Loading summary
Transcript59 lines
- [00:01]
A
This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome, that's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50 page restoration block. Or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it, ready to make anything online make sense. There's no place like Chrome. Check responses set up, required compatibility and availability. Various 18 plus.
- [00:27]
B
Introducing Meta Glasses. You have questions. They've got answers.
- [00:32]
C
Hey Meta, what's the capital of Peru?
- [00:34]
A
Lima.
- [00:35]
B
How do you say where's the restroom in Spanish? Dondes talbano. Hey, Meta is a hot dog a sandwich? Technically, no. Spiritually, yes. Hey Meta, what should I do with my life? That's one of life's biggest questions. What do you think? Ask anything. With the new Meta glasses, the bitcoin miners I know are all shifting to AI. I think you've covered this as well. And it's because they have access to power and that is the bottleneck. And you just make more dollars per kilowatt hour going into an AI compute than you do in bitcoin ultimately. And B, because power is the constraint, the value flows to what's scarce.
- [01:19]
C
Bang, bang. Today, guys, we got a great conversation with Ramez Nam. He's an investor at Planetary vc. And in this conversation we're getting deep into the weeds of what's going on in the artificial intelligence space. We know that energy and power are a massive bottleneck and Ramez is here to explain exactly what the potential solutions are, who's likely to win, and what are the various companies doing so that we can get more power to get to super intelligence. On top of that, we talk about how super intelligence is actually going to be applied. It's going to be general or narrow. What is it going to mean for you? And, and what are some of the companies that he's excited about that are actually helping to usher in a world where super intelligence makes you and I happier, healthier, wealthier and richer. Sounds good to me. And Ramez is here to explain it all. Here's my conversation with Ramez Nam. All right, Ramez, there's a massive problem in the AI industry. Energy and power are the bottleneck at the moment. Can you please help us understand why is this problem persist and what are the potential solutions as to how we're going to get around this?
- [02:12]
B
Yeah, absolutely. Like AI is hugely compute intensive. The way that we've been making AI better is the scaling laws. And the scaling laws look awesome at the beginning, but they're brutal. It means you have to keep doubling the amount of compute to get linear gains in AI performance. So there's this huge race to do so and we're just not used to expanding the grid at this pace. In the 50s we built out the grid 5, 7% growth a year, but for the last two decades it's been zero growth. So we're just not accustomed to putting on like new power or new sources of demand at this pace. So the, the poles and wires themselves are the bottleneck right now. Chip production is about twice the pace that we can hook things up to the grid and that's the bottleneck everyone is trying to work around.
- [03:00]
C
Now when you think about that bottleneck, can you walk through, there's different companies with different potential solutions for this. They're all kind of racing to see not only what solution is going to be the one that is the winner, but also how do they fund it, how, you know, there's a lot of stuff, Supply chain issues.
- [03:14]
B
Yeah.
- [03:14]
C
Where maybe what are the ones that you're most excited about or the ones you think are most noteworthy as a solution?
- [03:20]
B
Well, let me walk you through, like the overall set of options people have. Option one is you just, you're going to build a data center. You put in a request to the grid to get hooked up and a few years ago that was fine. Now it's maybe a five to seven year wait. Option two is behind the meter power. So people are trying to buy natural gas turbines to build on site at their data center to run their load. The big turbines made by companies like GE are sold out five to seven years. So now they're turning to smaller turbines that are maybe a tenth, a twentieth the size. And a whole bunch of these things on the back of semi tractor trailers, basically. And you have companies, boom. Supersonic was trying to build a supersonic jet a few years ago or as recently as six months ago. They've pivoted to taking their engine and, and making it a power plant for behind the meter power for data centers. Option three, one that I really love is batteries. Just being really clever, we build the grid out for peak demand for, you know, late afternoon in the summer when the AC is on. But most hours of the day the poles and wires, which again are the bottleneck, are not loaded. So you take a battery, install it at your data center, fill it up at night, don't need to touch the grid in the afternoon, and suddenly you can get powered on a lot faster. We've Just passed some regulatory changes for that. And then you get into the sci fi stuff. You get into Elon talking about space based data centers. Nothing about the physics says that can't work, but the cost of launch has to come down a whole lot. Or one of a startup that I'm invested in called Pantalassa making floating ocean data centers that get power from waves. And the, the benefit of both space and ocean is it's going around regulatory, it's no longer permitting and local opposition and wait times that connected to the grid.
