All right. I was born in 1957. Bora, you could have hit it. You could have hit it. You know, nobody was doing their inference stage. No, it's the Mark 1 Perceptron. The first, you know, it took a perceptron, put it in a computer, and it could actually do a little bit of image recognition. My God, 1957 image recognition. My point is AI is not new, and so much of the public in all of our various countries think it's new because of generative AI. It's not new. And in finance, it's been used for several decades.
Finance is a network of networks. I don't think of finance as a toolkit, but it's a network, an intersecting set of networks that basically prices and allocates resources and risk. We'll come back to that. So why is finance such a fertile ground for AI? You think about it, it allocates and prices risk. Risk is probabilistic. AI is probabilistic, stochastic. It's why it also hallucinates. But it's probabilistic, it's stochastic, and a lot of finance is probabilistic, stochastic. It's about risk. But it's more than that.
Finance is data-intensive, and it has been since antiquity. This is not a new thing. It's a remarkable thing about finance. It's so data-intensive. Will Jasmine—sorry, I thanked Jasmine earlier—but will Jasmine repay that automobile loan that she took out? Hopefully Jasmine will repay the automobile loan she just took out, right? Or maybe not. So, it's data-intensive.
But here's the thing: beyond its reliance on math, which we all know—it's very reliant on math; in fact, it invented a lot of math. A lot of math was invented over the thousands of years because of finance. But we've also digitized so much of it. In the last 40 or 50 years, we've fundamentally digitized. You cannot get a physical paper sovereign debt instrument. You can't get a piece of paper that's the Chinese government's debt or the US government's debt. It's digitized. We trade digitized more in the US than other countries, but it's so highly digitized. It's math-intensive. It's data-intensive.
And the other thing, since antiquity, it's symbiotic with technology. I stood up a fintech course at MIT, and I start with like, what was the first customer interaction? And it was not an app. At first, it was in tents, you know? Then you went to bricks and mortar, then somebody figured out what an API was, and my God, the last 30 years. But finance is what invented money as well, by the way. Invented double-ledger bookkeeping and so forth. So technology and finance are symbiotic, and AI is just the latest wave. John Hennessy talked about the waves of disruption in technology, and that was a wonderful piece that he put up.
Remarkable waves of disruption in finance as well. The internet came along, and Goldman Sachs, JP Morgan, Morgan Stanley, they thought they really owned equity trading. And within 20 years, equity trading really changed, and the Citadels, the Virtus, and the Two Sigmas stood up and disrupted that whole space. Now, Goldman and JP Morgan still are pretty darn strong, don't get me wrong. The incumbents are still there, but a lot of incumbents that were there at the beginning of my financial career are gone. So, it's a fascinating period, and technology is awfully disruptive, but this is as well.
So, where's AI in finance? Where is it now? Probably 50-plus percent of the applications—if Jamie Dimon said, "Hey, I want to know all the places we're using AI at JP Morgan," and they give him a list of 500 to 1,000 places, and maybe it's even more than 1,000, I don't have any confidential information—probably half of it is somewhere around customer interactions and compliance right now, and I'll come back to that in a second. I think the fun meat of finance is this middle bit: asset management, investing, and so forth. And then chief financial officers' lives are changing as well, and of course, coding and data management. But I really think the meat of finance is still not quite there yet.
By the way, finance has been using AI in some way at least for 30-plus years. And by the 2010s, the pattern recognition—I will call it traditional AI if you want, predictive AI if you want, pre-transformer model AI if you want, whatever you want to call it—is awfully good at pattern recognition. So in this space, it's great for pattern recognition. In investment research, we've been using sentiment analysis for years before transformer models. Sentiment analysis: Jay Powell speaks, even sometimes when Gary Gensler, I guess, spoke, a computer was listening. 500 earnings calls happen a quarter for the S&P 500; a computer is listening for sentiment analysis, meaning how did the CEO and the C-suite talk? Did they pause a little longer? Did they sweat a little bit more?
Whatever it is, if I was in investor relations at a big C-suite, I would say we should run our whole presentation through sentiment analysis: the slide deck, the talking points, the words, because they will matter as to how our stock moves in the next three hours. And I would be doing this sell-side research to understand how others will do it. I didn't do that at the SEC. I didn't really. But if I was Jay, I might—Chair Jerome Powell. I've known Jay since the 1980s when we were kicking around Wall Street.
Customer side: I was fascinated by Alice's talk on the panel earlier, but in finance, again for five to, I'd say, at least 10 years, AI's been used. Why? Natural language processing. Most of you take for granted now—my God, our postal services use natural language processing to tell what our handwriting is, right? Our chicken scratch, right? We just take that all for granted. We take photographs of checks, and the checks get processed. But natural language processing is already incorporated in marketing, customer engagement. But I really do think recommender systems are going to be personalized, particularly in the brokerage side: the Robinhoods of the world, and so forth, the crypto exchanges as well, that want more engagement, want you to engage more.
