# Arvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse All-In · 2026-10-21 · 40 min · https://www.youtube.com/watch?v=yzk44lz-RGI ## Transcript Speaker 1 I know you’ve been using the T word a lot in your conversations over the last couple of days: tokens. But perhaps we could move away from AI a little. I suppose this is somewhat rooted in AI, but there’s been a lot of talk over the last few days about cryptography and some of the risks to it. I think it took root after the OpenAI math release, but now there are more conversations about this quantum initiative and how it could lead us into an era where we need to start thinking about adopting post-quantum cryptography. Arvind, maybe we could start by talking a little bit about where things are today, from your point of view. Arvind Krishna Sure. Something David just mentioned is that DARPA has been organizing a competition called QBI, or the Quantum Benchmarking Initiative. A couple of years ago, about 200 companies joined together. I think there are about 13 left now, so that gives you an idea of what might be real. I would say perhaps it’s not entirely real, though part of it is. The goal is to find out who can build a working quantum computer. Let’s leave the technological aspects aside. The debate is: will it be in 2028, 2029, or 2033? If we choose an intermediate point, let’s call it 2030, it will break current cryptography. That’s not an assumption; it’s just a reality. If that’s going to happen, what kind of cryptography should you use to protect yourself? What’s even worse for some of you is that if an encrypted hard drive gets stolen today, someone could keep it for 5 years and then break it. The good news is that post-quantum cryptography is understood, it’s well known, and the code is written. But people have been too slow to adopt it. I believe the time to wake up is now. Speaker 1 Are the solutions we understand quantum-proof or quantum-resistant? They’re based on our current understanding of what quantum mechanics is likely to be. Arvind Krishna They’re based on a theoretical understanding of what quantum mechanics could do. They’re quantum-proof with the same level of rigor as current cryptography, which can’t be deciphered by current classical computers. That’s not a theorem-proof-level answer. It’s based more on complexity theory in computer science, which is ultimately based on some assumptions. Speaker 1 If we look at where we are today and at that midpoint in 2030, can you explain the evolutionary moments we need to see—the points where we all need to start paying attention and asking, “Okay, how do we move forward?” Arvind Krishna Today, you have toy quantum computers. I’ll call them toys because they can solve certain problems, but they’re riddled with errors. They still don’t have a way to fully scale up. So the first question will be: “Who is the first to demonstrate—not a simulation, not in parallel—but some level of error correction or mitigation in a working quantum computer?” That’s step one. I predict you’ll see someone making claims next year, so you should get third parties to evaluate whether those claims are true. You’ll find that one or two will make progress by 2027—not on a massive scale, but some error correction nonetheless. The second question will be: “Can you run these things at some level of scale?” Not yet at full scale. I call them gate operations, for lack of a better word. Today, most are close to 10,000, which, if you think about a conventional computer, isn’t a big deal. Your iPhone does a couple of billion per minute. So, are any of the gate operations scaling from 10,000? Are they adding a zero every year? That’s the second milestone I’d mention. If you add those two steps and then add a zero each year, you could reach 10 million by 2029. That’s extraordinary. It can do things that will surprise everyone. Speaker 1 Does this refer to the notion that a logical qubit is an operational qubit with error correction? Is this a term you’ve adopted for qubits, so everyone has the same milestone, the same metric? Arvind Krishna I don’t love the term, but I think it’s fair to explain it to people this way: “Hey, if I say I’m building a computer with 10,000 qubits and they say, ‘Yeah, but they’re all noisy,’ you’re right. If I start fixing errors, it kind of reduces the machine’s effective capacity. So a logical qubit says, ‘Okay, what’s the effective capacity of this machine?’” I think that’s a reasonable way of looking at it. It’s a term that, while I don’t love it, isn’t like I’ve just taken these four things and turned them into a DRAM cell and that’s it. The whole machine has error correction, but it has some overhead. So, 10,000 physical qubits, maybe 200 logical qubits. Nikesh, you have red teams that think about what’s on the horizon and take it further. Let’s say you start to see these errors, these orders of magnitude, start to accumulate. If you had to guess, where would your red team say the obvious honeypots are? Where would you