The Titanic Lessons of VC with Josh Wolfe

Danny Crichton:

Hey, it's Danny Crichton and this is the Riskgaming Podcast by Lux Capital. Every quarter, we launch another letter into the stratosphere with Josh Wolfe writing to our limited partners, a riff on investment themes, global, macro, local investment around venture capital, et cetera, et cetera. To come an institution, we do this every quarter. So Josh, welcome back to the program again.

Always great to be here for this. What I love this quarter, is there was a tweet that went out on the internet that was like, "Everything in this letter is great. All I do is I put it into ChatGPT and I'm like, 'Rip out all of the Greek gods and the Roman myths and all this history and all this past,' and there's a ton of insight." And I thought it was both funny, but peculiar because the reality is, is that looking towards our past, the lessons we've learned as humanity over hundreds of years, thousands of years, to me, is intrinsic to what our success has been, is to take the lessons from the past, apply them to the present so that we have success in the future. And so, in some ways, yes, I know we do talk about Roman gods and Greek gods, and you did again in this letter this time around, but nonetheless, those lessons are super important to the success of the firm.

Josh Wolfe:

Yes, I mean, the history of literature is rich with the nuances of all the gray areas that aren't simple black and white. We never want to just say, "Future AI robotics, biotech past has been bad. Good, move forward." Everybody obviously has shorter attention spans and wants the quick sugar hit, but reading the greats of the past, I think is really important. Having a very diverse set of inputs so that you can find analogies, particularly the timeless analogies. And when you go back thousands of years and you look at some of the parables, many of these things are in fact timeless. The players change, the settings change, the scenes change, but it is a guidepost when you are going through a future, which is always and never-endingly uncertain. People always say, "Oh, we have heightened uncertainty." No, it's just heightened levels of confidence about the uncertainty.

But the world is uncertain. Having guidebooks about how mankind in the broadest sense adapted and reacted to things and anticipated things and planned for things. The hubris, the follies, all of that is critically important in the drama and tragedy and comedy of life and business. So in this particular letter, we started with this parable about these four brothers, Prometheus and Atlas and Manetius and Epimetheus. Epimetheus, he's the impulsive fool. He's the one that's got the short-sighted decisions that get unleashed and unforeseen consequences.

Danny Crichton:

Right.

Josh Wolfe:

And so there, you would think about the Pollyanna, techno utopians that are not really thinking about, what are the implications of this? What are the moral implications? What are the compounding implications when everybody has a particular technology? As we all know, it's very hard sometimes to predict the social implications when a technology comes onto the scene. So, that's Epimetheus. Manetius was the reckless titan.

He's the one that is brought down by his hubris and a warning for us all against overreach. And this was something that we talked about in prior letters as well, about not just being pure techno optimists and not being Cassandra or Pollyanna, but thinking about, what are the implications in doing this in a thoughtful way? Atlas, of course, the image with the globe burdened on his shoulders is the burdened explorer carrying the weight of the world while charting new frontiers. And in many cases, that's the plight of the entrepreneur. More so than any of us venture capitalists, we're given the fuel and the money and the belief to allow people to go on these journeys. But the hero of the Atlas journey is the entrepreneur who has to deal with the burden of coups internally and people wanting to join you and then quit or start competing endeavors and having competition, which you never want to say is validating.

You always want to say, "I want to crush," just like a Greek god. And then you have Prometheus. And Prometheus was the rebel who defied limits, gifted fire to mankind, and ignited to human progress. And so the Promethean promise, which is something that we talk about in the letter, is particularly encapsulated and captured in this idea of another word that we coined extensionalism, not existential about being existence, but extending our senses. And when you really look back at the annals of technology from the microscope that let us see very near and the telescope that let us see very far, that was extending the sense of our sight. When you think about hearing, as I'm talking now, a microphone that is able to record my voice and play it back through endless speakers, endless forms, most beautiful of speakers, there are all kinds of new technologies that we're backing to be able to introduce AI and cutting-edge digital technologies for next generation, hearing aids for being able to promote and produce speech de novo.