The compliance side is interesting to me. Anti-fraud: Visa, Mastercard, everybody's been using this. Again, this is pattern recognition—traditional ML, deep learning, pattern recognition, GenAI layered on top of it, pattern recognition for anti-money laundering, fraud detection, and the like. But what's fascinating about this to me, rather than customer engagement—what's fascinating to me is the threat actors. The threat actors are using AI in this whole space. This is an arms race.
If you're Visa or Mastercard, it's the threat actors on the other side. And it's not just the cybersecurity threat. I mean, that would be in and of itself something, and the ransomware cottage industry of LockBit and all those folks in Eastern Europe, that's one big thing, but it's also like how I break through sanctions. So it's kind of an adversarial—you know, whether some GANs and a bunch of AI that you can think through to protect yourself, to think like this. And maybe there's a commercial opportunity for somebody out here to be an entrepreneur and create something. Hopefully not on this side; I am not recommending—I want to make sure everybody knows I'm not.
Look, the AI tech stack, you know all so well. This is not for me to give you a course on the tech stack; I do that back with the MIT students. But I just want to say, where does finance compete? Big finance is data, and obviously everything about data: the orchestration layer, the app layer, that's where they are. That's where they have to be. That's where big insurance, big finance is—BlackRock, Blackstone, and so forth. But finance is also funding these other areas: the hyperscalers, the cloud, the Nvidia chips. So it's not just sticking to its AI API side. Again, these are two great systems that are intersecting with each other.
So what does finance do for AI the other way around? Well, it provides money. Think about the money that's pouring into AI that maybe some of you have already monetized personally and done well on. And I applaud you; I applaud you. It's also a bunch of risks. But I would note something else about this money: money and risk. Finance allocates and prices money and risk. Money and risk. In 2025, there's at least a negative funding of $300 billion. Now, the US economy is $30 trillion, so the math gets really easy—and we're at Stanford, and I taught at MIT, and this is an AI conference—$300 billion divided by $30 trillion is 1%. So, we're spending 1% of our GDP, the US, or more negative funding all of this. Call it burn rate, call it Silicon Valley thoughts, whatever; to me, that's a risk.
And this is just a little breakdown that came out in the last day or two, so it's just a fun little slide. The four big hyperscalers: $298 billion of capex this year, estimated to grow to like $400 billion next year. So the question is, how sustainable is this? And I would say we have a disequilibrium. Finance has disequilibriums all the time. I'm not throwing a red flag on the field, but disequilibriums have to come into balance at some point in time. And so I would say for any of you thinking about your cost of compute, all these hyperscalers have to figure out some way to fill this hole.
Do they have two years to do it? If you're a VC, how long would you give them? Would you give them three years? Would you give them five years? How long before they either figure out a revenue model on the inference side, probably—how do they figure out a revenue model, or they have to go something like Google did and figure it out through ads? Does OpenAI basically create an ad revenue model? I'm not saying they will; I'm not predicting anything. But somewhere they've either got to figure out a revenue model that fills this hole, or this spending and this data center boom is going to change, or all of you are going to pay more for inference. I don't know. I just think it's kind of like math, and it's just how long does it stay?
Now, you can look at other technology waves. You can go back to the 19th century and talk about railroads in the 1840s and the 1860s in the UK, in the US in the 1860s and '70s. You can look at electrification in the 1890s to the 1920s. And obviously John Hennessy took you through a bunch of other waves. I was not alive in any of those, and neither was John, but I'm saying they go out of balance. Big, consequential, general-purpose technology like AI—and I believe it is a general-purpose technology like the internet, like electricity, like the railroads and so forth—generally have overspending in the early years or somewhere in this cycle.
Electricity actually took a little longer to have the overspending. Remarkably, the first electric company of the US was in New York, Edison Electric in the 1880s, but it really took till the 1910s and post-World War I in the 1920s to just boom. And it took Henry Ford to figure out how to actually reinvision the factory. So, who's going to reinvision the factory with AI? Doing things beyond automation and reinvisioning the factory—who's the Henry Ford of the 2020s or 2030s, or does it take to the 2040s to reinvision so much of the factories around the globe?
Economics of AI: productivity is a big thing. People debate this. I'm going to throw a slide up in a second on the wide debate. It's the stages of automation all the way to transformation. You heard on the tech panels—Emmett was so articulate earlier today. He said, "Listen, you can automate things." Emmett, remember, on the medical panel, he said, "You can automate things, it's fine, but I'm really interested whether you're transforming something, and you've got to create value." And then on the same panels, it was like, "No, no, it's all right just to automate." So it's not like a right or wrong, but true productivity growth of gross domestic product will probably come from transformation and augmentation; automation might not lead to much GDP growth.
The question is, what's going to happen to market concentration, jobs and wages, growth, inflation, all of these things? Financial stability: I wrote a paper back in 2020 on financial stability. How many people in the room know the movie Her? What can you tell me about Her? Just shout it out.
