Theory in the Age of AI: What Changes, What Doesn’t? | Christensen Institute Innovation Summit 2026

My name's Michael Horn. This is Rich Alton. We have the pleasure of finally having the AI conversation in the room, if you will, like the elephant in the room. We're the AI conversation in the room. I won't speak for Rich, but maybe I will. We come to this conversation with a lot of anxiety for a few reasons. One of them is not that AI is going to destroy us tomorrow. But instead, AI is something— it's a little like education. Everyone has an opinion about it. And so we know that depending on where you are in your own journey, how you use it, if you think about the different tiers of customers on the disruptive innovation diagram or your jobs to be done, you all have different experiences with AI, and it's a fast-moving field. And secondly, I'll be super blunt and speak for me. There's a lot of people from the Christensen family here who have lived and breathed these theories for far longer than we have. And so if we say something— what's the word, Bob? Dumb. You know, you all will hold us to account. So with that said, I hope you— Clay had a line that he used to say, well, I don't have an opinion on your problem, but I bet the theory does. And so we're gonna try to apply that today, and so I hope you'll judge us not on what we maybe predict, but how we use the theory itself, to go back to a line Anne had this morning in the keynote around teaching us how to think. So, with that said, let's dive in, Rich. You started working, just to ground people, I co-founded the institute, I started working with Clay in 2005, '06, You started working, my recollection is 2011, 2010, something like that?

2009.

2009, first fellow of the Forum for Growth and Innovation. That's right. Okay, Paul LeBlanc's been cited a number of times. You helped organize that higher ed conference that brought Paul. Right. I remember this email from you still where you said, you're not responding to Paul's emails, he thinks you hate him. We fixed that, but the question I wanna start with is, like, let's set the table. Everyone's talking about how revolutionary AI is, Is it so different in your viewpoint that we need a different playbook of theories?

Oh gosh, not at all. I think the theories are super well calibrated to analyze AI in all of its developments, both now and I think into the future too. Very fast moving, and that always complicates things. But just as an example, so 3 years ago, I was teaching one of our exec ed courses, that the institute does. And so I was standing in front of business leaders from Brazil who had gathered, and one of the first questions they asked me was, I took them through the disruption material, is, well, is ChatGPT gonna disrupt Google? Hmm. And this was a topic that was on everybody's mind. You know, Google's share price was being weighed down by this. It was all over the media. And depending on what headlines would come out related to this topic, Google's share price would go up and down. And so I just led them through the analysis to find out what the theory's opinion was. And when you apply the theory, at that time, you could— it was very clear that you could see that LLMs would be a sustaining innovation relative to Google's traditional search, right? And it aligned so well with Google's existing business model. What do they do? They run large numbers of computers to produce information upon request for people. And then they monetize that through ads. And so this was just gonna help their existing customers do the job even better than they had before, and Google could monetize it in the exact same way. And so in that class, they predicted, not me, but they predicted that ChatGPT would not disrupt Google's traditional search business. And in fact, that's exactly what happened. It's exactly what, it just, the theory nailed the prediction perfectly. To the point where it was just, I think, just last year Google announced that they were monetizing their AI summaries at just as good, if not better rates, than their old, just purely link-based search results. So I think it's very well calibrated to make predictions about AI.

It's super interesting also, 'cause I think sometimes we also have momentary reactions. The other theory, the other thing that the theory did in that case, allow you to predict is Google would struggle to put it in a disruptive application that upset their existing business, that they would definitely use it as a sustaining innovation. And you see that right now with Muse and Meta having launched this, and undoubtedly people have speculated Google has the same thing, these personal assistants that sit alongside you in development, and they're not rushing it out because they have to figure out how is it with the existing business model, not blow it up. And so how do you do it in an ad-supported way as opposed to having a personal assistant that is ignoring all the advertising and upsetting the business model is actually a really interesting puzzle. And the other thing that occurs to me out of that, and I'd love your reaction on this, is a lot of the reaction that I see in terms of AI is very point in time, whereas the disruptive innovation notion, if you will, is over time and dynamic. And it's not necessary, just because the technology curve is moving very quickly doesn't mean that all of the competitive dynamics of a given industry are gonna flip over within a month, or that people are gonna nail the jobs to be done. Your story was 3 years, right? Right. Despite a technology enabler that allowed you to do a lot of building quickly, Understanding people, getting the systems aligned right, and so forth takes time. And so these disruption stories of, you know, it's how did you go bankrupt? Gradually and then all at once. I think that still exists as well. There's a lot of upmarket movement that could occur, and there could be some disruptive entrants underneath still that maybe change search in a much longer time horizon. But I'm sort of curious your take on that.

