lot
29 stories and discussions about lot, aggregated from every source we track.
[This is a guest post by Amit Sahai. This blog post was initially written in a different file format and converted using AI. — T.] When I was an undergraduate student, I remember talking with…
<p>The question of shell scripting (Is it always a bad idea? How to do it well?) comes up a lot. There should be a tag for it.</p>
I finally hit 1,000 days of doing math daily! Early on in my journey I was a lot more aggressive with my XP targets, but settled into a low-volume rhythm as my goals evolved. I worked from MF1 (Math…
When I look at why I expect the world to change a lot in the next few years, and why other people expect slower changes, I think a big component is disagreement on the extent to which AI will affect non-computer work.…
[This is a guest post by Amit Sahai. This blog post was initially written in a different file format and converted using AI. — T.] When I was an undergraduate student, I remember talking with…
LLM agents can write a lot of code for you very fast, but they also take away opportunities to learn and grow.
Mocking comes up a lot in discussions of testing effectful code in Haskell.One of the advantages for mtl type classes or Eff freer monads is that you can swa...
A lot of the perspective on all the AI incidents has been shared from the outside in, and little has been said from the inside looking out, through the lens of a security person living through it.
A top-down parking sim about the last ten minutes of a fourteen-hour day. The lot is nearly full, the one open stall is between two sleeping rigs, and you have a 53-foot trailer to put in it.
Satoshi Nakamoto I'm not a huge crypto guy, but I do think Bitcoin is a fascinating project, both in the technical implementation and the origin story. Like a lot of people, I've had a lot of...
Hi everyone! It’s been almost a year since the last release, and now a new one is finally here. Unfortunately, things took longer than expected, so this is a pretty big release with a lot of new fe...
My last post on the Rust compiler’s performance was two months ago and a lot has happened since then.
AWS keeps customer costs down in part through networking advances
I’m old, and I’ve worked with a lot of companies, large and small.Some employed me, while others were clients.
Kernel developers do a lot of kernel builds. Since the kernel is not a small program, those bu [...]
In June of 2025, I was messing around with a large language model, basically doing what a lot of people were doing at the time.
As an industry, we like to shout a lot about experiments when we kick them off, all confident swagger that this time our superior intellects have landed from first principles on the correct reimagining of this or that…
Everyone in the world-models space is sitting on a pile of cash and a ton of buzz, but good luck getting anyone — from the founders to their own data suppliers — to tell you what they're actually building.
With a lot of help from ChatGPT I generated a little interactive animation to help visualize what is going on in the classic light-clock ...
<p>I'm currently reading "JavaScript for Kids" and it's a whole lot of fun. Though marketed for children I think that humans of any age could get a lot out of it if they wanted to start learning JavaScript. If technical books were more like this, with fun illustrations and frivolous use of exclamation marks, I'd read a lot more of them!</p>
Today, I’m talking with Mike Cannon-Brookes, who is cofounder and CEO of Atlassian. Atlassian is one of those companies that every other company runs on — it makes important platform tools like Jira and Trello that allow people to organize and manage big teams, create shared databases of company information, and generally allow work to happen. As you’ll hear Mike say, all of Atlassian’s products are actually different expressions of a single core platform, which really shapes how Atlassian itself is structured and how those products are built. All of this means Atlassian is also right in the middle of the way AI is changing how all these companies work — AI tools might be able to look at all the different tools and systems you have and just read them for you, making a big migration to Atlassian’s platform less enticing. Or, if you buy the idea of the so-called SaaSpocalypse , AI might just build all of these tools for you, destroying this entire category of businesses. Verge subscribers, don’t forget you get exclusive access to ad-free Decoder wherever you get your podcasts. Head here . Not a subscriber? You can sign up here . Obviously, Mike had a lot of thoughts pushing back on this. So we spent some real time talking about the value of design, human users, how AI is actually increasing the usage of Atlassian’s tools, and what this all looks like a few years from now. Of course, AI has also changed Atlassian. Like so many other tech companies, Atlassian did a round of layoffs earlier this year , with Mike saying he felt the company needed a different “mix of skills.” So I asked Mike what he thinks that mix of skills is and how AI is changing what it means to run a software company. But Mike was also refreshingly direct about what AI can’t do, and he was willing to push back on his own friends like Cloudflare CEO Matthew Prince, who was just on the show talking about AI eliminating what he called “measurement roles.” This was a fun one — and shoutout to Mike, who recorded this from Atlassian home base in Australia, which meant he had a deep conversation about org charts at 5AM. Okay: Atlassian CEO Mike Cannon-Brookes. Here we go. Mike Cannon-Brookes, you’re the co-founder and CEO of Atlassian. Welcome to Decoder . Thanks for having me, man. I am really excited to talk to you. I believe you’re in Australia? Yes. It’s tomorrow