point if you were a bad actor? Nikesh Arora If you think about it, virtually everyone here has cryptographic keys running throughout their entire IT infrastructure. It’s not just one thing. Everything is encrypted using cryptographic keys. Each supplier in our space has keys at each end. So the first thing that’s going to happen is that everything that isn’t quantum-proof is going to be decipherable. Our view is that you can’t wait until 2029 or 2030, because it’s not like one day we’re all going to say, “Oh, crap. We’re going to burn $5 worth of IT infrastructure and rebuild it in 1 year because now we need secure, quantum-proof infrastructure.” You should be replacing your IT stack with quantum-safe infrastructure right now. There are technologies and capabilities already on the market that you can use to start wrapping post-quantum keys around traffic as an interim measure. You can do it, but if it’s 2029 or 2030, we’re already late. We’re 3 years away, and I’m pretty sure everyone in this room has a few hundred million dollars of IT spending earmarked for this year. I’m guessing 50% of that spending is on infrastructure that isn’t quantum-secure. Speaker 1 There was a fairly experienced AI researcher who said we don’t need it. 2029? Hmm? Well, I mean, his opinion was that current AI models can actually get us there before quantum computers do. Do you have an opinion on this? I suppose it’s rooted in the idea that, at the frontiers of number theory, AI is starting to develop models and tests that aren’t obvious to humans today, allowing us to rethink how we can crack codes effectively. Perhaps this week’s OpenAI math dump is an early indication that this could be a good prediction. Arvind Krishna Since you got a little excited about going down that road, you’ll excuse me if I get a little nervous. It’s not going to be possible. At the end of the day, this brings us back to things that people like Richard Feynman did. You can’t simulate something on a different architecture and expect a more efficient result. Speaker 1 Correct. Arvind Krishna If you take that as a thesis, quantum computers don’t operate on arithmetic, which is what all classical computers do, including GPUs. Could you get an AI algorithm to discover a new quantum algorithm? Of course. Could it run effectively and efficiently on current hardware? Yes, it’s not possible. You need quantum hardware to run these things and get the advantage that the quantum algorithm offers. Otherwise, it will be slower, by a factor of a thousand or ten thousand. Could you prove that it’s a viable algorithm? Yes. Could you get that benefit? Unlikely. Speaker 1 I think the argument is that, using classical computing, you can actually take a different path—an arithmetic path. For example, if you had a better predictor of the distribution of prime numbers, at the frontiers of number theory that we haven’t really solved, we could start to have classical models that can crack cryptography. Arvind Krishna It’s possible. The question is how long it will take, because that’s always the question. Speaker 1 Yes. Arvind Krishna I think it will be possible, but if it takes 10 years to crack a key, most people will rest in peace. That’s not a practical answer. That’s where I think the debate will be, but it won’t be proven until it’s done. ### Open Weight vs. Frontier AI Models Speaker 1 Let me bring the conversation to something that’s happening right now. I’d love to hear your opinions on the following concept. There’s open source, there’s closed source, there’s the frontier, there’s the strategy of going beyond the frontier, and now there are very reasonable alternatives that are at least open source. What’s your philosophy across the spectrum, and what gives us an idea of how we should think about models and capabilities in the coming years? Nikesh, how to begin? Nikesh Arora Sure. I think someone said earlier, “Don’t restrict your teams in terms of capability during the experimentation phase. Don’t start trying to optimize costs. Don’t start trying to create your own at that stage.” The teams working for us—20,000 people—have only just discovered AI in the last 12 to 18 months, and it’s not fully integrated into the organization. So if you make it harder for them to adopt it, if you make it harder for them to understand it, it will take us longer to be ready. Right now, what we’re saying is, “Use the best model you can find, the simplest way to use it, so you can understand and validate what these things can do, whether that’s from an efficiency, revenue, or product development perspective.” I’ll call that phase one. But in phase one, what we’ve discovered is that everyone has their own priorities, right? What people don’t understand is that there was a time when everyone thought cybersecurity was going to be consumed by AI models, and that our lives were over because Mythos was going to solve all our problems. Nikesh Arora Well, guess what? We inspected all the traffic on one endpoint at a cost of $20 