All of those kinds of things would be examples of extending the senses for both sight and sound. Then you've got touch and hear, you're thinking about prostheses and how do you extend through teleoperation, whether you're a robotic surgeon or a company like Physical Intelligence, which is developing the AI brains for robots. In a similar way, endless forms are most beautiful, lots of varying forms that the robots will take. But instead of just being humanoid, they will really be run by this common operating system of very complex AI that is able to take unstructured data and let them navigate terrain in the form of homes or workplaces that they've never been before.

And then when you think of one of the last senses that really has not been touched, you have taste and you have smell, and smell is the most salient of our senses. You go back to Proust in search of lost time, and he bites into the Madeleine and hearkens back the narrative, remembrances of all the experiences he had. We all live and experience that every day. When you smell a smell that reminds you of the fragrance or the cologne of a first kiss or a loved one, the smell of your grandparents' home, the smell of cooking that takes you back to a particular place. And sometimes you're like, "I can't remember what that is, but a memory just through the associative neural linking inside of our brains just brings you back to that moment." And it would be a beautiful thing if we can capture smell the same way that we can capture sight and sound through our technological devices. Our microphones capture our voices and sounds and things and things we hear, speakers play them back.

Our cameras capture a visual, both in static and dynamic videos, and we can play them back on our pixelated screens. We have not heretofore until our company Osmo, which spun out of Google, been able to effectively Shazam a smell. And these guys, as you know, just for the first time, teleported a smell, which is absolutely wild. They took the essence of a plum, were able to parse the molecular signature of that, create a precise digital fingerprint of that, so to speak, and then recreate it chemically using a combination of human and robots to orchestrate effectively a fragrance that was a perfect match to that original essence. So teleporting a smell, recording it, playing it back, absolutely wild. This is a full fulfillment of the Promethean Promise.

Danny Crichton:

Well, and we had Alex Vilsko on the show I believe last year, and what was amazing is I think we had an LLP moment a couple of weeks ago. And live in the audience, Alex is actually taking requests for smells. You can combine and create your own cologne live in the room, and actually delivered it two, three weeks later in a way that using his AI algorithm, he's actually able to create that signature and build all new scents. And you're like, "Well, I want to smell something rose, but not exactly rose." And today you're limited by certain materials, certain constraints, and to me, it's not just existentialism for humans, but also existentialism for this particular category of volatile organic compounds where you can say, "Look, now we can invent entirely new things you've never actually smelled before, but they're extremely enjoyable and we know that's going to be something that's pleasurable."

Josh Wolfe:

And some people call that psychophysics, it's almost in folks like the Foundation Series by Isaac Asimov, but it's this idea of like, "Okay, I'm on a beach in Coney Island, Brooklyn," which is where I grew up, "And there's a wafting smell of waffle cones, but the smell of Nathan's Hot Dogs, but this musky, salted beach smell." And you just describe that, and all of those things have correlates in what those fragrance and molecules are, and it's very difficult for us to describe these things, but computers now are being given and gifted the sense of smell. So I think it's absolutely wild and very grateful for people like Alex that are obsessed with this their entire life and have sought to be able to produce this, in a sense, delivering fire to humans.

Danny Crichton:

Well, and so Prometheus obviously delivers fires thousands of years ago, at least in Greek mythology, but over the last couple of thousand years, humans have done this themselves. They've built microscopes, they've built telescopes in the skies, we just launched a James Webb Space Telescope a year or two ago, capturing images basically from the birth of the universe, and we're learning so much and so quickly. But something we've talked about quite a lot over the last couple of months and the last couple of years here at Lux is there's so many constraints on entrepreneurs, on entrepreneurial energy, bureaucratic constraints, capital constraints, in some cases, team constraints, sometimes just the constraints of quantification that you're at a company that this is an unprofitable business up front and no one's willing to make that investment. And so we're trying to match something. You've dubbed the Promethean Promise to the Atlas answer, which is a couple of different strategies that you've devised. Maybe we talked about that in a next.

