Hut 8’s $5B Illinois Site, Kimi K3 Trained on H200s, GPU Rental Price Contango

Summary

---

Blockspace Podcast Episode Summary

Episode: Hut 8’s $5B Illinois Site, Kimi K3 Trained on H200s, GPU Rental Price Contango
Release Date: July 31, 2026
Hosts: Colin & Charlie Spears
Guests: Mike San Miguel (Luxor), Matt Black (Lygos Finance), Carmen Lee (Compute Exchange)

---

Episode Overview

This episode explores major developments across data center construction roadblocks, surging AI infrastructure demand, the shockwaves of a large-scale Bitcoin hardware wallet hack, and the intricacies of GPU and AI token markets. With several expert guests, the hosts dissect the ongoing standoff around Hut 8’s huge Illinois data center, the realities of supply and pricing in the GPU secondary market, and how export controls on semiconductors are playing out globally — culminating in a detailed look at market dislocations between GPU rentals and AI compute tokens.

---

Key Segments & Insights

1. Bitcoin Mining Market Update

• Timestamps: [02:29–05:32]
• Discussion: The hash rate index shows a slightly softer market for miners with minor negative difficulty adjustments. Bitcoin’s price remains in a depressed cycle (~$62.8k).
• Notable Quote:
  • “The more things change, the more they stay the same, my friend.” — Colin [02:29]
• Insight: Despite a cyberattack looming over wallet security (the cold card hack), fee markets and mining fundamentals remain subdued, with low transaction fees and minimal hash rate volatility.
• Memorable Moment:
  • “And I'm depressed, everyone's depressed.” — Charlie [03:58]

---

2. Hut 8’s $5B Illinois Data Center Stalled

• Timestamps: [06:12–12:40]
• Discussion:
  • Hut 8’s 500MW, $5B Logan Prairie data center project was stymied when the Village of Latham cancelled a critical zoning vote, citing insufficient support. The small town’s (~350 residents) concerns included loss of farmland, local energy prices, and noise. The application remains active but postponed, with the next possible action in September.
  • Economic stakes: Hut 8 proposed annual community benefit payments ($8–11M/yr, which dwarfs the area’s $291K annual revenue) and $65M/yr in property taxes for 30 years.
• Local Process Quagmire: Several layers of annexation and county procedural defects add complexity.
• Power Access Uncertain: No confirmed utility interconnection yet for the required 500MW draw.
• Notable Quotes:
  • “Like buying friends. I mean you have to do this community engagement.” — Colin [08:50]
  • “This will be a great case study of a company who is really proactive with trying to court the community and we'll see how the community responds.” — Charlie [12:41]

---

3. GPU Market Trends & Hardware Depreciation

Guest: Mike San Miguel (Luxor)

• Timestamps: [14:01–28:46]
• Major Points:
  • Nvidia’s Vera Rubin Rack-Scale Design:
    • Represents a significant shift toward integrated, ARM-based, rack-scale reference architectures, maximizing efficiency and targeting dedicated AI “token factory” use cases.
    • Quote: “Vera Rubin isn't just a faster GPU. It's a rack-scale reference architecture that changes what you buy.” — Colin [14:05]
  • Secondary GPU Market:
    • Prices for older models (A100, H100, H200, B200) are soaring due to tightening supply — H100s jumped from $150–$170k up to $260–$280k/node in just months.
    • Quote: “Every single index year to date is just up into the right; A100 is up 20%... H100 is up 25%...” — Colin, citing Carmen Lee [18:03]
    • Even “outdated” GPUs (e.g., A100s) find near-100% utilization due to lingering software support and strong demand for less cutting-edge workloads.
  • Lifecycle & Depreciation:
    • Analysts’ typical 5-year depreciation timelines are being challenged — hardware may be productive for 8–10 years for certain inferencing tasks.
    • End-user needs vary: “...most workloads for inferencing are probably not going to be ultra latency specific. It’s going to be more of a conversation about, ‘How do I get the most affordable tokens out that have the performance needs that I need?’ ...[so] five to seven years, that’s in line.” — Mike [23:31]
    • Differences in depreciation approach exist within hyperscalers (Meta, Amazon) depending on their custom ASICS and models.
  • Market Outlook:
    • Near-term: Prices may keep climbing for immediate hardware needs due to Nvidia lead times and near-total pre-allocation of new stock to the biggest clouds/hyperscalers.
    • Used market likely to plateau only when profitability on secondary nodes is exhausted.

---

4. Cold Card Bitcoin Hardware Wallet Hack

Guest: Matt Black (Lygos Finance)

• Timestamps: [30:00–45:55]
• Incident Summary:
  • A critical vulnerability was discovered in the seed/entropy generation for Cold Card hardware wallets (MK2–MK5 lines since 2021), enabling private key extraction. Over 1,000 BTC ($65M) swept—an attack enabled in part by advanced AI assistance.
  • “...if you have a Cold Card device, MK2, MK3, you generated that seed anytime 2021 or later ... you should be looking at moving your funds immediately.” — Matt [30:34]
  • Issue stemmed from a code bug related to random number fallback logic; reduced entropy makes brute-force attacks feasible.
  • The vulnerability was accidentally introduced when the Cold Card team moved away from earlier open-source code—ironically closing community oversight that might have caught the flaw.
• Industry Implications:
  • Larger debate is brewing about the future of self-custody, open source, and the rising capabilities of AI to rapidly find vulnerabilities.
  • The incident re-ignites arguments for multi-signature wallets and proactive Bitcoin protocol-level security features ("Covenants").
• Memorable Quotes:
  • “The irony is, you know, that’s really biting a lot of people in the… ass today.” — Matt [33:11]
  • “...anyone could have pointed an AI, you know, AI model at that... and found the vulnerability in no time on the MK3 code base.” — Matt [40:02]

---

5. US-China AI Export Controls & Kimi K3 Training Debate

• Timestamps: [47:21–56:57]
• Background:
  • Bloomberg reported Moonshot’s Kimi K3 was trained on up to 20,000 Nvidia H200 Hopper chips supplied by Alibaba—raising questions about the real effectiveness of US export bans.
  • U.S. policy: Blackwell chips are outright banned; H200s can be exported to China with White House approval. But “GPU rental” and third-country cloud workarounds (e.g. Thailand, Indonesia) allow practical sidestepping.
  • Both Washington and Beijing are posturing: US wants to keep cutting-edge silicon away from China, while China claims to encourage domestic compute development.
• Industry Reality:
  • Growing skepticism that these controls are effective, especially when model access is governed by API rather than through physical hardware.
  • AI security researchers already rely on Chinese models (like Kimi) when US models are restricted for “white hat” auditing.
• Notable Quotes:
  • “What even are these export bans functionally doing? ...kind of almost like the days of trading crypto with VPN on Binance... you weren’t actually prevented.” — Charlie [51:40]
  • “If the angle is to make sure China is always six months behind, it’s not working... K3 deployment right now... the most efficient chips deploy K3 is in the Nvidia chips.” — Carmen Lee [57:30]

---

6. Market Dislocation: GPU Rentals vs. Token Pricing

Guest: Carmen Lee (Compute Exchange)

• Timestamps: [56:57–73:08]

• Key Trends:

  • On-demand GPU rental prices (A100, H100, H200, B200) have risen 15–30% this year, even as AI “token” prices (cost per million output tokens) have dropped (–18% in a month).
  • Forward rental contracts show “contango”—longer-term GPU rental prices are often higher, not lower, than spot/on-demand, defying commodity market expectations.
    • “Sometimes in our world... the forward contracts are in contango rather than backwardation.” — Carmen [63:34]
    • “This is really reason why I think dated futures... different expiration will be precisely needed for the market. Because... you have a term structure risk.” — Carmen [63:56]
  • Bulk buyers are willing to pay a premium just to lock in resources amid persistent supply uncertainty.

• Enterprise Dynamics:

  • True enterprise AI adoption is still early; most demand is proof-of-concept, not at-scale production.
  • Commodity logic: As token (compute) prices fall, more applications and experimentation become viable, possibly increasing total demand (Jevons Paradox).
  • End users, especially application firms, mostly care about token price rather than underlying GPU models—mirroring how electricity consumers focus on the end price, not the power source.

