Google's AI Spending and a Bubble Temperature Check | Sharp Tech with Ben Thompson

Yeah. Building out infrastructure, building out this was part of the thesis why everyone was unsure if anyone would be able to respond to Amazon when it came to AWS. Why did no one really respond to AWS in a meaningful way for a long time? And I was like, "Oh, that's a good idea by Amazon. They're not going to make much margin on this, but no, Amazon, your margins, my opportunity, very clever." What happened with Amazon AWS was when they released the financials, they were making like 35% margins. It was like, "Wow, okay, that's doable. Well, we could." And then competition sort of flooded in. But just in general, that was a question.

For sure. For sure. And Microsoft had started Azure kind of before then, it was like Windows Azure, Windows servers in the cloud. It was just sort of muddled, but it really got on steroids and they realized, "Oh wait, no, it's not Windows, it's just servers in the cloud."

Well, that's part of the question. So you on the optimistic side of things, it's this idea that our business is infrastructure like Google's business is infrastructure. That is a lower margin business than what we've traditionally been, but we're going to make it up in volume such that the absolute amount of money that we make is going to be even greater than what we could have made from our asset-light sort of search business. And I think that is what Google is thinking, or I think that's the vision they sold to Berkshire because I think that fits the Berkshire investment approach, which is they appreciate these asset-light models, but the asset could always go away. They always said because they had Geico, a big part of their insurance business, they pay a lot of money for Google ads and they've always talked about how they probably should have gotten Google a long time ago given how much they were paying for it, but they never did. But this is what they understand: you put in money for infrastructure and the margins are relatively lower, but the absolute amount is very large. The question is how is it defensible? Like a railroad is. What is going to differentiate different infrastructure providers?

Which is where I come back to $10 billion. I think from Google's perspective, to the extent they want to shift the business or continue to shift the business to lower margin—which the cloud is inherently lower margin than search—having Berkshire endorse this shift is very valuable and is worth some amount of equity. And from Berkshire's perspective, putting in $10 billion, a lot of money but not that much money, to sort of feel this out and see if this is a place where you can deploy your capital also makes a lot of sense. So I think this is a signaling investment, but this will probably get much larger over time. But both companies have the chance to say that doesn't make sense. The market barely responded; Google was down a little bit. But I think it feels like the market gave Google permission to go further. By the way, there's a fourth reason: Google wants to soak up money before SpaceX goes to market, before Anthropic goes to market. Let's get in first. Why not?

Yeah, in the fullness of time once we have enough compute, I would expect margins to go down, which is part of the questions about what is defensible here. But this is where I do think the TPU bit is important because we've talked about in a commodity market, where you make money is by having superior cost structure. Assuming compute is a commodity you can go anywhere to get, then the price of the compute is set by the marginal contributor of compute who's providing the last bit of compute that is bought, and that's the market clearing price. That price would apply broadly. Now, is compute going to be like that? No. This articulation is like copper ore or something like that, a pure commodity fully fungible. Oil is another great example. In oil, the marginal producer of oil is making basically zero dollars on it. And then if you're Saudi Arabia who can get your oil very cheaply, the difference between that market clearing price and your cost of production is all profit. And because it's not a lump of labor fallacy thing here where because we get our oil very cheaply we have to sell it very cheaply—no, you get to sell it at the market clearing price. That is how you make profits in commodity markets, by having superior cost structures. Is there a thesis here? I think it's definitely plausible that TPUs in the long run, along with Google's networking and owning the whole stack, that they have superior cost structure for their data centers such that if compute is sold at the marginal price, they're going to make the most money. Now, will that maintain? That's all TBD. Nvidia will say, "Actually no, let's see the actual cost of ownership. We'll document ours. Where's yours?" So we don't know all this for sure, but it's definitely theoretically possible. Also, Google in a perfect world isn't in the commodity market. You're actually on Google because you're on Gemini or they actually figure out AGI with their world model approach, and that's the upside. You're paying for performance in addition to all the efficiencies you're getting. In that case, Google gets to charge higher margins than everybody else because they have the better product. That's the upside case. The downside case is compute ended up being a total commodity, no differentiation, models are all substitutable, but we do it better and at better scale than everybody else.

