Subscribe
Share
Share
Embed
The Slow Hunch explores how big ideas form over long periods of time. Big innovations are often characterised as single “eureka” moments, when in fact they're often the culmination of many smaller ideas coalescing over a long period of time. On this podcast, USV's Nick Grossman explores how those ideas took shape, and the nonlinear paths of the people behind them.
4K_TheSlowHunch Podcast- Meltem V3.txt
English (US)
00:00:00.120 — 00:00:46.200 · Meltem
In the first half of this year, people were just plowing money into anything in the eye space that looked or smelled like a Nio cloud. Everybody's lying their faces off. What I spend some Bitcoin people are just making shit up. A lot of these smaller, unproven NIO clouds who have taken on a lot of debt have never deployed a single GPU.
They're three months behind schedule, six months, 12 months. Right? I would not have known this until we went through developing and deploying our own cluster. But when people show us spreadsheets now of their financial models, I'm like, you're off by margin of like 50%. The anthropic S-1, I think, was really great because it's the first time we're really understanding the underlying economics of a frontier lab.
Guess what? Their business model is turning GPU hours into tokens. Of course, it's a refining business model. It's a commodity business.
00:00:46.240 — 00:00:46.840 · Nick
There it is.
00:00:46.880 — 00:00:51.000 · Meltem
And here's the beautiful thing. Commodities are what changes the.
00:00:51.000 — 00:00:51.560 · Nick
World.
00:00:51.560 — 00:00:56.040 · Nick
Its markets all the way down and maybe all the way forward.
00:00:56.080 — 00:01:20.910 · Meltem
Humans have been doing this since the dawn of history. There are scarce resources. We use information as a way to determine how those resources can be best allocated. I think it is so important to take a market driven approach, because markets are going to be the best mechanism for distributing the economic effects of AI as broadly as possible.
00:01:50.230 — 00:02:19.580 · Nick
Welcome to the Slow Hunch. Thank you so much for being here. I've been really looking forward to this conversation. As you know, the slow hunch. The premise is to talk to people who have been pursuing an idea for a period of time through twists and turns. Um, but like with a steady arc. So, so maybe for for folks who don't know you, please introduce yourself.
And then let's try and say or like put some words behind what we think that arc is or has been for you.
00:02:19.620 — 00:02:49.820 · Meltem
Yeah. Um, so, Meltem, um, what do you do with s'mores? You melt them. Is how is how I, uh, say it? It's an unusual name. The story of my unfolding. If you are my slow hunch. Um. I started my career in the oil and gas industry. I graduated college during the GFC. Great financial crisis. Um, what was so interesting at the time, it was the era of shale oil and gas.
So talk about major technological shift in production. Fracking.
00:02:49.860 — 00:02:53.020 · Nick
Sure. From right. A brand new mode of production.
00:02:53.660 — 00:04:15.800 · Meltem
That fundamentally shifted the cost curve. But it also changed the economics of production. So this area called GGP gas gathering and processing was actually an area I spent a lot of time on. At the time, there was vehement public backlash against fracking and against the oil and gas industry. Right.
This was the beginning of sort of ESG. So I worked in that industry. In 2013, I went to grad school in Boston, and I learned about startups. I learned about venture capital, and I learned about Bitcoin. And I was like, Holy shit, this is so cool. Um, forget about working in oil and gas. I'll do this. This is crazy.
And so I met, um, Ryan Sulcus and Barry Silbert, who at the time were working on something that would become Digital Currency Group, and I joined them. And so that was the next three years of my life. So from 2015 to 2018, I was at DCG helping build that firm basically from inception in the earliest stays um through to Bitcoin taking off, right?
20 1820, late 2017 I think was a really defining year. And if you'll recall, in 2017 when the bitcoin mining network right started to evolve and industrialize, bitcoin mining started getting a lot of backlash for not being environmentally friendly, for consuming energy.
00:04:16.120 — 00:04:17.120 · Nick
Sounds familiar.
00:04:17.239 — 00:05:48.910 · Meltem
I had lived this. Yes, in Shell Gas. Right. And I became obsessed at the time with this idea of that energy to compute value chain is what I called it. What to me was so interesting about Bitcoin is that we could produce verifiably scarce digital value using thermodynamics. Right. So this relationship between using computation, which is a physically intensive process that requires you to expend thermodynamically scarce resources to produce something that's purely digital and form, I thought was really novel.
Totally. So then I spent the next five years at a firm called CoinShares. It's an asset management firm that also had a trading business. And then I helped build our venture in sort of early stage innovating, if you will. ALM and I spent five years in that business. We took it public in 2021, which was a really interesting experience.
Um, when we were at CoinShares, we had a full time research team that did a lot of work on researching the economics, the energy draw of Bitcoin mining to try to counter, I think, some of the misinformation about Bitcoin's energy usage. And what I realized, and I think what's now very popular narrative is we see the evolution from Bitcoin mining to HPC is that Bitcoin actually was a great battery in soaking up excess stranded energy production and turning it into a fungible commodity in the form of Bitcoin.
00:05:48.950 — 00:06:07.980 · Nick
And I remember I wrote a blog post in 2021 called Bitcoin as Battery, which was built on this idea, which is part of what one of the things that has, you know, tied me to your work and also to this connection between energy and compute. And I remember getting absolutely skewered online.
