04:17.720 --> 00:04:28.930
This was also a big pillar for Josh Shapiro's executive order in Pennsylvania. He wants the data centers to pay for grid upgrades, transmission, et cetera, to their sites.
24
00:04:29.360 --> 00:04:34.000
And that is included in this contract that TerraWolf has with Big Rivers and K Energy.
25
00:04:34.720 --> 00:04:42.600
Quoting directly from the contract as well, quote, "TerraWolf must either prepay or provide and maintain credit support for Big Rivers' capacity payment obligations to MISO.
26
00:04:43.180 --> 00:04:47.280
TerraWolf must provide credit support equal to two times the estimated highest monthly bill."
27
00:04:47.320 --> 00:04:58.380
It's basically a contingency to make sure that they can actually pay, um, their, uh, their contractually obligated rate for this energy, making sure that if something goes wrong, they have credit support to back it up.
28
00:04:58.599 --> 00:05:03.080
[lip smack] That's the other thing that is important, uh, to understand about this.
29
00:05:03.700 --> 00:05:15.240
"Big Rivers' utility will procure the campus's energy and capacity through the MISO power pool rather than supply it from the utility's generating fleet." That's according to TerraWolf's energy plan.
30
00:05:15.300 --> 00:05:30.980
Those MISO costs will pass through to TerraWolf dollar for dollar for the full contracted term, ensuring, quote, "The energization or the generation to supply TerraWolf in the near term, medium term, and long term will be satisfied without imposing additional cost on other ratepayers."
31
00:05:31.040 --> 00:05:43.440
End quote. So Big Rivers will be sourcing this energy from within MISO. It won't come from their, their plant specifically, and that's to protect ratepayers in the area. [lip smack] In fact,
32
00:05:44.460 --> 00:05:52.040
one of the things that I thought was interesting about this, Charlie, is that this could potentially lower bills for ratepayers in the area.
33
00:05:53.060 --> 00:06:00.228
Quoting here from the contract or the, um, rather the summary of it by the utility...
34
00:06:00.238 --> 00:06:11.068
commission, the Public Services Commission, quote, "The K Energy adder will defer the need for distribution rate increases or otherwise be used for the benefit of ratepayers," end quote.
35
00:06:11.508 --> 00:06:26.558
That adder specifically is they ha- TerraWolf has a $15 million wholesale adder, which is basically a tariff that they pay annually to Big Rivers, and a $10 million transmission tariff that they pay to Big River, as well as a $1 million distribution adder to K Energy.
36
00:06:26.588 --> 00:06:28.508
Both of these escalate after three years.
37
00:06:30.068 --> 00:07:00.568
Uh, quote, going further here from the summary, "The Big Rivers adder and new transmission revenue will dir- directly benefit all ratepayers through Big Rivers' member rate stability mechanism, under which 60% of Big Rivers' margins in excess of the margin required to achieve its tier for, um, generation are used to pay down regulatory assets, amounts owed to the utility by ratepayers, and the remaining 40% is credited to all ratepayers through bill credits and other, and example, through rate reductions.
38
00:07:01.057 --> 00:07:05.468
TerraWolf is not eligible for any of these MS- MRSM credits."
39
00:07:06.368 --> 00:07:17.048
A big thing to remember about, like, why that would be possible, that this will potentially lower rates for folks in the area, this used to be an aluminum smelting plant.
40
00:07:17.548 --> 00:07:24.808
So Justify is more or less a brownfield on a former aluminum smelting plant that was decommissioned in 2022.
41
00:07:25.028 --> 00:07:38.268
600 jobs were lost as a result, and the county has missed out on $40 million in estimated tax revenue since 2022 when it was decommissioned. And now they also have not received...
42
00:07:38.288 --> 00:07:46.108
Or sorry, that's transmission, uh, transmission tariff revenue, so even more in tax revenue. So that's $40 million that could be going to the utility here.
43
00:07:46.728 --> 00:07:50.468
Now TerraWolf is stepping in and is going to start paying those tariffs.
44
00:07:50.908 --> 00:08:06.508
That means that grid infrastructure upgrades, or rather maintenance, can be satisfied by those tariffs, and also it means that there might be a chance for the utility to lower ratepayer costs with credits and other things like that.
45
00:08:06.588 --> 00:08:12.068
And so this is, I o- I've brought this up before, Charlie, and I had a conversation with my dad about it last week.
46
00:08:12.388 --> 00:08:22.348
It seems counterintuitive at first when you say a data center's gonna move in here, suck up a lot of energy, but it's actually going to lead to rate decreases for other ratepayers.
47
00:08:22.838 --> 00:08:30.548
The reason for that is, are, are these tariffs and the fact that now the utility can potentially max out what it is able to provide.
48
00:08:30.928 --> 00:08:44.508
Still has the same overhead cost, but if you're able to provide more power than le- uh, than less power for the same overhead cost, then you're gonna be able to pass that, those discounts down to your ratepayers, theoretically.
49
00:08:44.538 --> 00:08:53.608
And we've actually seen this in practice as you shared Nick Carter's Murmurations post on this, uh- Mm-hmm... last week. Few other notes here.
50
00:08:54.568 --> 00:09:06.968
In terms of tax revenue, local school districts will receive 3% utility gross receipts on electricity purchases for the site, which is equal to roughly $7 million annually according to the estimates.
51
00:09:07.008 --> 00:09:12.398
Kentucky, the state itself, will reap 6% from energy sales equal to $14 million annually.
52
00:09:13.548 --> 00:09:26.428
Um, so you have 15 million in estimated property tax on top of the utility, or on top of the school district tax, for a total of 22 million in annual tax revenue to Hancock County.
53
00:09:26.588 --> 00:09:44.008
The in- the total estimated tax receipts from the state and the local level combined for this data center will be $1 billion over 20 years for the Anthropic deal. Whoa. Yeah. Holy smokes. This county's gonna eat.
54
00:09:45.508 --> 00:09:57.188
It's pretty incredible, and I think that that- that's why when I opened this- What's the, what's the population of Hancock County? What, what's the population of Hancock County? Let me figure this out.
55
00:09:58.748 --> 00:10:09.008
[laughs] Oh my God, Colin, okay. Over/under 50,000- How many? Over or under 50,000 people. 50,000? Yeah. That goes a long way, man.
56
00:10:09.388 --> 00:10:20.668
And I, that's why when I opened this I said- No, I'm not, I, I, I'm saying as it's 9,000. I was asking over and under. Oh, over and under. Oh, yeah. 9,000 people, a billion dollars over 20 years? Holy smokes. Wow.
57
00:10:21.328 --> 00:10:41.928
Yeah, and, uh, you know, that's why when I opened this I wanted to kind of make the, uh, the data center propagandist point that, you know, that when you look at a contract like this, you actually have so many refutations to the litany of gripes that you typically see from the anti-data center folks and the people protesting these things.
58
00:10:43.088 --> 00:10:56.638
And, uh, a few other things here too. So the site will generate, uh, 800 to 1,000 construction jobs estimated, plus 95 full-time employees from TerraWolf on the site, plus Anthropic staff.
59
00:10:56.728 --> 00:11:03.928
Now, of course, like I don't know how many Anthropic will have on there. Let's, let's just like say maybe like ballpark 120 for the whole thing.
60
00:11:05.488 --> 00:11:15.388
That's nowhere near 600 for the smelter, but as I, I responded to a comment on our YouTube page today from what I believe was a bot, and I just said, "Well, what's the alternative?"
61
00:11:15.908 --> 00:11:25.008
Because manufacturing isn't competing for these sites, right? Like no- nothing else is moving in here, so it's either you take that or you just don't get anything.
62
00:11:25.448 --> 00:11:31.968
You don't get the jobs, as meager as they may seem considering how big the data center is. You also don't get the revenue from the taxes.
63
00:11:32.648 --> 00:11:42.468
The other thing that I thought was super interesting, T- TerraWolf did a lot of community engagement leading up to the approval for this site. They held m- uh,
64
00:11:43.448 --> 00:11:50.098
numerous local meetings, had PowerPoint presentations and kind of like town hall meetings to answer questions from the community,
65
00:11:51.808 --> 00:12:04.638
and th- the, the summary specifically mentions that this will be a closed loop cooling system. Do you wanna kn- uh, y- go ahead and guess how many gallons per day the smelter was consuming before it shut down.
66
00:12:04.728 --> 00:12:18.766
Oh, oh, am I... I, I would say, oh, I don't know. 20 million Oh, not, no, no, per day. Per day. Per day. I don't know. Give-- Tell me. Just answer. 550,000 gallons per day. Oh my gosh.
67
00:12:18.766 --> 00:12:27.836
And TeraWolf, the summary points out, will not draw from the Ohio River, nor will it draw from the municipal water system.
68
00:12:27.876 --> 00:12:40.176
So, you know, y- y- when you think about that, n- I doubt there was any backlash to the, to the smelter when it came in. You know, it creates jobs. But th- this, this data center will use
69
00:12:41.216 --> 00:12:44.696
n- nothing close to the same amount of water that that was using.
70
00:12:45.396 --> 00:12:55.816
And I, I just do, uh, want to point out that this is the energy plan that TeraWolf submitted to the Kentucky Energy and Environment Cabinet, and this is what is, was being summarized in this.
71
00:12:56.596 --> 00:13:02.076
Um, but it was reviewed by these regulators and by these officials, all that stuff. So, uh, the, the last,
72
00:13:03.136 --> 00:13:14.095
the last thing I wanted to point out, if you look at the tax revenue, you can back out an estimated power cost for this agreement. Now, this is estimated. TeraWolf has not, has not disclosed this.
73
00:13:14.216 --> 00:13:27.296
I just wanna get that out front. Also, it, it's not a fixed cost PPA, it'll be floating. So-- But if you look at the tax revenue, back that out, you're looking at roughly pre-tax 55 to $61 per megawatt hour.
74
00:13:27.846 --> 00:13:40.616
And the 61, uh, the higher end there, the $61 per megawatt hour is based on TeraWolf's 90% load factor used for the tax estimate. So, you know, pr- pretty decent rate considering...
75
00:13:40.936 --> 00:13:54.246
You know, that would've been an okay rate for Bitcoin mining, but, but for what they're using this for now, for hosting AI equipment, that's, I would imagine, pretty favorable with regards to the margins, so. Yeah.
76
00:13:54.255 --> 00:14:00.796
[clears throat] I mean, a lot of people think that pulling from the power grid is like pulling from a well,
77
00:14:01.976 --> 00:14:09.616
but it's more like driving a car, where if you're ramping the car up and down and gassing the pedal, you're really burning fuel inefficiently.
78
00:14:10.396 --> 00:14:28.016
But if you're able to run the car at a, like, 55 miles an hour at a s- high speed and make the engine bigger and more efficient and balance it, uh, then, then aggregate, the whole system performs better, your power prices go lower, net more benefit across the board.
79
00:14:28.616 --> 00:14:43.336
When you take away core pieces of the load-consuming part of the power grid, uh, the entire system gets less efficient, less optimal. When you put them back, we're back to working baby.
80
00:14:43.396 --> 00:14:56.666
So that's why a billion dollars to the county of Justified- Hancock... Hancock County, uh, well, more, more power to 'em. Uh- I love that analogy about a, about the car. Yeah. That's really good.
81
00:14:56.776 --> 00:15:07.576
I, I just, I came up with that just now. We gotta keep going. We gotta roll. We gotta... Let's talk about Morgan Stanley and Powered Shells. Our data center propaganda rolls on after a word from our sponsor, CleanSpark.
82
00:15:07.685 --> 00:15:22.616
[gentle music] We are CleanSpark, America's Bitcoin miner, a publicly traded company with the largest operating hash rate, powered entirely by self-operated infrastructure across four states.
83
00:15:24.456 --> 00:15:35.876
This is our proof of work, and we are setting the standard for what's next. Learn more about the intersection of energy and Bitcoin at cleanspark.com.
84
00:15:40.756 --> 00:15:45.736
All right, Charlie, let's see what the bankers are saying. Yeah. The bankers wanna buy your stocks.
85
00:15:46.276 --> 00:15:54.926
[chuckles] The bankers want you to buy their stocks on the sell side, and shockingly enough, they are uber bullish on Powered Shells.
86
00:15:55.536 --> 00:16:00.536
Headline here from Blockspace, Morgan Stanley expects Powered Shell deal wave by October.
87
00:16:00.736 --> 00:16:11.706
[lips smack] Morgan Stanley expects Powered Shell operators to sign a wave of data center contracts by October as AI customers race to secure projects that can energize before the end of 2027.
88
00:16:12.556 --> 00:16:26.216
The bank maintained an attractive industry view in a Monday note led by analyst Steven Bird. Morgan Stanley kept overweight ratings on Cipher, TeraWolf, Riot, Hut 8, and Galaxy.
89
00:16:26.236 --> 00:16:39.356
Based on August 21st closing prices, the bank's $62.50 TeraWolf target applies about 300-- implied about 300% upside. It's pretty rich. It depends on what your timeframe is though, I guess. Yeah.
90
00:16:39.396 --> 00:16:52.876
While its $43.50 Cipher target implies 176% upside, its t- $273 Hut 8 target implies 238% upside, and its $43 target implies about 159% upside.
91
00:16:53.556 --> 00:17:03.876
Morgan Stanley rates Applied Digital as equal weight and Mara underweight, although it sees contract-driven upsides for both. So obviously very bullish.
92
00:17:04.476 --> 00:17:20.356
They get into some of the evaluation modeling, or evaluation modeling that they use for this. Also, some of the modeling that they use for, um, [lips smack] for, uh, the deals themselves.
93
00:17:20.856 --> 00:17:33.876
Morgan Stanley's indicative model assumes a 20-year lease generating $2.38, uh, cents per watt annually, with an 85% margin and construction cost of $11.50 per watt.
94
00:17:34.596 --> 00:17:45.156
The model uses 80% loan to value for the site, 7.25% project debt, a 9% equity discount rate, and a 6.7% terminal capital rate.