a year. That’s hundreds of megabytes of traffic per endpoint, at a cost of between $20 and $40 a year, because machine learning runs on your endpoint, on your compute, and we run it on a payload of 20, 30, or 40 megabytes. If you want to take all that traffic and run it through OpenAI using tokens for every employee in your company, that sounds like a lot of revenue. OpenAI could probably generate a year’s worth of revenue from one company. That’s not a practical example of how models do it. You take a small language model, compress it, make it task-specific, and keep training it until it reaches 99.9% accuracy. You can say, “Instead of using machine learning algorithms, we’re going to use generative AI and small language models.” The point is that we’re trying to figure out what task requires what level of intelligence, and what that intelligence costs for that particular task. Suddenly, you discover that you don’t need a Ferrari for every task. And customers don’t want to pay for a Ferrari for every task. I’ll give you the other side of the coin. Suppose you’re suffering a breach. You’re being attacked, and you know you’re being attacked. At that moment, all the customers in the room say, “Don’t worry. Use whatever model you need to stop the data breach,” because they know the business impact of that breach is billions of dollars. At that point, I can run swarms of agents and do all sorts of things from a defensive perspective. You might want to use the most up-to-date, state-of-the-art LLM and say, “Think about it, reason about it, and figure out how to stop this data breach.” What we’re finding is that you’ll end up using a state-of-the-art LLM in one-off cases where the business value at stake justifies it. We can do drug discovery if we want. For most practical use cases in businesses, I think we’ll end up training some version of an open-source model. Even in peacetime, if we’re running a defense model, we’ll most likely take an open-weight model, use our own proprietary data, and train it. We’ll manage the weights, because that gives us auditability, explainability, and an open system where we can actually explain how something happened. Arvind Krishna I completely agree, but I’m going to take a slightly different perspective to arrive at the same conclusions. Let’s take the economics. Today, it’s $20 for a million tokens for probably generation minus one, and $50 if you want to go to the most extreme end. We could take one of the Chinese open-weight models, take a Kimi K2 with something like 800 billion parameters, and run it on our own GPUs with a little optimization. I’m talking about a week’s worth of engineering work, and you could reduce the cost to around $1 or $2 for a million tokens. What company in the world wouldn’t accept an answer that’s 10 times cheaper, or in this case, 40 times cheaper, for something that can effectively do 95% of the work? I think the economics are important. The second point is the chain of thought, not just the final tokens. In practice, you’re giving away your intellectual property if that line of thought can be captured, registered, and audited by someone else. The branding and intellectual property will make people say, “I need to run the models myself.” Well, you can’t run a frontier model on your own, because none of them will give it to you. So we’re back to open-weight models. The third point is geopolitics. The best open-weight models right now are Chinese. I think there’s very little danger if you put them in an isolated environment and run them yourself. Sure, if you run them from the cloud, it’s different. But I highly doubt that, in the next year or two, great models won’t appear in the West that are also open-weight and just as capable as the Chinese ones. The gap in these things is 6 to 12 months at most. Those three things lead me to conclude that I don’t know if it will be 75% to 90%, but that’s the range I’d put all token use in: open-weight models, for the reasons I mentioned, plus the reasons Nikesh mentioned. Speaker 1 And the remaining token consumption—is that because you actually need it, or because it’s the default behavior and people are too lazy to change it? Arvind Krishna I’m going to talk about developer behavior, and since I was one, I’m poking fun at my younger self. Developers are lazy. They’re going to use the most expensive answer unless you put a check on them. Speaker 1 Sure. Arvind Krishna Because it’s easy. Speaker 1 Yes. And that’s going to get us there. Nikesh Arora I would say that if, in the experimentation phase, you allow something to be proven to work, then moving to a cheaper answer for production is probably okay if it gets you there much faster. But doing that for production use is where I’m skeptical it’s worthwhile, except for a small number of use cases. Speaker 1 Nikesh, you were brought in as part of the small circle of people who got to preview Mythos, as a partner. I’m sure that’s true now, too. You’re doing it with OpenAI. When you’re