Steve Jurvetson | Moonshots @ Vision Weekend USA 2025

Well, thank you for having me. I love these events. I remember vividly some of the first ones I came to, learning about Dyson spheres and reconstituting the matter of Jupiter into computronium, Eric Drexler and the whole nano future. It's just been a brain spa as described, and so I love it. I'm glad to be back.

So, the topic is maybe more proact than I normally like to speak about at events like this. This is an investor group, and so I normally like to talk about entrepreneurs and technology trends, not so much what investors focus on. So, I'm going to try to sort of pivot into here's what we like to invest in, but really focus on the themes of AI investing, not venture capital. And I like to call it moonshots because these formerly discrete domains may actually be unifying in the not too distant future.

So, we are a venture capital firm called Future Ventures. We invest in startups that are trying to change the world for the better in a major way. And if you think about where that happens, it's always in a field where there's been disruption. Big picture, this is why democracies and deregulated environments tend to achieve more changes because you have less regulatory capture, less of the big getting bigger, less authoritarian kind of regime protection and more disruption. So, we thrive in disruption. We look for it. We find it sometimes in new channels of distribution like the internet and mobility. We find it sometimes in deregulation, not so often unfortunately. But most of all, we find it in the nonlinear exponential change in progress in technology. So, that's really going to be my focus and segue to AI.

Just a snapshot of what we've done the last couple years since I last was being here. So, it's been some AI investing, which I include nuclear energy in that or base load power, and I'll get to why in a moment because that may be the gating factor for AI's future. We've avoided generally speaking the application layer because we don't invest in enterprise software that much. Find it kind of boring and we don't really think history books should be written about any enterprise software application layer.

I mean, in a strange way, literally 20 years from now will we even remember Oracle, or will kids care, or Salesforce.com, or pick any? Maybe that's just me, but I just find it terminally boring. Instead, we're investing in weird things like analog AI, which we think is the ultimate computational substrate for AI, infrastructure layers like trying to improve open models or long context windows or edge AI. Some applications, but not much. Sphere Semi, for example, a recent investment, is a very strange one doing generative AI for analog circuit design, which is synergistic with analog AI chips, but basically trying to go perhaps in a vertical that you wouldn't imagine OpenAI going into directly, like making the chips themselves. But let me move on to something more interesting now.

I show this slide in every talk I ever give for over 20 years, but I always ask how many people have seen it. Ray Kurzweil's version of Moore's Law. I know it's a higher percentage. There's no way this audience doesn't know Ray Kurzweil's Age of Spiritual Machines. This is the Bible, if you will. So he's maintained this curve for a while. It has a huge influence on me. It still does. For those who don't know, it's a logarithmic scale. How much computation does a dollar buy in inflation-adjusted dollars? What's astounding is that over 100 years, people have been fitting to a curve they didn't even know they were on for most of that time period. And this was Ray's version of it. To make it more interesting for everyone else, I've added some more data points. And it's kind of interesting. A straight line would be an exponential. Clearly starting to look like a double exponential. And just for a sense of scale, this covers a 10,000 billion billion x improvement in price-performance computing. This is the substrate of AI futures. This is what allowed Ray to predict when a human would be beaten by a computer in chess, when we'll have AGI, which he still believes his original forecast of 2029 is still true today, which is it's coming in the next couple years.

Oh, shoot. You can't read this, can you? Okay. Well, the little things are noise. Imagine each of these is the best price-performance computer of their day, roughly speaking. But interestingly, a substrate handoff from Intel to Nvidia, and Nvidia chips are becoming more like ASICs. So, the distinct custom circuits, if you will, the TPU v7 from memory, the Tesla Dojo, which then got cancelled, the Tesla AI5 chip, and these two orange ones are interesting. Those are analog AI chips from Mythic that vastly outperform. Remember, this is at least a 10x, if not more.

Yeah. You can buy this one. This one was started shipping in 2019. Two of these are not yet shipping. The Tesla chip is in final tape-out and the Mythic M2 is not yet available. That'll be 2026. So there's these two data points, these two that are forward-looking. The rest are not. But good point. TPU v7 is not forward-looking. It's shipping now.