• Product Innovation:

  • Compute Exchange is rolling out token-price forwards to allow enterprises to hedge their costs at the API/service level, matching the true economic exposure of deploying AI models at scale.
  • Key product challenge: Need for standardized benchmarking and SLAs to ensure contract and price integrity.

---

Notable Quotes & Highlights

• On GPU Market:

  • “We're seeing older generation GPUs at close to 100% utilization. So this isn't going to be a rip and replace of the whole infrastructure.” — Mike San Miguel [17:09]
  • “The H100 life cycle is probably 8 to 10 years out.” — Mike, citing Coreweave CFO [23:31]

• On Bitcoin Security:

  • “It's been a thousand bitcoin that's been affected ... that's just from the folks that are the most at risk... But now, and now it's a free for all for getting the rest of them.” — Matt Black [33:32]
  • “Anyone can point their model at something and find a vulnerability. ...Now everyone's going to be running a security analysis on every piece of code before they go and put that out in the public. So I don't know. That feels like the new paradigm to me.” — Matt Black [40:02]

• On US-China Tech Tensions:

  • “Physical location on that note of these Alibaba GPUs is unclear and, and that matters legally because the rent restricted chip offshore workaround hinges entirely on whether where the hardware physically sits.” — Colin [53:16]

• On Market Contango:

  • "This is wild to me. ...Nothing crystallizes the everyone is short compute narrative more than that." — Colin [73:35–73:38]

---

Timestamps for Major Segments

• Hashrate & Mining Market Update: [02:29–05:32]
• Hut 8 Data Center Standoff: [06:12–12:40]
• GPU Trends & Depreciation ft. Mike San Miguel: [14:01–28:46]
• Bitcoin Cold Card Hack ft. Matt Black: [30:00–45:55]
• US-China GPU Controls, Kimi K3 Training: [47:21–56:57]
• GPU vs. Token Market, Contango Explained ft. Carmen Lee: [56:57–73:08]

---

Conclusion

A rich, fast-moving episode spanning the state of Bitcoin infrastructure (and its risks), a behind-the-scenes look at the intersection of regulatory policy and compute supply, and the unforeseen but telling economics of the compute market as AI demand explodes. From stymied $5B rural data centers to GPU rack form factors and the wild west of crypto-security and AI model exports, the big takeaway is: infrastructure, risk management, and market structure are anything but settled.

Find more Blockspace content at Blockspace Media.

---

Transcript

[Mike San Miguel]

What's going on, y'?

[Matt Black]

All?

[Colin]

Welcome back to Blockspace Live presented by Clean Spark. Packed docket today, specifically with guests. We will kick off with A Story of Hut 8 Logan Prairie Five Billion Dollar Data Center Build hitting a stalling point with a local commission. Following that, we have Michael San Miguel of Luxor on to talk about the GPU secondary market and what they are seeing from their GPU compute desktop as we see rental rates continue to skyrocket, Charlie, and demand for models as old as the A100. After that, we will have Matt Black on to talk about the recent sweep of funds from cold cards. I saw figures anywhere from 600 to 1,000 bitcoin stolen in this Charlie. So taking a little bit of a divergence from the compute route, but we felt like this was an absolutely necessary thing to cover. So Matt Black, co founder of Lygos Finance, is going to give us a breakdown and our fearless co host Charlie Spears will be furnishing some hot takes and questions about this incident. After Matt Black, we have a note on how Kimmy K3 was trained, reportedly trained on 20,000 Nvidia hoppers and what this means for the current export control regime. We're going to untangle some of the differences between what's actually banned for Chinese companies and where they have wiggle room. And then we will have Carmen Lee of Compute Exchange on to talk about compute forwards, rental prices and the token forwards that they are working on over at Compute Exchange.

[Charlie Spears]

That's right. Blockspace goes live every single weekday at 1pm Eastern. We are Compute's daily livestream, featuring quick hits on AI data centers, emerging technology and markets. If you like the live stream this turns into a podcast. You can find that anywhere podcasts are found. And if you can't get enough of the podcast or the live stream, you can continue discovering more of the Blockspace content powerhouse at our website, Blockspace Media. The URL is Blockspace Media. Head there. We have a lot of written content too. This show is brought to you by CleanSpark. Nasdaq listed ticker CLSK. More on CleanSpark listed later on in the show. Before we go to the news, Kong, gotta kick it off. Little hash rate index.

[Colin]

Gotta do a little hash rate index update from Luxor here. And you know what, Charlie? The more things change, the more they stay the same, my friend. We're looking at having a roughly negative 7.0.74% difficulty adjustment. We're almost halfway through the current epoch and we're coming off of a negative adjustment of roughly 1% from the last epoch. And this comes at a time, Charlie, where hash price looking a little better. Because of these difficulty adjustments, we're holding the line at $30 bread hash per day. It's currently $31.59 per PETA hash per day. But bitcoin taking a hit right now. Don't want to read too far into the tea leaves. I don't think that this hack with the cold card has breach containment yet. Maybe some folks are selling because of this, but we are starting to see more volatility in the markets. Bitcoin has held up remarkably well since cratering to below 60,000 in recent months. But Bitcoin down to 62,840 hash rate currently stands at 886.91 exahashes and new week same story. Charlie. I don't see much wiggle room for hash rate right now with 4 CP season in full swing, with bitcoin's price still relatively depressed and hash price depressed as a result.

[Charlie Spears]

And I'm depressed, everyone's depressed.

[Colin]

Funnily enough. It takes a tragedy to bring people together. Bitcoin Twitter has been popping off over the last 24 hours with this cold card hack. No one talking about hash rate though, because quite frankly there's not too much to unpack here. Just a very subdued environment. That being said, if you have low enough power cost, high efficiency machines, you're kind of liking what you're seeing right now. There's really nothing.

[Charlie Spears]

Or maybe you're okay with just burning money to keep the lights on because you're a true believer of which there are increasingly few left. But I guess my last comment on hash price is calling no. With this cold card hack, which we will talk about later on in the show, there is cause for people sweeping their funds to move them to more secure places. This could result in sweep transactions, more coin joins, very urgent high fee transactions. However, we really have not seen that on bitcoin yet and it's unclear how much those could those could contribute. The other one I'll put on your radar is due to the exposure of the public keys or the private keys. There could be a scenario. Again, I'm talking gobbledygook to a lot of the miners here, but this is important. Put this on your radar that we could see RBF sniping wars. Now if that happens and it becomes a major theme, we'll cover it because there's very few people who can. But other than that, very quiet. It's too quiet, one might say.

[Colin]

All quiet on the block space front. I like that angle there, though, Charlie. We might see a pop in transaction fees for admittedly a very sad reason, as potential victims try to sweep their funds to new addresses and attackers try to front run that by rbf those transactions. But so far, you know, transaction fees not really doing too much. About 1% of the total of total block rewards in the last 24 hours. Okay, let's go ahead and move on to our first story of the day.

[Charlie Spears]

Charlie, let's do a quick story because we got multiple guests here today.

[Colin]

Yeah. So this is an update on a Data center for Hut 8 that we've covered in the past. They have had. They've been stonewalled by this local council. This is hut 8,500megawatt Illinois data center that was having some trouble with approving. Earlier in the year. We covered the local council's meeting on the data center. And right now there has been a cancellation for the vote for zoning this data center. That doesn't mean that it is dead. I want to get that out in front of the story, but it is stalled. And I believe the earliest time that it might be reconsidered is in September. So we've got about a month or so before this might re enter the docket for zoning approval. But details are as such. Hut 8's proposed 500 megawatt data center, $5 billion projected spend for the Logan Prairie project, hit another permitting wall when the village of Latham, Illinois canceled an Aug. 3 vote on annexing and rezoning the site. The officials pulled the vote because they determined the proposal lacked the votes to pass. Again, crucially, the application has not been withdrawn. That application was filed by Logan Prairie Data Center LLC, a subsidiary of HUD 8. The attorney. The village attorneys called it postponed with a new date, tbd. So it's stalled, not dead. What's really interesting about this to me, Charlie, is the fact of the matter that there are very few people in this township, roughly 350 residents, and that their complaints were enough to stonewall this data center in a trend that we have covered extensively here at Block Space. The local pushback to some of these sites and the residents have concerns which don't override, according to them, the the perceived benefits that Hut 8 is pitching for their locality. So Hut 8 says that they plan to spend roughly 8 to 11 million dollars per year for 15 years for community benefit programs. That's against the annual revenue of this township of $291,000. So anywhere from.