95
00:17:45.656 --> 00:17:58.856
The TeraWolf data center, I didn't mention this during that segment, but it is estimated to cost 4 to 4.5 billion at a cost of about 10 to $12, uh, per watt, 10 to $12 million per megawatt.
96
00:17:59.186 --> 00:18:06.556
And that's specifically for the Powered Shell itself. So that doesn't include the, the, the GPUs, right? They're just building the data center.
97
00:18:06.616 --> 00:18:13.096
[lips smack] But on the note of other models for the valuation here, Charlie, I thought this was interesting.
98
00:18:13.736 --> 00:18:28.236
They were looking at lease rates, and Morgan Stanley says that lease rates have risen from about o- $1.80 per critical IT watt in September 2025 to $2.30 to, to $2.40 per watt currently.
99
00:18:28.926 --> 00:18:40.546
Morgan Stanley estimates that a 100% basis point increase in average yield on cost could create another $3.50 per megawatt, more than offsetting the valuation effect of 100% basis point increase in interest rates.
100
00:18:41.086 --> 00:18:53.506
That being top of mind, uh, you know, we'll have Jim Bianco on tomorrow to talk about this, great macro analyst, but 30-year yield is really running away from the, from officials right now.
101
00:18:54.546 --> 00:19:13.366
Treasury has decided to up its total ceiling for, uh, interventions for buybacks from two billion to four billion per operation, and so there's, there, there's a, a decent amount of fear right now within the bond market that the long tail of these durations are going to continue to rise.
102
00:19:13.866 --> 00:19:24.146
And if that is the case, then debt for these data centers, as it already has been, is going to get more expensive, um, over the next few years if that continues to float upward. Yeah.
103
00:19:24.346 --> 00:19:33.606
Um, kinda looking at this Morgan Stanley note, it's pretty much if you wanted a one-liner, it's which powered shells have existing megawatts that are not yet leased.
104
00:19:34.146 --> 00:19:45.745
Those are the people they are bullish on, and that list is, as you said, what? Cipher, TeraWolf, Riot, Hut 8, Galaxy as, like, the, the leading ones who have not announced a lease.
105
00:19:45.866 --> 00:19:58.926
Uh, there's another dimension here, and I thought I'd tap into DI Metrics, our, uh, data center intelligence tool, dimetrics.ai, link in description. Um, and
106
00:20:00.166 --> 00:20:11.726
it's kind of interesting to look at the political risk 'cause you have some of these companies have, within their portfolio, multiple sites. Some of the sites are, like, good to go.
107
00:20:11.786 --> 00:20:26.286
They don't have permitting or, like, procedural, uh, issues they're waiting on, and a couple of them do. Specifically with, like, TeraWolf, who's, uh, and there's actually a political risk, uh,
108
00:20:27.206 --> 00:20:39.766
uh, number or rating assigned to each company according to DI Metrics here. I'll actually share it. And, um, uh, it's really interesting.
109
00:20:39.886 --> 00:20:51.526
So, I mean, let's look at, uh, TeraWolf, who has the highest political risk of all of the, uh, companies that Morgan Stanley talks about here, with a political risk on DI Metrics side of 41.
110
00:20:51.586 --> 00:20:57.146
And that's because out of their five sites, Lake Mariner, Muskie, Hawesville, those three are good to go.
111
00:20:57.726 --> 00:21:09.066
Morgantown has an elevated permit risk, as Lake Hawkeye is, uh, has a permit denial risk, um, according to DI Metrics.
112
00:21:09.846 --> 00:21:23.606
Um, if we look at the companies in, uh, in Texas with Batch Zero, it looks a little bit different. So that would be Cipher, Galaxy, and Hut and Mara.
113
00:21:24.366 --> 00:21:41.166
Um, because Texas, uh, at the state level forbids counties banning data centers, uh, outright, um, they don't, they, they forbid data center moratoriums so the counties have to then handle these at the, like, you know, more, uh, lower level permit issue.
114
00:21:41.766 --> 00:21:56.466
You have companies like Hut 8, Cipher, and Galaxy have political risk scores of very low. In Cipher's case, um, their, seven of their eight sites, Colches, Apollo, Barber Lake, Mikeska, McClellan,
115
00:21:57.466 --> 00:22:14.566
Milsing, and Black Pearl, those have very low political risk, whereas Ulysses does have a higher permit risk as it's not in Texas. It's in Ohio. Um, so some pretty interesting, uh, insights here. Uh,
116
00:22:15.506 --> 00:22:23.386
I have some more data. And I, I think you already said this, but it's- Okay... worth noting for TeraWolf specifically that, you know, for the,
117
00:22:24.466 --> 00:22:33.526
for the Lake Mariner campus, they already have approvals for what they're building out right now. Exactly. The moratorium doesn't affect them. Exactly, yeah. Moratorium doesn't affect them in New York.
118
00:22:33.846 --> 00:22:48.326
Uh, they're ob- you know, grandfathered in, you could say. Uh, but, um, you know, I would say that m- as Morgan Stanley just looks at unleased megawatt, um, among a portfolio, uh, they should,
119
00:22:49.366 --> 00:23:02.386
I, if, you know, if I were Morgan Stanley, I'd be very, very reluctant to consider megawatts which are still subject to approval in hostile counties as absolutely not done deals. Yeah.
120
00:23:02.426 --> 00:23:09.186
Because- And that's a fair point to make, and one of the things they say towards the end of the research note is they're actually kind of in support of what you're saying. Yeah.
121
00:23:09.196 --> 00:23:20.816
They expect Cipher to contract 700 to 1,000 megawatts gross by October and Riot to lease the full 1,000 gross megawatts for Corsicana by the end of September. Curious to see what they know. Yeah.
122
00:23:20.846 --> 00:23:28.726
I mean, usually these analysts have an inside look at, like, how some of these things are progressing. So, um, again- Yeah... would be, be curious to see.
123
00:23:28.766 --> 00:23:41.066
But again, a caveat, they're on the sell side for all these companies. So, that, so, you know, those, those, um, those upside targets, take 'em with a grain of salt. They're trying to get me to buy more Cipher stock.
124
00:23:41.116 --> 00:23:53.846
Trying to build- Joke's on you. I already own Cipher stock. [laughs] So, uh, uh, we, we're, we'll keep covering the data center moratorium story, uh, a lot, obviously, over this fall. But before
125
00:23:55.306 --> 00:24:08.166
we keep co- uh, we, but, but rather, we have in the wings, we have Ethan Vera, uh, backstage. We're gonna talk about demand response and the inaugural GPU Demand Response event in Texas this past month.
126
00:24:08.686 --> 00:24:15.716
Before we bring up Ethan, we gotta do a s- a word from our sponsor. Colin, our sponsor, Luxur.
127
00:24:16.266 --> 00:24:28.066
[static sound] This episode of Block Space Live is brought to you by Luxur's Commander, Bitcoin miner management software for enterprise operations.
128
00:24:28.606 --> 00:24:34.346
Luxur's Commander gives you real-time fleet monitoring, bulk remote commands across your fleet, and Intelligent Miner.
129
00:24:34.914 --> 00:24:41.804
That's an automated profitability engine that runs every five minutes and adjusts your fleet's power settings to live energy and hash rate markets.
130
00:24:42.334 --> 00:24:48.384
In fact, ERCOT back tests show 10% improved profitability with intelligent mining versus old-fashioned binary mining.
131
00:24:49.154 --> 00:24:57.754
Commander Pro is roughly half the cost of competition, $100 per megawatt, or a 25 basis point pool fee adder. But you can also try it free for 60 days.
132
00:24:57.874 --> 00:25:11.094
So if you'd like to learn more, go to luxor.tech/commander to get started. All right. We got Ethan. Let's bring him on up here. Ethan Vera, Luxor, welcome to the show. Gentlemen, great to see you both.
133
00:25:11.934 --> 00:25:25.654
Good to see you too, man. Great. Thanks for joining. Absolutely. So you guys had a big announcement, uh, an hour ago. Luxor Energy and BentHouse... Am I saying this correctly? BentHouse? Uh- Spot on. Okay.
134
00:25:25.874 --> 00:25:38.694
Uh, complete first AI GPU response to ERCOT 4CP signals. Give me the, the high-level breakdown of what happened. Yeah. I mean, there, there, there's a lot of themes here that we can, we can touch on.
135
00:25:39.014 --> 00:25:48.354
Um, and so what, what it comes down to is really we're in the business from an energy perspective of sending signals to data centers to participate in energy programs.
136
00:25:49.034 --> 00:25:57.354
We, we've been playing in that space in the Bitcoin mining side, which we've talked a lot about on, on this podcast, and we think that's vital for the data center space.
137
00:25:57.874 --> 00:25:59.654
Um, it probably goes to a lot of themes you talked about.
138
00:25:59.694 --> 00:26:07.814
I don't know if you talked about it earlier today on this pod, but certainly there's been a lot of pushback against data centers, uh, both domestically as well as globally.
139
00:26:08.374 --> 00:26:17.244
And we think that a, a big part of that is data centers' inability as of today to participate in the grid servicing programs that are really vital for, for the grids.
140
00:26:17.814 --> 00:26:25.834
And so when we think of, you know, different grids around the United States, they're trying to solve for supply and demand to make sure that things like brownouts and blackouts don't happen.
141
00:26:25.934 --> 00:26:32.754
Um, ancillary service programs and demand response is crucial for that, and that's why really Bitcoin miners have been helpful for the grid.
142
00:26:32.794 --> 00:26:44.814
We think that it's, it's inevitable that they really push data centers in that same direction, and that data centers consuming their dozens and soon hundred gigawatts of power within the United States are participating in these programs.
143
00:26:45.394 --> 00:26:57.074
And so really what we did was we ran an H100, uh, and a B200 through, uh, 4CP. So 4CP is the four coincidental peaks in, in, in ERCOT, and we ran it through these dispatches.
144
00:26:57.634 --> 00:27:10.063
The, the B200 was live curtailing based on, on these dispat- dispatch signals and basically throttling down to 25% of its power consumption. During that, it did not lose its, like, inference workloads.
145
00:27:10.114 --> 00:27:21.834
It was doing checkpoints. Um, and, and so we're, we're basically furthering the test of, uh, AI GPU's ability to go through this. Next up for us is to go through some of the ERS programs in the fall.
146
00:27:21.854 --> 00:27:37.184
And so we procured a node of H100s we're shipping to a data center just inside, uh, kind of Dallas-Fort Worth area, and we're gonna be enrolling it in, in these various, uh, ERS programs to, uh, continue to do some research development here with the eventual goal of helping data centers do demand response.
147
00:27:37.994 --> 00:27:49.134
So Ethan, you said that the data centers have been slow to implement these kind of things, or rather that it just, we haven't seen it at scale yet. What are some of the reasons for that?
148
00:27:49.174 --> 00:27:55.364
Is that simply because a lot of these compute contracts have, like, mandatory uptime clauses?
149
00:27:55.414 --> 00:28:06.674
You know, if you think about building a tier three-plus data center, you're committing to a much higher threshold of uptime versus a Bitcoin mine, and you have to be delivering that compute or else you'll be in violation of the contract.
150
00:28:06.734 --> 00:28:16.734
Is it that, is it that part of it? Is it that the systems just haven't caught up, or the tooling hasn't caught up to allow this to happen? What are some of the obstacles? I see it in two ways.
151
00:28:17.274 --> 00:28:24.114
Fir- first is the economics. So when we talk about Bitcoin mining, that's become a commodity business over the years.
152
00:28:24.364 --> 00:28:33.494
And, and over the years what we've seen is that it's very crucial that Bitcoin miners are participating in grid servicing programs because it ends up impacting their profits in a very positive way.
153
00:28:33.994 --> 00:28:45.694
If you look at an example like Riot Blockchain, for example, or now Riot Platforms, in some cases they were making more money off of their energy business and, and, and participating in these programs than they were Bitcoin mining itself.
154
00:28:46.214 --> 00:28:54.154
And that really speaks to the, the type of, like, commoditization that happened in the mining space. That has not happened in the AI data center space as of today.
155
00:28:54.633 --> 00:29:04.534
Margins are still 95%, and really it does not make sense to be curtailing your machines for any grid servicing programs out there or any financial incentive that is currently offered by the grid.
156
00:29:05.574 --> 00:29:10.934
Um, so that's like the first obstacle, which I think will, will start to change over time. The second is the technology stack.
157
00:29:11.514 --> 00:29:24.494
Um, there's historically been, uh, very little interest in, in finding ways to, uh, curtail like a training workload, for example, or even inferencing. Um, that being said, there's dozens of teams working on this.
158
00:29:24.554 --> 00:29:27.734
We partnered with BentHouse, uh, Zeoni Software on this one.
159
00:29:27.874 --> 00:29:36.974
Um, but there are dozens of really smart technical teams that are working on, on the ability to curtail GPUs, but not give up and be detrimental to its workloads.
160
00:29:37.234 --> 00:29:47.054
So those are like the two obstacles I see, both of which are being addressed right now and I think will, will, will look a lot different in a, in a few years from now. I am kind of curious.
161
00:29:47.174 --> 00:29:54.734
Um, I, uh, I could ask a question about maybe more like the hardware, like how, how does the curtailment work?
162
00:29:54.974 --> 00:30:07.674
But I'm actually more interested in like the decision-making process because in a Bitcoin mine, uh, it's, it's a little more straightforward, like who makes the call to turn off, uh, the, the mining rig?
163
00:30:08.094 --> 00:30:25.214
But in these, uh, much more complex dynamic contractual arrangements of tenants of these, uh, data centers, uh, I'm curious like, uh, while this business does not maybe exist or the, or ra- these relationships don't exist today at scale,
164
00:30:26.214 --> 00:30:37.454
how do you envision like the decision-making of how and when to turn off, uh, going down the stack? Is this a c- a tenant decision thing? Is this a operator decision thing?