for you. What’s going to happen tomorrow? It’s tonight. It’s a great day here. It looks beautiful. It looks like it’s going to be an amazing day. At this point in the AI news cycle, I feel like I can just demand to know what’s going to happen tomorrow, and everyone understands the urgency behind that question. There’s a lot going on with Atlassian. There’s a lot going on with your products. There’s a lot going on in the very concept of SaaS businesses and business processes generally. Let’s start at the very start. I think people know Atlassian. They know Jira. I started my career as a Trello person. Tell people what you think Atlassian is today. Atlassian is a platform that enables businesses to collaborate and manage work across their teams. We connect their business teams and their technology teams to handle the most challenging work problems in a singular platform across any organization that is technologically driven. For any organization that has software and technology as its core competitive advantage, we make a broad platform that allows them to collaborate on content, manage projects, and unleash the knowledge of their teams and organizations across their business, their strategy, operations, all the way through to their service teams in all areas of their business. It’s a very large platform now that we provide. Just in terms of what work is today and what it might be in the future as we add more and more AI to these enterprises, maybe this is too reductive, but I have always thought of Atlassian as a company that makes tools to help teams figure out what they’re going to do tomorrow. Jira is a perfect example of this. You don’t do the software engineering in Jira, but Jira helps you organize large teams of software engineers and file tickets and prioritize problems and tasks. I was a Trello person a long time ago. My goal in life now is to never use any enterprise software. I feel like that’s a sign of true success if you’re just an iPad person in first class. I’m working on it. I’m not saying I’m there yet, but that’s one of my goals. But there was a time when I was the managing editor of The Verge and my job, straightforwardly, was to look at Trello every day and make sure everyone was doing the right thing at the right time. And I’ve always thought of Atlassian as that class of products. We’re going to organize the processes of the company. There’s something about AI that’s changing that. Has that conception of Atlassian changed for you? Is it changing faster because of AI? Understanding what work you have to do at an individual level like Trello, at a team level like Jira, or at an organizational level like our Strategy Collection for large-scale strategy and operations — that is definitely a big part of what we do. Understanding where processes are at — It’s less about work to do than processes. If you think about a business as a system of processes that are put together, everything in a business is some sort of system. It’s a process in the collection of all these, and how well we execute them is what your business is doing. The fundamentals of that haven’t changed. The way those processes run, the number of them running, and how to figure out what’s going on hasn’t changed. We think about Jira as a human reference to work. When we say that, it’s a really interesting term because, as you said, the work is not done in Jira. Developers don’t live in Jira. Marketers don’t live in Jira. Finance people don’t live in Jira. Jira is a workflow engine with highly collaborative parts that enables teams to understand what they, their colleagues, and their broader organization are doing. That’s why agents and AI and other things are going to increase the speed of those processes. In a lot of ways, they will increase the reliability, the quality, and the consistency of output of certain processes. They also put more emphasis on where the humans, the judgment, the intuition, the human bits that are unique, and the initiation come from. It’s our job to make sure that that’s all still understandable so someone can just look at any level of their business and know what the fuck’s going on. What are people doing? Are we successfully doing whatever it is that we’re trying to do, whether that’s service function in a finance sales deal, exception service desk to work out how many exceptions did we give out in sales this quarter, or it’s an engineering team who’s trying to work out, “There’s stuff happening all over the place. What did we build this month? Or what are we going to build next month?” All of those things are incredibly hard to do at scale and volume in a business. One of the more important features of the tools you’ve built is legibility. It’s user experience. You have a big database of exceptions given out by salespeople, and you need to look at a dashboard and see how the company’s doing. Or you have a whole bunch of tasks, and Jira will let you see at various different levels of abstraction how many tasks are being completed at any given time. That was something that was very important for human managers. A lot of the capability unlock of businesses as they added software like the software Atlassian makes is to just be able to see and coordinate more things at a higher scale. The turn that’s coming is that maybe those databases aren’t going to be looked at by human beings anymore, or they’re going to be fed into more synthesized databases, or AI is going to make some other kind of information at different levels of abstraction. I know you’ve got a product called Rovo which feels like the new face of Atlassian in some ways. What’s the path here? What does the end state look like? There’s no doubt AI is going to help with a lot of those processes. There’s no doubt it’s going to pick up some steps in a process. Most businesses at the