sitting with these guys, part of you thinks they’re looking at your business and saying, “Wow, this is a big business.” Do you think they’re like that? What are the actual technical boundaries you expect to be respected? I ask because we’ve spoken to a couple of other CEOs. Marc Benioff is the best example: We’ve talked about ZDR, and he’s been incredibly aggressive about making sure there’s no data retention on behalf of his customers. The interesting thing about ZDR and all of this is that none of the models say they won’t generate synthetic data. They’re literally not there. You should look. There are all these loopholes and trapdoors. So I’m wondering what you’re thinking when you’re in the room. Do you come out and say, “These guys aren’t going to mess with my business. They’re not going to be able to mess with my business in any way”? Nikesh Arora Yeah. If you think about what happened when Mythos came out, we’ve been paying billions of dollars to coding agents. Well, guess what? Not everyone has ZDR in coding. There are tons of people who use cloud-based code and have no problem with that data training the models. So we train these models with end users who code, and the models just keep getting better and better. If you teach a model what good code looks like, it can also figure out what bad code looks like. The first thing Mythos did was figure out what bad code looks like. If you unleash it against your code, it will tell you what’s bad. The fascinating thing is that even then, it was hallucinating 30% of the time. But the problem is, it’s perfect for attackers because they only need one vulnerability. They don’t need 10. If you’re a defender, you need to find all the vulnerabilities and patch them. Arvind Krishna I think everyone in this room has gone through this process in the last 12 months. Everyone here has tested them, whether with Mythos, some cousin of Mythos, or some open-source model they discovered. Every piece of malware needs to be tested, understood, and cracked. I’m guessing that if I raised my hand in this room, people would say that roughly 60% to 80% were real, and 20% were hallucinations. From that perspective, models are great at identifying vulnerabilities. But I don’t think anyone in this room would tell you they’d use Claude Code to patch a vulnerability without retesting it. Speaker 1 ### AI and Cyber Warfare I have a very random question. You guys are in the business of defending companies, but now it basically feels like a war. As you both know, it’s like a war. Is there anything you can learn from how a war has been or is being fought and use to inform how technology should behave? I don’t know if this question makes sense, but I think you can improve together and collaborate. Since Arvind runs a program, we’re now saying that we’re all finding vulnerabilities, so why don’t we pool them and find the defensive ones? Instead of every company in the room testing all the open-source vulnerabilities they have, figuring out how to patch them, and patching them individually, you could take some trusted partners, give us your vulnerabilities, and we’ll patch them and distribute the patches so everyone can be secure at the same time. It’s sort of like how cybersecurity works today. If you find an attack technique in a different part of the world, we can take that attack technique, write machine-learning code, incorporate it into our product, and make sure it’s available to all customers. Is AI-generated code bug-free? Nikesh Arora No. Arvind Krishna No. Speaker 1 Okay, explain that to me. Arvind Krishna Well, first, about your question about war: What Nikesh talked about is that you can have powerful alliances where the alliance becomes more powerful than the attacker. That’s technique one. Technique two in warfare—and the most effective one—is asymmetry. Very often, an attacker is cheaper than the defender, so the defender runs out of money and, at some point, gives in. Can you make the defense equally cheap? In that case, you can basically have a symmetrical war, and then you have a much better chance of success. The third, which is how all wars have been won, is when you achieve technological superiority on one side. So let’s go back to World War II, Korea, Iraq, or any of those conflicts. If you achieve technological superiority on one side, you tend to win. Our goal is to see whether we can achieve technological superiority, and that’s what Nikesh is referring to. Can we find vulnerabilities, patch them, and put limits on them that make it very expensive for the other side to find a way out? That’s what I would answer. Speaker 1 Will agentic code ultimately be superior to human-written code? Arvind Krishna Absolutely. We can find out. We now use agents and coding agents for probably half the code written in our environment, and it’s superior to what was written before. The reason it’s superior is that agents can also perform evaluations and ensure that everything they can find is fixed before it’s released. That