165
00:30:37.514 --> 00:30:45.862
Like, how does this, how should, how do you think this, this will work in the future? Yeah. And, and this is a bit future dated, so all of this is kind of like assumption of future state.
166
00:30:46.542 --> 00:30:51.642
Um, the, the majority of partners we work with are-- they, they separate on the value chain.
167
00:30:51.902 --> 00:30:58.382
So there's one company that runs a data center, there's another company that owns the GPUs, there's another company that rents the GPUs.
168
00:30:58.442 --> 00:31:05.772
I mean, the clearest example is like a core scientific, core weave, like anthropic mix, right? Where it's not a single company that, that is doing everything.
169
00:31:06.362 --> 00:31:12.762
And so it, it's a really good question, which is like, who's actually dictating the participation in these programs and making the decisions?
170
00:31:13.602 --> 00:31:25.182
Ultimately, the, the end user, whoever is renting the compute on bare metal or owns the comp-- like hardware themselves and is using the hardware, they're, they're gonna have to sign off on participating in these programs.
171
00:31:25.782 --> 00:31:52.062
And so, um, that's gonna be reflected in, in SLAs and, and MSAs that are being signed between them and the owner of hardware and the co-location, and they'll need to get them across the finish line of saying, "Hey, look, we'll, we'll offer you this co-location space, you know, twenty megawatts worth of rack-ready power capacity at, you know, twenty cents a kilowatt hour, but you're gonna have to be able to curtail one percent of the year because we're participating in these programs."
172
00:31:52.562 --> 00:31:59.152
And maybe, you know, tho-those energy benefits will pass through in some capacity to that end user, so they'll benefit from it in some ways.
173
00:31:59.962 --> 00:32:06.782
Um, as of today, we haven't seen a lot of those deals get struck for that s-similar reason we're talking about, which is the economics just aren't there.
174
00:32:06.802 --> 00:32:14.262
It's better right now to run your GPU workload a hundred percent uptime than ninety-nine percent and get a bit of energy like ERS revenue.
175
00:32:15.002 --> 00:32:24.842
But in the future, that is either probably gonna change as the industry gets commoditized and margins come down and/or regulation and it, uh, basically forces their head.
176
00:32:25.382 --> 00:32:37.922
And that, that's kind of one of our thesis is that there is, um, in, in the interconnect, uh, demand, both ERCOT and other markets, there's dozens and soon hundreds of gigawatts of demand of new AI data centers.
177
00:32:38.002 --> 00:32:46.042
Clearly, uh, the grids are looking at ways to support that through cogeneration, through storage, um, through any means possible.
178
00:32:46.082 --> 00:33:07.642
We think demand response participation is gonna be a big part of that, and so it's gonna be almost a forcing function where it's like, okay, if you want a hundred megawatts on grid opportunity in say NYISO or MISO or PJM, you need to be able to participate in some of these programs that, uh, helps the grid stabilize and you become an asset to the grid rather than necessarily, you know, a, a harm.
179
00:33:09.342 --> 00:33:27.082
Uh, you mentioned the Emergency Response Service in ERCOT, ERS, and the, uh, uh, you know, that's more of a shutoff to keep the, the power grid on while a winter storm rolls in. As AI data centers proliferate in ERCOT,
180
00:33:28.002 --> 00:33:38.002
what does the ERS landscape look like? And is this like a business opportunity? Is this, uh, goodwill? Is this to jump through the permitting hoops? Like,
181
00:33:39.162 --> 00:33:46.422
what does this look like as, as Luxor Energy, uh, kind of views the whole system? We, we view it as a mix.
182
00:33:46.482 --> 00:33:54.282
Like, in some ways it is like maybe a way to speed up your, your approval to get, you know, uh, power access in ERCOT.
183
00:33:54.562 --> 00:34:15.822
And so if, if you can come to the grid and say, "Look, we have an application for, let's call it a hundred megawatts of power, and we're willing to participate in all these programs," in our view, the grid is gonna look at that much more favorably in ERCOT will than somebody that wants a hundred gigawa-- or a hundred megawatts of power is not participating in these programs and is not, you know, providing storage or cogeneration.
184
00:34:16.262 --> 00:34:25.262
And so in some ways, like for us, it, it is like the path forward for data centers to actually achieve, uh, and, and acquire the power access that they are looking for.
185
00:34:26.162 --> 00:34:33.622
Um, and so that, that's, that's kind of like the, the main view that we have regarding like regulation and AI, uh, demand response.
186
00:34:35.022 --> 00:34:50.562
As you talk to the engineers, um, who have, who actually have to build these integrations to get the H100 to turn off, um, it's, I assume it's, it's much more complex than shi-shi-shutting off entire buses of ASICs.
187
00:34:50.602 --> 00:35:03.522
I mean, 'cause now you have, uh, whole like systems, modular systems, integrated GPU, CPU, Vera Rubin racks, seventy-two of these things, and the GPU here is not the same as the GPU further down the stack.
188
00:35:03.582 --> 00:35:13.762
Like, what are you seeing from the engineers about the engineering problems that have to be solved to roll this out at scale? Yeah. I mean, you hit the nail on the head.
189
00:35:14.002 --> 00:35:26.362
Th-this is extremely more complex than demand response for ASICs. Um, in ASICs, like for Bitcoin mining, it's very clear the economic signals to curtail, um, all the machines have very similar APIs to call on.
190
00:35:26.982 --> 00:35:34.402
Um, you can quickly prioritize like the, the dropping down the least efficient equipment. I mean, that's what we do through intelligent mining, um, and Commander.
191
00:35:34.702 --> 00:35:40.322
So like that logic is, is a lot easier to think through in Bitcoin mining than it is AI. Um,
192
00:35:41.502 --> 00:35:49.982
I don't have all the answers yet, um, and I think this is where like a lot of the smart teams and companies getting funded are gonna address, including NVIDIA's spending a lot of time on this too.
193
00:35:50.522 --> 00:35:58.182
I mean, the key thing too is like what is the, what is the workload that the GPU is running? Some of those workloads might not be able to be interrupted.
194
00:35:58.802 --> 00:36:09.692
And so like having key insights into the, into the workload, uh, side of the GPU is a unique problem that like Bitcoin mining didn't really face because we all knew what the workload was. It was SHA-256.
195
00:36:10.142 --> 00:36:16.442
And so like there's that distinction here that still needs to be worked out. And so that's why I kinda think a lot of this like software is like long dated.
196
00:36:16.882 --> 00:36:20.262
Over the next like few years, um, we're gonna see a lot of innovation in the space.
197
00:36:20.282 --> 00:36:32.462
There's gonna be dozens of different software providers, including probably, you know, in partnership with NVIDIA, some of the big OEMs like Supermicro and others that are gonna be directly integrating the software into the hardware builds.
198
00:36:32.502 --> 00:36:43.922
And it sounds like to me, Ethan, I think in the, for the pilot that you were talking about at the beginning, the workload was reduced to like twenty-five percent So should we basically expect that to be the standard?
199
00:36:43.962 --> 00:36:53.602
You're not gonna shut off the entire deployment, you're just going to kinda curtail its capacity so that it's working maybe a quarter as much as it would otherwise. Exactly.
200
00:36:53.942 --> 00:37:00.662
Um, t- 25%, uh, like, made sense for the B200 in this certain, like, workload circumstance.
201
00:37:01.482 --> 00:37:08.602
Um, I think directionally that's probably right in terms of what other GPUs running inferencing workloads could be throttled down to.
202
00:37:08.622 --> 00:37:18.722
But my understanding as of today is that you're not gonna wanna throttle these down to 0%. Um, from a hardware longevity standpoint, from a workload standpoint, it just doesn't make sense.
203
00:37:18.782 --> 00:37:25.462
That being said, even throttling it down, like 50, 75%, is, is insanely valuable to the electricity grids.
204
00:37:25.662 --> 00:37:37.101
So if you can bring a 100 megawatt facility down to 50 megawatts when you're called upon and there's an emergency, like a cold winter storm or, you know, a heatwave in, in, in Texas, that's a really valuable asset.
205
00:37:37.142 --> 00:37:51.932
And now the grid looks at you as a value add rather than a detractor. Um, you mentioned it, training versus inference. It's, like, very difficult to know where, what this landscape will look like in five-plus years.
206
00:37:52.502 --> 00:38:07.442
But could you just speak to some of the, you know, insights that you and the team are having about, like, the different types of workloads, even just a simple high level of training versus inference, but, uh...
207
00:38:07.692 --> 00:38:14.342
and, and, like, w- how, how those might fit into these demand response, uh, si- signals? Yeah.
208
00:38:14.402 --> 00:38:21.202
So this is k- kind of where our partner BentHouse came in on it, so we'll, we'll probably have to arrange another, uh, one with Bob on.
209
00:38:21.302 --> 00:38:32.362
Um, Luxer in this partnership, we're the QSE, right, in our cloud, so we're setting the actual dispatch signals to this and, and, um, sending the signals to the GPU when to curtail based on four CP.
210
00:38:32.992 --> 00:38:41.362
Um, their software is the one, uh, working with the GPU to kind of assess the workload, curtail, uh, well, throttle down the power consumption.
211
00:38:42.002 --> 00:38:55.462
Um, based on my understanding talking with them, and they're really, like, the experts on this topic, is that as of today they haven't found a way to, um, throttle GPUs running training without it being detrimental to that training effort.
212
00:38:55.472 --> 00:39:07.802
And so my understanding, like from a very high level perspective, if, if you throttle a fleet or a cluster of GPUs that are i- like, actively training a model, you're gonna, like, destroy a lot of work that's been occurring.
213
00:39:07.922 --> 00:39:17.402
Um, it may not, like, reset it to zero, but it's gonna set it back significantly. Imagine, like, throttling a, say a steel mill, where you turn off the power immediately and there's stuff in production.
214
00:39:17.682 --> 00:39:21.802
Like, it, it's really hard to fire it back up and, like, there's just cost there.
215
00:39:21.842 --> 00:39:29.682
Whereas certain inferencing workloads, they're easier to interrupt and you can, like, kind of now send that inference workload to another data center to start working on.
216
00:39:29.702 --> 00:39:40.122
And so it, it, it's easier to imagine a future in which inferencing is a bigger participant in demand response programs than other workloads, but that's based on my understanding today. And like,
217
00:39:41.022 --> 00:39:51.362
again, the technology's changing so fast here that we could be speaking in two years and that could be totally different. And it's- Yeah, and I'm imagining, uh, getting surge pricing on my Claude queries, so.
218
00:39:51.382 --> 00:40:02.082
[laughs] Yeah. Well, you know, that's... Yeah, I was just gonna say, it sounds like it, it kind of surfaces a new engineering problem for networking these data centers together for, for clusters all around, right?
219
00:40:02.162 --> 00:40:10.782
It's like if one of them goes down or has to c- curtail capacity, then you have to ship that demand somewhere else. And then what does that do, Charlie, like you said, to on-demand token prices?
220
00:40:10.862 --> 00:40:21.301
Yeah, if it's too cold in Texas and a winter storm's rolling through, I'm just gonna not use AI for a few days [laughs] 'cause it'll blow out our budget. Yeah. So, you know, maybe... You know, I could see that future.
221
00:40:23.802 --> 00:40:38.862
Uh- Absolutely. Yeah. So, uh, so you s- you did this initial test. Uh, any indication on, like, what the next step is as far as demoing, scaling up, what, uh... You did one GPU, two GPUs. Are you gonna do 10 GPUs next?
222
00:40:38.902 --> 00:40:46.562
What's it look like? Yeah. Well, it, it's gonna be a lot of fun into the fall. So I think there's two things in our mind that we really wanna test. So we have this H100 node.
223
00:40:47.102 --> 00:40:58.871
We're gonna be doing, like, token factory, right? So providing tokens, um, as part of that. We're gonna match it up to see what we generate if we ran that 100% uptime as per, like, a tier three data center.
224
00:40:59.082 --> 00:41:01.342
Well, let's call it 99.9% uptime.
225
00:41:02.062 --> 00:41:13.332
Versus now running it at roughly 96, 97% uptime, but also getting the benefits of participating in these various energy programs, and we're gonna match up what the profitability looks like.
226
00:41:13.942 --> 00:41:22.122
And that's gonna basically key in and be, like, a key r- D effort to, to showcase what it would look like for NeoCloud to participate in these programs.
227
00:41:22.802 --> 00:41:31.472
My thesis is that it'll be better to run a higher uptime as of today, but if margins start to come down, that can flip similar to what we saw in Bitcoin mining.
228
00:41:31.562 --> 00:41:42.102
So, like, it, it doesn't necessarily disprove that people are gonna be doing demand response if it comes in lower, but it is an interesting insight to see for now. Um, so that, that's one test we're gonna be doing.
229
00:41:42.442 --> 00:41:48.782
The second test we're gonna be doing, because we have a compute trading business, so we, we buy and sell compute, uh, offtake contracts bare metal.
230
00:41:49.482 --> 00:42:06.422
After that, we're gonna be, uh, seeing if we can, uh, sell a, a cluster of GPUs that come with an SLA with downtime, and passing through the economic benefit of the energy program and see if there's, like, really, like, uh, buyers of compute that are interested in entering this type of contract.
231
00:42:06.982 --> 00:42:19.682
And so that, that is kind of another kind of, like, r- R&D effort for us to see, hey, is this actually something you can build and practice given the nature of the industry and all the different stakeholders along, uh, the value chain.
232
00:42:19.982 --> 00:42:28.102
That seems like where the synergies really lie, by those s- downtime SLAs. Ethan, thank you so much for your time.
233
00:42:28.462 --> 00:42:44.254
Uh, excited to see how Luxer Energy cracks the GPU DR nut, and, uh, yeah, we'll have you on again sometime. Thanks, gents. Have a good one. Thanks, Ethan. Cheers. Uh- Right... same old, it's the same story.