moment have a process, and they’re using AI to automate 80% of this step. It may not be 80% of the distance. It can be the width. If you think about that sales deal exception process that you just mentioned again there, if AI can take 80% of the exception — say this customer wants 45 days, not 30 days — and we can write a relatively readable document, AI is pretty good at reading the document and saying, “Yep, this one’s fine.” But there’s going to be a customer that needs a lot more complex exceptions. They say, “I want to pay in 90 days in this way, and I want this thing and that thing.” That needs to go to a human for intuition and judgment. So 80% of those may be easier done with AI now because it’s very capable at processing our rules in this particular part of our business, and that’s great. It’ll increase the speed of the process, the quality of it. It doesn’t take away the need to know what’s going on. Someone in the business will still need to know, “How many exceptions did we give?” What happens in most of these areas is businesses find way more ways to make themselves scale and efficient, and they find way more ways to make higher-quality products and services with AI. I see this all the time where they are utilizing AI in ways that people don’t expect. It’s not all about mechanistic efficiency. There’s a huge amount of quality that is able to be measured and understood that was never before. The deal exception process is a good example because it’s actually a very human process. It’s not usually rule-driven. The rule-driven parts are easy. It’s the non-rule-driven parts that are more complicated and require judgment. The business still needs to know, for that process, “Do we need to add more people to this process? Do we need to change it? How much is going through it?” You still have a lot of work to understand the flow of work around an enterprise, which gets more complicated as the flow’s volume goes up a lot. I’m asking because I’m headed towards a series of questions about the supposed SaaSpocalypse, and I’m very curious where you think tool providers like Atlassian fit next to the frontier models are just going to eat more and more capability. Eventually, Claude will just do this for you. You’ll say to Claude, “Run my business for me.” In the backend, Claude will burn a bunch of tokens, and it’ll just do it all for you, which is a promise that many, many people seem to believe in. You’re describing a somewhat different path where the tools get smarter or more capable because of AI, and you’re still adding a lot of human judgment, but at no point is a frontier model just coming and running your entire business for you. If a frontier model is capable of coming and running the entire business for you, if it’s truly running the whole thing, even assuming four or five years worth of increase, you have a relatively simple business. Most businesses aren’t simple. They are global conglomerations of rules, compliance, laws, staff in different places, human inconsistency — which is the same as saying human creativity — that they’re putting it all together to try to deliver products and services for their customers in their industry: healthcare, university, finance, automotive, or whatever the industry is. They’re trying to compete with some other business. They are still going to have a huge number of humans, I would argue more knowledge workers, more developers, in five to 10 years time than today because the ability to do tasks to compete will go up. The bar for competition will go up in almost all these industries. You will still have a huge amount of things to go do. Most of the things that we think are simple will get done for you. That is the history of technology. I’m a big fan of saying AI is just technology. What we can do with spreadsheets, what we can do with the internet, what we can do with ordering goods online… You think, “Once Amazon arrives, we’ll all just never have to go to the shops again.” It doesn’t happen. I heard someone talking about bank branches as an analogy, which I thought was a really good one. Around 2000 when mobile banking came along, everyone thought bank branches were dead. Bank branches started closing down. Everyone wasn’t going to go to their branch anymore. It was a pretty negative place. At least it was in Australia. I’m pretty sure it was in America as well. If you look over the last 10 years, bank branches are increasing. People are like, “Huh, this is a narrative violation. Why is this?” And the answer is bank branches did close down. People did go to internet banking. They did use their mobile app. They don’t go into the branch anymore with a passport to say, “I’d like to send $100 to Nilay, please. Here are his details,” and write them all down. But bank branches have adapted. The reason they’re growing is there is a huge customer service element to a bank branch. The services they provide today are totally different than what they provided 25 years ago due to the technology of the bank and the availability of things. They are much more about helping you with higher-level processes and services than they ever have been. That turns out to be, presumably, profitable for the banks, which is why they’re opening new branches. Bank branches are going up. The things those branches are doing are totally different to what they were 25 years ago. I think you’ll see the same thing with knowledge workers and everything else, that it will still be incredibly important. I don’t see that part changing. I am working my way up to getting Jamie Dimon on the show to find out what happens in all the Chase Bank branches in New York, because I don’t actually know, but we’ll set that aside. Jamie will come on the show. We’ll ask, and we’ll figure it out. One of the reasons I’m asking all these foundational existential