Yeah, when we were making the prediction at the time about Google and search, I thought it was pretty clear that you could only really extrapolate that prediction out for maybe 2 to 3 years at that point.

Interesting.

Because the technology was gonna keep, the technology was so applicable across a wide variety of jobs to be done, right, that it wasn't going to be the case that just because it wasn't disruptive relative to search, that there wouldn't be other disruptive applications that would be found. I do think one thing that we see in the AI, just conversation in general is that there are certain, especially those in the technologist kind of camp, they do think things are gonna happen so fast, just so fast, and that we'll have to worry about, we will have to go to universal basic income because everyone's gonna be laid off. And it's just that belies a misunderstanding about, I think, jobs to be done and how technologies are actually absorbed and adopted by real people in the real world.

I wanna come back to this point very hard in a little bit, but I do think that interaction with the physical world, with friction of organizations and how decisions are made and the like, is something that is continually discounted. The other thing I think that's going on that makes AI particularly well-suited for the theory of disruptive innovation, and then some of the theories that spill out of it, the corollaries and other theories, are 2 things I'd argue. Unlike the last 20 years we've been living in with zero marginal cost economics with a lot of the digital products that are out there, that has been made, it's made it harder arguably in that market to quote unquote overserve people because you didn't see cost increases. The economies of scale were so steep, right, that I could be packing in more and more functions and features, and yeah, I didn't use 90% of them, but the price point didn't fundamentally change that much. I was happy to keep spending using, that at the moment anyway, we'll see how much cost decline of compute continues, but at the moment, we're definitely not in a zero marginal cost world. And so that seems to me to have profound implications that actually feel much more similar to, say, steel or other industries that Clay studied with huge fixed costs, a variable cost equation, disruptors that could flip that around at some point and the like. And so in some ways, I actually think it may be more applicable than it has been over the last 20 years, not less. The second thing I'll just sort of observe, and then I'll let you react or disagree, but is that folks, and I love Ben Thompson's writing, for example, but a lot of analysts like Ben have made the observation, gosh, AI is such a weird technology. It seems to be both used for sustaining innovation for Google and disruptive innovation in these sort of AI-native companies. That's so weird. weird, except that when you go back and look at the evolution of almost every industry, almost every technology is used for both sustaining and disruptive innovation. So Bob mentioned photography in your talk with the iPhone. Turns out Kodak was— you talked to Willy Shih across the way, he was at Kodak. They were a pioneer of the digital technology, and it's not that they missed it, they actually used it quite successfully for a very long time, to get people to print, to sell photo paper. And then when it became purely digital and a sharing experience and so forth, the world flipped on them, but there was this hybrid phase where the technology was used to sustain the Kodak business for a long period of time. You see automobiles, electric vehicles, disruptive relative to gas. Well, there's a hybrid phase, right, of it. Clay loved to go to the 1800s, Sailing ships. Steam gets invented in the early 1800s as a way to disrupt. It's not particularly good for long transoceanic travel. It's unreliable. It doesn't power the long miles that you need, and every once in a while you start a fire, it turns out, on your ship. But it's great in non-consumption, right, in these small inland waterways where you have to navigate very quick, and wind is a very unreliable technology. And yet, steam is still used as a sustaining innovation initially for those sailships, just to— when the wind died down, you could just provide a couple hours of power to keep going on your way as you were crossing from Europe to America or vice versa. Online banking, like we could go on and on down these things. Almost every single technology that comes out has both a hybrid phase and a disruptive phase. And it's very much the business model in which it's used, the use case, and understanding the jobs to be done that really cements whether that particular case is sustaining or disruptive. Your thoughts on all that?

Yeah, no, 100% agree with what you said. And just an important reminder for practitioners is the theories don't— you can't even really accurately apply the theories until you've defined a business model, which is— or sorry, a true disruptive innovation or sustaining innovation. And by that, we mean a technology coupled with a business model that's applied to a job to be done, right? And unless you've got all 3 of those things, you can't even really apply the theories. So when people ask, is such and such a disruptive technology, you need to better identify just the context of what you're working in.