said, we’re only finding the bugs we know how to find. So we enter this strange loop of computer science. Is it totally verifiable? It’s not, because verifying it would be almost computationally unfeasible at this point. I’m not going to entertain the hope that it won’t turn into what you were talking about with number theory and all that. We might find a way to verify the correctness of these things, at least when the inputs are limited. If the input can be anything, then I think it’s very difficult. Does that make sense? Speaker 1 ### AI Regulation and the Global AI Arms Race Yes. Let me ask you about the implications. Of the points you just mentioned, it seems obvious to me that there isn’t a world where China and the United States, the West, or any other group of countries can join hands and say, “Let’s all stop the development of AI.” If that were the case, it would necessitate an arms race: a requirement for defensive or offensive superiority, a requirement for computing power—all these needs would arise to protect industry, our economy, and our jobs. Does that argument win out against everyone trying to stop AI? Should the political and social conversations about AI stop, and should we stop doing this? Does the inevitability of one side having it and the other defending against it change the conversation? I’d love for you to share anecdotes you’ve had with senior party officials who may not be in power today but could be, and whether they’ll change their minds because of this moment. Nikesh Arora I think it’s very clear that no one is stopping. Everyone seems to be working at a steady pace. It’s frantic: everyone is trying to get a better, smarter, faster model. It’s quite interesting. You can take the charitable view and say there’s no technology in the world that was created for people to build security and power. I don’t think the TSA existed when they built airplanes. It’s like the Wright brothers saying, “Great, we have airplanes. Let’s figure out how we’re going to take their shoes off. We’re going to pat everyone down before they get on the plane.” That wasn’t what they were thinking. Speaker 1 Keep the explosives out of your shoes. Nikesh Arora Yes, exactly. That wasn’t what they were thinking. Unfortunately, we’ve been shown the dark side of these models, and we’ve focused too much on it. Now the whole conversation is about the dark side and the dangers of AI, not its benefits. I think there’s a narrative problem. Despite that, I don’t think anyone is slowing down. I think the most likely outcome, which I hope is correct, is what you said—and what Arvind said earlier: if you build something, you’re responsible for it. If you build something and release it, you can’t say, “Whoa, I don’t know how to control it.” That argument doesn’t hold water. Speaker 1 Yes, because you wrote the software. Nikesh Arora Yes. And this whole argument that you didn’t write the software, that the software wrote itself, because the engineer is going to say, “Well, the agent wrote it”—“74565 of my harness”—it’s like, “No, you typed it.” The company typed it anyway, so there’s product responsibility with the company. You’re a business. You’re getting economic value in return for giving me something; you’re obligated to live up to the responsibility you agreed to. I think it’s that simple. There are enough regulations in the world—maybe too many—to hold people accountable for the things they put out on the open market. They have to take responsibility. Well, but, David, it’s actually impossible to control unless we think the whole world is going to sing “Kumbaya” and say, “Let’s slow down,” which Arvind said is unlikely. So, if that’s the case, fine. You put a deceleration here, maybe. Does that mean Japan slows down? Does that mean China is going to slow down? Does that mean India is going to slow down? No way. They’ll say, “Great, you’re slowing down. It gives me the opportunity to catch up and get ahead.” It’s impossible to really stop digital technology, and the European example tells us that. There, you’re afraid of all the possibilities. I’m not talking about the real dangers; you’re afraid of everything, and first you impose regulations against such a vibrant tech economy. Speaker 1 Yes. I mean, neither of you answered my question, so I’ll tell you about a conversation I had last week in Washington, D.C. I had drinks with a member of Congress, a pretty high-ranking guy, and the conversation had 2 parts. One was that one of the CEOs of the world’s largest companies is completely freaked out about cybersecurity, is going crazy, and doesn’t know what to do. I just had dinner with him, and this is all he said. Also, he thinks that by the fourth quarter we’ll see a data center ban in 50 states. That’s not likely, but 50 is a big number. Nikesh Arora A big number. Speaker 1 Have 50 states ever agreed on anything in this country? Nikesh Arora No. Speaker 1 Well, some things about the ’60s would be incredible. Nikesh Arora Yes, okay. Everyone