[Charlie Spears]

Get that bag. Yeah, I mean, get that they do it. I mean, look, what is it? You're $5 billion project. I think the size of this is what Asher or Hudd estimates. So if you, if it takes you under eight figures to, to make the county like you do that, you know,

[Colin]

like buying friends, I mean you have to do this community engagement. And Hut 8's actually been one of the standard bearers of this and their deals. They've made a huge point of trying to educate and engage the community and making legitimate financial stakes in the community to show that they are trying to be a good partner to these areas. But not only is there that 8 to 11 million dollars per year contingent, Charlie, hut 8 estimates that the property could generate 65 million dollars a year on average in property taxes for a projected 30 years. I guess that's the lifespan they're putting on this data center. Or at least that's how far they see it moving out into the future or in the midterm.

[Mike San Miguel]

Right.

[Colin]

So you're talking about 75/billion dollars that will enter this community of 350 people. And that's real money that they could use for enriching schools in the area, for infrastructure, for parks, all of these things. But that's falling on deaf ears because there are a few key worries that the locality has. First of all, or first up, they're really worried about Hut 8 gobbling up farmland. That's one of the things that was mentioned in this. There was less of a worry about water, but they're worried about farmland being occupied. They're worried about energy prices rising and they're worried about noise. So that those were the key pushbacks on this site and a few other things holding it up. There's an annexation dependency. The benefit agreement is contended on Latham annexing not just the project property, but intervening parcels whose owners would each separately have to petition for annexation. So there's like a multi party consent problem that's at the center of this too. There's also problems with a parallel county in terms of a data center moratorium. Logan county passed a 12 month data center moratorium on May 19th. But the county state's attorney later said that the moratorium was invalid because it skipped the required zoning process. The county officials also found technical defects in Hut 8's application. According to the reporting, Hut 8 fixes and resubmits, county hearings can't happen before December. So again there's no meeting on the books for this. But the earliest we could see any sort of re upping of this zoning push would be in September. There's also a question about the power that was needed for this site. The 500 megawatts would draw from a local substation. But the utility, Amerin nor Miso has publicly confirmed an approved interconnection. So according to the reporting that we saw, the power is not even guaranteed for this site yet. 500 megawatts, big chunk, especially for some of these rural areas. I would imagine that was probably the largest consideration with the local council with regards to whether or not they wanted to push zoning through. They might ask Cut eight, hey, you need to actually have interconnection approval before we can take this seriously. That's just speculation on my part though I'm not sure if that those conversations are underway. But that is the TLDR for this news story. The reason we bring it up. This site in Illinois is the largest new site in Hut 8's pipeline. It's the only one that they have disclosed or has been disclosed. The reporting is being worked on and it's equal to the phase, 500 megawatt phase that they are working on to extend Beacon Point. So a pretty notable piece, piece of the pipeline.

[Matt Black]

Yep.

[Charlie Spears]

Not a lot of comms except this will be a great case study of a company who is really proactive with trying to court the community and we'll see how the community responds. So because we've seen the up, we've seen the other side of being surprised by community pushback for good and bad reasons. So we'll eyes on this one before. So we're going to keep on going. We got Mike San Miguel from Luxor in the wings. We'll bring him on up here. We'll talk about GPUs and GPU racks specifically. Before we bring Mike up, a word from our sponsor, CleanSpark.

[Colin]

We are CleanSpark, America's Bitcoin miner. A publicly traded company with the largest

[Matt Black]

operating hash rate powered entirely by self

[Colin]

operated infrastructure infrastructure across four states. This is our proof of work and we are setting the standard for what's next. Learn more about the intersection of energy and [e-mail] Alrighty, we've got Mike San Miguel, the senior business development manager at Luxor's hardware desk. We are going to get him on the stage. Mike, welcome to the show sir. Thanks for joining.

[Mike San Miguel]

Hey Charlie and Colin, always a pleasure. Good to see y'.

[Matt Black]

All.

[Charlie Spears]

Likewise.

[Colin]

Good to see you too man. So I want to kick off today's segment with a piece that you co published with Ian Philpott on hash Rate index. Specifically Nvidia's Vera Rubin. Why the Rack is the new GPU the excerpt here, Nvidia's Vera Rubin isn't just a faster gpu. It's a rack scale reference architecture that changes what you buy. And so my first question with this, can you break down how this form factor is changing the consideration for how we view GPU rollouts and also how much demand are we seeing from buyers on the compute side of things for this new form factor?

[Mike San Miguel]

Gotcha. Yeah, yeah, great question. So by and large there's a couple trends that factored into this rack scale rollout and sort of this pivot towards rack scale. So it's not a new thing. In Nvidia's portfolio. Within the current Blackwell generation, there's the GB300, before that the GB200, which is another rack scale solution. The technical trend is that they started integrating ARM based processors into these. So as Opposed to your x86, your intel and AMD Prox for your CPU side, these are using the ARM based processors. It hasn't matured quite, you know, before to the point it is now. That has to do with software compatibility. But a lot more of these models are being made to run on the ARM based procs and servers. The, the other trend is a financial model. Right. So at GTC San Jose this year, Jensen announced sort of this move more towards the, the token factory. That's really what these are, are by and large designed to do. So you've got the first name Vera in the new generation, which is the cpu, the ARM cpu, Rubin, which is the gpu. The next thing with Vera Rubin is that they're integrating an AI accelerator and AI ASIC specifically for token generation on the inferencing side. So those are the two trends that are sort of pushing the way for these probably won't take over the whole portfolio. You know, x86 platforms are probably not going to go anywhere, but that's what's pushing this, this trend.

[Colin]

It seems to me like with each new advancement you can have incredible efficiency gains. But a lot of these neo clouds and powered shells are outfitting their data centers for a very specific architecture. And it's not very easy to just backwards make them backwards compatible.

[Matt Black]

Right.

[Colin]

And you have billions of dollars invested in these things, but ultimately you can't just rip all of that out without having a huge capital expense.

[Mike San Miguel]

Yeah, well, and especially for the Vera Rubin stacks. So not to get too technical, but there's essentially five different racks that go into these. It's 80 different manufacturers, the, the weight of these is completely, you know, a new scale that we, we haven't seen before, even in the previous Grace Blackwells. So just the cement slabs, the floors that these have to roll on has to be really built up. And the, the management for this even goes down to the grid level for managing power. So it's, it's a new ball game. But it, you know, to your point, it's not gonna replace previous installs. You know, eventually those will migrate down stream. But currently, you know, we're seeing older generation GPUs at close to 100% utilization. So this isn't going to be a rip and replace of the whole infrastructure.

[Colin]

I'm glad you mentioned that, Mike, because that kind of leads well into my next question. And I have a tweet here from Carmen Lee, who will be joining us at the end of the show. She says here, quote, we've seen 15 to 30% price increases year to date across our A100, H100, H200 and B200 Neo Cloud on demand indices. And if you look at this, every single indices year to date is just up into the right A100 is up 20% year to date, H100 is up 25%, B200's up nearly 30% and GPU index H200 is up 14.4%. Supply is clearly very tight right now. So I'm curious with regards to what you are seeing on the compute desk and in the secondary market, how much demand are we really seeing for those older models? And also how much is this pushing consumers down the curve towards, you know, little brother AMD rather than getting their hands on the latest from Nvidia.