- [05:20]
C
Now let's talk about a couple of these like going and just applying for power and kind of doing it the traditional way. Not very interesting. I think people understand that. And obviously there's those long wait times. You talked a little bit about these like almost mobile centers. One of the things I found really interesting is when Elon went to go build the original Colossus, my understanding is that they essentially just brought in a ton of generators and then he just lined the entire facility on both sides with these generators and kind of like hacked it together. It was kind of like a do it yourself project if you will. But it worked and they got up and running. And so how sustainable is that type of stuff where people are essentially just cobbling together different solutions and they're able to just get up and running. Is that something that could persist for years?
- [06:01]
B
It works in some places. About a third of data centers that we think will get built this year are doing some version of that. Maybe as many as half next year. But there's a lot of debate when you do that. You probably have to overbuild by 30, 50% to have sort of the spare capacity and you end up with problems. You end up with local air pollution issues like those power plants, Colossus, Colossus 1 probably exceeded their permits. There were lawsuits. You end up with noise when you hear people complaining about data centers making noise. Data centers themselves are quite quiet, but these generators behind the meter are really loud. So this works best in rural areas out in the middle of nowhere. And so you'll see people doing that. I think almost every data center really wants to have a grid connection. Ultimately nobody wants to manage their own power plant. But people, your time to power is everything. Let me put this another way, just economically for your listeners here. If you look at the cost of electricity going into a data center versus the revenue it generates, it generates 20, 30, $40 of revenue per dollar it spends on energy. So time to getting it hooked up is everything. And you'll Pay more for energy. You're paying more if you're doing these small generators on site than if you would go to the grid. But it's worth it because you're going to make so much money off of it.
- [07:25]
C
One of the aspects of the data center build out that I think most people have missed is I've seen a number of examples where these data center provide show up somewhere and they say, hey, I need to get more electricity. The electrical company is like, well, we just can't get it there. We don't have the infrastructure or whatever the thing is. And the data center company literally says, I'll build the infrastructure, I'll pay for all of this build out. I'm happy to have everyone else get the upgrade as well who uses this grid. But if it helps me get online faster, it is worth that expense. And so that seems to be a strategy that a lot of people are using as well.
- [07:55]
B
We're getting there. And all the data center builders have signed this voluntary pledge that they're not going to raise prices for their neighbors. And in fact, you know, contrary to the public perception, places where we've seen the greatest data center buildout in general have lower electricity prices in the country and have not seen them go up. But that said, there's actually like regulatory challenges here. The utility rates are set by the utility commission. There's actually work that has to happen to make it possible for, for the utilities to let the data center builders pay for all this stuff and not pass on the costs. So everyone's scrambling to figure that out now.
- [08:32]
C
Got it. Now let's go and talk about maybe the orbital data centers first. So the promise here is, if I was to regurgitate back Elon's thesis, if you will, is you need power. Well, where is power persistent at all times from the sun, as long as you are far enough away from Earth in space that you can capture 100% of the Sun's energy. And if you can capture that via solar in space, and then you can have basically a data center right there at the point of energy capture, you can then beam the data, the connection, etc. Back down to the Internet. He's proven that he can launch rockets, he's proven he can reuse the rockets, he's proven that he can beam the Internet down from these kind of low orbit satellites. Do you think it's possible to actually build the orbital data center? Like, I don't think there's actually one that exists yet, but it seems like the math works. It's possible.
- [09:20]
B
The math doesn't work on a price basis yet, but it could work with Starship. So look, a Starlink satellite doing Internet is kind of an orbital data center. It's using its compute to handle telecoms and compression and channel hopping and so on. Not for computing, AI, but it has some compute on it. And we've had the first example of a satellite launch that has a GPU on it doing that sort of thing. So nothing says it's impossible. When you do the math, you, you've got to bring launch costs down by somewhere between a factor of four and a factor of 10 from where they are now to compete against costs on the ground. Starship should be able to do that, but it's going to require that we have, you know, two stage reusability. They got to catch both the upper and the bottom and they've got to be able to launch a lot of starships. They've got. You got to get to the point of launching Starship multiple times per day for that launch cost to get cheap enough. But the flip side of that, here's a way to look at that. If you think demand for compute is basically infinite, if you think it's unlimited, eventually you just run into limitations on the ground. And so even if it's more expensive to go to space, if that's the only place from a regulatory standpoint that you can go, it might make sense. Now, I don't actually buy that argument. I think there's a lot of other places we can figure out to put compute, but that's part of the beta Elon's making. I'll make one more argument. Flip side of this. Elon's wanted to go to Mars, right? You can't. There's no business model for going to Mars. He wants to get Starship down to these incredibly low launch costs, a tenth of what it costs today. That only works if you're launching again, you know, 1,000 Starship flights a year or more. There is no demand. There's not enough demand for launch, for telecoms, for Starlink type services to actually scale Starship to the fleet size and launch cadence that he wants. So in a way, this is a godsend to Elon. It's really smart. Like, Elon originally wanted to go to Mars. He didn't want to build a satellite telecoms network. He built Starlink because it was the business model that worked for launch. And so now orbital data is a business model that can actually justify enough mass to make Starship make sense. Flip side is, you Know, we've got a whole lot of desert on planet Earth. And it's probably for the time being, until Starship is at that scale, it's probably cheaper to, you know, cover the deserts of the world with solar and batteries that work 247 to build these data centers. You see the first signs of projects that pencil like that in the uae. In Chile, we have solar plus battery projects that are baseload solar that are cheaper than behind the meter gas. So I expect the world to start doing some of that for data centers as well.