00:06:08.060 — 00:06:12.660 · Meltem
Um, that means you're doing it right. Yes. I've been skewered many times.
00:06:13.260 — 00:06:25.300 · Nick
Um, and but it really that this idea that you can connect, um, the physical world with the digital world in a in such a pure way, like totally stuck, stuck with me at that time.
00:06:25.340 — 00:06:40.580 · Meltem
After CoinShares, I took some time off. I think for me, I was very vocal. Right? I was always very vocal about Bitcoin and my views and why I thought it was important. FTX broke something in me. Um,
00:06:41.660 — 00:07:10.370 · Meltem
it was a really disappointing moment and so many people that I trusted, um, ended up being really disappointing. So I took a step back and then I just started kind of investing my own time and energy and capital into this energy to compute value chain. But the realization I had when GPT three first came out, it was like, Holy shit, the token isn't Bitcoin, the token is intelligence.
00:07:10.410 — 00:07:11.530 · Nick
Intelligence? Exactly.
00:07:11.570 — 00:07:15.530 · Meltem
Yeah, we're still working on the naming of things. Not great branding, but it's intelligence.
00:07:15.570 — 00:07:16.090 · Nick
Same thing.
00:07:16.130 — 00:08:47.150 · Meltem
It's an output of digital labor, right? So the product is digital labor. In the Bitcoin network we use computation to facilitate secure financial ledger write transacting. Ethereum extended on that with secure financial computation. And then AI extended on that even further with work. Right. And then what was so strange, as in the early 2024, I was approached by, um, an investment bank I knew, and they were like, you're doing all this writing and talking about the energy to compute to crypto value chain, right?
And at the time, I thought crypto and AI would intersect a lot. We can talk a bit about where that goes. They're like, what are you doing? Like, do you want to come work with us? Like, no, I don't really think I want a job right now. I don't know what I want to do yet. And then another family office in New York that I'd known for some time said, like, when are you going to start a firm?
Like, can we co invest? How do we work together more? So that was sort of the genesis of Crucible. Um, I landed my two anchors if you will. Um, with some bullet points that I wrote, I was like, I think this is what I want to do and what I want to focus on. And it sort of grew from there. So I think it was, again, an unfolding.
And the name Crucible was an interesting one. Um, a crucible is a vessel for transformation. Okay. But more importantly, I had become really obsessed with the small town in North Carolina called Spruce Pine. Okay, do you know what's made in Spruce pine?
00:08:48.990 — 00:08:49.710 · Nick
Um.
00:08:50.070 — 00:08:51.670 · Meltem
Take a guess. Uh, Caroline.
00:08:52.070 — 00:08:52.510 · Nick
Uh,
00:08:54.030 — 00:08:56.710 · Nick
Cole. No, that's too close. What?
00:08:56.750 — 00:09:39.190 · Meltem
Um. High purity quartz. Okay. Yeah. So there's one place in North Carolina, and it's the only place we know of in the Western hemisphere. China claims to have their own high purity quartz deposit. Oh, but, um, it's the only place we know of in the Western hemisphere where you find the specific high purity quartz that is used to make the crucibles in which you can forge silicon ingots go right.
Those silicon ingots are grown in these crucibles, and then they're sliced into wafers. And that is the start of the CMOS process. Right. Is the primary substrate upon which all chips are built, not just GPUs. The chips in your phone, in your car, in your fridge. And what was so interesting to me. I love stack graphs.
00:09:39.230 — 00:09:39.710 · Nick
Me too.
00:09:39.750 — 00:09:57.380 · Meltem
Right. And you look at the compute stack graph. It's like the entire compute stack StarCraft depends on this tiny place that nobody's ever heard of, that nobody can name or point to on a map that produces this one thing nobody has ever thought about. I thought that was interesting. So that's where Crucible came from.
00:09:57.460 — 00:10:03.340 · Nick
And give me the two bullet points on on what were what was the the initiation of Crucible then.
00:10:03.380 — 00:10:29.860 · Meltem
Energy compute crypto. I think that has since the I'll be very candid the crypto investing we tried it just it's not there. I'm not saying right now it doesn't mean it won't be there. I do think it's interesting that some of the most prolific founders in the AI space came from crypto. So Alex Atala, open router, Prime Intellect New research, produces her muse agent.
Um, together AI Crusoe core. We write there's a laundry list goes on.
00:10:29.900 — 00:10:30.420 · Nick
Makes sense.
00:10:30.460 — 00:11:14.159 · Meltem
Yeah. And so, um, now I think what we've evolved to is if I think about categories, it's energy, compute materials and markets. Right. The view is the twin engines of progress are technology that unlocks fundamental cost curve shifts, making production cheaper. And I do believe that we are a decade or less away from incredible abundance in energy compute materials, coupled with this incredible engine that America has built called capital markets.
Right. And these two together are the twin engines of progress. And I didn't really see a lot of firms in 2024 that were operating at that intersection. And so that's really the genesis of Crucible. And then if we look at the team, right, many
00:11:15.240 — 00:12:30.230 · Meltem
of our views and my team were traders, our LPs, our traders. It's a very capital markets focused lens, which I think is interesting. And now Silicon Valley I think is very much waking up to it. We think about how Bitcoin was traded in 2015, right. It was all bilateral DVP delivery versus payment. So the business is in 2015 where basically, hey, we're a trusted third party that can cross agency trades, right?