Josh Wolfe:

Yeah, so if you take the symmetry here of the four, Epimetheus and Manetius and Atlas and Prometheus, there's four strategies that we're employing here, is what we call the Atlas Answer, just to be poetic about it. There's Lux Labs, which is de novo company creation, solving part of the problem of what you described. There's corporate spinoffs which are related to this. There's tactical global opportunities, and then there's what we have termed for this new theme of maintenance, X-Wear. And until we come up with or someone else comes up with a better name, we're sticking to it. But let's start with Lux Labs. Lux Labs, probably 20 years of work that we've done where we identify a principal investigator at a laboratory or post-docs that are working on the cutting edge breakthroughs that they have, probably half the time, they're biotech or life science breakthroughs, but the other half they might be electrical engineering, computer science, AI, machine learning, pretty much anything.

And we will go and take that kernel of intellectual property and the kernel of know-how in that original founding scientific team. They may not have an entrepreneur yet to commercialize this. And so we will help to midwife that, put together the capital, the team, the strategy, and start these companies from scratch, serve as a co-founder. We typically will create the company de novo, a true zero to one. This thing didn't exist, now it exists. We love doing that. We've done over 20 of them in everything from nuclear waste remediation, identifying a problem that existed and how do you solve the nuclear waste problem? No company start existed. So we started one from scratch. That was a company called Kurion that we would later sell to Veolia after helping with the Fukushima cleanup, which was a very significant thing. Being early founding investors in a company like Auris Surgical Robotics, which we would sell to J&J for $6 billion dollars, giving new capabilities, again, in extension of human senses, but born in the need first for belief and then backing an entrepreneur with cutting edge tools and AI and capabilities.

We've done this in biotech with certain professors where we've started five or six different companies with them, like Charles Zuker at Columbia, focused on all kinds of neural control of different parts of the body. The latest, which is going to come out of stealth soon is using the nervous system to actually control the immune system, which is going to be a very big deal. And so this genre of creating companies from scratch is something that most venture capitalists say, "Oh, we have proprietary deal flow and we're value at investors." This is something that we had to do in our early years out of necessity. Now we do it out of love, but it's a playbook that we have, that every year, I don't know if it's going to be one, two, three, four, five companies. Typically in any given fund, which we're raising every two or three years, we do about 10% or 15% percent of the portfolio construction by starting companies de novo with scientists and founding entrepreneurs.

Danny Crichton:

And when I think about this whole category, I mean, to me, in robotics, there's this idea of this uncanny valley where you go and you get so close. And I think there's something very similar to this in startups where you have people who are maybe energetic, they want to go do this, they want to see their research or this discovery that they found in the real world, they want billions of people to use it, and that's a hard problem.

And not only that, but there's 1000 ways it could go wrong. And to me, a part of the solution here is not just saying, "Look, we can do everything right." We can also help to avoid people making mistakes either because they're starting their company for the first time, or they're in a new space they haven't done before and they're switching fields from the first company to the second company. But there's so many of these tactical areas where it's like, "That's a mine. That's something to avoid. That's something that [inaudible 00:11:58] solve," that you sort of forget that half the companies failed just for one of those reasons, and if you take that out of the equation, the probability for success goes up by a lot.

Josh Wolfe:

Amen. And some of that is just removing the bureaucracy and the friction that exists in this. Sometimes those frictions exist because of intellectual property licensing and somebody not knowing how to spin something out. We are quite versed at being able to come into a tech transfer office, be able to work to strike a license, a royalty and equity deal, be able to transfer the team. That's something that came from a lot of hard work in doing that many, many times, different universities or varying degrees of ease with which to do that, but we're able to come in and give the playbook for that. Same thing if we're spinning out a team from a big company, and this is one of our second themes, which is adjacent because we will sometimes come in as a co-founder, but it's an amazing time when the cost of capital is rising.

And this is a broad theme, largely something we derive from risk gaming and understanding the rising cost of capital. Regardless of what the Fed does in an election year, a rising cost of capital means that just capital is more expensive, it's scarcer. Why is it going to be scarcer? Lots of funds out there with purported to have dry powder. As you know, we call it wet powder because they're going to have to fund some of that cash towards companies that were under-reserved or under-capitalized. In some cases, it's because big companies with higher interest rates are looking and saying, "Geez, these things that we were doing, these far-out 20 year projects that became these 20 month frenzied initiatives, we have to cut." And so Google's other projects and things at Meta and things at Apple, they might've been speculating on doing further out things, and now they have to focus on contribution margins, near-term profitability, and their shedding divisions.