234
00:42:44.434 --> 00:42:53.254
It's just demand response, energy, just different types of compute. You gotta love it. The, the highly transferable skills. I just have to know how GPU works now. [laughs] It's the- Yeah.
235
00:42:53.334 --> 00:43:04.724
You, you just have to figure out how something infinitely more complex than SHA-256 works. No big deal. Yeah. At least we've- No big deal... identified the problem. Um. Uh, quick pivot from our schedule.
236
00:43:05.234 --> 00:43:12.634
Unfortunately, Craig Deveras has had to cancel yet again. Gonna be talking to Hive's PR to see about when we can get him on later this week.
237
00:43:13.454 --> 00:43:26.494
But we are going to cap the show off with a little bit of a roundup of some of the data points on Dimetrics. But first... A word from our sponsor. A word from our sponsor, Ligos.
238
00:43:26.934 --> 00:43:40.014
[static] Bitcoin is gunning to reclaim 80,000, but don't get complacent just because the price is rising.
239
00:43:40.074 --> 00:43:52.054
Look, if you're using a Bitcoin-backed lender, do yourself a favor and check out Ligos Finance before you take on the risk of counterparties that may be playing funny games with your coins.
240
00:43:52.554 --> 00:43:56.454
With Ligos Finance, you don't have to worry about that. They use Bitcoin native smart contracts.
241
00:43:56.494 --> 00:44:07.074
And for that reason, they are our preferred lender here at Blockspace because with those contracts, you always hold the keys to your Bitcoin when you are using a loan with Ligos Finance. You heard that right.
242
00:44:07.094 --> 00:44:19.844
With Ligos Finance, you always hold your coins, self-custody, when you are using their Bitcoin-backed lending program. There's no rehypothecation, no bridging, no wrapping, no funny business.
243
00:44:19.934 --> 00:44:35.734
Just competitive rates as low as 8.5 to 10% APR. Go to ligos.finance to learn more. All right, Charlie, you gave us a good overview of the, specifically the political concentration risk of some of these- Yeah...
244
00:44:35.814 --> 00:44:48.154
companies earlier. What else have we got on Dimetrics that's showing us a little bit of a- Yeah, well-... trend going on? I think it's probably worth a, a stab at taking a look at what the, what, what are the GPUs doing.
245
00:44:48.294 --> 00:45:06.174
So, um, pulling up the compute, uh, pricing, spot pricing for GPU data here. We got A100s up on the day. H100s up on the day 4.5%. H200s down 1.5%, and B200s up 1.5%.
246
00:45:06.554 --> 00:45:23.114
A100s renting at 128 an hour, which by the way is, is nuts to me. Um, because just, um, if I were to zoom in here, just last, in, in June, they were in the high 80s, low 90s.
247
00:45:23.254 --> 00:45:32.354
And in April, they were renting out at $.70, so 70 cents, uh, per GPU hour.
248
00:45:32.374 --> 00:45:48.994
And now this nearly s- this six-plus-year-old, uh, card is renting out at $1.28 an hour. We've said it before- Kind of nuts... but GPU depreciation bears an absolute shambles. Absolute shambles.
249
00:45:49.054 --> 00:45:59.464
The other one, which is a doozy anyway- Can someone check in on Ed Zitron please? [laughs] Yeah. See what he has to say about this. Someone check in on, uh, who's, uh- Michael Burry... Michael Burry?
250
00:45:59.664 --> 00:46:08.394
Yeah, [laughs] Michael Burry. And Michael Burry. I think Michael Burry flipped bullish. I think he reconsidered his priors. Oh, no. Oh, it's over now. Yeah. [laughs] Well, it's over.
251
00:46:08.494 --> 00:46:20.994
[laughs] Well, don't, don't quote me on that. Uh, but I mean, look at this, look at this, uh, B200 chart here. Um, this is kinda nuts because if we go B200s, April of this past year, $3.80 per GPU hour.
252
00:46:21.034 --> 00:46:37.354
Now over, now double that, 7 bucks 30 to rent a, a B200. Um, actually peaked in July at, uh, $8.38. Um, yeah, GPU spot prices be wildin', man.
253
00:46:37.634 --> 00:46:42.134
Uh, they're, uh, all over the place, but up and to the right.
254
00:46:42.594 --> 00:46:54.794
And, and mind you, this is while Neocloud's pulled back while, uh, during the month, uh, during the end of June and July as Leopold Astrenburner got liquidated. But now the- Well, and- Yeah.
255
00:46:54.854 --> 00:47:05.514
And, and something to point, uh, to kinda just con- contextualize that too, Neoclouds obviously have exposure to these spot prices, but only so much as their contracts allow.
256
00:47:05.574 --> 00:47:18.313
Like, I believe CoreWeave and their Q2 update said 98% of their offtake agreements are for multiple years into the future, and only 2% are actually exposed to s- the spot market.
257
00:47:19.514 --> 00:47:23.514
Uh, a- and you know, obviously they're gonna wanna take those multiple year offtake agreements.
258
00:47:24.214 --> 00:47:35.114
But you do wonder, with prices marching up and to the right, how much do they, do they weigh getting more into the spot rental market over the course of the next few years?
259
00:47:35.894 --> 00:47:46.894
I, I would imagine that they'll probably prioritize those longer term g- agreements 'cause it really helps with credit risk, and it helps them, uh, engage with, with banks for more favorable terms for financing.
260
00:47:47.794 --> 00:47:56.464
But it, it does just, it, it does raise, raise the question, you know, how much of a spot mix do you want right now with stuff ripping?
261
00:47:56.994 --> 00:48:07.034
Or are we about to get a local top, and then we're gonna see this drop over the next few, few months? 'Cause it seem, it's really hot right now. And- It's really hot, but I- Go ahead.
262
00:48:07.194 --> 00:48:21.804
I wonder if I can pull up compute indexes, uh, GPU forwards. I don't think I can here live. But like, uh, GPU forwards are c- still in Contango. Um, you pay more for a GPU in six months than you do today.
263
00:48:21.824 --> 00:48:30.974
That's pretty incredible. So I, I would say like, I'm bullish GPU for the next- The bulls you're marching on. Yeah. So that's, that's what I would say.
264
00:48:32.494 --> 00:48:42.794
Um, we could probably take a stab at like another chart on here if you, unless you have a point you wanna make. Um, not really. I just wanna, you know, just a quick shout-out and chill.
265
00:48:43.794 --> 00:48:55.224
The, the DI Metrix dashboards beau- I mean, look at those charts, man. Absolutely beautiful. A- Yeah... incredible- Incredible resource. And getting better by the day.
266
00:48:55.304 --> 00:49:00.984
Brandon Bailey's really putting in some hours to improving it, uh, a little bit every single day, so.
267
00:49:01.024 --> 00:49:11.644
Yeah, we've done a couple interviews with Brandon Bailey, the spearheaded, the spearhead guy behind this dashboard, so strongly recommend everybody check those out.
268
00:49:11.844 --> 00:49:26.164
Yeah, and check out, if you're interested in learning more, signing up, seeing what the plans are like, dimetrics.ai. That's D-I-M-E-T-R-I-C-S dot A-I. Importantly, you get the dashboard, which is great.
269
00:49:26.544 --> 00:49:34.744
You also get the MCP for it, so you can plug it into your LLM of choice and draw directly from the database when you're doing your research.
270
00:49:36.364 --> 00:49:44.404
Um, I think maybe the last, uh, chart here, and this is sorry to the, the podcast listeners here, um, this is a pretty high level view.
271
00:49:44.564 --> 00:49:51.224
I think a lot of times when we talk about, um, the landscape of data- data centers, we don't, like, talk about...
272
00:49:51.264 --> 00:49:59.024
Like, we talk about where they are, what county they're in and what state, but sometimes it helps to actually zoom out at North America and see where these things actually are.
273
00:49:59.054 --> 00:50:15.524
And you really do get a whole story of over 1,000 words in a single picture, which is these data centers are going to Texas, PJM, and smaller ones in Tennessee and Georgia. So, yeah.
274
00:50:16.044 --> 00:50:30.364
The Southeast, Midwest, and Texas will inherit the earth, brother. Yeah, look at that. The, the great, the West, the land of opportunity. Uh, the, basically data centers don't cross the Rockies. [laughs] They go...
275
00:50:30.384 --> 00:50:41.454
Yeah. Well, and, you know, to be fair, most of that is, is kind of hostile for them in the sense that you can't build, you know, 10, nine- 9,000, 10,000 feet or 7,000 feet mile high, um, in the Rockies. Yeah.
276
00:50:41.484 --> 00:50:48.544
That being said, I will, uh, one thing I will say is these, a lot of these are Bitcoin miners, formerly Bitcoin miners, now powered shell operators.
277
00:50:48.984 --> 00:50:56.624
So there are a number of companies that aren't included on this map yet. Like, there are dozens more, if not hundreds more, that are building data centers to power AI.
278
00:50:56.904 --> 00:51:06.684
Yeah, these are really just companies that are in our universe that we cover. Yeah. This doesn't in- this doesn't include data center alley, you know, where the most data centers are. Right.
279
00:51:06.924 --> 00:51:08.384
But the trend- They're in Virginia...
280
00:51:08.424 --> 00:51:22.804
if, if you look at development and the, the arc of development in the United States, the Southeast and sections of the Midwest are leading the way behind Texas, which is undeniably the lo- the state with the most data center development going on- Yeah...
281
00:51:22.824 --> 00:51:38.964
currently. Um, notably, my, the state that I currently reside in, Oregon, which used to have, and maybe still does have one of the highest densities of data centers per capita, uh, is pretty underrepresented here.
282
00:51:39.724 --> 00:51:45.164
I mean, it, it, it's- Yeah... getting impossible to build in the Pacific Northwest, man. Just between- Yeah...
283
00:51:45.224 --> 00:51:57.024
the red tape and the political hostility and the costs too, it, the, you know, it's, I don't really foresee anything happening here. Washington is the only place where you see some development.
284
00:51:57.404 --> 00:51:59.254
Yeah, I mean, you see- California, good luck, man...
285
00:51:59.254 --> 00:52:11.664
a small Bitdeer site and that small Keel site, Moses Lake, which we talked about with Ben last week, um, which they're not, you know, not really a priority for them, as he said on our show the other week. Um,
286
00:52:12.604 --> 00:52:16.664
not a priority to announce a thing there. But yeah, I mean, crazy here.
287
00:52:17.084 --> 00:52:30.844
It really is the story of Texas, and my home state of Oklahoma, which is holding on for dear life so that when Texas secedes, they, uh, they take us with them. [laughs] The Red River rivalry is only extends so far.
288
00:52:31.164 --> 00:52:39.524
[laughs] Yeah. So, uh, thanks for listening to the show, y'all. Um, hope you enjoyed it. We do this every single weekday at 1:00 PM Eastern.
289
00:52:39.544 --> 00:52:50.824
We are Compute's daily live show featuring quick hits on AI, data centers, power, land, markets, and emerging technology. If Bitcoin goes back up, gosh darn it, we're stock, talking about Bitcoin again.
290
00:52:51.424 --> 00:52:59.964
Can't wait for the day. If you like the pod- if you like the live stream, make sure to listen to the podcast if you miss a live stream. Anywhere podcasts are found.
291
00:53:00.024 --> 00:53:05.584
Also, hit our website where you can get our newsletter, newsletter.blockspace.media, website at blockspace.media.
292
00:53:05.644 --> 00:53:17.704
Newsletter, uh, hits your inbox every single day, summarizes everything on the live stream and more, plus written content. Make sure you don't miss it.
293
00:53:18.864 --> 00:53:35.324
This show is sponsored by CleanSpark, uh, NASDAQ listed at ticker CLSK. I'm Charlie. I'm Colin. We'll see you tomorrow. [outro music]
1
00:00:00.080 --> 00:00:07.860
What's up, y'all? Welcome back to Blockspace Live, presented by CleanSpark for another beautiful week of compute, powered land, and energy.
2
00:00:07.900 --> 00:00:16.950
Terawatt just received approval for 482 megawatts of energy load at its Justified data center in Kentucky.
3
00:00:17.480 --> 00:00:36.300
And Charlie, the deal shines a very useful light onto exactly what goes on under the hood when one of these data centers comes in, starts soaking up power from the grid, specifically their taxes, who's really paying for the transmission, in this case, TerraWolf, and also the design of the data center.
4
00:00:36.400 --> 00:00:43.620
A lot of good ammo here for refuting some of the most basic FUD against data centers. So we'll be covering that for our lead story.
5
00:00:44.000 --> 00:00:56.920
Following that, we've got a research note from Morgan Stanley where they expect a flurry of powered shell deals to hit throughout the rest of the year, and they've got some incredibly bullish price targets on some of our favorite stocks.
6
00:00:57.000 --> 00:01:00.320
Now, pretty sure Morgan Stanley's on the sell side for all of these.
7
00:01:00.740 --> 00:01:17.080
Let's take it with a grain of salt, but still some pretty good analysis here, especially with regards to evaluating these companies and also taking a look at what lease terms they might expect as we see these powered shell leases continue to move it up and to the right in terms of dollars per megawatt.
8
00:01:18.000 --> 00:01:32.410
After that, we have two interviews. First, with Luxur COO Ethan Vera on their new BentHouse announcement, plus flexible AI load. You didn't think it was possible, but Luxur and Ethan are gonna tell us how it is.
9
00:01:32.720 --> 00:01:45.420
Then we will wrap up the show with Craig Taveras, the lead of Hive's Buzz HPC segment, to talk about their Q2 results, specifically as it relates to timelines on their latest AI builds and their strategy going forward.
10
00:01:46.620 --> 00:01:59.440
Blockspace goes live every single weekday at 1 p.m. Eastern. We are Compute's daily live show, AI, data centers, powered land, markets, and emerging technology.
11
00:01:59.560 --> 00:02:10.039
If you like the live stream but miss it, it's on podcast form. Anywhere podcasts are found, you can search Blockspace in the search bar of whatever app you listen to podcasts on.