So there was no principle risk because like, you want to buy, you want to sell. We're a trusted intermediary. We facilitate the transfer. And then from there in 2017, right. Perp started to take off because of BitMEX, right? Then we saw this whole ecosystem explode. I was a very early investor in Poly Market, right.
This idea that we would build at the time the view was information markets. Right. Which became prediction markets. But what was really cool about what happened when the Bitcoin capital markets eco system started to blossom? It was one of the first times in recent history, or only times in recent history, where we saw the creation of a net new market structure that was entirely outside of the traditional Wall Street capital market structure.
00:12:30.350 — 00:12:33.350 · Nick
In terms, not in terms of concepts, but in terms of.
00:12:33.630 — 00:12:35.230 · Meltem
In terms of the perp was new.
00:12:35.270 — 00:12:36.390 · Nick
Yeah, I guess the perp is new.
00:12:36.430 — 00:12:45.140 · Meltem
Models for funding, right? In traditional banking, There's no model where you preferred a transaction, but because Bitcoin settles with finality.
00:12:45.180 — 00:12:45.420 · Nick
Right.
00:12:45.620 — 00:14:04.130 · Meltem
Right. Exactly. So it was fundamentally different. Um, and so that was really interesting to have a front row seat to. And at DCG we invest in over 200 companies. While I was there, many of them were really about building this ecosystem of global exchanges where you could on and off ramp dollars for crypto, and at the time was really just Bitcoin.
Um, and I think, again, as I looked at what was happening in the compute space, I was like, we're going to need more energy markets, more instruments. Even if we look at compute markets, like, I think compute markets are going to be in massive surface area of expressivity, expressivity pardon. In terms of what financial instruments.
There's the asset backed credit side which is GPU backed credit. There's probably credit default swaps for large customers. Right. There's a huge circularity of revenue and guarantees issue. Then you have the off takers. That whole value chain is probably going to have dozens of different financial instruments that are traded by specialists.
And so I think we're at the early stages of the creation of this incredibly massive market. And the most incredible thing to me is we have new traders in the form of agents. And I do believe agents will be economic actors to determine what that looks like. But it's such a fun thing to think about.
00:14:04.170 — 00:14:27.770 · Nick
To add a wild twist. That was that was amazing. And that was the slow hunch in fast mode. And it's you can see it all laid out just thinking about compute as a new asset class in the AI world, without thinking about agents as actors in that market. Like that's a big enough new idea. Yeah. Not to layer on an even bigger, maybe newer idea on.
00:14:27.770 — 00:15:03.120 · Meltem
Top on par. Right. It's like that unfolding. And then as we think about agents, as economic actors, the innovation of crypto is computers that can make commitments. Financial commitments. So these pieces all fit together. We also built we financed, um deployed and now manage our own compute cluster as well.
It's a one megawatt B 300 cluster. And so I think it's been again this unfolding where we've learned a lot by actually doing the thing. Um, and there's a lot more we're going to learn. I'm just going to keep on folding.
00:15:03.840 — 00:15:29.520 · Nick
Well, well, so I kind of want to like, go back a little bit because what first kind of got you excited about the idea that financial instruments and financial engineering, let's say, were essential to making new commodities markets work? When did you have the kind of first aha, that that this idea was important to you?
00:15:30.160 — 00:15:48.030 · Meltem
I don't know that I ever had that aha like light bulb moment. Um, I think I'm kind of dense sometimes. And then it hits me in the face. Like, it just recently hit me in the face that the three technology shifts and capital market shifts I've lived through were also deeply unpopular.
00:15:48.110 — 00:15:48.750 · Nick
Yes.
00:15:48.750 — 00:16:07.390 · Meltem
With a general population, um, because of ESG concerns. And. Right. And so I'm like, wow, I have this penchant for, like, finding the most hated thing. Yeah. Like facing it. Yeah. Publicly. And everyone's like, wow, that lady sucks. I'm like, okay, great. I guess this is just what I like.
00:16:07.430 — 00:16:08.030 · Nick
Well, so that.
00:16:08.070 — 00:16:10.470 · Meltem
That masochist, I think that is that.
00:16:10.470 — 00:16:28.550 · Nick
Is true. And that's that's kind of why I was saying at the beginning, you have conviction and courage, because I think of you as someone who has been able to try to come up with like a very vocal and consistent and principled view on markets that have been deeply unpopular.
00:16:28.830 — 00:17:13.339 · Meltem
I think my love of commodities comes from my dad. My dad was a research scientist at a chemical. At a chemicals company. Okay. And my dad worked on plastics. That was his great love. Polymer chemistry. Right. And so as a kid, like, most kids, would do ABCs or, like, play fun games, but I was like, okay, today we're going to memorize covalent bonding.
Looks like so. Can't wait. Um, so I think learning about the material world, learning how things worked. And I really got into reading sci fi that's, I think, been a really important part of the story arc. I think what's interesting about sci fi and reading fiction is, um, sci fi is a great way for us to play out our hopes and fears of what technology may make us.
00:17:13.380 — 00:17:14.060 · Nick
Absolutely.