So when all my students right now at Harvard say, is AI disruptive? And I say, Maybe. Tell me more. Right. Okay, let me ask you this. Interestingly enough, you know, there's this, all this obsession over the frontier models right now, and there's a lot being made around distillation from the open-source models that are coming from China and the like. How are you thinking about the interplay between the frontier models? There's conversations around the pause, all these things that are playing out. playing into this, these open models that are coming somewhat, it seems, from below. What would you say the theory, put it through the various tests, what does it predict? And maybe let's start to bring in some other theories as well.

Yeah. So for a theory nerd, this is the Super Bowl right now, right? Because we're about to experience a point in history with AI and LLMs that really only happens once in an industry, Right? So get ready, get your popcorn, okay? And that's what we call the crossover point, right? And we don't talk about this a whole lot sometimes in theory land because it just doesn't happen that often. But the crossover point I'm talking about is the point at which, on the disruption graph, where the functionality of a new technology or an innovation starts to exceed or shoot past the needs of the average consumer, right, or the average customer. With me so far, right? That's the point at which disruption becomes possible, right? Sure. And it also unlocks other things, which we'll get into in a second, but just basically everything changes in an industry once you reach that crossover point, right? And again, I hate to reason by anecdote, and I know we all have different experiences with AI, but personally, I have not seen seen much effective difference from Opus 4.0 to now Opus 5.5. It hasn't really changed my work at all. And that starts to lead me to believe that we are either rapidly approaching or have already reached the good enough point, right? And so once you've reached that good enough point, it enables what? A good enough competitor to be able to come in from the low end and start to sort of steal off customers that don't have as demanding of requirements. for their use cases, right? So the open-source models that are free or super cheap aren't going to be able to solve cancer or figure out protein folding and stuff like that, but they will be able to help you proofread your email, right? Or help you write your resume. And that's the vast majority of actual use cases. OpenAI actually released this data, right? And so I think there's a very clear possibility of, um, low-end disruption starting to happen where the frontier models will start to be disrupted by, by open-source models. And maybe that'll be them running in the cloud. Maybe that'll even be them running, uh, on local machines because the, the, the computing power that will be required to generate the vast majority, maybe 80% of the use cases of the average person, can be served on the local machine. Siri AI, Apple's AI that actually works, supposedly, now runs a lot just on device, right? And then it can send other things up to the cloud. And we're already starting to see, interestingly enough, we're starting to see flee-upmarket behavior by the frontier labs. Say more. Yeah, where— well, so it's not that they're trying to come downmarket to compete with the open-source models. They're releasing Fable, Mythos, and, you know, sort of these super high-capable models that the Defense Department can use and Amazon can use to defend themselves from cyberattacks and things like that. That is, that's the fleeing upmarket behavior that you see in the steel mills.

Super interesting. I want to, I want to double-click on one thing just to bring out a point, which is a lot of people will say, well, the open-source models in terms of compute Per— The cost per compute, actually higher in some cases than the frontier model, so maybe it doesn't apply, but I think this gets to one of Clay's favorite things to look at was this fixed cost fallacy and the sense of how much these frontier models, I suspect, are investing in the frontier to continue to go upmarket, and that maybe is the capture that prevents them from dropping down to serving people that are increasingly perhaps overserved by their existing models. But I'd love to hear your take on it because you're looking at this much more.

Yeah, so I actually don't know the engineering in detail at that level, but I guess my confidence in engineers worldwide is that they will relieve constraints, right? And so if inference is a constraint, you have Google and Amazon, they're already designing specific inference chips that are specifically designed to be able to handle that workload. And so Moore's Law would tell me that these constraints aren't gonna remain constraints forever.

All right, so I wanna get into another part of this. You mentioned that it starts to unlock a bunch of other questions once you get that overshoot point. Among those is a theory of interdependence and modularity, and maybe I'll just tee it up like that and let you talk through what you see as the dynamics, maybe. started to come out from there?