12
00:02:10.759 --> 00:02:22.880
Leave us a review and hit our website, blockspace.media, for expanded written articles, a newsletter that hits your inbox every day, and everything Blockspace content-wise.
13
00:02:23.320 --> 00:02:33.160
This show is brought to you by CleanSpark, Nasdaq-listed ticker CLSK. More on CleanSpark later on in the show. Colin did a deep dive on TerraWolf.
14
00:02:33.320 --> 00:02:46.740
Big announcement this morning, and it's bullish TerraWolf and great ammo for beating that anti-data center FUD. Colin, what are we looking at? Yeah. So this hit the wire this morning.
15
00:02:47.480 --> 00:02:57.620
TerraWolf getting approved for 482 megawatts for a power agreement to power its Justified data center in Kentucky. Reading here from Blockspace,
16
00:02:58.710 --> 00:03:06.880
"The Kentucky Public Services Commission approved electric services for up to 482 megawatts for TerraWolf's Justified data campus in Hancock County.
17
00:03:07.320 --> 00:03:19.040
The August 21st order authorizes Big Rivers and K Energy to implement the agreement." Big Rivers is the utility, and K Energy is one of the owners of that utility.
18
00:03:19.090 --> 00:03:31.180
[lip smack] Um, I believe the exact term is an owner member of Big Rivers. [lip smack] "The services agreement runs for an initial 15-year term and then extends in rolling five-year periods.
19
00:03:31.190 --> 00:03:40.030
And TerraWolf, importantly, will bear the market transmission, delivery, and infrastructure costs specific to, um, the data center for Anthropic."
20
00:03:40.560 --> 00:03:53.020
So just to back up, this is a data center for Anthropic on a 20-year deal. [lip smack] I believe we talked about this earlier in the summer. I'm not-- I don't remember quite exactly when the deal hit.
21
00:03:53.600 --> 00:04:02.780
But that last note about transmission and all of that, Charlie, I think is useful to point out with regards to we're st- we're seeing this kind of across the board at this point.
22
00:04:03.300 --> 00:04:17.320
ERCOT in some sense is leading the way here with the PUC, um, and ERCOT changing rules on data center operators and developers paying for transmission to their sites rather than the utility doing that, as has been kind of standard in the past.
23
00:04:17.720 --> 00:04:28.930
This was also a big pillar for Josh Shapiro's executive order in Pennsylvania. He wants the data centers to pay for grid upgrades, transmission, et cetera, to their sites.
24
00:04:29.360 --> 00:04:34.000
And that is included in this contract that TerraWolf has with Big Rivers and K Energy.
25
00:04:34.720 --> 00:04:42.600
Quoting directly from the contract as well, quote, "TerraWolf must either prepay or provide and maintain credit support for Big Rivers' capacity payment obligations to MISO.
26
00:04:43.180 --> 00:04:47.280
TerraWolf must provide credit support equal to two times the estimated highest monthly bill."
27
00:04:47.320 --> 00:04:58.380
It's basically a contingency to make sure that they can actually pay, um, their, uh, their contractually obligated rate for this energy, making sure that if something goes wrong, they have credit support to back it up.
28
00:04:58.599 --> 00:05:03.080
[lip smack] That's the other thing that is important, uh, to understand about this.
29
00:05:03.700 --> 00:05:15.240
"Big Rivers' utility will procure the campus's energy and capacity through the MISO power pool rather than supply it from the utility's generating fleet." That's according to TerraWolf's energy plan.
30
00:05:15.300 --> 00:05:30.980
Those MISO costs will pass through to TerraWolf dollar for dollar for the full contracted term, ensuring, quote, "The energization or the generation to supply TerraWolf in the near term, medium term, and long term will be satisfied without imposing additional cost on other ratepayers."
31
00:05:31.040 --> 00:05:43.440
End quote. So Big Rivers will be sourcing this energy from within MISO. It won't come from their, their plant specifically, and that's to protect ratepayers in the area. [lip smack] In fact,
32
00:05:44.460 --> 00:05:52.040
one of the things that I thought was interesting about this, Charlie, is that this could potentially lower bills for ratepayers in the area.
33
00:05:53.060 --> 00:06:00.228
Quoting here from the contract or the, um, rather the summary of it by the utility...
34
00:06:00.238 --> 00:06:11.068
commission, the Public Services Commission, quote, "The K Energy adder will defer the need for distribution rate increases or otherwise be used for the benefit of ratepayers," end quote.
35
00:06:11.508 --> 00:06:26.558
That adder specifically is they ha- TerraWolf has a $15 million wholesale adder, which is basically a tariff that they pay annually to Big Rivers, and a $10 million transmission tariff that they pay to Big River, as well as a $1 million distribution adder to K Energy.
36
00:06:26.588 --> 00:06:28.508
Both of these escalate after three years.
37
00:06:30.068 --> 00:07:00.568
Uh, quote, going further here from the summary, "The Big Rivers adder and new transmission revenue will dir- directly benefit all ratepayers through Big Rivers' member rate stability mechanism, under which 60% of Big Rivers' margins in excess of the margin required to achieve its tier for, um, generation are used to pay down regulatory assets, amounts owed to the utility by ratepayers, and the remaining 40% is credited to all ratepayers through bill credits and other, and example, through rate reductions.
38
00:07:01.057 --> 00:07:05.468
TerraWolf is not eligible for any of these MS- MRSM credits."
39
00:07:06.368 --> 00:07:17.048
A big thing to remember about, like, why that would be possible, that this will potentially lower rates for folks in the area, this used to be an aluminum smelting plant.
40
00:07:17.548 --> 00:07:24.808
So Justify is more or less a brownfield on a former aluminum smelting plant that was decommissioned in 2022.
41
00:07:25.028 --> 00:07:38.268
600 jobs were lost as a result, and the county has missed out on $40 million in estimated tax revenue since 2022 when it was decommissioned. And now they also have not received...
42
00:07:38.288 --> 00:07:46.108
Or sorry, that's transmission, uh, transmission tariff revenue, so even more in tax revenue. So that's $40 million that could be going to the utility here.
43
00:07:46.728 --> 00:07:50.468
Now TerraWolf is stepping in and is going to start paying those tariffs.
44
00:07:50.908 --> 00:08:06.508
That means that grid infrastructure upgrades, or rather maintenance, can be satisfied by those tariffs, and also it means that there might be a chance for the utility to lower ratepayer costs with credits and other things like that.
45
00:08:06.588 --> 00:08:12.068
And so this is, I o- I've brought this up before, Charlie, and I had a conversation with my dad about it last week.
46
00:08:12.388 --> 00:08:22.348
It seems counterintuitive at first when you say a data center's gonna move in here, suck up a lot of energy, but it's actually going to lead to rate decreases for other ratepayers.
47
00:08:22.838 --> 00:08:30.548
The reason for that is, are, are these tariffs and the fact that now the utility can potentially max out what it is able to provide.
48
00:08:30.928 --> 00:08:44.508
Still has the same overhead cost, but if you're able to provide more power than le- uh, than less power for the same overhead cost, then you're gonna be able to pass that, those discounts down to your ratepayers, theoretically.
49
00:08:44.538 --> 00:08:53.608
And we've actually seen this in practice as you shared Nick Carter's Murmurations post on this, uh- Mm-hmm... last week. Few other notes here.
50
00:08:54.568 --> 00:09:06.968
In terms of tax revenue, local school districts will receive 3% utility gross receipts on electricity purchases for the site, which is equal to roughly $7 million annually according to the estimates.
51
00:09:07.008 --> 00:09:12.398
Kentucky, the state itself, will reap 6% from energy sales equal to $14 million annually.
52
00:09:13.548 --> 00:09:26.428
Um, so you have 15 million in estimated property tax on top of the utility, or on top of the school district tax, for a total of 22 million in annual tax revenue to Hancock County.
53
00:09:26.588 --> 00:09:44.008
The in- the total estimated tax receipts from the state and the local level combined for this data center will be $1 billion over 20 years for the Anthropic deal. Whoa. Yeah. Holy smokes. This county's gonna eat.
54
00:09:45.508 --> 00:09:57.188
It's pretty incredible, and I think that that- that's why when I opened this- What's the, what's the population of Hancock County? What, what's the population of Hancock County? Let me figure this out.
55
00:09:58.748 --> 00:10:09.008
[laughs] Oh my God, Colin, okay. Over/under 50,000- How many? Over or under 50,000 people. 50,000? Yeah. That goes a long way, man.
56
00:10:09.388 --> 00:10:20.668
And I, that's why when I opened this I said- No, I'm not, I, I, I'm saying as it's 9,000. I was asking over and under. Oh, over and under. Oh, yeah. 9,000 people, a billion dollars over 20 years? Holy smokes. Wow.
57
00:10:21.328 --> 00:10:41.928
Yeah, and, uh, you know, that's why when I opened this I wanted to kind of make the, uh, the data center propagandist point that, you know, that when you look at a contract like this, you actually have so many refutations to the litany of gripes that you typically see from the anti-data center folks and the people protesting these things.
58
00:10:43.088 --> 00:10:56.638
And, uh, a few other things here too. So the site will generate, uh, 800 to 1,000 construction jobs estimated, plus 95 full-time employees from TerraWolf on the site, plus Anthropic staff.
59
00:10:56.728 --> 00:11:03.928
Now, of course, like I don't know how many Anthropic will have on there. Let's, let's just like say maybe like ballpark 120 for the whole thing.
60
00:11:05.488 --> 00:11:15.388
That's nowhere near 600 for the smelter, but as I, I responded to a comment on our YouTube page today from what I believe was a bot, and I just said, "Well, what's the alternative?"
61
00:11:15.908 --> 00:11:25.008
Because manufacturing isn't competing for these sites, right? Like no- nothing else is moving in here, so it's either you take that or you just don't get anything.
62
00:11:25.448 --> 00:11:31.968
You don't get the jobs, as meager as they may seem considering how big the data center is. You also don't get the revenue from the taxes.
63
00:11:32.648 --> 00:11:42.468
The other thing that I thought was super interesting, T- TerraWolf did a lot of community engagement leading up to the approval for this site. They held m- uh,
64
00:11:43.448 --> 00:11:50.098
numerous local meetings, had PowerPoint presentations and kind of like town hall meetings to answer questions from the community,
65
00:11:51.808 --> 00:12:04.638
and th- the, the summary specifically mentions that this will be a closed loop cooling system. Do you wanna kn- uh, y- go ahead and guess how many gallons per day the smelter was consuming before it shut down.
66
00:12:04.728 --> 00:12:18.766
Oh, oh, am I... I, I would say, oh, I don't know. 20 million Oh, not, no, no, per day. Per day. Per day. I don't know. Give-- Tell me. Just answer. 550,000 gallons per day. Oh my gosh.
67
00:12:18.766 --> 00:12:27.836
And TeraWolf, the summary points out, will not draw from the Ohio River, nor will it draw from the municipal water system.
68
00:12:27.876 --> 00:12:40.176
So, you know, y- y- when you think about that, n- I doubt there was any backlash to the, to the smelter when it came in. You know, it creates jobs. But th- this, this data center will use
69
00:12:41.216 --> 00:12:44.696
n- nothing close to the same amount of water that that was using.
70
00:12:45.396 --> 00:12:55.816
And I, I just do, uh, want to point out that this is the energy plan that TeraWolf submitted to the Kentucky Energy and Environment Cabinet, and this is what is, was being summarized in this.
71
00:12:56.596 --> 00:13:02.076
Um, but it was reviewed by these regulators and by these officials, all that stuff. So, uh, the, the last,
72
00:13:03.136 --> 00:13:14.095
the last thing I wanted to point out, if you look at the tax revenue, you can back out an estimated power cost for this agreement. Now, this is estimated. TeraWolf has not, has not disclosed this.
73
00:13:14.216 --> 00:13:27.296
I just wanna get that out front. Also, it, it's not a fixed cost PPA, it'll be floating. So-- But if you look at the tax revenue, back that out, you're looking at roughly pre-tax 55 to $61 per megawatt hour.
74
00:13:27.846 --> 00:13:40.616
And the 61, uh, the higher end there, the $61 per megawatt hour is based on TeraWolf's 90% load factor used for the tax estimate. So, you know, pr- pretty decent rate considering...
75
00:13:40.936 --> 00:13:54.246
You know, that would've been an okay rate for Bitcoin mining, but, but for what they're using this for now, for hosting AI equipment, that's, I would imagine, pretty favorable with regards to the margins, so. Yeah.
76
00:13:54.255 --> 00:14:00.796
[clears throat] I mean, a lot of people think that pulling from the power grid is like pulling from a well,
77
00:14:01.976 --> 00:14:09.616
but it's more like driving a car, where if you're ramping the car up and down and gassing the pedal, you're really burning fuel inefficiently.
78
00:14:10.396 --> 00:14:28.016
But if you're able to run the car at a, like, 55 miles an hour at a s- high speed and make the engine bigger and more efficient and balance it, uh, then, then aggregate, the whole system performs better, your power prices go lower, net more benefit across the board.
79
00:14:28.616 --> 00:14:43.336
When you take away core pieces of the load-consuming part of the power grid, uh, the entire system gets less efficient, less optimal. When you put them back, we're back to working baby.
80
00:14:43.396 --> 00:14:56.666
So that's why a billion dollars to the county of Justified- Hancock... Hancock County, uh, well, more, more power to 'em. Uh- I love that analogy about a, about the car. Yeah. That's really good.
81
00:14:56.776 --> 00:15:07.576
I, I just, I came up with that just now. We gotta keep going. We gotta roll. We gotta... Let's talk about Morgan Stanley and Powered Shells. Our data center propaganda rolls on after a word from our sponsor, CleanSpark.
82
00:15:07.685 --> 00:15:22.616
[gentle music] We are CleanSpark, America's Bitcoin miner, a publicly traded company with the largest operating hash rate, powered entirely by self-operated infrastructure across four states.