00:17:14.100 — 00:18:22.970 · Meltem
And so, um, some sci fi has been really prescient. The Diamond Age, I think, is like one of my top picks for most prescient book. But Kurzweil. Right. Even in the 1970s, was writing things that have been incredibly prescient. And so I think reading a lot of sci fi, having a scientist dad. Um, and then the last piece.
I didn't really learn to appreciate capital markets until I worked in the oil and gas industry. My last role in the oil and gas industry. I worked for the treasurer at ExxonMobil. So what was so I say oil and gas is the original deep tech, super ambitious, incredibly technically complicated, fraught with challenges.
If you look at sub-Saharan Africa. Right. Like people were building wells in the Niger Delta and gorilla tribes would come in and hack the pipeline like it is an adversarial environment. It's huge sums of money. And so it's just such a fascinating industry. And just to see people planning projects on a 20 year time horizon and taking that level of risk, it was like one of the last frontiers where we were like real cowboys.
And I have a lot of respect for that era of oil and gas because it was wildly ambitious.
00:18:23.010 — 00:18:30.320 · Nick
Yep. And and in service of producing a commodity that was Underpinning the global economy.
00:18:30.320 — 00:19:57.910 · Meltem
And everybody loves to hate the oil and gas industry. What I thought was so ironic is, you know, you had Greta at the time was the face of like the anti oil movement and people were flying around the world on their jets to go to Davos and talk about how oil was bad. And I'm like, you flew in a jet like it consumes C6 jet fuel.
Like what are you doing. Like make it make sense. So I think, um, it's part of why I've been so outspoken is oftentimes I feel like I'm taking crazy pills because I look at the narrative that's out there. Right? People are anti data center. People are anti this, anti that. And I'm like do you understand that the abundance you enjoy.
Do you understand. You live in America. It's a country with incredible abundance. That is all the product of the fact that 20 years ago some people stuck their necks out, did a thing that was very unpopular, that made this country an energy producer instead of an energy importer. That is why we have abundance.
That is why your gas is $3 a gallon. That is why we have so many products in our grocery stores. We're flying in tomatoes from Argentina. We're flying in, you know, irises from Malaysia. Like, you have this abundance in this choice because there are a group of people who have done really hard things over really long time periods that were deeply unpopular, have commanded immense sums of capital to build things that should not be possible.
00:19:57.950 — 00:19:59.910 · Nick
Right. And that's the same.
00:20:00.070 — 00:20:00.550 · Meltem
The American.
00:20:00.590 — 00:20:02.270 · Nick
Same day. It is the American dream.
00:20:02.390 — 00:20:17.950 · Meltem
That's how I feel about AI. I'm like, what people are trying to do, um, is create things that we believe were impossible, right? So let's think about making sci fi real. Um, it's hokey, but I do think we're living in unprecedented time.
00:20:18.070 — 00:20:24.990 · Nick
As we sit here looking out at the AI build out, which is deeply unpopular.
00:20:26.100 — 00:20:41.140 · Nick
For a few reasons. It feels like it may be like a especially complicated intersection of technology and financial markets. And as someone who has, you know, championed these kind of intersections of technology.
00:20:41.140 — 00:20:42.700 · Meltem
And unpopular things.
00:20:42.740 — 00:20:50.220 · Nick
How do you how do you think about doing that in, in this moment with all the things that are ahead of us?
00:20:50.860 — 00:21:44.490 · Meltem
One of the things that I think is really misunderstood about AI is if we look at the amount of value that AI has created for public investors versus private investors, the majority of value creation today has gone to public equities, right. The Mag seven plus Broadcom AMD micron created $14 trillion of value over the last two years.
Wow. Right. The QS were up 25% sorry in 2024 22% in 2025. So there's this immense capital markets rally that is primarily taking place in 401ks and retirement portfolios. Right? So the benefits of AI, of AI actually have flowed to everybody who's a participant in American capital markets, not just VCs, which in a way sort of makes it like crypto, which is not a point, right?
It was not a insider or like financial industry led movement.
00:21:44.530 — 00:21:52.890 · Nick
Right. And or at least it, it, it broke out of those walls a lot sooner. And in the case of AI, it actually a lot of it started outside of. THE HEDGIES GOT IT WRONG. Leopold was the one who nailed the call.
00:21:52.930 — 00:22:40.440 · Meltem
Because the recipients the value recipients and the value chain. Right. The OEMs, the energy providers have seen a lot of the early capital. It's like reverse capital flow from out from Silicon Valley, right? VCs dump money in. And then it goes to traditional companies that historically haven't been particularly attractive.
So I think it's been really interesting because that doesn't get talked about as much as I think it should. Um, and I do think AI has also created an incredible amount of consumer surplus, of course, because consumer tokens are the most heavily subsidized. It's why many of us have many. We were like, should we get an enterprise, a ChatGPT plan?
And I was like, it's not. It's not as good. No. Not yet. It's actually better to have multiple personal accounts. It's just funny. But there's an immense amount of consumer surplus.
00:22:40.480 — 00:22:42.720 · Nick
There sure is. That's going to keep increasing.
00:22:42.760 — 00:22:43.360 · Meltem
I hope.
00:22:43.360 — 00:22:50.400 · Nick
So. This gets, you know, further deployed. I still think we're at the very, very early innings of, of this touching people like day to day.