- [12:15]
C
Now, when you're thinking about this stuff in space, the two things that I've heard people critique this is the cooling is still highly debated as to like, is it easier or harder in space? Maybe you have some opinions there. And then the second thing is, how do you fix the GPUs or the machinery if there's some sort of issue with it?
- [12:35]
B
Yeah, cooling is just physics, right? So space is an insulator. There's no, you know, convective cooling. There's no conductive cooling. You touch space and no heat leaks out. But you build giant radiators, big aluminum fins with special coatings on them. They radiate heat. So it just comes down to sort of a mass equation, and that comes down to cost. You just have to launch a lot of aluminum fins to be able to radiate that heat away. And that just adds to the weight and adds to the cost. You can do it. Now you do have some other complications. You're going to have to have these cooling loops, you know, filled with maybe something like ammonia or some of the coolant and pumps and circulators and compressors. And that adds to the moving part count, which again, adds to the possibility of things breaking. And so we don't know. The maintenance question is one of the most unanswered. We can do the math on cooling and say, okay, that just means launch cost has to be so much lower. I don't think we have good data on how often things will break there. And in particular, if you look at Elon's plans, the hotter you run a chip, the easier it is to radiate away the heat. But the hotter you run a chip, the more failures you have. So he's talking about running these chips, you know, 25 degrees Celsius, 50 degrees Fahrenheit, hotter than they run in data centers on Earth. And that's a challenge, ladies and gentlemen.
- [13:52]
C
I'm a software guy, but every once in a while I come across a piece of hardware that I just simply can't live without. And that's PLOD for me now. Now this whole thing is basically a hedge against me losing my memory. Because when I have all these conversations, I talk to founders, investors, all kinds of people every single day. There's three things that I always want to take away from the conversation. What I learned, what are the next steps, action items, and then what notes do I need to remember for some period in the future? You can even mark important moments during the discussion so that you can quickly reference them later, whether it's a founder meeting, an investor call, a podcast interview, or maybe even just want to sit down and talk to your spouse. Or I could instantly go back and forth and find key insights, decisions, action items that I talked about in that conversation. Over time, every conversation all of a sudden becomes an asset instead of a forgotten moment. And in a world where information is abundant, the ability to retain and leverage knowledge is a very real competitive advantage. That's why Plaudnote Pro caught my attention recently. It's a credit card sized note taker, slim portable, perfectly integrated with my phone. And the part that I love about it is that it automatically captures conversation, turns it into board ready summaries. You can have structured action items and it does. Flawless transcription is really magical and you should try it out. So go check them out. This hardware has completely changed my life and I think that you'll end up finding it very valuable. So if you want to use PLOD and get the same value out of it that I've gotten out of it, go check them out at plaud AI/ pomp. You can go to plaud AI/ pompomp and get 15% off if you use the code POMP or just go click on the link in the description. So go to Plaud AI slash POMPOMP today. Yeah, it's fascinating to me because I guess part of this is also if you can start to shrink the size of the machines that need to be up there. Like everything is just compression, Compressing timelines, compressing cost, compressing size, compressing, you know, all of this that really feels like the world is about to spend, I don't know, trillions of dollars just compressing this entire thing to make it as efficient as possible. Right?
- [15:42]
B
I mean, that's what we've been doing, you know, like we're making sand sentient, as they say, right. So we've been taking information and compressing it in physical stuff, the ability to compute into smaller and smaller packages. But that has meant that while total energy use Per unit of computation goes down, the density of it goes up. So the temperatures of these things go up a lot. I'll pitch again. One of my startups, Pantalasa, they do this at the ocean, Right. So the ocean guys.