83
00:15:24.456 --> 00:15:35.876
This is our proof of work, and we are setting the standard for what's next. Learn more about the intersection of energy and Bitcoin at cleanspark.com.
84
00:15:40.756 --> 00:15:45.736
All right, Charlie, let's see what the bankers are saying. Yeah. The bankers wanna buy your stocks.
85
00:15:46.276 --> 00:15:54.926
[chuckles] The bankers want you to buy their stocks on the sell side, and shockingly enough, they are uber bullish on Powered Shells.
86
00:15:55.536 --> 00:16:00.536
Headline here from Blockspace, Morgan Stanley expects Powered Shell deal wave by October.
87
00:16:00.736 --> 00:16:11.706
[lips smack] Morgan Stanley expects Powered Shell operators to sign a wave of data center contracts by October as AI customers race to secure projects that can energize before the end of 2027.
88
00:16:12.556 --> 00:16:26.216
The bank maintained an attractive industry view in a Monday note led by analyst Steven Bird. Morgan Stanley kept overweight ratings on Cipher, TeraWolf, Riot, Hut 8, and Galaxy.
89
00:16:26.236 --> 00:16:39.356
Based on August 21st closing prices, the bank's $62.50 TeraWolf target applies about 300-- implied about 300% upside. It's pretty rich. It depends on what your timeframe is though, I guess. Yeah.
90
00:16:39.396 --> 00:16:52.876
While its $43.50 Cipher target implies 176% upside, its t- $273 Hut 8 target implies 238% upside, and its $43 target implies about 159% upside.
91
00:16:53.556 --> 00:17:03.876
Morgan Stanley rates Applied Digital as equal weight and Mara underweight, although it sees contract-driven upsides for both. So obviously very bullish.
92
00:17:04.476 --> 00:17:20.356
They get into some of the evaluation modeling, or evaluation modeling that they use for this. Also, some of the modeling that they use for, um, [lips smack] for, uh, the deals themselves.
93
00:17:20.856 --> 00:17:33.876
Morgan Stanley's indicative model assumes a 20-year lease generating $2.38, uh, cents per watt annually, with an 85% margin and construction cost of $11.50 per watt.
94
00:17:34.596 --> 00:17:45.156
The model uses 80% loan to value for the site, 7.25% project debt, a 9% equity discount rate, and a 6.7% terminal capital rate.
95
00:17:45.656 --> 00:17:58.856
The TeraWolf data center, I didn't mention this during that segment, but it is estimated to cost 4 to 4.5 billion at a cost of about 10 to $12, uh, per watt, 10 to $12 million per megawatt.
96
00:17:59.186 --> 00:18:06.556
And that's specifically for the Powered Shell itself. So that doesn't include the, the, the GPUs, right? They're just building the data center.
97
00:18:06.616 --> 00:18:13.096
[lips smack] But on the note of other models for the valuation here, Charlie, I thought this was interesting.
98
00:18:13.736 --> 00:18:28.236
They were looking at lease rates, and Morgan Stanley says that lease rates have risen from about o- $1.80 per critical IT watt in September 2025 to $2.30 to, to $2.40 per watt currently.
99
00:18:28.926 --> 00:18:40.546
Morgan Stanley estimates that a 100% basis point increase in average yield on cost could create another $3.50 per megawatt, more than offsetting the valuation effect of 100% basis point increase in interest rates.
100
00:18:41.086 --> 00:18:53.506
That being top of mind, uh, you know, we'll have Jim Bianco on tomorrow to talk about this, great macro analyst, but 30-year yield is really running away from the, from officials right now.
101
00:18:54.546 --> 00:19:13.366
Treasury has decided to up its total ceiling for, uh, interventions for buybacks from two billion to four billion per operation, and so there's, there, there's a, a decent amount of fear right now within the bond market that the long tail of these durations are going to continue to rise.
102
00:19:13.866 --> 00:19:24.146
And if that is the case, then debt for these data centers, as it already has been, is going to get more expensive, um, over the next few years if that continues to float upward. Yeah.
103
00:19:24.346 --> 00:19:33.606
Um, kinda looking at this Morgan Stanley note, it's pretty much if you wanted a one-liner, it's which powered shells have existing megawatts that are not yet leased.
104
00:19:34.146 --> 00:19:45.745
Those are the people they are bullish on, and that list is, as you said, what? Cipher, TeraWolf, Riot, Hut 8, Galaxy as, like, the, the leading ones who have not announced a lease.
105
00:19:45.866 --> 00:19:58.926
Uh, there's another dimension here, and I thought I'd tap into DI Metrics, our, uh, data center intelligence tool, dimetrics.ai, link in description. Um, and
106
00:20:00.166 --> 00:20:11.726
it's kind of interesting to look at the political risk 'cause you have some of these companies have, within their portfolio, multiple sites. Some of the sites are, like, good to go.
107
00:20:11.786 --> 00:20:26.286
They don't have permitting or, like, procedural, uh, issues they're waiting on, and a couple of them do. Specifically with, like, TeraWolf, who's, uh, and there's actually a political risk, uh,
108
00:20:27.206 --> 00:20:39.766
uh, number or rating assigned to each company according to DI Metrics here. I'll actually share it. And, um, uh, it's really interesting.
109
00:20:39.886 --> 00:20:51.526
So, I mean, let's look at, uh, TeraWolf, who has the highest political risk of all of the, uh, companies that Morgan Stanley talks about here, with a political risk on DI Metrics side of 41.
110
00:20:51.586 --> 00:20:57.146
And that's because out of their five sites, Lake Mariner, Muskie, Hawesville, those three are good to go.
111
00:20:57.726 --> 00:21:09.066
Morgantown has an elevated permit risk, as Lake Hawkeye is, uh, has a permit denial risk, um, according to DI Metrics.
112
00:21:09.846 --> 00:21:23.606
Um, if we look at the companies in, uh, in Texas with Batch Zero, it looks a little bit different. So that would be Cipher, Galaxy, and Hut and Mara.
113
00:21:24.366 --> 00:21:41.166
Um, because Texas, uh, at the state level forbids counties banning data centers, uh, outright, um, they don't, they, they forbid data center moratoriums so the counties have to then handle these at the, like, you know, more, uh, lower level permit issue.
114
00:21:41.766 --> 00:21:56.466
You have companies like Hut 8, Cipher, and Galaxy have political risk scores of very low. In Cipher's case, um, their, seven of their eight sites, Colches, Apollo, Barber Lake, Mikeska, McClellan,
115
00:21:57.466 --> 00:22:14.566
Milsing, and Black Pearl, those have very low political risk, whereas Ulysses does have a higher permit risk as it's not in Texas. It's in Ohio. Um, so some pretty interesting, uh, insights here. Uh,
116
00:22:15.506 --> 00:22:23.386
I have some more data. And I, I think you already said this, but it's- Okay... worth noting for TeraWolf specifically that, you know, for the,
117
00:22:24.466 --> 00:22:33.526
for the Lake Mariner campus, they already have approvals for what they're building out right now. Exactly. The moratorium doesn't affect them. Exactly, yeah. Moratorium doesn't affect them in New York.
118
00:22:33.846 --> 00:22:48.326
Uh, they're ob- you know, grandfathered in, you could say. Uh, but, um, you know, I would say that m- as Morgan Stanley just looks at unleased megawatt, um, among a portfolio, uh, they should,
119
00:22:49.366 --> 00:23:02.386
I, if, you know, if I were Morgan Stanley, I'd be very, very reluctant to consider megawatts which are still subject to approval in hostile counties as absolutely not done deals. Yeah.
120
00:23:02.426 --> 00:23:09.186
Because- And that's a fair point to make, and one of the things they say towards the end of the research note is they're actually kind of in support of what you're saying. Yeah.
121
00:23:09.196 --> 00:23:20.816
They expect Cipher to contract 700 to 1,000 megawatts gross by October and Riot to lease the full 1,000 gross megawatts for Corsicana by the end of September. Curious to see what they know. Yeah.
122
00:23:20.846 --> 00:23:28.726
I mean, usually these analysts have an inside look at, like, how some of these things are progressing. So, um, again- Yeah... would be, be curious to see.
123
00:23:28.766 --> 00:23:41.066
But again, a caveat, they're on the sell side for all these companies. So, that, so, you know, those, those, um, those upside targets, take 'em with a grain of salt. They're trying to get me to buy more Cipher stock.
124
00:23:41.116 --> 00:23:53.846
Trying to build- Joke's on you. I already own Cipher stock. [laughs] So, uh, uh, we, we're, we'll keep covering the data center moratorium story, uh, a lot, obviously, over this fall. But before
125
00:23:55.306 --> 00:24:08.166
we keep co- uh, we, but, but rather, we have in the wings, we have Ethan Vera, uh, backstage. We're gonna talk about demand response and the inaugural GPU Demand Response event in Texas this past month.
126
00:24:08.686 --> 00:24:15.716
Before we bring up Ethan, we gotta do a s- a word from our sponsor. Colin, our sponsor, Luxur.
127
00:24:16.266 --> 00:24:28.066
[static sound] This episode of Block Space Live is brought to you by Luxur's Commander, Bitcoin miner management software for enterprise operations.
128
00:24:28.606 --> 00:24:34.346
Luxur's Commander gives you real-time fleet monitoring, bulk remote commands across your fleet, and Intelligent Miner.
129
00:24:34.914 --> 00:24:41.804
That's an automated profitability engine that runs every five minutes and adjusts your fleet's power settings to live energy and hash rate markets.
130
00:24:42.334 --> 00:24:48.384
In fact, ERCOT back tests show 10% improved profitability with intelligent mining versus old-fashioned binary mining.
131
00:24:49.154 --> 00:24:57.754
Commander Pro is roughly half the cost of competition, $100 per megawatt, or a 25 basis point pool fee adder. But you can also try it free for 60 days.
132
00:24:57.874 --> 00:25:11.094
So if you'd like to learn more, go to luxor.tech/commander to get started. All right. We got Ethan. Let's bring him on up here. Ethan Vera, Luxor, welcome to the show. Gentlemen, great to see you both.
133
00:25:11.934 --> 00:25:25.654
Good to see you too, man. Great. Thanks for joining. Absolutely. So you guys had a big announcement, uh, an hour ago. Luxor Energy and BentHouse... Am I saying this correctly? BentHouse? Uh- Spot on. Okay.
134
00:25:25.874 --> 00:25:38.694
Uh, complete first AI GPU response to ERCOT 4CP signals. Give me the, the high-level breakdown of what happened. Yeah. I mean, there, there, there's a lot of themes here that we can, we can touch on.
135
00:25:39.014 --> 00:25:48.354
Um, and so what, what it comes down to is really we're in the business from an energy perspective of sending signals to data centers to participate in energy programs.
136
00:25:49.034 --> 00:25:57.354
We, we've been playing in that space in the Bitcoin mining side, which we've talked a lot about on, on this podcast, and we think that's vital for the data center space.
137
00:25:57.874 --> 00:25:59.654
Um, it probably goes to a lot of themes you talked about.
138
00:25:59.694 --> 00:26:07.814
I don't know if you talked about it earlier today on this pod, but certainly there's been a lot of pushback against data centers, uh, both domestically as well as globally.
139
00:26:08.374 --> 00:26:17.244
And we think that a, a big part of that is data centers' inability as of today to participate in the grid servicing programs that are really vital for, for the grids.
140
00:26:17.814 --> 00:26:25.834
And so when we think of, you know, different grids around the United States, they're trying to solve for supply and demand to make sure that things like brownouts and blackouts don't happen.
141
00:26:25.934 --> 00:26:32.754
Um, ancillary service programs and demand response is crucial for that, and that's why really Bitcoin miners have been helpful for the grid.
142
00:26:32.794 --> 00:26:44.814
We think that it's, it's inevitable that they really push data centers in that same direction, and that data centers consuming their dozens and soon hundred gigawatts of power within the United States are participating in these programs.
143
00:26:45.394 --> 00:26:57.074
And so really what we did was we ran an H100, uh, and a B200 through, uh, 4CP. So 4CP is the four coincidental peaks in, in, in ERCOT, and we ran it through these dispatches.
144
00:26:57.634 --> 00:27:10.063
The, the B200 was live curtailing based on, on these dispat- dispatch signals and basically throttling down to 25% of its power consumption. During that, it did not lose its, like, inference workloads.
145
00:27:10.114 --> 00:27:21.834
It was doing checkpoints. Um, and, and so we're, we're basically furthering the test of, uh, AI GPU's ability to go through this. Next up for us is to go through some of the ERS programs in the fall.
146
00:27:21.854 --> 00:27:37.184
And so we procured a node of H100s we're shipping to a data center just inside, uh, kind of Dallas-Fort Worth area, and we're gonna be enrolling it in, in these various, uh, ERS programs to, uh, continue to do some research development here with the eventual goal of helping data centers do demand response.
147
00:27:37.994 --> 00:27:49.134
So Ethan, you said that the data centers have been slow to implement these kind of things, or rather that it just, we haven't seen it at scale yet. What are some of the reasons for that?
148
00:27:49.174 --> 00:27:55.364
Is that simply because a lot of these compute contracts have, like, mandatory uptime clauses?
149
00:27:55.414 --> 00:28:06.674
You know, if you think about building a tier three-plus data center, you're committing to a much higher threshold of uptime versus a Bitcoin mine, and you have to be delivering that compute or else you'll be in violation of the contract.
150
00:28:06.734 --> 00:28:16.734
Is it that, is it that part of it? Is it that the systems just haven't caught up, or the tooling hasn't caught up to allow this to happen? What are some of the obstacles? I see it in two ways.
151
00:28:17.274 --> 00:28:24.114
Fir- first is the economics. So when we talk about Bitcoin mining, that's become a commodity business over the years.
152
00:28:24.364 --> 00:28:33.494
And, and over the years what we've seen is that it's very crucial that Bitcoin miners are participating in grid servicing programs because it ends up impacting their profits in a very positive way.