So that curve is super important because it has obviously disrupted a number of industries. In the 80s, the obvious ones: computing itself, and then as networking became data-driven, packet-switched networks, and of course what was analog before became digital. That really drove advances in networking and telecom. The current era, it's perhaps obvious we think with Tesla and SpaceX that the companies that matter in automotive are Waymo and Tesla, the ones that matter in space are SpaceX. There's been a major shift in those sectors, and Anduril and other drone-based defense companies in that same bucket. That's sort of obvious and happening. And in the not too distant future, we think a lot of other industries, most importantly AI, which we'll talk about for the rest of today, are radically going to change because of this.

I won't belabor this, but you guys all probably know this pivotal event that occurred with the ImageNet competition in 2012, when neural nets and then what they call deep learning, but basically everything having to do with these connectionist, bottom-up kinds of models took over in training and compute. And it's been going more than the 2x per year Moore's Law. So by the way, that curve I showed in Moore's Law, the main takeaway is 2x per year, no matter what Intel tells you, largely lately through architectural advances. But then AI compute per model, and then even spending per model, it's been going up roughly 10x per year for 10 years.

So what are some unmet needs? We don't try to invest in the heat of the herd, if you will. With perhaps the exception of xAI with Elon's company, because someone has a shirt, don't bet against Elon. That's been my life motto. To give you a sense of what Mythic can do, this to your question, 2019 chip, 2019 architecture, major advance. Their next generation, a 40-nanometer process, by the way, this one, so really archaic process node. If you switch to a modern node, they get like another 65x advantage on top of this, which is what they're in the midst of doing. And it was really inspired by the brain itself. A lot of AI has been... let me check my watch as we go. And the idea of Mythic, and I think they're now still the only one really pursuing this, although IBM is doing a lot of research on it, is that you can do in-memory compute. You can do the 8-bit multiply and add that is the fundamental workhorse of all AI in a single transistor. And for anyone in EE, that's like, how can you do a multiplication operation on eight bits of data and add all in a single transistor? It's a flash memory cell where the gate voltage multiplies by the floating gate voltage that has been stored there. So you store your weights, the activations are coming in, and it just drops a current on a wire. It's the product of those two as the current, and they've shown this to work, and it gives you somewhere between 100 to 1000x power advantage, as well as about 10 to 100x compute advantage.

Now, Jensen over at Nvidia, really, this is not a quote he made related to any product they were developing. He thinks this is other people's chips, like Mythic and others, that are going to put these little jelly bean, 40-cent, less than a dollar price point chips into everything. Every video camera, obviously, every humanoid robot, every car will be an intelligent edge device. And that is, we think, the big untapped AI opportunity that is just latent and sitting out there and has not happened yet. Industrial sensors, all these things are still way behind in terms of rolling AI into everything, and it will be into all kinds of things.

Yeah, I won't even belabor that kind of obvious point. But why is it sort of interesting? So some of the reasons, maybe at a high level, is you can think of it as this reactive sensory cortex. Not just the big thinking thing, but things that are more akin to pre-processing in the visual cortex or reflexes, where you can basically first win in a product category if you're smarter. So we think people will only purchase cars or choose cars... how does that mean? One minute left. Okay. Based on how much intelligence they have inside. The only purchase decision for a car will be the AI software stack for autonomous driving. You can obviously have lower latency for voice interface on everything. You can have robust offline use when you don't have an internet connection. And then this privacy local data thing relates to earlier questions we were having. We think that's obviously going to be Apple's late play into AI when they really dropped the ball for a while. But they're waiting for the moment when you would trust your phone, not Apple, but your phone to do all of your executive assistant functions, health knowledge, and what have you, and be sort of your trusted assistant.