- [16:09]
C
Yeah. What do they do?
- [16:10]
B
Well, they're, we just announced a big round of theirs a little while ago led by Peter Thiel among others, they build these floating ocean data centers that bounce on waves. And the biggest waves on the planet are around Antarctica. So these are floating data centers that go up and down in giant ocean swells and they generate electricity from the water being sort of forced up a tube. And we think they're super cheap. And the cold ocean waters give you basically cooling for free. So it's another bet in many ways it's a bet like space based solar. Let's get past the limitations of the grid, let's get past local permitting and zoning. Let's get back local gets past local pushback and go to our frontier where there's a lot of energy and where there's not a lot of obstacles from a regulatory standpoint.
- [17:02]
C
Let's talk about this a little bit more. So there's basically like a big giant ball. I almost think of it like, you know that movie like Bubble Boy or whatever it was, Right. But inside of it you have a data center. How as the ball is being pushed on the waves, what is the connection to the electricity generation? Is there some sort of like wheel or something that is like capturing the motion?
- [17:21]
B
Yes, it's amazing. So what you see in the picture is, is a sphere at the top, but then you have a tube that's gently tapering. It's like a football field long, 85 meters long, that goes down, the waves send it up and when it comes down, water is forced up that tube and that water turns a turbine and the way they've got it, some very just clever systems of tubes and pipes, if you will. You can have 90, 92, 95% continuous power output from that system with almost no moving parts. And you don't even need a cooling loop. The GPU is just plugged into, you know, a heat sink that conducts electricity straight to the metal walls of this stuff that goes to this super cold ocean water and cools it.
- [18:12]
C
Now what happens if you get calm waters? Or are there no calm waters in Antarctica?
- [18:17]
B
In the Southern Ocean, a calm day is like the worst day on the, you know, waviest beach where people live.
- [18:28]
C
Got it. So there's always motion that's going on and that allows you to continue to do this. Now, if something is 85 meters deep in the water, like, I don't know, it's Titanic, what if it hits something? Or can these things break? Or how do you think about some of these edge cases when you're underwriting an investment like this?
- [18:43]
B
The ocean's really big. So I like these things, while enormous in some sense are tiny compared to the scale of what's out there. They have the ability to steer themselves to some extent. They mostly follow the current, but they can, they can set their course. And all of everything we build has some failure rate in space. We'll have risk of orbital debris, we'll have things break for various reasons, we'll have solar storms. But if you distribute the compute enough, losing one node is not catastrophic.
- [19:15]
C
And of these spheres that are in the water, do you have to build an entire network of them where they communicate with each other? Or can you just have, you know, one and there's really just like that thing is beaming connectivity or power or compute to. To an end customer?
- [19:30]
B
Yeah, that's a really good question. So when we look at AI compute, we're talking about data centers doing two different things. One is training the AI models and the other is inference serving up when you Go chat with ChatGPT or Cloud or whatnot. For training you need massive data centers with incredibly high speed interconnect like optical connections between all the racks. Maybe a gigawatt in one place is what people want. And so you're not going to use distributed systems for training, but for inference, when you're just asking a question to an AI model, you might be talking to one rack of servers, maybe talking to 20 GPUs at once. And so that can be pretty easily done with distributed model.
- [20:13]
C
Got it. Let's go back to the grid for a second. I want to throw a couple of different ideas. Some I've invested in, some I've looked at and not invested in, but I think are interesting. One is base power. They've got this kind of decentralized system that they've built. They're essentially selling. The way I describe it is like a Tesla powerwall which people are familiar with. They've got their own version, it's super cheap. You attach it to your home and then they are allowing you to hold power. Sell it back to the grid, you, you can use it if there's some sort of blackout. It's almost like a decentralized electrical grid is the way I think about it. Is that going to eventually be the default type of system for the entire US electrical grid or do you think that only is going to work in certain states?