Yeah, sure. So it's not as, um, you know, the, the frontier labs aren't as interdependent as, say, you know, the IBM of old because we just have a much more vibrant technology ecosystem. But in a sense, you can think of them as interdependent solutions. And what happens, um, because they pr— they, they're controlling the compute, they're running the prompt through their compute in the way that they choose to. Um, and you have to send them the, the data to them and then have them send back the answer to you. Right? Well, so the other thing that happens in an industry once you reach the good enough point is that industries tend to modularize, right? Interfaces become more standardized and you start to get the ability of customers to assemble their own solutions. And so, well, I came across a story the other day about Cisco, because they were very worried about sending their data to these other technology companies, they started— they've started to assemble their own internal LLM running on their own hardware that's based off of, um, uh, based off of an open-source model, right? So they're buying the compute directly from, from NVIDIA. They're, uh, sort of running it at their own, uh, data centers. Um, and that's, that's exactly the kind of behavior you expect to see in an industry that's, that's nearing or has crossed over the good enough point, and so is modularizing.

And then there's a commoditization effect from that that changes where money is made across the value chain, right?

Yeah, yeah, exactly. So once you get an industry that's disruptible and that modularizes, commoditization typically starts to set in. And again, that's, I say, in the absence of powerful sort of network effects or consumer brands or something like that.

And this remains a big question on the network effects piece. And I think you could imagine sort of this centralization, decentralization, observation that Clay often saw where industries would start heavily centralized, decentralize out. It's interesting because it was almost a— Clay loved to talk about law of conservation of profits in an industry. There's almost law of conservation of centralization as well, where Google greatly decentralized search and being able to look things up and the like, but obviously deeply centralized the market because of the network effects that accrued in that, and this is one of the big questions, I think, on this local versus cloud-based, is this thin versus thick? This has been a debate for decades in the computing industry, but this will be one of the things that really hinges on, I think, the answer to that question is how much are there network effects from really using these frontier models, or is it the case that actually it's much easier to swap places? I know I was telling you, again, danger of anecdote, I started heavily OpenAI ChatGPT in my own usage, went completely to Claude over the summer, and because of their compute limitations and ChatGPT catching up, have gone back completely the other way. And I find no trade-off in terms of the switching cost because of the ability to create all these APIs and links into the context that I'll choose to give it. And so that strikes me as like one of these variables that is still worth Tracking to be able to answer some of these questions.

Exactly. But if those switching costs are really that low and people don't have any loyalty between the models, as we're seeing some evidence of that, and we're also seeing evidence of price wars, right? I mean, there's just continuous slashing prices on this model, slashing prices on that model, and you have continual pressure from the bottom, from the low end. It's not as strong of a prediction that I can make there, but it does, I do lean towards the camp of the attractive profits in the value chain are probably not gonna be made at the end with the frontier labs.

Well, so where would the money be made then?

That's a good question. I think for sure that just the semiconductors, I think, will continue to be scarce. So the theory says that the profits will continue to accrue to the performance-limiting stages in the value chain, right? And so likely that's data. data providers, and I think the hardware that really constrains things. And then if people are able to avoid commoditization, right, the theory says that the way that you avoid commoditization is to integrate tightly around a job to be done, right? And so Clay's example of this was always IKEA. They make commodity furniture in China, and yet they have a super profitable business. And why is that? His explanation was always that it's because they were able to perfectly nail the job of help me furnish my apartment today, right, in one shopping trip. And so the question I have about the frontier labs is, and you're seeing some evidence in their behavior, can they design solutions around specific jobs to be done, or are they gonna maintain their generalist roots, right? Most of these labs were founded on the principle of we want to drive towards AGI, artificial general intelligence. Right? So that it can do anything. Well, if they, they continue on that path, I worry that maybe they'll lose out on the opportunity to, to, to, to create this sort of job lock-in, right? Mm-hmm. And you see some, you know, Claude Code is obviously designed around a very specific job. They've got Claude Design. They've got, you know, Excel and PowerPoint add-ins, things like this. But it's not, it's not quite direct enough. I feel like it might be a missed opportunity for them if if they don't really go out and attack specific jobs to be done.