00:22:50.480 — 00:24:03.990 · Meltem
I mean, I think what's interesting is we've lived through three computing paradigms already. We started with training. Right. Then it was post training, fine tuning and inference. Now we're going to untick swarms. So what's been so interesting? We've already lived through three eras of computing, right?
Each each era of computing drives new functionality, but also changes the functional requirements of what you have to deliver to facilitate that. I think the next era of computing is really interesting. I think we share similar views that we'll see a shift towards more modular, smaller scale data centers.
That introduces a whole new set of challenges and how you manage that in the physical world, how you manage that from a financial risk perspective and from just a financial model perspective, like you need on site hands operations and maintenance, the scale has to shift. And so I think we're just in the early innings.
The last thing I'll say is language has been the primary compression mechanism we've used for reality because humans. Right. We use language to communicate and site. What I think is also really interesting is thinking about how agents will communicate. I don't know that language is the right compression mechanism.
00:24:04.030 — 00:24:05.790 · Nick
It almost certainly isn't right.
00:24:05.790 — 00:24:08.710 · Meltem
And that's one of my out there ideas.
00:24:08.750 — 00:24:19.500 · Nick
It's really I don't think it's that crazy. I think it's sort of obvious that their agents will eventually invent their own language that will compress information in a different way.
00:24:19.540 — 00:24:24.860 · Meltem
Which dramatically will reduce compute needs. Right? Because at the end of the day, and this is where I'll bring it full circle.
00:24:24.900 — 00:24:27.380 · Nick
Thought about it for that reason. That's what makes perfect sense.
00:24:27.420 — 00:25:09.020 · Meltem
But like okay, an agent right understands physics and thermodynamics. The reality is energy and matter are scarce, right? Information needs to be coordinated and orchestrated, right? Which is why agents are great because they can contextualize like more than a human could. I love the way agents can just, like, contextualize an immense surface.
It's incredible. It's really cool. Yeah. Um, what I think is so interesting is when agents start to get involved in every step of the process, we now see agents doing semiconductor design EDA, right. We see agents doing by putting kernels and compilers like I did not think it would be possible to code a compiler, but it is.
Yep. That's so cool. Yeah.
00:25:09.020 — 00:25:14.940 · Nick
Um, well, and as they do this, they're going to think about the relationship between the scarce physical resources and, and.
00:25:14.940 — 00:25:26.330 · Meltem
They're going to build compression mechanisms because an agent's objective is to perform more work as efficiently as possible. Right. Right. An agent is like the most extreme version of a Benedictine monk.
00:25:26.370 — 00:25:26.850 · Nick
Mm.
00:25:28.090 — 00:25:43.010 · Meltem
Is a weird thing to say, but, like, they are like, what's dopamine for an agent? I imagine it speak highly productive. Yeah. The Benedictines, by the way, we can blame them for clocks. Okay. And being industrious is a virtue. Right. That was.
00:25:43.290 — 00:26:06.130 · Nick
Yeah. I like that. Um. Well, so. So it brings me to something I wanted to to to cover, which is. Right. Um, in the if we look at the capital markets around compute right now, uh, you made, you know, you were, I think early in calling out the, the the debt tightening up around, around the compute market.
00:26:06.650 — 00:26:09.890 · Meltem
People were very angry. I did not understand that.
00:26:09.930 — 00:26:46.720 · Nick
Well, I think you saw it. You see all of this, like a little bit ahead. Um, there's a lot of leverage in the system. A lot of it relies on high token. Like being able to sell tokens for a high price in order to pay for all the downstream infrastructure. And as it relates to, you know, sort of where we sit right at this moment, all of the opportunity of, of this capability is right in front of us in the form of a superintelligence.
There is enormous leverage in the system that, to an extent, is predicated on high token prices. And at the same time, we could have these sort of very deflationary or efficiency oriented forces.
00:26:46.760 — 00:26:53.760 · Meltem
And we're seeing that now. Right. Um, token prices are falling, GDPR prices are increasing, right.
00:26:53.800 — 00:26:54.880 · Nick
Because we're doing more.
00:26:55.000 — 00:26:55.240 · Meltem
Right.
00:26:55.280 — 00:26:56.160 · Nick
With less right.
00:26:56.200 — 00:26:57.240 · Meltem
And suppression is.
00:26:57.680 — 00:27:14.910 · Nick
Right. So what does that what's your view on what that what that looks like. You know, over the next 12 months as we see the anthropic IPO come as we see like I think there's going to be a lot of reality hitting this market in the next 12 months.
00:27:15.190 — 00:29:35.250 · Meltem
So let's go back to the credit comment. It was a Saturday afternoon. I was at a wedding. I think I was like going for a walk before the wedding ceremony. And I was chatting with a friend of mine on the phone about how many banks were taking a step back. Because if you think about many private equity funds and banks, right, in 2024, asset backed lending for GPUs was maybe 5% of their book.
Today it's 30, 40, 50% of their book. It's growing immensely. And there's a lot of risk and circularity within the AI value chain today. And what I think is also happening at the same time is I don't think that the criticism is deployed. Compute won't be valuable. I think the reality is in the first half of this year, people were just plowing money into anything in the AI space that looked or smelled like a NIO cloud.