questions is to understand how you feel about Atlassian’s relationship to businesses and how businesses might change as we add new technology. I agree with your general framing that automation technology arriving to a business is a pretty familiar phenomenon, but there’s something else going on with AI that is either making that faster or more dramatic. It’s obviously happening to your business as well. You’re the CEO of Atlassian, and you’ve made some changes around AI. I want to ask you the Decoder questions now. How many people is Atlassian today? And how is it structured? We are, I think publicly, 12,000-13,000-odd people around the world. We are structured globally, I would say. We are “ Team Anywhere .” Employees have the choice to come into an office or not. About 60 percent of people come in three days a week or more. About 25 percent of people don’t come in a single day per week. We run as a globally distributed company as we have done since the start. When you start in Australia and San Francisco, you kind of get used to the Pacific Ocean being a little bit of a distribution. This was before Zoom. We used to have these giant Polycom systems in meeting rooms. We don’t have those anymore. Technological progress is nice. We have customers all around the world. We have staff all around the world. We are structured functionally, I guess you would say, with a lot of matrices internally. I don’t know if that’s the answer you’re looking for. No, that’s absolutely the answer I’m looking for. What do you mean by “functionally with a lot of matrices”? We have a CRO. We have a CFO. We have two CTOs at the moment. We have two large product groups, each of which has a CTO and a chief product officer, one in the enterprise and emerging side and one in the future of teamwork side. Our products are organized into collections. We have collections of apps that are all built on a single platform. Well over half of our R&D is on the platform. The apps are increasingly a smaller amount of the proportional total investment in building. More and more is on the singular Atlassian platform. Customers hire us as a platform across their business to get work done, not for a single application. That requires a matrix internally. We don’t have salespeople per product, if you want to think about it that way. The sales, customer success, and the FDEs are all customer-related motions. They’re organized as per the customers and their geographies and their size and scale. Then we obviously have to have designers and product managers on particular products who care about the purpose of a product, so you end up instantly with the matrix between how the customers are organized and how the products are organized. We have a giant platform. We have the technical platform that runs across all the products. It’s a very large platform team. We then have finance and talent and other things that are also broad processes. Naturally, you end up with collaborative areas that you have to get together, and we manage that through our operating processes. Do you use the Atlassian platform to run Atlassian? We do. We run entirely on the Atlassian platform. I used to say we’re the number one user — we’re not anymore — of our strategy and operations products to understand what our business is focusing on, what our large-scale goals are, how the people, the investments, and the technical systems all come together, and what is going right and wrong across all of those. We are a big believer in our Strategy Collection to operate a large-scale business. One of my favorite questions to ask enterprise software CEOs is how much they personally use the enterprise software. How much do you personally use your products? An awful lot. What’s your number one feature request for your teams? My number one feature request is usually continual UI congruence. I guess that’s the category that people would put it in. I’m a bit of a stickler for design and consistency. We want to achieve eventual consistency in design, in which we’ve come a huge distance in the last five years. In the last two years, we’ve been truly world-class at doing it. That’s usually my biggest area where I use a lot of our products on a regular basis. As such, I’m moving around the platform, and I’m looking for ways where this thing over here happens this way. They do it better over there. We’re trying to achieve that eventual consistency across a large product surface area. I use Dia, our web browser. I think 96 to 97 percent of Atlassian now uses Dia on a regular basis. I mean, that’s a daily basis, so I literally spend many, many, many hours of the day on it. We are building a web browser for ourselves because we’ve reached the point that we feel, “This thing should just be better for knowledge workers.” We’ve already built, I would argue, the best browser for knowledge workers in the world, and it’s getting much, much, much better. Keep watching this space. So, yes, I would say there’s no 24-hour period that goes by where I don’t use one of our products. I doubt there’s a six-hour period that goes by where I don’t use one of our products at the moment. I want to talk about Dia in a second, because I’m very curious about that acquisition. Obviously, we covered The Browser Company very deeply at The Verge . That was a big acquisition for you guys. Setting that aside for just one second, one of the things I’m particularly curious about here is the distinction between the platform and the products that run on the platform. Again, maybe this is just reductive, but it sounds like there’s a big layer of capability, and the products are expressions of those capabilities in different ways, but they’re all running on the same core platform. Is that how you think about it? Yes, absolutely. Yes. We have to build a singular R&D platform, whether that’s