get together. Speaker 1 Yes. Well, when everyone agrees on something, they know it’s not true. But I think it would be like a miracle. Both statements are being made by someone. So where do you think this gets resolved? How do we work out this path forward? He’s good at this. Try to answer his direct question. Arvind Krishna So, look, between us, every single member of your group—we’ve had different conversations with different members of this administration. The previous administration, too. The Senate lasts longer than a single administration; Congress tends to change much faster. This question has been there since 2022. The Senate had closed hearings in 2022, they had open hearings in 2023, Congress had them, and the White House had them in both administrations. They keep going around and around this question. The issue of AI regulation and data centers is driven by populism, not by any reality. If I can get more votes, should I look like I’m trying to do those things? Politics is politics. I’m not going to say that’s not real. That said, a lot of people also care deeply about the long-term economic advantage, and that’s the balancing act taking place. I think right now we’re in the right balance, but I recognize that one election can change our position in that balancing act. On data centers, I think all 50 states don’t have a chance of banning them. Is it possible that 30 will? Yes. The remaining 20 will be rubbing their hands together with glee at that point. Speaker 1 If we look at geopolitics for a second, it looks like we’re trending toward a world of haves and have-nots. I can say that everyone in this room is rich. You’re a Western market company, and you have the advantage of being able to work with people like yourself. But there’s a price of entry that might not make sense for a randomly chosen company anywhere else in the world. What happens then when these models are so capable and effective? What’s the actual manifestation? I’m sure you’ve had to think about this in terms of how you want to respond. How do you think about the gap between the haves and the have-nots? Nikesh Arora Look, I think Arvind said that you have to make technology so cheap that it can be easily accessible to everyone. I think we have an opportunity to make that happen. I don’t think it’s going to be that expensive. If you’re able to run an economical enterprise, it should be possible to find a way to protect yourself within that budget. I don’t think we’re going to end up in a have-and-have-not scenario. I think the biggest challenge, which is more relevant to the people in the room, is that I’m afraid we’ll see rapid advances in computing development in 2 major countries in the world. We may not see that distributed evenly across the rest of the world. Computing is very expensive right now, and it’s very difficult to get computing to run anything at scale. The challenge is going to be that the further you are from the epicenter, the less likely you are to have access to this computing. I don’t know if it’s due to economic reasons, or simply supply constraints. I think that’s the biggest challenge, isn’t it? You could be a company in Europe with the right intentions and motivations to innovate, optimize, and take advantage of it. You might suddenly find that, because you’re far from the epicenter, you don’t have as much access as companies in the Western world, in the U.S., or perhaps in China. Arvind Krishna I think we have to look at history because it gives us some insight into how these things play out. This is the classic labor-capital trade-off happening. It happened in the 18th century, again in the 19th century, and again in the 20th century. People who spent more on capital than on labor tended to win in all the previous 7 or 8 cycles like this. So that’s what’s going to happen this time. What Nikesh pointed out is that the 2 nations here are also quite willing to be capital-intensive. Speaker 1 Yes. It’s not as simple as that. Ultimately, they’re willing to admit that labor will be less important. That’s how it’s going to happen. Nations that become obsessed with social and leftist philosophies insist that labor must be protected at all costs. It’s not that there have to be opportunities for labor, but that all existing jobs must be protected, or the jobs that aren’t protected will be lost. I think the jobs that are there, as Nikesh said, are cheap enough that someone in a country can use that capacity. But what they’ll think is, “Hey, does that mean there’s a loss of jobs or a displacement of jobs?” You should accept job displacement because it can lead to a better job. I think that’s the social philosophy that will determine the winners and losers here. ### Vertical Integration and Software Consolidation I have one last question before I give it to you for the final question. This is the last question for both of us—you mentioned it here, which was great. Can you tell us the story? I’m a big fan of vertical integration, a huge fan. Markets