But the ability for people to execute effectively orders of magnitude apart, right? There are some companies, like the causeways and the crusoe's and the metas and the hyperscalers, who are good at deploying compute. Right. There are other companies, a lot of these smaller, unproven NIO clouds who have taken on a lot of debt have never deployed a single GPU.
They're three months behind schedule, six months, 12 months. Right. Um, what's so interesting? I would not have known this until we went through developing and deploying our own cluster. But when people show us spreadsheets now of their financial models, I'm like, you're off by margin of like 50%. Nothing's going to happen on time.
Everybody's lying their faces off, right? It's like what I experienced in the shale boom, what I experienced in Bitcoin. People are just making shit up, right? They they couldn't operate a paper bag, but they're like, yeah, we've got kola. We've got liquid cooled kola. It'll be live March, you know, March 1st, 26 and you go and you look and you're like, you don't even have power from the utility.
And it's February like it's absurd. And so I think what just happened is people who were extending credit started to see this divergence. And so just like we have a K shaped economy, I think we have k shaped execution capabilities within these neo clouds to turn capital into profit. Right. And to meet their debt obligations and debt covenants.
So what happened is very natural. It's a repricing of that market structure. That's it. Right.
00:29:35.330 — 00:29:37.410 · Nick
And there'll be winners and losers in that market.
00:29:37.450 — 00:30:02.610 · Meltem
And there are companies who will continue to get access to low cost of capital. And there are companies who won't. The same way Exxon could get capital at 3%. Right. A wildcat driller was borrowing at like 30%, as it should be, because one has proven execution capability and the other doesn't. One has a balance sheet.
The other, like this, is intuitive, but somehow people are really angry. I was like, I don't know what's going on.
00:30:02.930 — 00:30:11.560 · Nick
I mean, I think it's probably healthy and it's part of the moment of of all of this becoming more real that there will be a shakeout of some kind.
00:30:11.600 — 00:30:21.440 · Meltem
The anthropic asks, one I think was really great because it's the first time we're really understanding the underlying economics of a frontier lab. Guess what their business model is?
00:30:22.000 — 00:30:23.960 · Nick
Um. Good question.
00:30:24.000 — 00:30:29.400 · Meltem
Turning GPU hours into tokens. Of course it's a refining business. It's it's a commodity business.
00:30:29.440 — 00:30:38.400 · Nick
There it is. Yes it is. It's it's refining. It's right there on the chart that we were looking at yesterday is the dollars of revenue per, per dollar of compute cost, right. Which is.
00:30:38.440 — 00:30:38.840 · Meltem
Modest.
00:30:38.880 — 00:30:40.600 · Nick
It's that. Not exactly that.
00:30:41.000 — 00:30:47.160 · Meltem
Yes. And here's the beautiful thing. Commodities are what changes the world.
00:30:48.480 — 00:30:56.040 · Nick
Indeed. And there's a commodity for each generation of technology for each each era of technology. A lot there's a lot.
00:30:56.080 — 00:30:57.920 · Meltem
But you and I need compression.
00:30:58.040 — 00:31:06.800 · Nick
Yeah, exactly. Um, let's look ahead to the future. Right. There's this sort of new asset class that's forming around compute. The core commodity of the AI era.
00:31:07.150 — 00:31:11.510 · Meltem
Um, there's one more commodity. What do you need to make the chips?
00:31:11.670 — 00:31:12.150 · Nick
Sure.
00:31:12.230 — 00:31:13.630 · Meltem
The electrons. Right? Yeah.
00:31:13.670 — 00:31:19.670 · Nick
Yes. Energy. Energy and compute. Are the are the the commodity pair? Yeah. That's underlying.
00:31:19.790 — 00:31:45.990 · Meltem
And I think the reason the energy component is important and you know this because you've done a lot of energy investing. We have energy markets today. We have oil and gas markets. We have electricity markets the ISO market. But I think we will see the creation of new energy markets for behind the meter.
Right. A new generation of IPS. So I think energy will be a really interesting space with some new markets.
00:31:46.350 — 00:32:06.210 · Nick
And that was actually my question, which is what other markets, what other commodities or other markets may form. Right. So it's like pockets of energy and new ways of trading energy. We've got the compute side. Like have you. Are you thinking ahead to the next set of commodities. Yeah. That may be part of the sort of The AI era.
00:32:06.610 — 00:33:15.360 · Meltem
Absolutely. So we did invest in a plastics marketplace for resins. Um, I think even in the traditional world of commodities, there are many opaque markets that don't have any market structure. One of the reasons I wanted to invest in a company fabricating nuclear fuel for Smrs, there's no established market structure for uranium, whether it's U.
308 Halo Tri. So that market is like completely illogical, right? The way it's price is really bizarre. Um, there's a great podcast by the gentleman who runs, I believe it's Section Cove on uranium market microstructure so wonky. It's 30 minutes. Highly recommend. It's like such a niche thing. These are the things we love.
Materials don't have marketplaces. We invested in a company that facilitates hedging of. They started with metals, but they just booked the first ever tungsten hedge because tungsten prices have increased fivefold over the last few months, and they did that using a synthetic derivative created by them.
Listed on Kalshi with Susquehanna, a large market maker, taking the other side.
00:33:15.560 — 00:33:21.400 · Nick
So in the relationship between these materials markets, there's all inputs to the to the AI.