how we talk to AI gateways, whether that’s how chat works, whether that’s how automation and identity and logging and governance and compliance and content classification work, whether that’s how our home and our search engine work. There are many, many more technical capabilities that have to be shared. Then you have all of the next layer up, and customers want that to work the same. It doesn’t matter if I’m in Confluence or Jira or Loom or the Service Collection, automation shouldn’t be different per place in the world. That’s very expensive. It’s very hard. It takes a lot of time, but it delivers a huge amount of customer value and loyalty. Most vendors don’t try to do that. They literally don’t try. Then you have a layer up, which is important, which is the consistency of operation and how content can be mixed. Operations and processes can be mixed across applications. If I open a Confluence page inside of Jira, I want that to feel like a Confluence page, but I want it to be inside of Jira. I don’t want to leave Jira to open it. I just want to read it, close it, keep going. There’s a lot of UI layer, a sort of recomposability, if you want to think about it that way. Lastly, each application has to do its job. If you blend enterprise software together, you end up in a world of hell. People have tried this many times. You want a tool to do a job. You want to pick up a screwdriver or a hammer or saw and feel like that thing’s going to do the thing you hired it to do. The interface is designed to help you do that task. That’s where there is a difference in the interface layer and the interaction layer, between those tools. The combination of those is where a lot of our design challenges live. For example, our chat layer, Rovo, is amazing. I would argue it’s one of the best, if not very close to the best enterprise chat products in the world. Our customers continually say that. “How come you’re better than insert-famous-brand-here?” There are a lot of reasons why we perform better in evaluations and testing with Rovo. It’s not because we make a foundational LLM, I think, which is the flaw of thinking. However, using Rovo inside a product, if you’re going to Rovo for the sake of Rovo, is awesome. You have a question to ask, you want to search across everything, fine, you go to a particular interface. If you’re inside of Loom or Jira or the Service Collection and you just want to open up chat, ask a question, you want it to understand the context you’re in, but you want that operation to be consistent across the platform. That’s what we’re striving to deliver to customers. One of my constant tropes on the show is that if you describe to me your org chart, I can tell you 80% of your problems. The structure of the company leads to some kind of natural politics in some kind of way. The tension between “we have products that are designed to do a job” and “we have a platform that contains the core capabilities” is pretty well known. Mm-hmm. Maybe your product teams are going to desire platform capabilities that don’t exist yet or they will desire different platform capabilities from one another. How do you break those ties? How do you make those calls? We have a number of ways that we are different and we try to do that. First, I would say most importantly, we have a platform-centric CEO and co-founder who thinks in terms of platforms. Thinking in Systems is my favorite book, man. I have a copy on every desk. You have to put the platform first collectively, and then you go solve a bunch of problems. That does not mean genericism wins, but it does mean when you make trade-offs, you have to spend the extra effort and time to make it platform centric, to make it work. And that is a forever task. It helps to have a CEO, if you think about the referee across all the functions, that’s platform centric. They know where I’m going to land and so we’re going to end up there. Secondly, we have design as a very, very senior role. We have a massive investment in design. Our chief design officer reports to me, sits on our executive team. We now have three, four, five former chief design officers in the business from other very large businesses, some larger than us. We have a huge investment in design and experience. Especially in the AI era, that’s getting ever more important actually, much as Claude Design or something would tell you the opposite. I don’t think so. I actually think [when it comes to] design and the experience of software, if the cost of building goes down, the quality and differentiation is on the experience and the design, and that’s not something that’s easy to do. That requires a huge amount of taste and judgment. Lastly, I think we aren’t afraid to put things together in unusual ways in the organization where we need to solve problems. The latest example of that probably is our internal IT and engineering function now reports to our chief people officer and I’ve combined those two roles, which is unusual. “Why is HR running IT? That doesn’t make any sense.” It actually does, I would argue, at the moment because of the things we need to do to transform our business internally continually, which is that if I look at AI and becoming more AI native as a business, I end up with two piles of projects. A pile of projects that relate to talent, how I need to change and grow the people we have, get them using these technologies, get them thinking in certain ways. I have a huge talent problem. Hiring, growing, training, changing people is very difficult to be competitive. I have a huge system problem. I need to change our internal systems to have MCP servers. I need to make sure that everything is operable. I need to make sure that we have all of these system changes so that AI rolls out. How are we managing our token spending? There’s a huge amount of