- [20:51]
B
Base is amazing. Very, very smart team. I'm acquainted with them. They've done some really clever stuff. Their model works because Texas, where they had started, is a deregulated energy market. So they're, you know, giving you a battery for. I think the current quotes I've seen are nine bucks a month and you get backup for your house. Why does that work? It works because they're using that battery to buy and sell power on the grid. To buy power when it's cheap, sell it back when it's expensive and they become your electricity retailer. And so that business model all around pencils for them like you're paying them nine bucks a month, but in some sense they're using your house to build a small power plant. And it's a battery that, you know, pulls in and out power and they're taking over. They become your utility that you're paying. That business model works in eight US states, Texas being the largest. So ercot, the Texas power grid is a deregulated market. And it's kind of amazing because it's a very free market system where both on the retail side you can pick who your power company is and they compete on prices. And on the wholesale side, on the generation side, anybody can build a power plant and just sell power onto the grid. It's incredibly competitive. In fact, Texas is the number one state in the US for solar, wind and batteries without any specific policies to advance them. It's just a very, very competitive market. And so these things win on cost. So I'd love to see more of the US adopt a model like that.
- [22:28]
C
Now there's companies like Giga Energy in Texas. What they are doing is these kind of modular. They started with bitcoin mining. Now they've really went hard at AI infrastructure and think is interesting about them is they really started with building the hardware, but now they're doing powered land and really saying, well, maybe we need to kind of eat the full stack when it comes to the development of the data center in this way is that the natural progression for a lot of these players is to eventually go from hey, I've got one kind of corner of this supply chain or one piece of this and then now I've got to go and be able to deliver kind of an entire project.
- [23:00]
B
Look, I think it makes sense for a lot of reasons. I think we will still see multiple business models make sense. But a the bitcoin miners I know are all shifting to AI. I think you've covered this as well. And it's because they have access to power and that is the bottleneck. And you just make more dollars per kilowatt hour going into an AI compute than you do in Bitcoin ultimately. And B, because power is the constraint, the value flows to what's scarce. So the ability to get things hooked up to power or the ability to route around power limitations of the grid, that becomes a limiting factor where the value flows.
- [23:41]
C
Now another company is American Consolidated Energy or Electric. I'm sorry. And so when you think of, of that business, a friend of mine runs it and I think American Consolidated Electric started looking at, well, turbines, switchboards, you know, these various different components in the supply chain, those are in very high demand, very low supply. There is a promise of yeah, sure, maybe you'll get it in three years or something like that. And so in a weird way, it almost feels like a business that is going after the supply chain and saying that you're going to sell actual hardware components to these data centers or to these land developers. That's like selling the picks and shovels back in the gold rush. I'm surprised that we don't see more companies stepping in to solve those bottlenecks.
- [24:24]
B
We see some of it, but you know, transformers classically are in short supply right now. So all of these components are in short supply. These come, you know, often from companies that are not public, that are sometimes family owned, old companies that are a little risk averse to building out a new assembly line or increasing their supply chain. But the prices are just going so high that we're seeing more and more investment both from the existing players and from new entrants that are saying, well, I can supply that. We're also seeing things are changing. Electrical system, the power output from a solar plant, for example, is DC. The grid is AC. At the end of the day, the GPUs are consuming DC. So we have all this conversion back and forth. So we're seeing people talk about entirely DC based data centers. We're seeing people talking about increasing the voltage in a data center because it reduces how much copper wiring you need and makes the whole thing cheaper and simpler to assemble. So the whole world is routing around and trying to orient itself around these data centers that are, they're not yet an enormous amount of our energy consumption, but they're incredibly concentrated load. They have like the highest power density of any consumption of energy. How do we make that all work better? How do we move to these high power systems. How do we solve the bottlenecks of components that people just weren't buying in such volumes a couple of years ago?
- [25:55]
C
Now we've been talking a lot about the electrical grid and some of the bottlenecks in terms of power and energy. Let's talk about where this is all going. I think there's a lot of folks who believe that there is going to be some sort of like general super intelligence and that general superintelligence. You know, we're close, we're going to hit it, we're going to take off and the world's going to be amazing. It sounds like you're a little bit bearish on that being the potential path that we take. But you do believe in this kind of like narrow super intelligence. And so describe the difference between general and narrow. And why do you think NAR is the more likely path?