00:33:21.560 — 00:33:58.480 · Meltem
Correct. So think about copper. Right. There's an immense amount of copper in a data center from the rocks, which have an immense amount of copper in the bus spars to the wire and cabling GPUs themselves. There's now a lot of investing going into creating more supply of, um, SIC and Gans. So silicon carbide, gallium nitride.
Right. Uh, new substrates for silicon, um, new substrates for computing, uh, superconductors. Right. All of these things require inputs for which there are no markets. If you traffic in that right, if you're building, for example, solid state PD use and you need gallium,
00:33:59.640 — 00:34:06.710 · Meltem
how are you going to hedge that. How are you going to guarantee that you can profitably produce this? You need a market, right?
00:34:06.750 — 00:34:08.070 · Nick
So has derivatives.
00:34:08.510 — 00:34:32.070 · Meltem
So historically right. People have just discovered Marc Rich and Trafigura because there's this great book, The World for sale, which is so well written. Um, but commodity traders. Right? Historically. Trafigura. Mercurial. Vital. These firms have played this role. Yep. So in the age of agents, everybody can have a mercurial at their disposal.
Because an agent can trade options and construct.
00:34:32.070 — 00:34:43.190 · Nick
Hedge all day long. Looking ahead, trying to get more wild on the future. So agents as traders. Right. So so markets for every new commodity agents as 24 over 7365 traders and market makers.
00:34:43.190 — 00:35:05.020 · Meltem
I want to just, um, I want to provide some important nuance here. I think we're probably a few years away from agents, quote unquote, traders. I think they'll be optimized, they'll suggest, but they'll still require human in the loop, most likely, or they'll be given very strict parameters, I think, until we have a better idea about whether or not agents can in fact be contained, or if they.
00:35:05.060 — 00:35:05.620 · Nick
Are sure.
00:35:05.660 — 00:35:06.420 · Meltem
Right. I think there's.
00:35:06.420 — 00:35:08.020 · Nick
A lot going to give them a big checkbook.
00:35:08.060 — 00:35:31.540 · Meltem
I think there's a lot of concern within trading firms, right, about how to manage and sandbox, of course, agents. But I think as optimizers, right, as you know, an agent can proactively monitor even Sarah go from conviction had a great example where she's like, I keep my money in a bank account that has very low interest.
Instinct suggested I move it to a high yield checking account.
00:35:31.580 — 00:35:36.780 · Nick
That's an easy yes. Correct. Right. And as a suggestion, you know, with a human in the loop, that's.
00:35:36.860 — 00:36:25.730 · Meltem
But for your data. Right? Even if I'm financing a factory, if someone, if my agent could go and shop for equipment lending rates, then I don't need to raise as much venture capital. I can get asset backed credit if I need trade financing. Right. Somebody could go out. They know what I'm trying to do. I voice prompt an agent.
I say, here's what I'm trying to do. It goes in and shops for credit with 70 different shops. Right. In crypto, by the way, there's a company that did this called paradigm. They created a, um, trade workflow platform for large block trades. So they allowed for people to RFQ or request for quote block trades of options.
And they had this dealer network where they brought on all these institutional counterparties for the first time. You could say, hey, I want to buy a block of this and you got a bunch of quotes back all at once to me, like, that's the unlock with agents is every firm can have agents that are quoting in real time.
00:36:25.970 — 00:36:33.890 · Nick
Um, if agents may invent their own language, um, do you think they'll also start inventing their own financial instruments that we can't imagine today?
00:36:35.850 — 00:36:54.370 · Meltem
I would assume, yes, but I think I'm going to say something maybe unpopular, I don't know one. One thing I think about is thermodynamics, right? There are these fundamental laws of the universe that we have not been able to defy in our known reality.
00:36:54.370 — 00:36:57.120 · Nick
Sure, physics always wins, isn't that? Yeah.
00:36:57.160 — 00:37:45.480 · Meltem
That was sort of my mantra when I started. Crystal riders, like the laws of physics are really just there. Yeah. Um. What I learned about the last, I guess, working in oil and gas and then crypto, there was a lot of irrational exuberance in crypto and this attempt to do financial alchemy. It doesn't work.
Those things haven't worked. All of those tokens go to zero because finance has its own fundamental laws. Sure as well. Right. And what I love about tech is like every ten years or so we're like, no, no, no, we can defy the laws of finance. But you can't. Right. Right. Yep. Um, so while I do think agents will potentially create new instruments, finance also has its own physics.
And I don't know that agents will break that. They are still bound by those same constraints of how markets function.
00:37:45.920 — 00:37:47.080 · Nick
Okay. So to try to take.
00:37:47.120 — 00:37:47.920 · Meltem
That to be a fair.
00:37:47.960 — 00:37:59.710 · Nick
Even. Well, I like that. Let's take it to a weirder place. If agents may not create their own financial instruments, Maybe. Could they create their own version of an economy?
00:37:59.750 — 00:38:02.870 · Meltem
They'll have their own markets for sure. And I was so curious.
00:38:02.990 — 00:38:04.230 · Nick
Different laws of physics.
00:38:04.390 — 00:38:22.590 · Meltem
Um, maybe. I think what is interesting, what I've been thinking about lately is, um, interfaces for agents, right? Agents don't need an interface. One thing I've been contemplating is currently my website is built for humans, but most people will use agents in the future, right? So I'm like, my website should just be a series of markdown files.