153
00:28:33.994 --> 00:28:45.694
If you look at an example like Riot Blockchain, for example, or now Riot Platforms, in some cases they were making more money off of their energy business and, and, and participating in these programs than they were Bitcoin mining itself.
154
00:28:46.214 --> 00:28:54.154
And that really speaks to the, the type of, like, commoditization that happened in the mining space. That has not happened in the AI data center space as of today.
155
00:28:54.633 --> 00:29:04.534
Margins are still 95%, and really it does not make sense to be curtailing your machines for any grid servicing programs out there or any financial incentive that is currently offered by the grid.
156
00:29:05.574 --> 00:29:10.934
Um, so that's like the first obstacle, which I think will, will start to change over time. The second is the technology stack.
157
00:29:11.514 --> 00:29:24.494
Um, there's historically been, uh, very little interest in, in finding ways to, uh, curtail like a training workload, for example, or even inferencing. Um, that being said, there's dozens of teams working on this.
158
00:29:24.554 --> 00:29:27.734
We partnered with BentHouse, uh, Zeoni Software on this one.
159
00:29:27.874 --> 00:29:36.974
Um, but there are dozens of really smart technical teams that are working on, on the ability to curtail GPUs, but not give up and be detrimental to its workloads.
160
00:29:37.234 --> 00:29:47.054
So those are like the two obstacles I see, both of which are being addressed right now and I think will, will, will look a lot different in a, in a few years from now. I am kind of curious.
161
00:29:47.174 --> 00:29:54.734
Um, I, uh, I could ask a question about maybe more like the hardware, like how, how does the curtailment work?
162
00:29:54.974 --> 00:30:07.674
But I'm actually more interested in like the decision-making process because in a Bitcoin mine, uh, it's, it's a little more straightforward, like who makes the call to turn off, uh, the, the mining rig?
163
00:30:08.094 --> 00:30:25.214
But in these, uh, much more complex dynamic contractual arrangements of tenants of these, uh, data centers, uh, I'm curious like, uh, while this business does not maybe exist or the, or ra- these relationships don't exist today at scale,
164
00:30:26.214 --> 00:30:37.454
how do you envision like the decision-making of how and when to turn off, uh, going down the stack? Is this a c- a tenant decision thing? Is this a operator decision thing?
165
00:30:37.514 --> 00:30:45.862
Like, how does this, how should, how do you think this, this will work in the future? Yeah. And, and this is a bit future dated, so all of this is kind of like assumption of future state.
166
00:30:46.542 --> 00:30:51.642
Um, the, the majority of partners we work with are-- they, they separate on the value chain.
167
00:30:51.902 --> 00:30:58.382
So there's one company that runs a data center, there's another company that owns the GPUs, there's another company that rents the GPUs.
168
00:30:58.442 --> 00:31:05.772
I mean, the clearest example is like a core scientific, core weave, like anthropic mix, right? Where it's not a single company that, that is doing everything.
169
00:31:06.362 --> 00:31:12.762
And so it, it's a really good question, which is like, who's actually dictating the participation in these programs and making the decisions?
170
00:31:13.602 --> 00:31:25.182
Ultimately, the, the end user, whoever is renting the compute on bare metal or owns the comp-- like hardware themselves and is using the hardware, they're, they're gonna have to sign off on participating in these programs.
171
00:31:25.782 --> 00:31:52.062
And so, um, that's gonna be reflected in, in SLAs and, and MSAs that are being signed between them and the owner of hardware and the co-location, and they'll need to get them across the finish line of saying, "Hey, look, we'll, we'll offer you this co-location space, you know, twenty megawatts worth of rack-ready power capacity at, you know, twenty cents a kilowatt hour, but you're gonna have to be able to curtail one percent of the year because we're participating in these programs."
172
00:31:52.562 --> 00:31:59.152
And maybe, you know, tho-those energy benefits will pass through in some capacity to that end user, so they'll benefit from it in some ways.
173
00:31:59.962 --> 00:32:06.782
Um, as of today, we haven't seen a lot of those deals get struck for that s-similar reason we're talking about, which is the economics just aren't there.
174
00:32:06.802 --> 00:32:14.262
It's better right now to run your GPU workload a hundred percent uptime than ninety-nine percent and get a bit of energy like ERS revenue.
175
00:32:15.002 --> 00:32:24.842
But in the future, that is either probably gonna change as the industry gets commoditized and margins come down and/or regulation and it, uh, basically forces their head.
176
00:32:25.382 --> 00:32:37.922
And that, that's kind of one of our thesis is that there is, um, in, in the interconnect, uh, demand, both ERCOT and other markets, there's dozens and soon hundreds of gigawatts of demand of new AI data centers.
177
00:32:38.002 --> 00:32:46.042
Clearly, uh, the grids are looking at ways to support that through cogeneration, through storage, um, through any means possible.
178
00:32:46.082 --> 00:33:07.642
We think demand response participation is gonna be a big part of that, and so it's gonna be almost a forcing function where it's like, okay, if you want a hundred megawatts on grid opportunity in say NYISO or MISO or PJM, you need to be able to participate in some of these programs that, uh, helps the grid stabilize and you become an asset to the grid rather than necessarily, you know, a, a harm.
179
00:33:09.342 --> 00:33:27.082
Uh, you mentioned the Emergency Response Service in ERCOT, ERS, and the, uh, uh, you know, that's more of a shutoff to keep the, the power grid on while a winter storm rolls in. As AI data centers proliferate in ERCOT,
180
00:33:28.002 --> 00:33:38.002
what does the ERS landscape look like? And is this like a business opportunity? Is this, uh, goodwill? Is this to jump through the permitting hoops? Like,
181
00:33:39.162 --> 00:33:46.422
what does this look like as, as Luxor Energy, uh, kind of views the whole system? We, we view it as a mix.
182
00:33:46.482 --> 00:33:54.282
Like, in some ways it is like maybe a way to speed up your, your approval to get, you know, uh, power access in ERCOT.
183
00:33:54.562 --> 00:34:15.822
And so if, if you can come to the grid and say, "Look, we have an application for, let's call it a hundred megawatts of power, and we're willing to participate in all these programs," in our view, the grid is gonna look at that much more favorably in ERCOT will than somebody that wants a hundred gigawa-- or a hundred megawatts of power is not participating in these programs and is not, you know, providing storage or cogeneration.
184
00:34:16.262 --> 00:34:25.262
And so in some ways, like for us, it, it is like the path forward for data centers to actually achieve, uh, and, and acquire the power access that they are looking for.
185
00:34:26.162 --> 00:34:33.622
Um, and so that, that's, that's kind of like the, the main view that we have regarding like regulation and AI, uh, demand response.
186
00:34:35.022 --> 00:34:50.562
As you talk to the engineers, um, who have, who actually have to build these integrations to get the H100 to turn off, um, it's, I assume it's, it's much more complex than shi-shi-shutting off entire buses of ASICs.
187
00:34:50.602 --> 00:35:03.522
I mean, 'cause now you have, uh, whole like systems, modular systems, integrated GPU, CPU, Vera Rubin racks, seventy-two of these things, and the GPU here is not the same as the GPU further down the stack.
188
00:35:03.582 --> 00:35:13.762
Like, what are you seeing from the engineers about the engineering problems that have to be solved to roll this out at scale? Yeah. I mean, you hit the nail on the head.
189
00:35:14.002 --> 00:35:26.362
Th-this is extremely more complex than demand response for ASICs. Um, in ASICs, like for Bitcoin mining, it's very clear the economic signals to curtail, um, all the machines have very similar APIs to call on.
190
00:35:26.982 --> 00:35:34.402
Um, you can quickly prioritize like the, the dropping down the least efficient equipment. I mean, that's what we do through intelligent mining, um, and Commander.
191
00:35:34.702 --> 00:35:40.322
So like that logic is, is a lot easier to think through in Bitcoin mining than it is AI. Um,
192
00:35:41.502 --> 00:35:49.982
I don't have all the answers yet, um, and I think this is where like a lot of the smart teams and companies getting funded are gonna address, including NVIDIA's spending a lot of time on this too.
193
00:35:50.522 --> 00:35:58.182
I mean, the key thing too is like what is the, what is the workload that the GPU is running? Some of those workloads might not be able to be interrupted.
194
00:35:58.802 --> 00:36:09.692
And so like having key insights into the, into the workload, uh, side of the GPU is a unique problem that like Bitcoin mining didn't really face because we all knew what the workload was. It was SHA-256.
195
00:36:10.142 --> 00:36:16.442
And so like there's that distinction here that still needs to be worked out. And so that's why I kinda think a lot of this like software is like long dated.
196
00:36:16.882 --> 00:36:20.262
Over the next like few years, um, we're gonna see a lot of innovation in the space.
197
00:36:20.282 --> 00:36:32.462
There's gonna be dozens of different software providers, including probably, you know, in partnership with NVIDIA, some of the big OEMs like Supermicro and others that are gonna be directly integrating the software into the hardware builds.
198
00:36:32.502 --> 00:36:43.922
And it sounds like to me, Ethan, I think in the, for the pilot that you were talking about at the beginning, the workload was reduced to like twenty-five percent So should we basically expect that to be the standard?
199
00:36:43.962 --> 00:36:53.602
You're not gonna shut off the entire deployment, you're just going to kinda curtail its capacity so that it's working maybe a quarter as much as it would otherwise. Exactly.
200
00:36:53.942 --> 00:37:00.662
Um, t- 25%, uh, like, made sense for the B200 in this certain, like, workload circumstance.
201
00:37:01.482 --> 00:37:08.602
Um, I think directionally that's probably right in terms of what other GPUs running inferencing workloads could be throttled down to.
202
00:37:08.622 --> 00:37:18.722
But my understanding as of today is that you're not gonna wanna throttle these down to 0%. Um, from a hardware longevity standpoint, from a workload standpoint, it just doesn't make sense.
203
00:37:18.782 --> 00:37:25.462
That being said, even throttling it down, like 50, 75%, is, is insanely valuable to the electricity grids.
204
00:37:25.662 --> 00:37:37.101
So if you can bring a 100 megawatt facility down to 50 megawatts when you're called upon and there's an emergency, like a cold winter storm or, you know, a heatwave in, in, in Texas, that's a really valuable asset.
205
00:37:37.142 --> 00:37:51.932
And now the grid looks at you as a value add rather than a detractor. Um, you mentioned it, training versus inference. It's, like, very difficult to know where, what this landscape will look like in five-plus years.
206
00:37:52.502 --> 00:38:07.442
But could you just speak to some of the, you know, insights that you and the team are having about, like, the different types of workloads, even just a simple high level of training versus inference, but, uh...
207
00:38:07.692 --> 00:38:14.342
and, and, like, w- how, how those might fit into these demand response, uh, si- signals? Yeah.
208
00:38:14.402 --> 00:38:21.202
So this is k- kind of where our partner BentHouse came in on it, so we'll, we'll probably have to arrange another, uh, one with Bob on.
209
00:38:21.302 --> 00:38:32.362
Um, Luxer in this partnership, we're the QSE, right, in our cloud, so we're setting the actual dispatch signals to this and, and, um, sending the signals to the GPU when to curtail based on four CP.
210
00:38:32.992 --> 00:38:41.362
Um, their software is the one, uh, working with the GPU to kind of assess the workload, curtail, uh, well, throttle down the power consumption.
211
00:38:42.002 --> 00:38:55.462
Um, based on my understanding talking with them, and they're really, like, the experts on this topic, is that as of today they haven't found a way to, um, throttle GPUs running training without it being detrimental to that training effort.
212
00:38:55.472 --> 00:39:07.802
And so my understanding, like from a very high level perspective, if, if you throttle a fleet or a cluster of GPUs that are i- like, actively training a model, you're gonna, like, destroy a lot of work that's been occurring.
213
00:39:07.922 --> 00:39:17.402
Um, it may not, like, reset it to zero, but it's gonna set it back significantly. Imagine, like, throttling a, say a steel mill, where you turn off the power immediately and there's stuff in production.
214
00:39:17.682 --> 00:39:21.802
Like, it, it's really hard to fire it back up and, like, there's just cost there.
215
00:39:21.842 --> 00:39:29.682
Whereas certain inferencing workloads, they're easier to interrupt and you can, like, kind of now send that inference workload to another data center to start working on.
216
00:39:29.702 --> 00:39:40.122
And so it, it, it's easier to imagine a future in which inferencing is a bigger participant in demand response programs than other workloads, but that's based on my understanding today. And like,
217
00:39:41.022 --> 00:39:51.362
again, the technology's changing so fast here that we could be speaking in two years and that could be totally different. And it's- Yeah, and I'm imagining, uh, getting surge pricing on my Claude queries, so.
218
00:39:51.382 --> 00:40:02.082
[laughs] Yeah. Well, you know, that's... Yeah, I was just gonna say, it sounds like it, it kind of surfaces a new engineering problem for networking these data centers together for, for clusters all around, right?
219
00:40:02.162 --> 00:40:10.782
It's like if one of them goes down or has to c- curtail capacity, then you have to ship that demand somewhere else. And then what does that do, Charlie, like you said, to on-demand token prices?
220
00:40:10.862 --> 00:40:21.301
Yeah, if it's too cold in Texas and a winter storm's rolling through, I'm just gonna not use AI for a few days [laughs] 'cause it'll blow out our budget. Yeah. So, you know, maybe... You know, I could see that future.
221
00:40:23.802 --> 00:40:38.862
Uh- Absolutely. Yeah. So, uh, so you s- you did this initial test. Uh, any indication on, like, what the next step is as far as demoing, scaling up, what, uh... You did one GPU, two GPUs. Are you gonna do 10 GPUs next?
222
00:40:38.902 --> 00:40:46.562
What's it look like? Yeah. Well, it, it's gonna be a lot of fun into the fall. So I think there's two things in our mind that we really wanna test. So we have this H100 node.