- [26:28]
B
So I'm one of the few people in the west coast who doesn't think we're about to hit a singularity. And, and asi, that just does everything for us magically. Utopia or dystopia. But look, AI is limited in a large sense to the intelligence founding its training data. And so you can train AI. Today we're training AI off of basically all written human work. We're close to that. Where we might be in that now, maybe 20, 28, we will have trained off of basically all the books, all the Internet, all the scientific articles. And, and AI can only really get as good in some sense as what exists there. It's also limited by how much noise and error there is in the data. So you crawl everything on the web, you're going to find some stuff that is amazingly insightful, you're going to find some stuff that's utterly crap. The places where we see AI become truly super intelligent are these formal domains that we think about as highly verifiable. Right. So the first things that AI beat us at were games. Why was AI machine learning models able to get so good at chess or the game of Go? Well, you can play an infinite number of games against yourself, so you have no limit on how much data you produce to train on. And two, you know, instantly in like with 100% reliability. This was the right answer. This is the wrong answer. Did I win the game or lose the game? There are very, very few domains like that, domains that are like that. Formal math. So we're going to see AI actually help the world's top mathematicians solve math problems. And coding is kind of like that. Coding is A little bit more messy. The requirements for coding are still spoken in natural language, but you can at least tell like did it compile, did it work, how fast did it run, and so on. But other things like write me a novel or run my company or figure out the right foreign policy decisions for the US and Iran, those are incredibly messy. You don't have a way to judge quickly in a million simulations did I get the right answer or the wrong answer? And so it's just much harder to produce this totally formal, totally precise and nearly limitless amount of verified training data, which is what makes AIs truly superhuman.
- [28:45]
C
Makes Makes sense. Now how much of the narrow super intelligence is going to be based on specialized workflows and you know, a lot of the like reinforcement learning that comes from experts inside of a specific industry or the feedback and, and a lot of the stuff that we're starting to see now where I think there's two bets, there's like let's go build general purpose models and chase that general superintelligence and then these narrow superintelligence pursuits really to me feels like it's a bet just on specialized workflows with industry based knowledge or experts. And that's kind of where it sounds like you believe the value is.
- [29:18]
B
Great question. Super insightful. So up until a few years ago, all of the work was just training on the Internet and books and papers and so on. But a lot of the gains are coming from synthetic data and reinforcement learning, right? So pre training is what consumes all of this massive amount of text. And then we do, even in pre training, we put in synthetic data. Like a simple example is you find some code example in one language, you translate it into every other programming language that exists and you make that additional synthetic data that goes in or you solve a bunch of math problems that you know you can solve formally, you train on that data and then reinforcement learning takes that further. And so this domain, A, that is how we're getting a lot of the gains. We think it's secret sauce for all the labs, but that's what the, the what we are figuring out from observing them and B, generating that data is becoming itself a massive market. So we think the sort of specialist human data generation market is something like $10 billion a year. Right now you have companies out there like Mercor is one that are valued private startups at multiple billions of dollars who basically produce this data in the right manner with the right software and so on to help the labs or sometimes to help enterprises improve their AI. Models. And so I think you're going to see more and more and more of that a related domain There you see leaders in AI have said things like AI will cure cancer, AI will double the human lifespan or whatnot. These are overblown expectations. It's not like that at all. But if you look at what's happening in AI and biology, again, AI is limited by the quality of its training data and that data has to be experimentally derived if you're talking about curing disease. So you see AI companies partnering with or acquiring or licensing data from biotech companies that have platforms to do experimentation and get real world data that can feed back into these AI models. Data is the bottleneck on AI Today's
- [31:36]
C
episode is brought to you by Simple Mining. Bitcoin mining has a reputation for being complicated, risky and hard to evaluate as a real investment investment. If you're considering mining in 2026, what actually matters isn't headline profitability, it's uptime repairs and whether the operation is run like a real business. That's why I've been using Simple Mining. They're based in Cedar Falls, Iowa and they run a white glove hosting operation where you own your miners, you choose your own pool and you have Bitcoin sent directly to your wallet. They were featured on the Inc. 5000 list as the fastest growing company in Iowa with over 40,000 machines under management. What stands out to me is execution. They have the number one rated ASIC repair center and for the first 12 months repairs are included. If mining margins get tight, you can pause with no penalties. And if you want to resize or upgrade your fleet, there's a marketplace to resell equipment instead of being stuck. To help people think it through whether mining actually makes sense right now, they put together a short resource called the 2026 Bitcoin Mining Blueprint. It will walks through the five mistakes investors make when allocating to mining and they also explain how to avoid them before deploying capital. If it sounds interesting to you, you can get it for free at simplemining IO Pompom. That's Simple Mining IO Pompom. Go check it out today and see if you should get into the mining