Yeah, this signals to me Greg Abel is going to drive Berkshire into the ground.

Yeah, and the collective action where everyone has to participate even if they think it's going over the edge because if they don't, it's a game theory problem. If you drop out of the game, that might be good for everyone as a collective, but you just killed yourself because you're no longer in the game. So everyone ends up overinvesting because if they don't invest, they're out. And thus everyone is collectively out. Maybe Google is worried about that, and it's like, "Well, if this is going to go off the cliff, better for it to be equity than us to be left holding this."

Yeah, I mean, we're clearly getting all the power thing that we were theorizing a couple years ago. There are power projects everywhere. That is 100% happening. That was always, by the way, one of my signals of when to know we're in a bubble: when we're building as much power as possible, because power is a massive long-term investment where the margins are not really screwed. Lots of debt goes into building power by and large. So power is definitely happening. I would argue the whole SpaceX thing—there's a whole thesis about going to market with data centers in space that might end up being BS. I think in the fullness of time it will probably work, but it might not work in the medium term. But they have a story to tell because of the bubble, which makes their valuation higher. They get to build more rockets. So you could argue AI is giving us better rockets in the long run. That is an output of the bubble. You're getting the models—to what extent they're open is a good question—but you have all this investment in new memory. I think the memory guys are kind of screwed up because number one, they're opening the market for Chinese memory in a huge way. Number two, no one in tech has ever really worried about memory ever. There is almost certainly huge amounts of low-hanging fruit in terms of improving memory usage that no one's ever bothered to pick because you could just go add more memory. That's not the case anymore. So they're in some respects creating the conditions for memory companies to cannibalize, forcing people to algorithmically improve their products such that they need less memory, ultimately hurting their market in the long run. But they're going to build more memory. Intel Foundry is going to end up being a thing at some point just because TSMC didn't invest sufficiently in 2023 and 2024. They slowed things down, and they also ensured they have more competition in the long run because money is going to go to Intel or Samsung.

I'd say we're already well down the road in terms of reaping the bubble benefits, which is also another way of saying setting bombs all over the economy.

I do have a question on that. The most famous company I think doing that was Uber. Do we know what Uber does?

I was very curious. I was like, what were you using AI for in the first place? What was the plan there? But do you have any reaction to that genre of story right now?

It's like doing, getting a computer to do the things the human did. We're just trying to recreate humans.

And do the job the way they did them. This kind of sounds like how should we monetize content on the internet?

Well, in newspapers, next to the content, so let's put ads next to the content. And it wasn't until you had internet-native content delivery methods, i.e. the feed, fully customizable and infinite amounts of content, which ties into that whole TikTok thing we were talking about before, where you get an explosion in content and you sort and filter and deliver to you exactly what you want. So it's all good even though the percentage of goodness on the platform is testable. That is when the monetization actually happened. Now the idea of comparing print advertising to digital advertising is absurd in the opposite direction. The feed was invented in 2007, I think, i.e. seven years after the bubble burst and after the whole dot-com era, much of which was predicated on trying to do jobs but make it online without rethinking how it was done. That's almost certainly how the AI thing is going to go.

There's going to be a lot of time and energy spent trying to sort of do what we do now onto old companies. I mean, that's part of the issue. Of course AI is going to be expensive, and it's a grim thought, but unless there are mass layoffs, layoffs are ultimately going to work. There's so many jobs—we've talked about this.

Tesla.

And the bubble is going to be a lot of existing companies realizing the AI spend is not worth it, and it's going to become a meme. "It doesn't work." So the Uber stuff is a precursor, I think, of how the bubble will burst. When it becomes a widely accepted meme that actually this stuff doesn't work, then it's not transforming our business. We're still paying all our employees and we're not really adding to the top line.