223
00:40:47.102 --> 00:40:58.871
We're gonna be doing, like, token factory, right? So providing tokens, um, as part of that. We're gonna match it up to see what we generate if we ran that 100% uptime as per, like, a tier three data center.
224
00:40:59.082 --> 00:41:01.342
Well, let's call it 99.9% uptime.
225
00:41:02.062 --> 00:41:13.332
Versus now running it at roughly 96, 97% uptime, but also getting the benefits of participating in these various energy programs, and we're gonna match up what the profitability looks like.
226
00:41:13.942 --> 00:41:22.122
And that's gonna basically key in and be, like, a key r- D effort to, to showcase what it would look like for NeoCloud to participate in these programs.
227
00:41:22.802 --> 00:41:31.472
My thesis is that it'll be better to run a higher uptime as of today, but if margins start to come down, that can flip similar to what we saw in Bitcoin mining.
228
00:41:31.562 --> 00:41:42.102
So, like, it, it doesn't necessarily disprove that people are gonna be doing demand response if it comes in lower, but it is an interesting insight to see for now. Um, so that, that's one test we're gonna be doing.
229
00:41:42.442 --> 00:41:48.782
The second test we're gonna be doing, because we have a compute trading business, so we, we buy and sell compute, uh, offtake contracts bare metal.
230
00:41:49.482 --> 00:42:06.422
After that, we're gonna be, uh, seeing if we can, uh, sell a, a cluster of GPUs that come with an SLA with downtime, and passing through the economic benefit of the energy program and see if there's, like, really, like, uh, buyers of compute that are interested in entering this type of contract.
231
00:42:06.982 --> 00:42:19.682
And so that, that is kind of another kind of, like, r- R&D effort for us to see, hey, is this actually something you can build and practice given the nature of the industry and all the different stakeholders along, uh, the value chain.
232
00:42:19.982 --> 00:42:28.102
That seems like where the synergies really lie, by those s- downtime SLAs. Ethan, thank you so much for your time.
233
00:42:28.462 --> 00:42:44.254
Uh, excited to see how Luxer Energy cracks the GPU DR nut, and, uh, yeah, we'll have you on again sometime. Thanks, gents. Have a good one. Thanks, Ethan. Cheers. Uh- Right... same old, it's the same story.
234
00:42:44.434 --> 00:42:53.254
It's just demand response, energy, just different types of compute. You gotta love it. The, the highly transferable skills. I just have to know how GPU works now. [laughs] It's the- Yeah.
235
00:42:53.334 --> 00:43:04.724
You, you just have to figure out how something infinitely more complex than SHA-256 works. No big deal. Yeah. At least we've- No big deal... identified the problem. Um. Uh, quick pivot from our schedule.
236
00:43:05.234 --> 00:43:12.634
Unfortunately, Craig Deveras has had to cancel yet again. Gonna be talking to Hive's PR to see about when we can get him on later this week.
237
00:43:13.454 --> 00:43:26.494
But we are going to cap the show off with a little bit of a roundup of some of the data points on Dimetrics. But first... A word from our sponsor. A word from our sponsor, Ligos.
238
00:43:26.934 --> 00:43:40.014
[static] Bitcoin is gunning to reclaim 80,000, but don't get complacent just because the price is rising.
239
00:43:40.074 --> 00:43:52.054
Look, if you're using a Bitcoin-backed lender, do yourself a favor and check out Ligos Finance before you take on the risk of counterparties that may be playing funny games with your coins.
240
00:43:52.554 --> 00:43:56.454
With Ligos Finance, you don't have to worry about that. They use Bitcoin native smart contracts.
241
00:43:56.494 --> 00:44:07.074
And for that reason, they are our preferred lender here at Blockspace because with those contracts, you always hold the keys to your Bitcoin when you are using a loan with Ligos Finance. You heard that right.
242
00:44:07.094 --> 00:44:19.844
With Ligos Finance, you always hold your coins, self-custody, when you are using their Bitcoin-backed lending program. There's no rehypothecation, no bridging, no wrapping, no funny business.
243
00:44:19.934 --> 00:44:35.734
Just competitive rates as low as 8.5 to 10% APR. Go to ligos.finance to learn more. All right, Charlie, you gave us a good overview of the, specifically the political concentration risk of some of these- Yeah...
244
00:44:35.814 --> 00:44:48.154
companies earlier. What else have we got on Dimetrics that's showing us a little bit of a- Yeah, well-... trend going on? I think it's probably worth a, a stab at taking a look at what the, what, what are the GPUs doing.
245
00:44:48.294 --> 00:45:06.174
So, um, pulling up the compute, uh, pricing, spot pricing for GPU data here. We got A100s up on the day. H100s up on the day 4.5%. H200s down 1.5%, and B200s up 1.5%.
246
00:45:06.554 --> 00:45:23.114
A100s renting at 128 an hour, which by the way is, is nuts to me. Um, because just, um, if I were to zoom in here, just last, in, in June, they were in the high 80s, low 90s.
247
00:45:23.254 --> 00:45:32.354
And in April, they were renting out at $.70, so 70 cents, uh, per GPU hour.
248
00:45:32.374 --> 00:45:48.994
And now this nearly s- this six-plus-year-old, uh, card is renting out at $1.28 an hour. We've said it before- Kind of nuts... but GPU depreciation bears an absolute shambles. Absolute shambles.
249
00:45:49.054 --> 00:45:59.464
The other one, which is a doozy anyway- Can someone check in on Ed Zitron please? [laughs] Yeah. See what he has to say about this. Someone check in on, uh, who's, uh- Michael Burry... Michael Burry?
250
00:45:59.664 --> 00:46:08.394
Yeah, [laughs] Michael Burry. And Michael Burry. I think Michael Burry flipped bullish. I think he reconsidered his priors. Oh, no. Oh, it's over now. Yeah. [laughs] Well, it's over.
251
00:46:08.494 --> 00:46:20.994
[laughs] Well, don't, don't quote me on that. Uh, but I mean, look at this, look at this, uh, B200 chart here. Um, this is kinda nuts because if we go B200s, April of this past year, $3.80 per GPU hour.
252
00:46:21.034 --> 00:46:37.354
Now over, now double that, 7 bucks 30 to rent a, a B200. Um, actually peaked in July at, uh, $8.38. Um, yeah, GPU spot prices be wildin', man.
253
00:46:37.634 --> 00:46:42.134
Uh, they're, uh, all over the place, but up and to the right.
254
00:46:42.594 --> 00:46:54.794
And, and mind you, this is while Neocloud's pulled back while, uh, during the month, uh, during the end of June and July as Leopold Astrenburner got liquidated. But now the- Well, and- Yeah.
255
00:46:54.854 --> 00:47:05.514
And, and something to point, uh, to kinda just con- contextualize that too, Neoclouds obviously have exposure to these spot prices, but only so much as their contracts allow.
256
00:47:05.574 --> 00:47:18.313
Like, I believe CoreWeave and their Q2 update said 98% of their offtake agreements are for multiple years into the future, and only 2% are actually exposed to s- the spot market.
257
00:47:19.514 --> 00:47:23.514
Uh, a- and you know, obviously they're gonna wanna take those multiple year offtake agreements.
258
00:47:24.214 --> 00:47:35.114
But you do wonder, with prices marching up and to the right, how much do they, do they weigh getting more into the spot rental market over the course of the next few years?
259
00:47:35.894 --> 00:47:46.894
I, I would imagine that they'll probably prioritize those longer term g- agreements 'cause it really helps with credit risk, and it helps them, uh, engage with, with banks for more favorable terms for financing.
260
00:47:47.794 --> 00:47:56.464
But it, it does just, it, it does raise, raise the question, you know, how much of a spot mix do you want right now with stuff ripping?
261
00:47:56.994 --> 00:48:07.034
Or are we about to get a local top, and then we're gonna see this drop over the next few, few months? 'Cause it seem, it's really hot right now. And- It's really hot, but I- Go ahead.
262
00:48:07.194 --> 00:48:21.804
I wonder if I can pull up compute indexes, uh, GPU forwards. I don't think I can here live. But like, uh, GPU forwards are c- still in Contango. Um, you pay more for a GPU in six months than you do today.
263
00:48:21.824 --> 00:48:30.974
That's pretty incredible. So I, I would say like, I'm bullish GPU for the next- The bulls you're marching on. Yeah. So that's, that's what I would say.
264
00:48:32.494 --> 00:48:42.794
Um, we could probably take a stab at like another chart on here if you, unless you have a point you wanna make. Um, not really. I just wanna, you know, just a quick shout-out and chill.
265
00:48:43.794 --> 00:48:55.224
The, the DI Metrix dashboards beau- I mean, look at those charts, man. Absolutely beautiful. A- Yeah... incredible- Incredible resource. And getting better by the day.
266
00:48:55.304 --> 00:49:00.984
Brandon Bailey's really putting in some hours to improving it, uh, a little bit every single day, so.
267
00:49:01.024 --> 00:49:11.644
Yeah, we've done a couple interviews with Brandon Bailey, the spearheaded, the spearhead guy behind this dashboard, so strongly recommend everybody check those out.
268
00:49:11.844 --> 00:49:26.164
Yeah, and check out, if you're interested in learning more, signing up, seeing what the plans are like, dimetrics.ai. That's D-I-M-E-T-R-I-C-S dot A-I. Importantly, you get the dashboard, which is great.
269
00:49:26.544 --> 00:49:34.744
You also get the MCP for it, so you can plug it into your LLM of choice and draw directly from the database when you're doing your research.
270
00:49:36.364 --> 00:49:44.404
Um, I think maybe the last, uh, chart here, and this is sorry to the, the podcast listeners here, um, this is a pretty high level view.
271
00:49:44.564 --> 00:49:51.224
I think a lot of times when we talk about, um, the landscape of data- data centers, we don't, like, talk about...
272
00:49:51.264 --> 00:49:59.024
Like, we talk about where they are, what county they're in and what state, but sometimes it helps to actually zoom out at North America and see where these things actually are.
273
00:49:59.054 --> 00:50:15.524
And you really do get a whole story of over 1,000 words in a single picture, which is these data centers are going to Texas, PJM, and smaller ones in Tennessee and Georgia. So, yeah.
274
00:50:16.044 --> 00:50:30.364
The Southeast, Midwest, and Texas will inherit the earth, brother. Yeah, look at that. The, the great, the West, the land of opportunity. Uh, the, basically data centers don't cross the Rockies. [laughs] They go...
275
00:50:30.384 --> 00:50:41.454
Yeah. Well, and, you know, to be fair, most of that is, is kind of hostile for them in the sense that you can't build, you know, 10, nine- 9,000, 10,000 feet or 7,000 feet mile high, um, in the Rockies. Yeah.
276
00:50:41.484 --> 00:50:48.544
That being said, I will, uh, one thing I will say is these, a lot of these are Bitcoin miners, formerly Bitcoin miners, now powered shell operators.
277
00:50:48.984 --> 00:50:56.624
So there are a number of companies that aren't included on this map yet. Like, there are dozens more, if not hundreds more, that are building data centers to power AI.
278
00:50:56.904 --> 00:51:06.684
Yeah, these are really just companies that are in our universe that we cover. Yeah. This doesn't in- this doesn't include data center alley, you know, where the most data centers are. Right.
279
00:51:06.924 --> 00:51:08.384
But the trend- They're in Virginia...
280
00:51:08.424 --> 00:51:22.804
if, if you look at development and the, the arc of development in the United States, the Southeast and sections of the Midwest are leading the way behind Texas, which is undeniably the lo- the state with the most data center development going on- Yeah...
281
00:51:22.824 --> 00:51:38.964
currently. Um, notably, my, the state that I currently reside in, Oregon, which used to have, and maybe still does have one of the highest densities of data centers per capita, uh, is pretty underrepresented here.
282
00:51:39.724 --> 00:51:45.164
I mean, it, it, it's- Yeah... getting impossible to build in the Pacific Northwest, man. Just between- Yeah...
283
00:51:45.224 --> 00:51:57.024
the red tape and the political hostility and the costs too, it, the, you know, it's, I don't really foresee anything happening here. Washington is the only place where you see some development.
284
00:51:57.404 --> 00:51:59.254
Yeah, I mean, you see- California, good luck, man...
285
00:51:59.254 --> 00:52:11.664
a small Bitdeer site and that small Keel site, Moses Lake, which we talked about with Ben last week, um, which they're not, you know, not really a priority for them, as he said on our show the other week. Um,
286
00:52:12.604 --> 00:52:16.664
not a priority to announce a thing there. But yeah, I mean, crazy here.
287
00:52:17.084 --> 00:52:30.844
It really is the story of Texas, and my home state of Oklahoma, which is holding on for dear life so that when Texas secedes, they, uh, they take us with them. [laughs] The Red River rivalry is only extends so far.
288
00:52:31.164 --> 00:52:39.524
[laughs] Yeah. So, uh, thanks for listening to the show, y'all. Um, hope you enjoyed it. We do this every single weekday at 1:00 PM Eastern.
289
00:52:39.544 --> 00:52:50.824
We are Compute's daily live show featuring quick hits on AI, data centers, power, land, markets, and emerging technology. If Bitcoin goes back up, gosh darn it, we're stock, talking about Bitcoin again.
290
00:52:51.424 --> 00:52:59.964
Can't wait for the day. If you like the pod- if you like the live stream, make sure to listen to the podcast if you miss a live stream. Anywhere podcasts are found.
291
00:53:00.024 --> 00:53:05.584
Also, hit our website where you can get our newsletter, newsletter.blockspace.media, website at blockspace.media.
292
00:53:05.644 --> 00:53:17.704
Newsletter, uh, hits your inbox every single day, summarizes everything on the live stream and more, plus written content. Make sure you don't miss it.
293
00:53:18.864 --> 00:53:35.324
This show is sponsored by CleanSpark, uh, NASDAQ listed at ticker CLSK. I'm Charlie. I'm Colin. We'll see you tomorrow. [outro music]