Well, and it feels like that's the opportunity for the builders, right? I mean, I think a lot of times, particularly in the education space where I spend most of my time, people say, oh, Claude has Claude Teacher now. We don't need to do X. Well, there's a reason people aren't rushing to adopt it, perhaps, because it hasn't very clearly done all the painstaking work of identifying that job, understanding the struggling moment, understanding the circumstance, and then tightly integrating around that, and I think the most enduring products and services in all these industries will be those that actually spend a lot of that time up front really understanding that struggling moment and purpose-build around it and knowing what and how to integrate. It's, again, it goes to this cavalier sense that I keep getting out of the messaging from these folks, which is just, oh, software inherently, or the AI will inherently assemble these components together, But the one thing we consistently see, Bobby alluded to it this morning, is that they consistently miss the context wrapped around people coming into these spaces. One part that I'll flip to, Rich, as we start to sort of wind up our conversation, we're going to try to get people back on time here. So we're going to do a couple more minutes and then we're going to release you into the 1:40. You can keep, by the way, as a reminder, the QR code is literally up there. fire away questions, not just on things that we've said up here, or that Bobby said earlier, that Anne said earlier, even stuff in the panels that went unanswered that you'd love us to respond to in the last session. Literally, take your anomalies, your questions, your disagreements, your tomatoes that you wish you had thrown at us, or me, you can put there. But I guess one other thought that relates to this commoditization/decommoditization question, is the Silicon Valley, I will call it, era of unemployment is upon us. My sense is that it ignores 2 very fundamental things about this space. One is comparative advantage and cost. The laws of economics and what Adam Smith taught us several hundred years ago have not disappeared, and particularly with marginal cost still being a factor, There is comparative advantage, it turns out, in humans for lots of things. And the second thing relates to this, which is that whenever something becomes abundant in a value chain and you see it get commoditized, the things next to it become relatively scarce. They weren't scarce yesterday because we didn't need them, we didn't know we needed them, but they become relatively scarce. That creates the opposite of the decommoditization, where there's value made and so forth. And frankly, there's a reason we all showed up today. I'm gonna echo Julia here, colleague. There's a reason you all showed up here and didn't just talk to an AI agent about all your questions about Clay Christensen. We wanted to be together in person because there's something different that we don't always know how to measure, I recommend Julia's Substack. There's a great piece you wrote, sorry to put you on the spot, several months ago about the ineffable connection of humans to humans that we don't always know how to quantify, but there's something about that becomes even scarcer and I think more valuable, and it ripples through, I think, lots of industries in terms of the physical world and how we connect, and I think there'll be lots of value we can't even imagine But we can predict will be there from the theory. I'm curious your thoughts.

Well, I guess I have a slightly more maybe street-level take on that, which is it just spurred the idea. I was coming back to Bob's session, Bobby's session from this morning about Intercom, which is that in a world where software itself is— the creation of software is completely commoditized, so anybody can assemble their own CRM. Well, what is it that still is not commoditized? It's the outcome, right? It's the outcome that we want, right? And so whether that's humans that need to be involved in that or whatever it might be, I think there will always be room for businesses and solutions that create outcomes, specific out— and actually outcome-based pricing is actually becoming quite a trend in Silicon Valley nowadays.

Yeah. Well, and interesting because it's been the opposite of the—

In the future.

Yeah, it's been the opposite of the solution shop businesses. And all of a sudden you may be able to do this in very particular use cases or jobs to be done, right, as it were, in very specific moments. As we wrap up here, I just wanna leave you with 2 thoughts. First, I wanna thank Saquon, Aiden, and Connor who have been doing the AV for us all day. We don't always get to say thanks to those people, but if you wouldn't mind. Thank you. And the second thing I would say is to be a Clay Christensen acolyte is actually to be somewhat of a Luddite. You all don't totally know what I mean by that, but because I'm thought of as like technology all the time in education, even though, by the way, I sent my kids to a Waldorf school for a couple years. But Clay was always very suspicious when the hype outreached, what new technologies would do, and he was very interested in the questions Rich asked before around what really is the job to be done, is there a business model, do we understand why people will pull this into their lives, and I think he would have treated AI no differently. He would have asked a lot of questions to put context around it and not jumped to a conclusion just because it's self-evident. So I might remind us of all that, that it does enable some unbelievable, almost magical things it can seem, and we should be careful in jumping to proclamations, and the theories can help guide us into that future. So with that, I'll just thank you very much. I hope this sheds a little light on how to use these theories, and thanks so much to Rich for being up here with me in this conversation. Appreciate you all. Thank you.