system change. I ended up refereeing a bunch of, “Is this a talent problem or a system problem?” As I say, that doesn’t make any sense. Usually the answer is this one’s 60 percent talent, this one’s 40 percent system, and this one’s the opposite way around. I’ve put them all together inside the business so that our system changes to apply AI internally and our talent changes to apply AI internally are in the same function right now, which is a little odd. The people who deliver the laptops and the people who manage compensation work for the same group, but it actually makes logical sense based on what we need to change about our business at the moment. What is this three-year epic that needs to change? We organize around the problems we need to solve and the goals we have as an organization when we need to. This is a brilliant segue, because I wanted to ask about this executive you have. I believe the formal title of your HR person is now “chief people and AI enablement officer.” That’s quite a title. And the point you’re making about the systems of the business needing to change leading to some change in your talent mix is coming true. In March, you laid off about 10% of the company. I watched the video you made in Loom about that. Your point there was you’re not necessarily replacing people with AI, but you had identified a change in the mix of skills that the company needed and you were going to make these changes proactively. Walk me through this. What made you say, “AI is here sufficiently that the mix of skills is different”? And what changes did you make to your systems to make losing 10% of your people effective or worthwhile? We have to start with the environment. The environment we’ve lived in in technology for a while is continual hiring and growth. The markets have changed in their view on that. Because one of the ways that we’ve managed changing businesses over time has been hiring people in different areas. If you have red dots and blue dots and you say, “We need more green dots,” you just say, “We’re going to hire all green dots for a while until we have a blend.” You manage it that way, that is no longer a feasible path, firstly. Secondly, we have certain areas of the business, our AI roots and area- Wait, actually, can I just ask one question about that? You’re kind of describing the COVID era over-hiring that every tech company was doing, and you’re saying that’s kind of over? Yeah. People hire at the rate of the best decisions they know at the given time and what they’re encouraged to do by others. Hopefully they don’t look at everybody around them, but I know that a lot of CEOs do look at the people around them and say, “I should do that.” I always think those are poor choices. Over-hiring, I don’t know. Pre-COVID, we were a few thousand people. You know what I mean? So we are four times, I don’t know, three times the size we were pre-COVID probably. Again, we’re accelerating as a business. We grew around 30% last quarter on a $7 billion run rate, which is the fastest we’ve grown in about two years. Customers are opting for more acceleration. I’m just trying to understand the “you can’t just hire a bunch of green dots” comment a little more clearly because I’ve heard from many of your peers, “During that period we just hired everybody that we could.” It sounds like that’s not specifically what you were experiencing, but it rhymes in some way. It does rhyme, but there are periods where that is the way to go. That’s not the current three- or four-year period of time. That’s the first thing I say because then you need to work at how we change our skill mix. We have a huge amount of internal programs to try to do that. We also have certain areas, our AI products and services specifically, and our enterprise sales areas are growing very, very fast. They’re doing well and we want to invest further in those areas. When you have certain areas you need to invest further in, you run into very difficult choices where you say, “I need to invest further in these two areas. I can’t grow the overall in a significant way. I have to manage how to transition the organization between that.” Which leads to a lot of very difficult choices. We are still hiring really hard in those areas in order to change the skill mix, but you also have a question of how fast we need to do that. It doesn’t, I always say, take away from any “internal training programs.” I say that in air quotes in terms of how AI is adopted and used. It’s more around learn, play, share, and mutual learning. It’s less like an L&D, in terms of “I’m going to sit down and watch a class on how to do this,” but that doesn’t take away from that. It’s a combination of all these factors as we’ve massively changed the business to be truly world leading at AI product development in the last two years. There’s no doubt that’s been a rapid journey and we’re trying to move as fast as we can to get there and to stay ahead. One of the things I’ve heard from a lot of CEOs in your position is, “The value of our senior people has increased.” There are more senior software engineer job listings out in the market right now because as you’ve said, that judgment and that taste is very important and maybe the tools can just build that stuff for you. I’m curious for your view on design. The idea that all of our designers will not be enabled to just ship code is very tempting to a lot of people. There’s a third view that says product managers roles do everything and all these roles are going to come crashing into each other. What mix of skills specifically did you identify and how are you prioritizing in this sort of classic triumvirate as more and more people get the ability to do all the other jobs? It’s very easy for people to have very binary takes on these things. I find it amusing that if you put an