Will they regret that in a few years? Possibly, probably. You're seeing that in some cases, researchers that have left and gone and started companies and they get reacquired for billions of dollars, so there's an opportunity there. We have done five spin-outs in the past few years. Two from Google and Google teams, two from Apple, one of which has not yet been publicly disclosed, and one from Meta. The one from Meta, we partnered with the founding team that became evolutionary scale, that was the frontier models for biology and helped to capitalize that and put that together. We did out of Google, two companies, one which was Osmo that we just talked about, one that was the founding team from physical intelligence partially out of Google, partially out of Stanford, developing the AI for robots. We did one out of Apple that became Aeva, publicly traded, focused on 4D LiDAR, and another that hasn't yet been announced.

So this is an exciting opportunity probably for the next two years where there are teams of 10, 12, 15 people that exist, they worked together, they had a big corporate budget, and now that budget is failing or the prioritization for their team is falling, and so there's an opportunity to spin them out, equitizing them, creating that structure, taking them out of the abundant budget and removing the bureaucracy of a big company that they might've come from is a sometimes daunting challenge for these folks, and so being able to give them a playbook to do that is also super useful.

Danny Crichton:

Well, and I think you look at these companies, obviously a very successful and attracting talent over the last 10 years when a couple trillion in some cases, in Apple, one to three trillion in all these. And what they basically did, is they aggregated so many smart people, they were working on so many projects, you were in this ZERP world of 2020 to 2022. As you pointed out, the fed rates went up, and so everyone is now hiring, the CFO's office is where the power is centered here. And people are saying, "Look, do we want this project or that project?" Okay, the McKinsey consultants are walking through and are saying like, "Okay, you have a portfolio of 500 things, these 100 need to go." And I think in the past, you would've just fired these folks, you would've just let them go. And I think fortunately for us, and fortunately for the teams that are working here, in many cases, people are realizing there's real value here that can both benefit their own companies.

So I mean, you see this with Physical Intelligence for instance, like Google may be a beneficiary of robotics down the road, but so can other companies. And so there's a win-win cooperative moment where you can say, "Look, there's an equity, people can win on both sides. But this is going to languish if you don't put the right budget behind it. Why would you allow such great talent to just sit in a closet basically and not be able to explore and really build something excellent going forward?" And so it's really, to me, an interesting positive opportunity, and frankly, the corps have been easier to work with, I think, than the universities in many cases because of the technology transfer offices, which can oftentimes be the hardest people in the world to work with.

Josh Wolfe:

Well, they're already primed by bias to be commercially minded because they're commercial enterprises.

Danny Crichton:

Yeah, right.

Josh Wolfe:

But this initiative, corporate spin outs, Lux Labs, the bottom line for us is it helps to unlock value by giving these breakthrough teams and the technologies that they're developing, the freedom and the focus to entrepreneurial attention that they can have to thrive, and that's it, and it works.

Danny Crichton:

Let's go over to the global. So we mostly focused on domestic, but even in the last couple of years we've seen this zooming up in Asia and Europe, in Latin America, all kinds of new startups and entrepreneurs, and we've adjusted our lenses as well to look more overseas. What are your thoughts on there?

Josh Wolfe:

Well, so if you again, zoom out, you've got Lux Labs, you've got global opportunities and the corporate spin outs, and then the last one we'll talk about is fixware. But global opportunities for us, is how do you find these brilliant teams? This doesn't mean necessarily we're going to have geographic [inaudible 00:16:50] and where you're just going to pick a region, but it's how do you find a brilliant team wherever they are? Now, it turns out that way, there was a brilliant team, one of the authors on the Transformer paper that came out of Google, another team that'd been really obsessed with complexity theory and work at the Santa Fe Institute and thinking about evolutionary algorithms and a novel take on AI, that became our investment in Sakana, and we were founding investors in the company. It is now one of the leading AI players in Japan.