[Mike San Miguel]

Yeah, great question. So on the, the used front we're seeing a lot of movement there. So H100 back in April price point depending on the configuration in the host, probably anywhere from about 150 to 170k per node in US dollars. Now they're moving similar specs probably 260 to 280 and that's a good deal. Occasionally you'll find new old stock that for some reason hadn't been deployed that are almost 300k for an H100. So it's all over the place. H200 is even rarer, but low to mid-300s per HGX node. And then for the A1 hundreds those are hitting the market a little bit more probably in the 70s to 90s depending on the specification per hardware node.

[Charlie Spears]

So I'm thinking about like GPU depreciation cycles. And maybe I haven't been dialed into high computer at this level for a while. But am I correct in assuming that previous upgrade cycles have not required full stack replacements of the entire node, the entire system? And then with the migration to the rack we have this logistics challenge I imagine of when we want to upgrade. Now we have to pull the entire rack or do we? And in light of this one, am I right about this? And two, what does this mean for like GPU depreciation or like the upgrade cycle of this high performance compute?

[Mike San Miguel]

Yeah, yeah, good question. So there's a couple things that I think are happening here. Number one, we're starting to see software users, AI software users actually starting to understand what they actually need. This is sort of a broader trend that was happening with you know, data data center IT infrastructure before then where most users aren't actually consuming all the hardware that they buy or that they rent. I spent seven, eight years at Dell running sizing on companies data center infrastructure and by and large their compute consumption was maybe 10% of the CPU that they had in there. So not all inferencing is built equally. And I think that's what is really happened with the secondary market is that users are seeing that they can get more bang for their buck with a older gpu. The other thing that usually drives hardware refresh cycles has to do with the software compatibility. So that will probably push more of these to phase out rather than any actual demand. Now if you're doing like a full AI factory and not just bare metal nodes, everything within that AI factory reference architecture has to be compatible. So that would be more of a rip and replace. The shared storage, the networking, the management nodes, head nodes, GPU servers, backend networking, all that has to be compatible. So when you go to upgrade that you're going to have to rip and replace.

[Colin]

Mike kind of building on that. With regards to depreciation, it's one of the key tension points when analysts and investors look at some of these NEO clouds and hyperscalers and anyone running this compute. I'm curious your thoughts on whether or not you think the depreciation schedules roughly five years. If that's too aggressive for what we're seeing right now. And it's kind of hard to say because ultimately even a 100, the oldest model that's still in the conversation is like what just from like 2022, 2023. So we don't even really have enough data yet I think to have a clear answer to this question. But I would love your opinion on whether or not you think that five year average that analysts typically use is too aggressive or too conservative. Yeah, I agree with you that there's

[Mike San Miguel]

probably not enough data to really feel 100% one way or another. You know, if what we've seen so far continues, then it's definitely not accurate. You know, earlier this year GTC CFO of Coreweave said that H100 life cycle is probably 8 to 10 years out. You know something that, that I saw in more traditional hardware infrastructure is that people run hardware until wheels fall off and the software stops working. Most workloads for inferencing are probably not going to be ultra latency specific need the latest and greatest for every single workload. It's going to be more of a conversation about how do I get the most affordable tokens out that have the performance needs that I need rather than how do I always rent a Ferrari? And to that effect I do think that probably five to seven years that's in line with pre LLM hardware. This is just another workload. It's a little bit more complex. It probably won't be that different going forward

[Charlie Spears]

I guess. Last question is recently again on the GPU life cycle. Amazon shortened their GPU useful life cycle predictions I believe and meta extended them. And there seems to be like disagreement across the industry. Chime in here. How do you reconcile everybody kind of searching for their own depreciation life cycle timelines?

[Mike San Miguel]

It comes down to use case which I know is kind of a great non answer to me. I would be perfectly fine having a 1950s Land Rover Series 1, but I'm not rock crawling up the side of Denali anytime soon. Getting around New Orleans is just fine and something nice and old. Now if I'm a serious rock crawler I'm going to get the latest and greatest. It's no different with these GPUs and what they're being used for. The other thing to understand with those hyperscalers is that they've been developing their own dedicated training and inferencing ASICS in house for the past 10 years and they've been in production for about as long. So you know their use case and the way that they utilize their GPUs, their ASICs, all of their AI hardware is very different from the way you know, a typical NEO cloud or a minor converting will use them. I don't necessarily think that, you know, it's, it's a apples to apples comparison. There's

[Colin]

Matt one one or two more questions from me. I'm asking you to look into the crystal ball here. So please bear with me and. Sorry, Mike, I. I'm scrambled today. You got too many M's on the show.

[Charlie Spears]

We got a lot of M's on the show today.

[Colin]

So quick question with regards to where you see hardware prices throughout the rest of the year, given the demand you're seeing, given the deal flow that we're seeing on the operator side, in terms of CapEx, do you expect prices to compress, expand or flatline throughout the rest of the year?

[Mike San Miguel]

I believe that due to Nvidia lead times, you're going to see higher prices for hardware needed immediately. I think you're going to pay a premium for what's immediately available. And if you're willing to wait six months, which is the average lead time right now for a bulk amount of B3 hundreds from any Nvidia certified OEM, you know those prices will probably flatten a bit, especially as people wait to see how Vera Rubin actually gets deployed. Currently Vera Rubin is in the testing, prototype stage, marketing starting to go out to the OEMs. These things haven't hit the open market, so we don't really know what demand is going to look like for Vera Rubin. There'll probably be a market hiccup when Vera Rubin finally launches. But keep in mind, any of those newest generation chips, they're already pretty much spoken for, right? Your hyperscalers are going to get them in the first six months,

[Colin]

maybe, you

[Mike San Miguel]

know, some, some big Neo clouds after then and it'll probably hit the public in about 12 months or eight depending on how friendly you get with your Nvidia rep. Now with used market, if there's anything I've seen it keeps going up, I think there will be a break even point around profitability with those and that's going to cause a flat line. Every time I see it, I think we're there, but they keep going up.

[Colin]

Well, we'll be keeping an eye on that. Have to have an update towards the end of the year, see where things are at. Mike, thank you for joining, man. Really appreciate it and love the insights on hash rate index. Keep it up and have a great

[Mike San Miguel]

weekend y' all too. Thanks for the time.

[Charlie Spears]

Thank you. All right, we're going to roll on and we haven't totally abandoned it. We're going to have a bitcoin conversation because it's all hands on deck for the cold card hat. We have Matt Black in the audience. We'll bring him on up here right after a word from our sponsor, Luxor.

[Colin]

This episode of Blockspace Live is brought to you by Luxors Commander Bitcoin miner management software for enterprise operations. Luxor's commander gives you real time fleet monitoring, bulk remote commands across your fleet and intelligent miner 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. In fact, ercot backtest show 10% improved profitability with intelligent mining over binary mining. And you need every single SAT with hash price compressed at where it is now. Commander Pro is $100 per megawatt or a 25 basis point pool fee adder. But you can also try it free for 60 days. So if you'd like to learn more, go to Luxor tech forward slash commander.

[Charlie Spears]

All right, let's talk bitcoin about time and under duress. Do we do this time? I think let's bring up Matt Black. Matt, welcome to the show.

[Matt Black]

Thank you gents. Thanks for having me. My goodness, this is, this is crazy to see what's going on. I can't, I can't believe it.

[Charlie Spears]

I, yeah, it's Matt. What explained to me what is going on again? Like our audience has kind of shifted to more data center REITs. AI folks, what is roiling the bitcoin ecosystem right now?

[Matt Black]

So there was, for those that haven't been listening, there was a vulnerability found in the seed and entropy generation of cold card devices. MK2, MK3. First off, if you're listening right now and you have a cold card device, mk2, mk3, you generated that seed anytime 20, 21 or later. And actually the later devices too are affected. You should be looking at moving your funds immediately. The entropy is the process of, you know, how does bitcoin work? Bitcoin has private keys. It's a random number in a very large search space. But if you don't generate that random number properly, someone can find your private keys and they can take your bitcoin. So, so yeah, that's what's going on.

[Charlie Spears]

You know, maybe talk through like the implications of this because I, I don't think the average person realizes like that it's not bitcoin that broke, but it's like a major piece of hardware from like a prominent manufacturer.

[Colin]

And the implications, and not just prominent, but like gold stamp Bitcoin Maximus.