engineer, a designer, and a product manager in a room, they’d all sit there and they’d all say, “Oh man, it sucks to be you. Your job’s going to go away.” I say, “Ah, that probably is your answer right there.” We have a lot more product managers and designers who are writing code from prototypes to actually shipping code than we ever have before, that is for sure. We have a lot more marketers and finance people and HR people who are doing the same thing, by the way. The code that they’re writing and shipping builds on two things. One, a platform that has the engineering robustness to be able to scale and deliver the customers a European-compliant data resident solution that is necessary, the legal constraints of country X, et cetera. You have to build a technical environment that is able to get out to a customer with the required degree of customer safety. That is a non-trivial problem. That’s not something you’re just going to vibe code. If I’m vibe coding my own solution to something for my household, fine, it can just work well enough. I’ll go fix it if it doesn’t. If I’m doing that for a large organization, I need some other controls and rules around it. Secondly, the specialization of those areas all doesn’t go away. The human judgment required in different areas is like the edges of the roles blending. They’re able to do a little bit more of someone else’s job. I have a lot of designers who are able to ship prototypes that are working code prototypes instead of a picture or a fake prototype in terms of a mock-up. That’s great. I still see them sitting in Figma a lot and doing that. I have a lot of engineers who are sitting in Cursor or Claude Code or whatever on a daily basis, writing features and building things, and they’re able to do more design than they were before. That doesn’t replace the need for the other job. It increases the quality of the overall software. It increases the speed of collaboration between people if you want to think about it. If they each understand each other’s role, they can have a much more fulsome conversation more quickly. The specialized nature of what they do goes up. It becomes more spiky, which is really, really important. The ability of building platforms — that’s going to get overused here — becomes more important. Just this week, for example, one of our designers wrote an interesting blog post about how we now have an MCP (Model Context Protocol) and a CLI (command line interface) for our design system internally. We have an award-winning design system, atlassian.design. It’s very public. We publish all of our design system. I’ve always said in the role of design, 50% of their job is to make our design system and to keep it scalable, flexible, and buildable. 50% of their job is to solve the hardest 5% of problems in design because the half where they’re building the platform, what used to be templates and design documents that are now CLIs and MCPs to do design internally in the way that we want design to work so it’s consistent and coherent, that is a really hard problem to build that. However, they’re unleashing thousands, if not almost tens of thousands of engineers to build better-designed things when they do that. Their leverage is a lot higher in building a design system than it is in solving, “Help me with this screen.” It’s the same in security. It’s the same in customer support and customer service actually. If you think of bug fixing as the platform on which you reduce customer service, more and more of that role is required on building those scalable templates, platforms, ways that the other functions can build on top of so they can do that thing. If a product manager is vibe coding a feature, as in they’re just stabbing around some sort of code generating area and it’s an area where it’s not infrastructural, it doesn’t need to scale, it’s fine. It should leverage our entire design system. Turns out, out of the box, that’s not very workable. We’ve just shipped MCPs and CLIs internally, which are on V4 that are amazing at understanding our design system so the designers can evolve that system and the features that are being built will continue to match what customers expect in that congruence and consistency. That gets us huge leverage overall, but it’s a non-trivial thing. It doesn’t mean the designers are going to go away, the engineers are going to go away, the product managers are going to go away. I don’t think any of those roles disappear. I’m always joking that Decoder is fundamentally a show about org charts, which means I think I know what the next five years of the show is going to be about because we’re on the cusp of some of the weirdest org charts in history in AI. Matthew Prince from Cloudflare was just on the show . We talked a lot about his Wall Street Journal op-ed where he said, “I’m going to fire all the measurers at the company because I can replace them with AI. They can look at all of my systems and do all the auditing for me and we’re going to have fewer middle managers because what they’re really doing is measuring and auditing things and AI can do that.” There are reports that Mark Zuckerberg wants managers to have teams of 50 empowered by AI somewhere inside of Meta. This strikes me as not a great idea, but we’re going to see how it goes. Where are you? Are you in classic org chart? Are you at wild ideas? How is this coming together for you? I would say I know both of those people very well. I’m certainly more measured than that and I think the reality of how the world will be. I think it’s easy to point to extremities. Is someone somewhere a manager with 50 people? Sure. Are they effective? Maybe. People always talk about Jensen Huang. He’s obviously very effective. He’s a good dude. He runs things in a very weird way. Is that scalable and repeatable? I’m not sure it’s