Nvidia & AI Bubble Questions | Sharp Tech with Ben Thompson

Yeah, I mean this is a subject of hot debate right. Like, is the Transformer model the end-all be-all? We just need to throw more compute at it, or do we actually need a fundamental algorithmic breakthrough that does stuff differently to actually make decisions well, right, as opposed to, you know, and eliminate this problem? I think that's an open question. I certainly think that big tech generally is betting on just throw more compute at it, which makes sense, they have the capability of throwing more compute at it. And Nvidia is the kingmaker right now, and yes, it absolutely warps the market. And Nvidia, there is one company that sort of dominates the space, making decisions based on are you going to bundle with all our stuff, are you going to do XYZ, oh by the way our usual license agreement says you can't translate CUDA, right? They're using legal agreements, they're using bundling, they're using all this sort of stuff that is theoretically illegal, but we're going to... just a real classic case. I can't wait for the Nvidia antitrust case in 2040 that doesn't matter anymore because actually it turned out there was a completely different approach. Tell you one thing, I know I'm newer to the tech space, Nvidia never happens without the Microsoft antitrust trial of 2001. So I'll just pre-write that complaint for 2040. But the other argument though is that a new approach actually is good for Nvidia because GPUs are relatively programmable. I think actually the question is, when you specialize in this particular sort of approach and getting specialized chips and all this sort of thing, once you start streamlining because you need efficiency, particularly from an energy perspective but also speed, latency, bandwidth, all these sorts of things that are often trade-offs against each other, that means you specialize. The more you specialize, the more rigid you become and the less able you are to adapt to a new paradigm that comes along, whereas GPUs are better and more efficient than CPUs, as we talked about a couple weeks ago, yet also relatively programmable and can be adjusted to new workloads if what you need generally is parallelism, but how that parallelism is leveraged can differ. All these are very much open questions. This is where the energy thing does matter though, because it is a real forcing function that continues to drive pushes to do this different, right? And so I think if there is a better way to do it, someone's going to figure it out because there is tremendous economic incentive to figure it out. And just as there's tremendous economic incentive to keep building this out, and you know, maybe it's going to be a situation... I mentioned the Nvidia event made me feel a little bubbly, right? Like there's some of this stuff that is starting to feel like it's getting a little crazy, like what exactly are we using this stuff for? Right, like how useful is this? But you can imagine a world like after the era where you had all this bandwidth that was built out, all this dark fiber that actually was completely economically ruinous but also created the conditions for the actual sort of Web 2.0, the online world. Will we get a scenario? It is interesting, people still find A100s pretty useful, they're still buying H100s. Are we going to get to a world at some point where there's a bunch of GPUs out in the world that were already paid for and they're just sitting there and, boy, wish someone would use this, and that actually creates the conditions for new use cases and creations where it's actually free, not because we solved energy, because it turned out not that many people wanted to use autocomplete? Yeah. A lot of people went bust. No, I mean, at the top of that Nvidia episode a week or two back, the thinking made a lot of sense to me, or at least the possibility that this is a bubble but it's directionally correct, it's just everyone's a couple years early. And so there's this huge infrastructure buildout and the actual use cases won't materialize at least in time for everybody to monetize their investments now, but it could still get there a couple years down the line or even further down the line. Um, I mean, there's a very famous model that is Carlota Perez, a Venezuelan sort of economist, she lives in the UK now, but she wrote this book called Technological Revolutions and Financial Capital, and this is sort of the core thesis. Going back and tracing through these major technological revolutions, how it impacted society, in all of them in her framework, an essential portion is there's this frenzy period that causes a bubble and the bubble bursts, and it's actually the bursting, it's the frenzy that inspires irrational investment that actually creates the conditions for what comes next. Because, like, you know, it was irrational. We go back to the Apple thing, the Apple Vision Pro, right? There's a bit where Apple and what they missed out on was the irrational building of great experiences by developers. What they needed was to excite developers to build an amazing experience on the Vision Pro even though it was stupid to do so. Why did they do it? Because they were so excited about the Vision Pro, right? And so they missed the irrational era, so they're now in the rational era, they need... so which means they need to pay developers. That's right. And so there's a bit where irrational exuberance, it sucks for the people that buy into it because they all go bankrupt, but it's actually societally speaking, it's very beneficial because you build up infrastructure that never made sense to build, but it got built, and so now that enables sort of what comes next. And I do think, you know, there is some aspect here, I'm not sure that we've... like we are, you know, at what point is this GPU buildout, is it actually being leveraged, is it actually being used? The reality is, I think the safe bet is it will become bubbly. Is it bubbly now? To be debated. You know, it's bubbly now when everyone agrees, "No, this is not a bubble, this is real." The moment people stop saying it's a bubble is the moment you know it's a bubble. And I don't think we're in a bubble yet, but it is feeling a little frothy. Um, but they estimated that the AI industry spent 50 billion on Nvidia chips used to train models and brought in only $3 billion in revenue. So make of that what you will. That was in the Wall Street Journal this past weekend. God bless them. Thank you. Appreciate it. These... this is the frenzy that we're counting on to transform society. This is step one. All right. For...