[Charlie Spears]

We've given out cold cards to listeners

[Matt Black]

so, well, you know, and I've, I've gone and I've recommended cold card to so many, so many folks over the years. It's like, okay, well, you Know, I want to go and I want, I'm buying bitcoin and I want to go and I want to self custody. How am I going to go do that? The recommendation has always been cold card. It's always been, hey, like coldcard is the gold standard. You know, you could, you could you generate your seat on it, you use, you can have the device so that it's, you know, not even, not even, you know, attached to your computer. It's air gapped. And it's just really sad to see that this is what happened. And so I don't know, like for me, you know, I've always told, you know, and I guess so people understand a little bit more like the actual issue was just one line of code, right? It was one line of code that was affected which was basically like, hey, how are we actually generating the seed? And for, you know, and I think this is a larger discussion too, which is like, okay, what is the right way to actually go and secure your bitcoin? Like what is the, the safest way to do that? It's always been multi sig is too complicated. You should use a single sig. Well, the irony is, you know, that's really biting a lot of people in the, you know, in the, in the

[Charlie Spears]

butt Today you can say ass because we're getting, yeah, it's, it's rough today. Okay, before we get into like the existential reflection on the ecosystem, let's touch a few more points about what we know ballpark. What is the scale of the attack and is it ongoing? Toss it to you.

[Matt Black]

Yeah, so far I, so what I've seen so far is that, and just looking at the mempool, it's been, I at least a thousand bitcoin that's been affected. What is that? You know, $65 million in fiat dollars we've got. And, and that's, that's just the start because now, you know, Pandora's box is open. So that was likely folks that had MK2s, MK3s, which were the most vulnerable. So generally when you generate a seed, you need a certain amount of entropy. And you know, if you have a 12 word phrase, that's usually 128 bits. If you have a 24 word phrase, that's 256 bits. But the problem there is what the issue with the cold card specifically was. There was a fallback. So they didn't realize that one of the values in the code was basically just checking does this value exist? Not is it zero? And so it ended up Having a fallback, which means the difference between having 128 bits of randomness of entropy and 45 bits is, you know, forever and never being able to find it and you're screwed. So, so it's been a thousand bitcoin that's been affected and that's just from the folks that are the most at risk. But now I think it's likely going to be a situation where we have, you know, the folks that have Mk4s, Mk5s, well, they probably haven't got all the Bitcoin of the MK3s, to be honest. Right. Like if you look at the blockchain, there's, there's people who had like half of their wallet stolen. Well how does that happen? Well, it's probably because it was someone using AI that was like trying to, trying to grab those funds and they only grab, they only grab this first x number of UTXOs and, and grab them. So now, and now it's a free for all for, for getting the rest of them. But the thing is it doesn't just affect people who just did a, you know, a single sig wallet. You know, if you set up a, say, say you set up a cold card multi sig, right. You maybe you did two of three, you had three cold cards, you have two of three and you set this up in 2022. Well now you're affected and now people are going to be looking at the blockchain to see you spending. So you know, it's, it's a really, really tough time. We haven't seen the end of where this is going to go and how many people are going to be affected.

[Colin]

So Matt, a quick question for me. I saw a lot of jawing on social media about this may have surfaced and I'm not as plugged into this as Charlie, so just caveat up front. This may have surfaced from the fact that Cold Card used to base some of their design off of foundations model and then they came under a lot of flack for that and then. Well, I mean allegedly, but they came under a lot of fun.

[Charlie Spears]

It wasn't. It was. I mean, declare. I'll clarify. Yeah, use some open source licensing from Cold Card. It caused the whole thing. I don't know if you want to speak to that Matt, but it, yeah,

[Colin]

but, but my question is they changed something when all of this came to a head and some people were insinuating that this may have been the root cause of the flaw that we're seeing here. Can you speak to that at all? Am I just totally off base or misunderstanding?

[Matt Black]

So, so my understanding was that in 2020, you know, foundation Devices was basically, you know, they're just getting off the ground. And originally the licenses of ColdCard were open source. And so obviously, you know, MIT license, you're able, you know, with attribution to the original author, you're able to use that code for commercial purposes. So that's, you know, that's what foundation did. And basically Cold Card coinkite was not happy about that. They were not happy that someone else was basically taking their code and then going and making money, money off of it. And so, you know, around 2021, they started having the conversation of, hey, we're gonna go and, you know, rewrite, you know, large parts of our code base in order to, and then put it under a different license, such that people weren't able to go and, you know, use that for commercial purposes.

[Colin]

And then they closed, forced everything. Right after that, Cold Card, no, it was still.

[Matt Black]

What, what's it, what is it? What's the term? Open viewable or like, it's where you can see the code but you can't like, use it. What's the term? Charlie, do you remember?

[Charlie Spears]

Oh, I don't know. But yeah, this is, we're, we're going to be too deep down the rabbit hole here. But. Yes, I know, yeah, but, but basically

[Matt Black]

like, so, and, and as part of those changes in 2021, they ended up like introducing this vulnerability. Whereas ironically, if they had just stayed like, hey, we're going to be open source and yeah, people are going to take our code and they're going to improve on it or they're going to change it. If they had been both using the same source code, well, it's likely the other company would have noticed a bug of this nature, which is kind of like the whole point of open source. So I don't know. There's a lot of contention now also in terms of the time of AI. Right. Because now, you know, it brings up the question of like, hey, we have open source, but you know, someone who's vibe coding can find vulnerabilities and, and hack things.

[Colin]

Yeah, and just a quick mea culpa for me, they're totally flipped. The relationship came out. Yeah, I appreciate that, Charlie, but I was just curious whether or not how much of this could be attributable to that spat. Because of the changes Cold Card made,

[Charlie Spears]

it certainly feels somewhat. We're still in the discovery period of this. There were spaces last night going well into the Wee hours and information is still coming out. Matt, you talked about AI and I actually think this is a major role to play because everybody who talks about AI and security and E cryptography knows that with meet those level capabilities and now with Kimmy being cheap and less constrained, we're in a new era for AI vulnerabilities or even like script kiddies, people who are maybe less experienced but maybe wildcatting able to start exploiting things. What does this mean? Is this the beginning? Is just this first shot across the bow? What does this mean for like the landscape of bitcoin and self custody?

[Matt Black]

Well, I think the like the reality is that it's just going to be that much easier for folks to be able, you know, your average vibe coder to be able to sit there, you know, whether they're in North America or they're in China using whatever model they're going to be able to look at all of this open source code and just run, you know, and what's, what's, what's really wild about this is with the cold card example it wasn't that hard to find the vulnerability. Anyone could have pointed an AI, you know, AI model at that. You know, one of the guys on our team that's you know, non technical ran codecs on this and found the vulnerability in no time on the MK3 code base. So it's, it's almost a thing of like you just need to point your code, you just need to point your model at something and you're going to find vulnerabilities. But I think the bigger question is like what does this mean for open source? Right? Is this the end of open source? Do people stop using it? But I think it's more likely that it's just a situation where we need to catch up, where you know, there's, it's a cat and mouse game where you know, now anyone can point their model at something and find a vulnerability. But you know, in two, three, four years time, you know, if we don't, unless we get another exponential growth in, you know, AI capabilities then, then it's likely that, okay, we're going to still end up, you know, building an open source land. But it's just that like now we're going to be checking, you know, everyone's going to be running a security analysis on every piece of code before they go and put that out in the, in the public. So I don't know. That feels like the new paradigm to me.

[Colin]

Final question here from me, Matt. What are the ramifications for the rest of the hardware wallet manufacturers here, I mean, right now this is a cold card issue. I think everyone who has their funds in a single sig hardware wallet like Trezor or Ledger is probably thinking,

[Mike San Miguel]

is

[Colin]

there a risk for my coins to be swept as well? Is it just a matter of time before other security vulnerabilities are found? It's impossible to say obviously what's going to happen in the future, but what are the shockwaves for the rest of the self custody ecosystem here and whether or not other users should be worried about similar attacks?