game. Today's episode is brought to you by uphold. Are you someone who's tired of juggling multiple apps just to trade, earn, stake or spend your crypto? Then listen up because Uphold does all of that in one single unified platform. You can access 300 plus crypto and fiat currencies with an interface that works for you whether you're A beginner or an expert. Uphold also features any to any swaps where you can swap crypto to fiat, fiat to metals and tokens to tokens directly in the app. If you prefer self custody, Uphold's vault gives you multi stake security, key recovery, direct trading access and peace of mind without giving up convenience. Uphold is also 100% reserved meaning no fractional practices and proof of reserves are updated every 30 seconds so you can verify your assets anytime for US users, you can even earn yield on dollars with a USD interest account. No fees, no minimums and funds are insured up to $2.5 million through the Atomic Cash Suite program. If you want one app for your entire digital digital asset life, check out Uphold today. Go to uphold.com to learn more. Uphold.com go check them out today when we start thinking about this data. Obviously there's a lot of companies so maybe I can just give you very real examples that I've seen. Let's take the media as an example. I have found it fascinating that there are many media companies across sports, business, pop culture, etc. That have been around for a really long time and a lot of people don't know but, but most of their archive is not digitized. It's just sitting in magazines or you know, newspapers or whatever. And so again if I was owning one of those businesses I would immediately digitize all of my archive and then I'm sure that there are people who would love to come and buy it because it's unique data that has, you know, real world application or whatever. That to me is like data that already exists and it's more of an access problem. How did the labs get access to it? And once they get access to it then it may not be as valuable on the go forward basis because kind of a one time thing that is very, very different than let's say for example I know of companies right now that are buying data as companies shut down. So they go in and they say hey, I want to come and I want to buy all of your Slack messages and Google Docs and just like how the company operates and I'm going to use that for trading. That to me is like, like current information but again it has some end life because there's no more data being generated.
- [35:15]
B
Yeah.
- [35:16]
C
Then the third step is like you or I, I don't know, let's say that we're a doctor and I'm just going to stream data on a daily basis to some model lab. Whether that is me writing stuff intentionally, whether it's Me somehow recording it through, you know, cameras inside of my office. But there's some like ongoing data. That third bucket to me feels like it is the most valuable going forward, but the most under explored. Is that how you view kind of this data bottleneck and like where some of this is going to get solved?
- [35:44]
B
Yeah, it's really interesting. So I think to your first example, I'm reminded of Andre Karpathy, you know, one of the OpenAI co founders led AI Tesla now at Anthropic talks about fossil data that we went out, we mined all of this fossil data like we mined fossil fuels. But now we have to move into a new domain of generating new data. I think this data, like ongoing data of how a business operates and so on I think is super interesting. I think again the more structured and more formal it is, the more precise it is, the higher value it is. The thing that everyone is looking for is proprietary data advantage. What do I have access to that no one else has access to? And it might be that the only way to actually generate proprietary data advantage is to be generating that data yourself. So again, like biotech being a clear example, the interesting companies in biotech that are being founded and funded right now are trying to figure out how to they speed up that loop. You look at a company like New Limit, founded by Brian Armstrong of Coinbase, working on some incredible stuff and cellular reprogramming and longevity. They have an amazing in house data pipeline where they do high speed experimentation on cells, capture that data, use that data to improve their AI models, which then hopefully allows them to do smarter experimentation going forward. So that sort of feedback loop or virtuous cycle, that's the real sustainable competitive advantage. And it's not common. We look at AI, there's a lot of fear of AI monopolies, AI conservation of power. But the reality is in the AI market overall there are no obvious network effects, there's no effect like Facebook. So it's just hyper competitive and the value accrues to us, the users today.
- [37:39]
C
I do wonder, you know, if we kind of take this a step further when you go and you look at the model. So there's two heuristics that I've been looking at. The first is inside of these companies and we have this business CFO Sylvia. It's like an AI cfo. And so one of the challenges with the business is that all model usage is user generated, meaning that you essentially have uncapped exposure to token price, right. Or a compute cost. Because as long as people are banging on that thing, you're paying for it. What we started to realize was we were very under optimized when it came to when do we hit the model, how do we hit the model, etc. And I think that at the end of Q1 or so we started paying attention to this. I started talking a lot of friends who were CEOs of companies, they all started being like, hey, we got to get more efficient at our token usage. We want the same output with lower token cost. I think that's pretty much now well understood like that is underway and is happening. But now what I see the conversation shifting to is okay, well we, you have open source that is starting to become very popular. A lot of it's coming from China. So what I see the public narrative being is closed sourced American models versus open source Chinese models. What I don't hear a lot of people talking about is what about open source American models? That to me feels like that is going to be a massive part of the market. But it just seems like we haven't really seen that the narrative take hold there yet. Why is that?