Just like you see different sovereign models, Mistral in France and maybe OpenAI here and others, that every country does not want to be dependent upon the US or China, and so they're developing their own models, this idea of sovereign models. Sakana for Japan has not only attracted foreign investors like us and some other US VCs now, but also some of the most important Japanese corporations. So when we did that announcement, we pinged Rahm Emanuel, who's a US ambassador, and he helped to amplify and highlight the great forge between US and Japan. We have many common interests, whether it's containment of China, thinking about regional support for each other, and then just going back to the cross-border trade between US and Japan, which are really two technological titans. So this was one of the first investments we ever made in Japan, and we expect that there will probably be another half dozen over the next few years as we meet the teams and the talent and feel comfortable with people on the ground that are able to help us navigate that.

Another area where we're doing this is in Israel. Israel right now is a fascinating space, in part not only because it's burdened by the war, but it is facing an existential threat that has forged these young men and women to really not just think about, "I'm going to develop a technology for ads or consumer products or something like this or a mobile app." They're really focused on this existential threat, and I believe that Israel is going to have the greatest generation as a result of this, just like the United States did after World War II. So people are teaming on all kinds of things, particularly in defense. The first two investments that we made are actually defense-focused. One has been leaked publicly, that us and Sequoia did together, focused on some of the white space and hardware and software. The other has not been announced, that us and Founders Fund and another group had done, and I'm very excited about both of those companies and particularly the teams, but it's certainly not the last investments that we're going to be making in the region.

And I think that the coming years as we start to see a change in the balance of power and the rise, particularly, with the new administration of an Abraham Accords 2.0 and seeing Saudi normalized with Israel, which I think is a very high probability and a detente coming between Israel and Lebanon and then the future of rebuilding Gaza, I think there's an incredible opportunity for Israeli entrepreneurs that we're very bullish on. We also did, in the UK, our first ever defense investment that's going to be announced soon. I just think that there's an amazing opportunity for US allies, cutting edge technologists, backed by US investors. It's going to create a great global ecosystem. They're going to be partnered and allied with each other at a corporate level just as the countries are at a country level.

Danny Crichton:

Well, I will say when you said the word detente, I got a push notification that there's a ceasefire that just got announced between Israel and Lebanon, so perfect timing on that one. But obviously, we talk about disruptive innovation all the time. Everyone wants to build new robots, new thing, and that is very exciting. But the other side of building all this new capital improvements, all these new devices is that they break.

Josh Wolfe:

Yes.

Danny Crichton:

And when they break, and as cost of capital increases, you have this challenge of you can't just buy new things all the time, you have to maintain the existing infrastructure that you've already purchased, and that is where this idea that you've dubbed fixware, which I like because it's quick and fast and dirty, it's better than our theory of maintenance, which was mine.

Josh Wolfe:

No, maintenance sucks, right?

Danny Crichton:

Yes.

Josh Wolfe:

And I appreciate you helping coin this. The idea of fixware is just like there's software and there's hardware, there's fixware. Well, fixware is focused on fixing the existing stuff we have, be it hardware or software. And why is that important? You just nailed it. Rising cost of capital, you look at CAPEX on income statement, balance sheet, financial statements. There's two parts of it, there's growth, which has basically been funded for the past 10 years because of ZERP and low interest rates, which means lots of people buying new stuff. Well, that stuff could be HVAC systems, it could be transport systems, it could be new planes, it could be new buildings, new infrastructure, digital, new software implementations. Now, rising cost of capital, what do you do? The other part of CapEx is maintenance, so now you need to maintain this. If I am a CFO or I'm on the board of a company, I am saying, "Ho, ho, slow down your spend and we'll get to the data centers."

Because that's the one thing on the AI where there's massive amounts of growth spending and CapEx spending, but you have to maintain the existing systems for longer. How do we make the existing stuff work so that we're not on a cycle every two years where we're buying new stuff? And I think that there is going to be heightened demand for companies that help to solve that. That could be software for early preventative maintenance and detection of aberrant anomalies inside of manufacturing facilities. That could be new systems to not have to constantly upgrade telecom infrastructure, but be able to have different kinds of infrastructure solutions that are more easy to maintain. That could be maintenance of military material, thinking about our ships and our subs where somewhere between 30% and 50% percent of the defense budget goes towards maintenance and operations and sustainment of the existing assets. So we think that there's going to be an existing base of assets that are going to have heightened demand for maintenance of them.