[Matt Black]

Well, I feel like, okay, to be honest, I feel like after the news came yesterday, likely what was occurring was every bitcoin company, whether they were a hardware wallet company, whether they were software wallet, mobile wallet, whether they're a custodian, was sitting there with every, every AI model they can get and double checking, hey, how is my entropy generated? So to be honest, I think, you know, this is a terrible, this is heart wrenching, this is a heart wrenching thing to occur. But it also means that everyone's like double, triple, quadruple checking their work now, which I think is, is incredibly important. But I, I think it also, I don't know, for me personally, it brings up a bigger question of like the multi sig discussion. You know, I remember those conversations back in 2020 where people would always say, don't do a multi sig because you're gonna screw up your setup and you're gonna lose your own funds. And that was the discussion. And it turns out the folks that, you know, were considered crazy at the time, who are, who are considered paranoid are the ones whose funds are now safe. And so I don't know, I think this. Well, there's two things. I think it brings up a bigger question of, hey, everyone's got to check their code with models now. And triple.

[Colin]

I'm. I'm. Why did you mute me? I'm just saying we need to wait to see what's going on with Matt.

[Charlie Spears]

Yeah, I didn't mute you. I mean, dang. Okay. He froze

[Colin]

one thing really quickly here, Charlie. I actually want to. I want to just flat. Oh, he's back.

[Charlie Spears]

Oh, he's back.

[Colin]

We're gonna bring him back up there

[Charlie Spears]

to just wrap up Matt.

[Matt Black]

Sorry, guys.

[Charlie Spears]

Yeah, you're good?

[Matt Black]

Okay. Yeah. You guys hear me? Okay.

[Mike San Miguel]

Perfect.

[Matt Black]

Yep. So, yeah, as I was saying, like, you know what, what does this mean for the larger ecosystem? We're gonna have folks that are quadruple, you know, quadruple all of their work number One for all of the hardware wallets out there. But, but I think the second thing is like the discussion around multisig, right? We're going to have a larger discussion around how can we make multisig simpler for folks. How can we set it up such that it's, you know, multi vendor, right? Like, and then how are we going to make it easier for folks to go and generate their own entropy, Right? People want to be able to, you know, now people know, hey, I should probably go and roll those dice so that I'm not at risk of this. But I think the final thing as well is the whole bitcoin community has been talking a lot about spam and I think it's time to get back to a discussion around Covenants about not only, you know, reactive security, but proactive security so that we have better ways of people being able to secure their bitcoin.

[Colin]

Go think of anything. Hopefully this reignites that discussion. Although I've been pretty black pilled about whether or not we're going to get, you know, fruitful discussion in the technical community over the last few months. Matt, before we leave, before we sign you off, I just wanted to flag this tweet from you. You had a great breakdown on Twitter about what you need to do to address this if you have a cold card and how you can protect your stack for the future. So go check out at Matthew J A black on X if you want to learn more about this vulnerability and what you can do to address it if you are potentially affected.

[Charlie Spears]

Matt, thank you so much for your time and insight. This is probably going to be an ongoing topic so we may touch on it. God forbid more bitcoin doesn't get hacked, but we know it's coming, so fingers crossed. Matt, thank you so much.

[Colin]

Cheers, Matt, Adios.

[Matt Black]

Thanks for.

[Charlie Spears]

Super sharp guy.

[Colin]

Very sharp.

[Charlie Spears]

Yeah, I, I'm glad he ended it with Covenants because that, that was like the beat we covered on the show for, like.

[Carmen Lee]

Yeah.

[Colin]

And if, if you're not that. Actually we need these things. What else? Yeah, you know what I mean, the whole, oh, it's just an education, dude. No one's going to use them. I don't know. Could have saved a lot of bitcoin here if we had them, so.

[Charlie Spears]

Well, you know, again, deep, complex topic. I'm unashamedly pro covenants and soft forks, but not the software coming in nine days. But we're not going to dwell on that because we are going to go back to AI, Colin, going to go back to AI and compute Specifically going to keep on the Kimi train. We'll cover the Kimmy Moonshot Nvidia story here in a moment after a word from our sponsor Lycos.

[Colin]

The sad fact is a black hat hacker with Kimmy K3 doesn't need to be the thing that takes down your Bitcoin stack. If you're using a Bitcoin backed lender that is not respecting your stack, you could become the next Celsius or Blockfi victim. But with Lygos Finance you don't have to worry about that. They are our preferred Bitcoin backed lender here at Block Space. They use Bitcoin native smart contracts to make sure you are always in control of your coins. You heard that right? With Legos Finance you always have control of your keys. When you are pledging collateral for a Bitcoin backed loan, there's no funny business, no rehypothecation, no bridging, no wrapping, just competitive rates as low as 8.5 to 10% APR. Go to Lygos Finance to learn more. All right Charlie, another headline in the export import wars for the heartbeats in the brains of the AI revolution. This one specifically is coming from Bloomberg and it is a report that claims that Kimike 3, specifically Moonshot, the lab behind it, had a computing power agreement with Alibaba for roughly 20,000 Nvidia chips, forming a substantial chunk of the compute behind its Kimi models. Sources say the chips are Hopper Generation H200s, though Alibaba disputes that specific claim. What's interesting is though, according to the reporting, Alibaba pushed back on the stipulation that it was Hopper generation, but they didn't necessarily totally dismiss that they were providing compute for Moonshot. What's interesting about this, number one, is that the United States doesn't want these Chinese companies to have cutting edge chips. We'll get into some of the export controls because actually Hopper generations are not totally barred as long as they have White House sign offs off. But Beijing also doesn't want them in their country either. Allegedly. That's their posturing because they want to show that they are protecting domestic producers of compute. So there's this weird push and pull here on both sides of the ocean in terms of the United States doesn't want Nvidia shipping GPUs specifically Blackwells to China. But China is also saying that they don't want their companies using them. That's what they're saying. But if this story is true, then clearly there's some wiggle room here. Or Alibaba was actually just caught with its pants down doing something that Xi Jinping does not want happening. So, Mike, there. A White House official, Michael Kratzios, accused Moonshot last week of illegally acquiring Nvidia's most advanced Blackwell chips. That's another curveball here. The government hasn't specified how many it believes Moonshot has. In the X post, he claimed that Moonshot distilled Anthropic's Fable model to develop K3, building a quote, sophisticated internal platform to conduct large scale distillation against US models. Critically, Bloomberg notes he provided no concrete evidence for the distillation claim. You know, maybe this is just, this is just posturing on our own government side. I would wager a guess that someone on the inside of the White House probably has better intelligence than Bloomberg. Call me naive, but I think there's probably a good chance that that's happening. But this distillation allegation is something that's been flagged across social media and across reporting. In fact, it's well assumed, if not well known, that the only way that models like Kimmy, K3 and Deepsea could actually have the rapid advancements that they have portrayed is if they're distilling frontier models here in the U.S. charlie, I'll toss it to you for second takes before I get into the legal nuances of what exactly the White House's export controls do to prevent the latest generation GPUs from being in Chinese hands.

[Charlie Spears]

Yeah, I mean, that was probably going to be my take, but I'll let you run with it. My kind of like subjective experience, given the fact that we just talked about this hack on Bitcoin hardware device, is that in just recent memory in the past 24 hours in some of these chats that I've had with developers who are performing security audits on their own hardware, they were unable to use largely the frontier models in the United States, specifically Fable, but also chatgpt soul level models, did not like want to play ball and do these audits from White hat hackers. And what they found is in multiple cases where they had to turn immediately to Kimmy. So this does like bring into existential question, what even are these export bans functionally doing? Are they functionally preventing the GPUs from being physically inside China, or are they simply just kind of friction for domestic American users? Kind of almost like the days of trading crypto with VPN on binance, like you didn't really stop anybody, you had to pay your taxes, but like you weren't actually Prevented. So some analogies there. I'll toss it back to you to get into the nuances of like what

[Colin]

are we doing here?

[Charlie Spears]

What, where can we run? According to obscure. I think this is U.S. export Control.