swing like pendulums. There was a world, even back in the ’60s, ’70s, and ’80s, where the conglomerate existed. It was an incredibly valuable enterprise, and in a turbulent time, it was a safe haven for capital because the conglomerate was the mechanism by which you organized yourself. Then, when there was peace and prosperity, people would break them up. I think we’re going back to the conglomerate world. You were just talking about IBM’s vertical integration at its peak, and I wanted you to tell that story. Then Nikesh asked you what you thought of vertical integration there. Arvind Krishna What Jensen and I have often talked about is the IBM mainframe. By the way, there are some people in the room who use it quite well and at a large scale. The IBM mainframe has lasted 65 years. Why? It was vertically integrated. We’re a little less vertically integrated now, in the sense that we don’t reach down to the ground. In the 1960s, we reached all the way down. We sourced sand, made our own wafers, built our own factories, built our own chips, and went right down to the metal surrounding the machine. That prevented any leakage of intellectual property, but it allowed us to keep the profits of the entire stack in-house. Today, the one part we don’t do is the manufacturing itself. Still, the chip design, the engineering design, the resilience, and most of the software are in-house. You have to decide where to vertically integrate. Of course, I’m not going to manufacture our own metal. You buy metal, and factories are about scale. But I think vertical integration is important in areas where there’s a significant advantage. Going back to the infrastructure point you both mentioned earlier, it’s incredibly important to understand where you want to be and where you shouldn’t be if it’s a commodity. Speaker 1 Yes, but you can see that the modern version of that is Nvidia. If you look at what Jensen is doing, it’s a modern version of the vertically integrated stack. Jensen went and bought the land. He’s embracing the space; he wants to. He understands there’s going to be a balance between frontier models and other models. He’s trying to fund everyone in the ecosystem to make sure they don’t stray from the vertically integrated stack he’s producing. The good news is that if you control the vertical stack, you can decide which part of it to commoditize. You can move profits between layers. If you see other people attacking a certain part of the stack and destroying its profit pool, you move the profits to another part of the stack. You can see that happening. Market leverage. Very powerful. So, to put you on the spot: If you look at the application layer today, what’s the obvious vertical integration strategy that OpenAI or Anthropic should adopt? And what’s the defensive mechanism for companies that might be on the receiving end of that? Nikesh Arora I think you’re already seeing it, and I think we’re trivializing this conversation when we start talking about open models and open weights. It’s literally an open stack versus a closed stack. I think at some point the notion of running something like an open router will become impossible, because there are more and more models, and each model requires more memory, more context, and more understanding. You can’t separate context and memory from reasoning. If you want to separate reasoning from context and memory, then the open router is only responsible for moving the reasoning tokens back and forth. Over time, I think you start to see that it doesn’t make sense to have an effective end-to-end stack. It’ll be like a CPA you talked about earlier when we were chatting: cost per action. The action will require you to go through all the logic from the prompt to the result. Memory will require that, too, and it will make things easier, but there will be too much friction to move your task from one stack to another because it has all your context and memory. Look at what’s happening in the personal agent space. If you start training a personal agent with all your context, all your APIs, and all your connectivity, you’ll have created a moat. Now I can’t take that every day and move it to the next stack and the next stack. So I think you’ll find that the application layer will also have a vertical stack that goes through the entire process of software, reasoning, memory, and context. Speaker 1 So the logical conclusion is that everyone will have their own. Nikesh Arora Yes, because it’s the right thing to do. And then brute-force distribution wins again in the enterprise space, right? You’re seeing that in the consumer space, too: brute-force distribution is going to compete. In the next 3 months, all the cutting-edge labs will want to push hard on the MCP concept. They tried it 9 months ago; it didn’t work. Now, with Facebook showing us how it can work, I think almost everyone has it. There will be APIs. It will be the biggest battle for consumer control. Every app on the iPhone has to become headless. Speaker 