Now, it gets kind of crazy when you look at the aggregate spend in this area. Just a few data points. Data center capex is now over 1% of US GDP. This is not AI. This is constructing the buildings and putting the HVAC systems and the power grid and buying the chips to put these data centers together. And it's exploding. McKinsey, one random group, believes it's going to go to $6.7 trillion by 2030. Again, this is the capex. But what it's being choked by is electricity. And you'll see this of course in people doing forward purchase agreements for fusion and fission and all kinds of different things to somewhere over 10% and growing of US electricity. But this problem is, it's not a generation problem. It's a transmission and local use problem. So you are seeing an incredible run on the market for natural gas-driven electricity turbines. They're sold out for eight years now. You have 146 natural gas-fired power plants in construction where it was zero just five years ago in their forward-looking forecast. And we can't build it fast enough. So there's a thesis around that that has us interested in where we might take that.

And one of the interesting investment theses that we're realizing is now because the latency has gotten low enough, and there was a big breakthrough this year just on a software update they did, oh wait, is this a build slide? Yeah, that Starlink basically allows you to have a data center for inference anywhere you want. So the latency is good enough with Starlink now that you don't need the fiber attached, which means if you had a nuclear power plant wherever they'll let you build one, or a fusion plant, or a hot rock deep geothermal well wherever the geothermal is best, put your data center there if you have the energy and use Starlink to serve inference wherever. An extreme example of this, not an investment of ours, but a really intriguing thought experiment is this company building floating autonomous data centers that will be in the southern hemisphere just at the place where Starlink is peak availability and there's almost no competition for bandwidth, and just run these autonomous things that power themselves off wave motion that are free-floating in the ocean. Well, that's an exciting opportunity.

But the really big one, of course, is well, if you can put them anywhere, how about put them in space where you can have perpetual sunlight in a sun-synchronous 1,200 kilometer orbit facing the sun with no occlusion, no battery storage because you're always in the sun. For a variety of reasons, somewhere between a 5x, maybe more, cost advantage in terms of watts per compute that you can put up there per dollar. And it's obviously the darling now of Elon and strangely at the same time Bezos and Eric Schmidt, who all really think this is the future. I was not a fan just a few months ago and then did a lot more thinking around and research with Grok actually of all things on what the size of the heat fin would need to be to make the heat dissipation work, and it's actually not as bad as I thought by far. It's very doable. And then it just becomes a chip reliability issue. But ASICs do better than GPUs. So don't assume these are going to be based on any Nvidia chips in my opinion for reliability reasons.

So let me end it there and go to Q&A. Just to say I think AI, you know, the bottleneck in this grand unification with space is very interesting. I mean, Elon's talking about like 300 gigawatts per year of data compute capacity being put up. This is a fairly large fraction of the entire US grid per year. And then to take it to the next level, of course, put a mass driver on the moon, build solar panels from the abundant silicon, iron, and oxygen on the moon, and just shoot them into orbit to get to much larger numbers of compute, with an ultimate goal to, of course, start the original foresight dream of computronium everywhere.

Exactly.

Stranded assets? Like data center capex not being maximized or usable if we're pushing so much of training and compute to the edge as you envision? Gotcha. Personally, I don't think so because I think there's no upper bound on how much compute we want. So what he's saying is when you have all these AI6 chips in the vehicles and you think about the number of vehicles that are out there, wow, wouldn't that be an incredible distributed compute resource that you could use? And you've seen precursors of this idea with SETI@home and other distributed, like you have stranded GPUs in the home and in certain businesses. That aggregation has a couple of challenges. One is that it's not coherent compute. There's periodic access based on where people are driving. So there's almost a bifurcation of things you could farm out, and you could probably piece together a fairly interesting inference. And for the macro hard product, which is what they believe that xAI was going to be their major revenue driver in the future, this idea of humans being like white-collar jobs being replaced by a virtual employee that you could imagine per compute cycle you could have several of those virtual agents working. So the agentic macro hard vision, yes, could probably certify a number of cars, but still training, you're going to want the huge coherent clusters like they have, the gigawatt one currently in Memphis. And I think with all the other efforts, there's still going to be overwhelming need. So I don't worry so much about stranded assets as much as a rapid obsolescence cycle, perhaps on training, less so on inference. And yeah, I do put it this way: we're not investing in that capex. I don't get that business model at all. But God bless them, someone loves that and they're plowing money into it.