[Colin]

I think this is important because kind of like with the tariff argument or the tariff discourse, there's no gray area in people's minds on these things. And the actual facts of what's happening get obscured by sensational headlines of, you know, Trump bans China from having access to GPUs or, you know, there are 100 tariffs on everything now. So for the specifics here, the US rules actually do allow for access to restricted chips via compute rental agreements, renting, not buying. And that's what Kratzios actually says Moonshot did through an unnamed party in Thailand. And this is a workaround that we've actually covered for other things like Bloom Energy's procurement of Scandium. There is a, there's kind of a plausible deniability element here where Bloom's Energy CEO can say we don't rely on China for Scandium because they're buying from offshore companies in Southeast Asia that are importing the Scandium from China. But there's enough of a layer there to where what you're saying is factually true, even if it's not directionally true. And so these, most of these Chinese companies do have access to compute in places like Thailand, I would imagine Indonesia and other Southeast Asian hubs for commerce. The physical location on that note of these Alibaba GPUs is unclear and, and that matters legally because the rent restricted chip offshore workaround hinges entirely on whether where the hardware physically sits. Now that being said, GPUs themselves aren't wholly banned from being exported to China, but there are big catches. The Blackwell line are completely banned. Again, the White House and Uncle Sam do not want Chinese corporations having access to bleeding edge tech, but the Hopper generation are conditionally allowed. These were recently unbanned in December 2025 and Trump announced that H2 hundreds could be sold to China on the condition that they get a rubber stamp from Washington. So if Washington approves the sale, they can be exported and purchased by these Chinese companies. So the older, less capable variants are actually somewhat approved. But there's this other consideration here, like I said at the front, Charlie, where the dynamics here are very odd because ultimately Beijing has postured that they don't want their companies purchasing these because they want to see as actually having leading GPU development on their own. Right. And they want their models to be trained and inferenced with Chinese hardware. Whether or not that plays out in reality obviously is up for debate. But the TLDR for the actual restrictions, rentals themselves. A Chinese company can rent GPU capacity. I mean the United States can't tell them that they can't and they're just going to do it right? If they're renting it outside of the United States. If it's outside of China and it's an Nvidia chip, the United States has thrown their hands up and say we can't really control that. But they are restricting black wells and bleeding edge units from being shipped to the mainland. So yeah, a decent amount of nuance here.

[Charlie Spears]

And I'll probably throw this topic as a first question to our next guest who by the way is in the backstage and we're going to roll to our final segment and guest of the stream Carmen Lee going to bring her on up here. All right Carmen, welcome back to the show.

[Carmen Lee]

Thank you. It's great to be back here.

[Charlie Spears]

Yeah. So we just talked about the Kimmy Nvidia deal. Where are these GPUs going? And it kind of gets into like the export control conversation. What do you think like is, is the US just kind of tying a bow on and pretending like they are functionally prohibiting Chinese companies from running, you know, using GPUs or is there actually, is it somewhat effective in a way? I know. What are your thoughts on this, this story and topic?

[Carmen Lee]

So I heard the last five minutes of your guys conversation so I came a little late so I didn't hear the whole argument. But in general I feel like what is the angle? Right? If the angle is to make sure China is always six months behind, it's not working kinda. It's not working right. If the angle is say hey we just want to make sure that people developing on Stubidia cluster Nvidia ecosystem system that's kind of working because right now the most efficient chips deploy K3 is in the video chips. It depends what's the angle which I don't want to speculate too much but you look at K3 deployment right now which just came out a few days ago, right? There's two, three providers. They public claim they can support K3 which is kind of exciting because it's very hard model to support. You need to have a lot of capacity obviously and it's completely different licensing model. If you guys chat about this already, take a look at terms conditions for K3. It's not like the MIT license which they had previously you have to freaking pay for licensing. So you can argue where in the way between open source and closed source models now, right? Obviously it's closed source models which game is off. Right. If you are token factory in your cloud you can host but you have K3 of the world where sure you can, you have all the weights, you have everything you need. But you had to pay Kimi for licensing fee for API and they set a price, they set a floor price, right. So even token factory you're bound, you can just charge whatever you want, right? There's a price for that. I'm not saying it's bad or good for the token backtrace, but obviously there's rules supposedly. But then if you are saying that I'm going to deploy your own cluster on premise, you can do whatever you want, you're not API based, Right. But again if you look at terms and conditions, depends what kind of enterprise you are you, who are you serving, what kind of products. Right. There's all those arrangements that people need to go through. That's not. I think things change, right? If you look at closed source model, other open with models, you look at llama, there's terms, conditions, your enterprise adoption, I think no enterprises I like to think want to do the wrong thing. So you have to be really thoughtful about how you want to use the model commercially. Will make sense. Does APM make sense or does only your clone cluster that gets someone managed for you? Makes sense, right?

[Charlie Spears]

Yeah. So you get into tokens and GPU prices and that's what we want to talk about here. And you had this tweet recently noticing saying we've seen a 15 to 30% price increases year to date across a 1/ hundreds, h 1/ hundreds h 200 b 2/ hundreds neocloud on demand indices. But at the same time we've also seen that token prices have been down, I believe down 18% in the past month or so. What's going on here? Is there a mismatch between GPU prices and token prices or is there some kind of market dislocation happening?

[Carmen Lee]

I think it's interesting, right? So again our token indices measured the willingness to pay for million dollar token doesn't measure total demand of the society. Right. If you look at GPU prices which is a reflection of the world, the total world supply demand curve x China where it meets so the demand so the reason why it goes up usually the demand rise faster than the supply. So meaning people want more GPU but they don't want to pay the premiums they were Paying for the closed source models back then. Right. So translate to people using different model selection. However, obviously whatever the cutoff today is not many people using K3 yet. K3 is not that cheap. If you look at the API prices. If we see huge adoption of K3 people moving from the Z AI, the price may come back up. I have no idea. It depends on when people adopted. Again, it's more expensive models. So this is all willingness to pay, right? Do you want to pay more? Less than the premium labs, but still more than the other open source models is available. So this can change.

[Colin]

So I'm curious from your perspective on that tweet, Carmen, you mentioned obviously rental rates are increasing and you said that y' all run like 17,000 reference points points for your research. What trends and insights are you seeing that you think other people are missing from these rental rate increases? Like what's the read between the lines? Look at what's happening here.

[Carmen Lee]

So one of the things I think will be interesting is look at forward curve, right? So everything I post on that Twitter is on demand. Really give people a what's going on right now Snapshot, right. If you're on demand, what's the pricing people paying? However, I argue majority of the world's capacity is sitting with reserve capacity. Meaning you reserve it and then that's it. Right. It's not part of on demand supply supply curve. So those are the prices what impact a lot. If you RFQ anything today for oneyear long contract or three long contract, you know, and then it's can be content or downward sloping. So you think about in general, you would like to think the longer contract you signed, the cheaper should be per GPU per hour, right. The more you signed up for, right? Buying, you know, 20,000 versus 2, you will think the more you buy, the cheaper it should be per GPU wise. It's really not true. Sometimes in our world where you see

[Colin]

the forward contracts are in contango rather than backwardation.

[Carmen Lee]

Exactly right. If you look at. I don't know, Ken, can I share screen?

[Charlie Spears]

Yes, you can.

[Carmen Lee]

Let me see if we can share screen here. Can you see this?

[Charlie Spears]

Yes, there we go.

[Carmen Lee]

If you look at H100, that's today or yesterday. We see a little contango here. This. Look at the blue line only. This translates to if you're entering contract today for 36 months long contract, you're paying about $2.36. But you entered 24 with two years, you're paying $2.30. So you can see all contango here, right? So you will this is in commodity happens, right? So now a few things. Number one is a lot more people want to pay they willing to pay premium just to lock in the resources. Then say hey you know we'll at the mercy of the blind supply curve. Anytime or the supplier has option have the market power saying that guess what? We only want to do three years. That's just we want to do it right. They can just move everybody three years long curve. So if you look at last November the price were going down last year across all different GPUs it was downward sloping. So essentially this curve is a market consensus every day tell you what people think about the future, right? The orange line basically is a forward rate that silicon data calculate is imply rate for short term rental beginning given term. So you can essentially similar that methodology as treasury futures curve. Right? So people think the price can come down versus keep going up. If you look at B200 curve again it's completely completely different. Right? And then you see a little even three months you have to see all different kind of variation. This is really reason why I think dated futures so have futures have different expiration will be precisely needed for the market. Because you every so there's like every other commodity you have a term structure risk right? Either your bank you have six long six years long exposure. You know, you have a client for first two or three years. What's going to happen year three, four, five. You're concerned like is that my rent going to going to my rental income going to pay down my debt that you have that duration exposure you need to hatch. That's precisely future can help you do so.