You look inside of our portfolio today, impulse space, cutting edge of delivering things at space, but going to play a key role led by Tom Mueller, Elon's leading Rocketeer for nearly 20 years at SpaceX now at Impulse space, being able to deliver things and help maintain existing systems that are up in space, be that satellites or other things. Nominal led by Cam Accord, same kind of thing, focused on logistics and maintenance and advancement of existing programs in aerospace and defense, Epsilon3, Hadrian, and maintaining the manufacturing base for the Arsenal of Democracy and Freedom's Forge, Kinetic Auto being able to do this for alignment and fixing and maintaining electric vehicles. So this is a growing theme with a growing part of our portfolio, but it's something that I would put the beacon light out there, that anybody that's working on cutting edge things in maintenance to maintain existing assets is something that's of great interest to us.

Danny Crichton:

And I'll also point out, I mean, obviously, hardware is a huge part of this, but there's also the software component as well. With the rise of AI coding agents and people helping and software engineers building code, there's two sides of this. One is you can write a lot more code than you could before, which actually increases the surface area, both for cyber security and also for maintaining all of that code. And so we're now seeing the next play of this, the 2.0, which is to say, "Well, how can I refactor my code? How do I protect it?" Because I can write thousands of lines, and just because you can write so much code does not actually mean that your budgets are going down. They actually increase because you have more work to cover and maintain.

Josh Wolfe:

MLook, we'll see what happens with Doge between Vivek and Elon and Trump's influence on government efficiency. I actually am quite optimistic about that and hope that it really takes off. But I think there was a tweet from Elon or one of them that was basically noting something that we put in a prior quarterly letter about how much of the government in our financial system runs on old archaic systems like COBOL and FORTRAN and the ability to upgrade these things to basic cloud-based things, and therein lies vulnerabilities as well, but the maintenance of existing code is really important, too. So absolutely, not just hardware, not just software, hence why we're calling it fixware.

Danny Crichton:

And then finally, we want to talk about AI. So this was the core middle of the letter focused on, I guess it was the Menacious moment as you dubbed it poetically, but the core of this was around a recent AI summit that we hosted in New York City with about 150 engineers. In fact, some of the engineers actually coded live on stage, which was a pretty impressive demonstration of the technologies we have available to us these days. But what were some of the lessons learned both from the summit and more generally that you had as a position in the AI infrastructure race and elsewhere?

Josh Wolfe:

Companies like Runway and tldraw were focused on the application layer and they were thinking about, "How do we democratize things that people can't do?" And so in Runway's case, it's video, which the pace of them shipping is just absolutely amazing. In tldraw's, it's these collaborative innovations along the lines of Canva and Figma, both superpowered for AI. Infrastructure inefficiency, I want to say that one for a second because I think that that is probably the most interesting thing that also leads to some contrarian insights. But the big one was Mark Warner came, he's senator on Intel committee and was talking about some of the threats that we face as a country, both internally and externally, particularly visa China, and thinking about AI's role in this and where regulation is going to help or thwart both domestically and overseas, I think is a pretty critical thing.

New York City itself, as a major AI hub, was one of the big themes. So you've got 20 new unicorns over the past year, I think from 2023, 10 just in the past 12 months. New York City is thriving. You have a ton of talent, in part because Meta is here, in part because Google's here, in part because Hugging Face is here, and you're seeing this diaspora of AI talent that are young, they are in many cases single, they want to be in a thriving, really vibrant city like New York. There was a big shift away from SF, that will sure come back, but you've captured a lot of talent in New York when a lot of people were like, "Oh, New York is just for advertising and the double clicks of yesteryear and fashion and media." But no, this has become a real tech hub for AI, so that has been a bright spot.

But turning back to the infrastructure piece, this to me is the one that's interesting because you heard it from Vipul who runs Together Compute, you heard it from Clem who runs Hugging Face. The compute demands today are for AI, the number one thing that I would say is consensus. And anytime we have a consensus, we at Lux are always trying to figure out what's the contrarian take. You go back, look at the largest AI clusters, most of them today, a few hundred thousand H100 chips and the next ones, and those assume 100, 200 megawatts of power, next year's cluster is going to be between 300 and 500,000 chips. There was news that when Elon was doing this for xAI, he couldn't even get enough grid power, so he had to bring in mobile gas generators to just add 35 megawatts. You've got Meta's new facility out in Arizona, a billion dollars.