1 Perhaps we can conclude by moving from vertical integration to horizontal expansion. It’s been a common business strategy for enterprise technology companies to expand horizontally through mergers and acquisitions, to drive synergies, cross-selling, and so on. Obviously, both of you have done this exceptionally well. How has your view been affected over the last year, or this year, in terms of valuations, multiples, and cost of capital? Let’s talk about mergers and acquisitions in the cyber realm, and then outside of the cyber realm. Nikesh Arora In cybersecurity, I’ve always maintained for the last 8 years that it’s foolish for our customers to have to buy 40 different solutions and bundle them together. We barely understand our top 10 vendors. Do you think you understand the other 10? You can bundle them together—God bless you. We’re working as fast as we can to take disparate parts of our industry and find a way to bring them together so that the data works together. I think it’s quite a journey. When I started my first job at Fidelity Investments many years ago, we had 23 systems that collectively would be called CRM today. We had master tables and tables that were concatenated with them. We had 23 different applications trying to maintain them. Today, you go to buy something off the shelf and you shudder to think about the cost of Salesforce, Microsoft Dynamics, Oracle, or whatever you choose. In 30 years, what’s happened is that it’s consolidated into a single platform. Customers no longer bundle solutions together, because these software solutions become consistent. They all become the same; they probably all do the same thing, and there’s no differentiation. I think the same thing is going to happen with cybersecurity. That’s what we’re trying to do. We’ve tried to find best-in-class solutions, put them together, and make them work so our customers don’t have to put them together themselves. Speaker 1 For some reason, the cybersecurity sector is enjoying much better valuations lately because of the AI mania, which is fine. That will rationalize over time because all these guys are going to spend unlimited amounts of money on cybersecurity. Outside of cybersecurity, I think it goes back to the conversation Chamath and I had earlier about how difficult it is to understand the terminal value of software. It’s not clear what software is going to survive the next 5 to 8 years, given where we are with AI and the notion that product-market fit is being altered for many of these companies. We all know how our current software should behave in an ideal scenario if AI were applied well. I know what I want from how I package the software I buy. I know exactly how it should work in a world where AI is first, if I could wave a magic wand and make it do things. Someone will do it. I think there will be huge disruption in the SaaS space in the next few years. Arvind, are you thinking something like, “Hmm, juicy multiples”? Arvind Krishna Over the course of the year, it was very attractive at the beginning, but not so much a month ago. We’ll wait and see how it goes over the next few months. But I really want to expand on the idea of Nikesh. This isn’t unique to cybersecurity. The average company uses between 500 and 5,000 different software vendors. In a world where you have to protect against everything, it’s not just about the cyber threat, but a huge attack surface. You have to ask yourself if you could replace 490 of them with smaller AI capabilities and whether you need to go back and leverage between 5 and 10 core platforms. That implies massive consolidation, because that’s where the spending is going. That’s what I think is going to happen. There will be consolidation because, if you want to keep distribution costs low, you can distribute everything. If the value of that software is falling, you can’t afford separate distribution for each one. I think that’s going to happen, whether it takes 6 months or 6 years. It’s somewhere in that equation. Arvind Krishna I think you also have to differentiate between software types. There’s software with a lot of interaction value, which, as we’ve clearly seen, can be replaced with a lot of AI. The 2 examples you used were things that were rich in interaction. Speaker 1 Correct. Arvind Krishna But if you ultimately get to the point where you need to be reliable to the power of 10 to the minus 15— Speaker 1 Correct. Arvind Krishna —which is a database lookup for someone doing retail banking or closing the books, I think that’s going to be difficult to replace easily with AI-generated software. So, I think there will be a distinction in how easy it will be to do that. Speaker 1 You’d need a digital twin for years before you switch. Arvind Krishna Correct. Speaker 1 And then there’s no cost advantage to doing that. Arvind Krishna Yes. Speaker 3 Guys, thank you very much. Arvind Krishna, Nikesh Arora, thank you. Speaker 1 Thank you. Speaker 3 Thank you to everyone. I’m going all out. I’m going all out.