00:38:22.630 — 00:38:24.350 · Nick
Sure. And right now.
00:38:24.670 — 00:38:26.870 · Meltem
I know that human. I've also reviewed my website.
00:38:26.910 — 00:38:41.070 · Nick
And a lot of websites are having an agent's MD or some other, you know, text version. That's the secondary version. Like, I think Matthew Prince from Cloudflare said the other day or recently that I think it's like maybe 40% of the traffic on the internet right now is agents. I think.
00:38:41.470 — 00:38:42.750 · Meltem
I heard over half.
00:38:42.790 — 00:38:43.790 · Nick
Maybe it's 60%.
00:38:43.870 — 00:38:44.230 · Meltem
Which was.
00:38:44.390 — 00:38:45.430 · Nick
Certainly that's going to go.
00:38:45.430 — 00:38:45.870 · Meltem
Up.
00:38:45.950 — 00:38:47.230 · Nick
Substantially into the.
00:38:47.270 — 00:38:48.110 · Meltem
So then the question.
00:38:48.110 — 00:38:48.670 · Nick
Is.
00:38:48.790 — 00:39:43.660 · Meltem
I think for capital markets, the question is like an agent doesn't need an order book. Does it need a trading interface? Um, the interesting thing to think about is abstraction layers, right? So even the way that agents interact with GPUs today, they need an abstraction layer. They need an inference code layer.
Right. A whole generation of companies, some of which you invested in, were like the pioneers of IAC, right? Think about containerization in terms of deploying a workload on CPUs. Right. Containerization is what make cloud possible. Similarly, the question I ask is what are the abstraction layers that need to be created that will allow agents to provision resources on command using some form of currency?
Right. Whether it's stablecoins or something else or IOUs, which isn't actually the substrate, does it matter that much? But it will be a very liquid market and you will need to be able to pay as you use. Right. You won't pay to provision, you won't prepay, you won't have a subscription. You'll spin it up, you'll spin it down.
Right. Today sandboxes I think, are.
00:39:43.980 — 00:39:44.300 · Nick
Close to.
00:39:44.300 — 00:40:05.130 · Meltem
That. Yeah. Or dedicated VMs. But that layer is really interesting to me. And thinking about what even is that abstraction? And then how do both humans and agents interact with it? We're like humans are not touching code as much, right? What are interfaces going to look like in capital markets?
00:40:05.170 — 00:40:15.130 · Nick
Right. For that are more agent first. The same way you're going to start to have compute environments that are maybe agent first and what are sort of capital market environments. But potentially agent.
00:40:15.130 — 00:40:34.770 · Meltem
Versus the linchpin of all of this labor is verifiability. Yes. And so you do need the ability to verify and have a human verify right. And have this. But I think this interface is question for me is like the big thing I wonder about is what does an agent market look like? Um, I don't know, but I'm enjoying exploring that concept.
00:40:34.810 — 00:40:39.130 · Nick
It's markets all the way down. Um, and maybe all the way forward.
00:40:39.330 — 00:41:44.280 · Meltem
If you think about it, though, humans have been doing this since the dawn of history, right? There are scarce resources. They tend to be energy materials. Right. We use information as a way to determine how those resources can be best allocated. But open markets, in my view, are the most egalitarian mechanism to diffusing the effects of technology throughout society.
And so as I look at what's happening now, um, I think it is so important to take a market driven approach because markets are going to be the best mechanism for distributing the economic effects of AI as broadly as possible. And so, um, it'll be interesting to see what happens. Like Crucible is a small firm.
I'm very excited by what we've been able to do over the last two years. I'm excited about what we might be able to do going forward. And there's a lot I don't know. I'm just going to keep unfolding, seeing what I find and following wherever my curiosity leads.
00:41:44.320 — 00:41:48.990 · Nick
That sounds amazing, and hopefully we can come back here in another 10 or 15 years. And.
00:41:49.030 — 00:41:49.830 · Meltem
Well, our agents.
00:41:50.070 — 00:41:51.990 · Nick
Right. Of course it won't be us. Yeah.
00:41:52.110 — 00:41:53.950 · Meltem
No, I hope it'll still be us.
00:41:53.990 — 00:42:08.590 · Nick
Yeah. Um. Thank you so much, Meltem. This has been really amazing. And every time we get a chance to talk, I'm. I'm kind of blown away by how deeply you think about all this stuff and the level of connection you have on the ideas.
00:42:08.830 — 00:42:19.350 · Meltem
Well, it's also new to me. I don't I don't have light bulbs. I just have like, oh, yeah. Oh, like, I didn't know that. So thank you for giving me the container to connect ideas.
00:42:19.350 — 00:42:19.670 · Nick
It's my.
00:42:19.670 — 00:42:20.870 · Meltem
Pleasure. It's so good to talk to you.
00:42:21.350 — 00:42:21.790 · Speaker 4
All right.
00:42:23.470 — 00:42:24.030 · Meltem
Nerds.
00:42:24.070 — 00:42:24.710 · Nick
All right.
00:42:24.750 — 00:42:25.190 · Meltem
Oh, nice.
00:42:25.470 — 00:42:26.790 · Nick
That's amazing. Yeah.