That alone is 12,000 tons of steel. I mean that's the equivalent of 30 fully loaded Boeing 747's, so huge, despite of irony here, huge infrastructure demands in steel and copper and water and natural gas when you think about the clouds being so light and airy and scalable on the zeros and ones and bits and bytes. So one of the big insights here is everybody will say, "We need more and more data centers, more and more power for that." The irony, of course, for the power, which I've been very pleased to see being a champion of nuclear power, we've attempted to rebrand it as elemental energy, is that elemental energy seems like it may arrive with greater zeitgeist support, not because of the greens that see zero carbon and a carbon driven push, but because of a silicon driven push in the insatiable demand for power.

Now, realistically, I think that nat gas is far superior. If I was building data centers, I would look at the Permian Basin where we have abundant natural gas and figure out the infrastructure for cables and transport from there, and I think that there are people that are focused on that, but we see two futures basically emerging here. One is less centralized, more distributed, and Together is a great example of that. They're one of the fastest companies in our portfolio to grow to a hundred million dollars. Truly NexGen Cloud provider for AI, and giving lots of people that shift from CapEx to OpEx and doing it very efficiently. The bigger thing, which I think is the contrarian insight is Apple. When you listen to the names I named before, Google and Meta and Salesforce and Oracle and all the people that are building data centers and spending billions of dollars on this, and in some cases aggregate hundreds of billions of dollars, the one company that isn't doing this is Apple.

Now, Apple a few weeks ago came out and Tim Cook, because he's a little bit behind in the consumer facing AI, has said, "We're not first, but we'll be best." But the interesting thing is not in the application layer for me, it's in the phone itself. They came out with a paper maybe 11 months ago, not widely talked about, which was being able to use memory on the device, on the phone itself to do small models, and it dawned on us in conversation here at Lux that maybe the future is not going to be these 100,000 clusters.

You go back to Sakana, their latest model was trained on eight GPUs, not hundreds of thousands of these things. But what if 30% to 50% percent of the inference is not cloud-based, but on device? That your emails, your texts, your photos, your videos, your health data, all of the things that you trust, on top of it with a layer of privacy from Apple, which you trust, is done on device and it's not going to the cloud. And then for the 50% to 70%, you're actually going to the cloud because you're pinging perplexity maybe instead of Google or ChatGPT, Faura web search or Claude or whatever might come next.

I think that that's a real possibility that ends up proving, in hindsight, as almost every infrastructure compute wave occurred, that we get overbuilt and that there's going to be massive failures and that the beneficiaries of that will always be the consumers and not necessarily the investors. Go back 25 years ago, the build out in the early internet for fiber optic cables, global crossing, ultimately went bankrupt, the distressed players picked up the assets for cents on the dollars and were the winners, as was to the Third World who got connected to the internet essentially for free in the largess of that build out.

Go back another decade forward from there, the shift from on-prem to cloud, you saw the Equinix's and various people that were going from on-prem to co-location to cloud, rocket ships on the way up, and then many of those guys ended up bankrupt on the way down. So I think that we will have an over-build, a glut on the data center side because what the company individually does rationally, collectively, they're doing irrationally, and then I think Apple sneaks in on the side and is like, "Wait a second. A significant portion of this can run on small models on device." Who would benefit from that? Next-gen memory players. Today, that's the pocket portfolio of SK Hynix and Micron and Samsung, but the high bandwidth memory and the next generation for that I think is going to be the next winner.

Danny Crichton:

Well, Josh, I know we're already up on time, so for anyone who is listening, we did miss a couple of big sections of the LP letter for Q3 2024. Please find it online. We have the full copy, I believe on your Twitter account, and presumably, we'll post it on the Lux account. You always beat us to the punch.

Josh Wolfe:

Well, just put it through ChatGPT and summarize it in 10 words.

Danny Crichton:

Yeah, as folks do. But [inaudible 00:30:44] lessons, Josh Wolfe, thank you so much for joining us.

Josh Wolfe:

Thanks Danny.

‍