[Colin]

That's a really crazy insight and kind of runs afoul of common sense for like you said at the beginning, for the whole reason why you enter into these forwards to lock in a potentially lower rate for revenue certainty. You kind of touched on this. But if you can unpack it a little bit further. Carmen, what do you think that's saying about the supply and demand story of compute? Because we're going to see gigawatts of capacity come online next year and into 2028. But to me it almost seems like the market is betting that that's not going to be enough. Especially if we see token prices fall. And then Jevons paradox kicks into where people just start using more and more because the token prices are so cheap that you actually don't have a supply buffer that people are expecting. What is the market telling us with that contango in Your mind.

[Carmen Lee]

So the market obviously think the price will not come down. You know, so again we all can calculate supply sides. You know how many data centers supposed to be online globally. You know how many chips Nvidia can turn out. You know how much megawatt gigawatt we can religiously put to data centers, which was the bottleneck that will that constrain a supply curve, right. In the long term, the next two years, whatever. But demand side then your speculation is as good as everyone else speculation, right. To your point, like when people going to see adoption, right? Because everything you see today, I would say enterprises really in the beginning piloting production phase. I don't mean even within us. Right. You see some companies are more sort of cutting edge and maybe mostly coastal companies. You look at Europe, right. Who is starting look at Japan, they haven't. Some of the businesses are not even digitized yet. Like when they are going to pick up the whole use cases. Right. So those are people trying to figure out when is the true enterprise ROI backed demand sustainably going to come in. Right. Obviously lower the cost will translate to people more willing to experiment. They're okay with not having gigantic ROI because they're spending, but not that much. They can try a hundred things and pick maybe two or three that make sense to them. Versus Holy. It's so expensive. I can only try to make sure one works. Right. Just by lowering the cost you allow a more piloting program to start without too much a threshold. So at the end. Yeah, you give people more optionality to adopt for sure.

[Charlie Spears]

You know, one interesting thing is that pretty much all the end users and even like many of the application companies just care about token price. They don't care what. They're not even aware what GPU is running underneath. Yeah. And they should. And well, what are the implications long term for the industry? Maybe you've already spelled it out but like the world's going to consume tokens. They don't really care about the GPU model. What does this mean long term.

[Carmen Lee]

So which is interesting, right. You look at our GPU indices, majority are Nvidia based. Right. We're agnostic, we do have Mi 300 indices. But essentially if you do token emu couldn't tell us what GPU is running under. Can be seminova, can be anything, can be different design houses. Right. So that essentially token, token indices, token price can be more encompassing of the whole infrastructure. Right. In system than just hey, this is particular chips. And the way I will think about this is we all need to use Electricity every day, but we don't really care. I mean aside from obviously, you know, being green and then the carbon footprint, right, we don't look at energy sources, how people we generate electricity, we care about electricity prices at the end. So similarly to this, right, we care about token prices. We don't care about what GP running underneath it. If you are average user, we don't need to hedge our electricity use at home because it's not that much money. We pay whatever on the spot, whatever they charge us. But if we were factory electricity, token is the essential productive mean. Can be 10%, can be 80% of my cost bank goods sold. Yeah, I will care a lot. I want to hedge the token prices because right now token prices are pretty much on demand. You're paying API for example, you're paying whatever they charge your API. You can't just say hey can we talk about it? Can we just say you are going to pay you this price for the next six months. You can do that. So that's why configuration is doing the token four packages now is you can lock in a price. So obviously even for the same model K3, you have all different parameters, different precisions, different quantizations and different sort of other latencies. People will charge you differently. They should, right? So whatever you decide on, whatever the specs you decide on for that particular model, you want to lock in price, say hey, EVA is dedicated to this particular open source models. This is the price that I want to pay for next six months at least I logged in, right? In theory you can do so because you know from token factory point of view, you know what GPU you have, you know what's the throughput for that particular cluster, you know in theory how much tokens you can generate in pull output tokens and then slap a margin on top of bare metal. And you know how much you want to charge people and you can lock in the prices without large enterprises and then people both side have peace of mind, you know you're going to be consuming consumed from token manufacturing point of view, from client, you know what price to lock in, you know exactly what you're going to get. Then rely on PI, which you don't even know the precision sometimes.

[Colin]

Carmen, a closing question for me. So as Compute Exchange is designing these token forwards necessary counterpart to the GPU hour, the GPU rental price forwards, it seems like there are more variables that would go into informing the the price of those token hours versus GPU rental rates. I could be totally wrong about that. But I'm Curious. When you're designing this product, what do you have to take into consideration for nailing how to specifically tailor these forwards for the token price model?

[Carmen Lee]

So similar to GPUs, right? Even GPUs, they are heterogeneous. So even the H100, even the same CPU, same RAM, same geolocation can have different things throughputs and different latencies and different memories and different flops. Right. So you need third party benchmark to say hey, this is what the specs is. That's the model itself. Here's the token, here's the throughputs variability latency we observe from third party point of view. And here's SLA again it's a hit or junior genius. You need third party to say hey, this is the deliverables category against physically settled products, either token or gpu. You define the metrics, you have to meet that metrics and you charge its prices. And that's enough story. So you have to make sure third party benchmarking have observable results.

[Charlie Spears]

Carmen, thank you so much for your time and your insights. Would love to have you on again as the contango continues, token pricing and GPUs. Carmen, thank you so much.

[Carmen Lee]

Yeah, thank you for having me to come back.

[Colin]

Have a great weekend, Carmen. Cheers.

[Carmen Lee]

Have a good Friday.

[Charlie Spears]

Cheers. Glad I know what contango is from hash price forwards. So

[Colin]

that is wild to me.

[Charlie Spears]

Yeah, it is wild.

[Colin]

Nothing, nothing crystallizes the everyone is short compute narrative. More than that.

[Charlie Spears]

Yeah, yeah. If you didn't. Yeah. And it's crazy what a forward curve can reveal. It basically cuts through like will they, won't they? Is your favorite blogger, right? Well just look at the market. So obviously asterisks. Look at volume.

[Colin]

Well, yeah, because that flat forward curve could collapse tomorrow. I mean not tomorrow, but you know what I mean. All it takes is some shifts in how people are viewing the capacity environment, more capacity coming online, usage dropping or another variable that I haven't mentioned and that could go to backwardation. Right. I mean we've seen this in the Luxor. Luxor's forward curve a lot. A change in Bitcoin's price, a change in difficulty can completely throw that regime on its head. But yep, anywho.

[Charlie Spears]

All right, thank you so much for spending the week with us, finishing the week out strong with Block Space Live. If you like our content, you can find all of it written and on demand at our website, Blockspace Media. That's our website, Blockspace Media. And if you got your bitcoin on a cold cart. Get it off. Never thought I'd be saying those words, but go do it now. Heart goes out to all my fellow bitcoiners who are panicking or have opened

[Colin]

up their wallets to chagrin Pannikins and control once again.

[Charlie Spears]

Yeah, I know. Anyway, but there's always a brighter future. Don't know when, but sometime in the future. Thank you for listening. This show is brought to you by CleanSpark. NASDAQ listed ticker CLSK I'm Charlie.

[Colin]

I'm Colin. And quick note. Next week, we are going heavy into earnings. We will be covering earnings exhaustively starting next week and into the following week as some of our favorite names release their quarterly earnings. We're talking hut. We're talking cipher. We're talking riot Mara Keel all the way down the line. We will be having coverage on the live stream of the numbers and we will also have executive interviews lined up. So stick around for all inclusive coverage of that. And with that, have a great weekend, y'.

[Matt Black]

All.

[Charlie Spears]

Have a great weekend, y'. All.