- [39:01]
B
Well, I think the benefit of open weight and most of these models are not fully open source. You can't get the source code, you don't know the training. But you have open weight. You can download the weights, you can tweak them, you can run them where you want to. That benefits the fast follower the most. If you believe that you are in the lead then releasing your model in a way that people can use freely or tweak themselves is just sabotaging your own margins. But if you think that you're not in the lead, releasing your model in an open way can give you a big publicity boost and it can help you get enterprise sales, enterprise contracts to implement this stuff to tweak the model for enterprise use. That's how a lot of the Chinese companies are, are making money. So I think it'll generally be that way. But that said, there are a couple places that open weight models come out from the us. One is Google Gemini puts out their, their Gamma or Gemma models which are actually really pretty good. And the other is that Nvidia puts out Nematron because Nvidia wants to sell compute. So to Nvidia it's actually quite important to have open rate models. And they don't just put out open weight, they put out the rec, the training data. Here's how we trained it, here's what we trained it on, here's the post training we did and so on. So that's the closest to open source that we have possibly. So, you know, there's some rumblings right now that China might start restricting the release of Frontier AI in response perhaps to, or maybe for the just similar reasons to why the White House, you know, took some steps to slow down the release of recent models from Anthropic and OpenAI. And so Nvidia is. I'm very glad they're out there with open weight and close to open source models themselves.
- [40:48]
C
Yeah, it does feel like that's a very big piece of it. Now the second thing that I'm starting to see is it sounds like people are starting to build model routing as well and this is becoming much more popular where even at Sylvia, we're starting to work on. Okay, well if you get a query, should it go to the highest powered, most valuable, you know, kind of highest cost model or should it be routed to something that is cheaper, faster, et cetera? There doesn't seem to be a lot of companies yet that have built these solutions. Almost everything I'm seeing is custom built. But I got to imagine someone's going to step into the market and say, hey, we can help you do this in a much, you know, kind of easier, more efficient manner, right?
- [41:26]
B
There are some startups working on this. It's pretty exciting. I'm not sure if I can talk about them yet. It's a, it's a very interesting space and they find not only that they can they save you money on tokens, but by being really smart they can sometimes get results that are better than any single frontier model on its own. If they just know when to ping, which model or sometimes the hardest questions, they'll ask multiple models and try to interpolate their answers. So I think there's an enormous opportunity there. At the same time, model routing is famously hard. I don't know if you recall like chatgpt around, you know, sometime after 4 oh, took away the model picker and just did automatic routing of people's chats to different models and it was a nightmare. Nobody liked it. So you have to have some degree of explicit control and some transparency for the user of what's happening while also being smart in this way. And I will say, you know a couple interesting people for your followers to read online. Brian Armstrong recently posted about what they're doing at Coinbase in this way. Aaron Levy of Box has lots of smart things to say say about real world usage in these cases of using multiple models. And then Jack Dorsey has just done some more radical stuff really at square block in terms of how they're using AI.
- [42:50]
C
Makes sense. Mez, what's your mission? What are you kind of going after here?
- [42:55]
B
I want to make the world a better place. And you see the solar and wind behind me. I've been a clean energy guy for a long time. My roots are software and AI, I think is the most important transition we have happening right now. And to me, the most important conversation in AI is making sure that AI stays democratized and not centralized and that we all have access to frontier intelligence and that no one company has a monopoly or lock on it. And that's my current obsession in the world of AI.
- [43:29]
C
Yeah, it's pretty interesting, I think, how somebody comes from the clean energy space and now it's like, hey, where is super intelligence? And the full stack understanding to allocate capital is. It's becoming harder, but it's also becoming kind of much more asymmetric in the returns. Right. I was talking to a friend yesterday and Anthropic is what, four or five years old?
- [43:55]
B
Yeah.
- [43:55]
C
And it's a trillion dollar company. If you had told somebody in 2010, you know, that people are going to build a trillion dollar company in five years, they wouldn't even believe that companies could be a trillion dollars. Right.
- [44:06]
B
Fastest revenue growth of any company ever in history. Right. It's not just a speculative valuation. There is speculation, but it is. The revenue growth is just off the chain. Yeah.
- [44:19]
C
It's pretty crazy. All right, where can we send people to find you on the Internet?
- [44:21]
B
It Planetary vc Ormez on Twitter.
- [44:27]
C
Amazing. All right, Ramez, thank you so much for doing this. We'll do it again in the future.
- [44:30]
B
Thank you, Pomp. Take care.