scalable and repeatable. I’m not sure if you took an entire set of large companies and made them all work that way, they wouldn’t have worse results as a collective. Part of business is scalability and consistency, and being able to be understood and managed that doesn’t require uniquely talented individuals to understand a very odd structure as they move around a business. I suspect you are going to get more role blending and people able to handle multiple types of skills. I suspect they’re going to be more generalist. The edges of the role will blur into the roles around it. A product manager is able to do a bit more marketing and a bit more engineering. That applies almost all across the way. Finance people are probably able to do a bit more legal than they were before. They’re able to understand the legal frameworks we have internally with agents and that’s great. I don’t think that changes the roles themselves. It may make teams a bit broader. I do agree with Zuck’s general direction that organizations will become a broader triangle, a fatter, wider triangle. The companies will get wider and hence less deep. Broadly that’s a true direction. Do you go to 50 person teams? Then the span of control is like three people. The organizational depth is like three or something. If you multiply them by 50 at each level, if there are two and a half thousand people with two levels of reports, I would question that. That’s probably not an average. I don’t know about this term, “measurers.” I don’t know that anybody could sit around in a knowledge work job and say, “My role, my purpose is to measure things. I just walk around with a ruler and I’m like, ‘That’s a foot.'” I know there are a lot of jobs where — I know what Matthew Prince is saying — that becomes the effective output of a role. We’re measuring it for a purpose, I would believe, and we are going to work out how to make that more scalable. My probable way of looking at it is a little bit different. There are two types of jobs. If you want to simplify an organization, there are supply-constraint and there are demand-constraint jobs inside a business. Input bound and output bound. But what is the constraint on a job? Look at a marketing or a technology function, a creative function versus a legal or a customer service function. Legal and customer service are input bound. If I deploy a lot of AI, I don’t suddenly get a lot more contracts. The number of contracts coming in, if I’m looking at NDAs or leases or trying to understand the company’s legal position in certain areas, I can potentially process that work faster with AI. Through systems, I can be more efficient in how I do that. That is one set of roles that you can look at almost everything in a business process and say that’s input bound. I’m going to manage that differently than things that are output bound. If I think about marketing or engineering, unless I run out of roadmap items — I’ve never heard any engineering team in my history of 25 years of doing this where they say, “Actually, we’ve reached the end of the roadmap and we’d like some more ideas of what to do next, please.” Said no one building software ever that has a large customer base. Their output, their creativity is bound by their own initiation and their ability to organize groups of people and processes and everything else, which is what the Atlassian platforms do. That is their limiting factor. It is not the human ability to create ideas. It may be that you have less input bound functions and areas as a proportion of your total employee base and you have more creative output bound functions. How much is this human ingenuity and creativity and imagination? A possibility. That’s likely the way things shift and trend over time. That’s going to take much longer than people think. We’re big in technology. “Next year everything will be like this!” and it takes 10 years to happen, but it will happen. I think I’m very fond of this term. I keep meaning to put up a rant online about it. I’m stealing another friend’s quote here, but we used to talk about information asymmetry and how much certain people know things and that is their source of power or advantage. I do think we’re moving to — he said it and it really hit me like a bolt — imagination asymmetry. We’re moving from information asymmetry to imagination asymmetry where your ability to create and think is going to be far more important as a competitive advantage for your team, your job, and your business than your access to information, your control of information, your understanding of information. That part is going to be cheap. That’s what LLMs are really good at. What they’re not good at is imagining things. They didn’t decide to start Atlassian. What business has been started by an LLM? Zero in the world ever. Anthropic has run a failed vending machine for several years now. Come on. But it requires a human to say, “Here’s what we’re going to do. Let’s fail at running a vending machine today.” They were probably sitting in front of a vending machine cursing the thing when it didn’t drop their candy bar and they said, “I’m going to do this better with AI,” which is probably the way that most of those companies think about every single problem as they’re staring at a vending machine. It required a human being to say, “Let’s go do this.” It required them to write a bunch of stuff to go create the vending machine operation code and try it. We all learned a lot as a result of this scientific experimentation. Let me ask you the other Decoder question, then I want to apply all of this to what is broadly happening to enterprise software and SaaS. The other Decoder question I ask everybody: How do you make decisions? What’s your framework for doing that? Someone told me you were going to ask this question. I said, “What a broad, open-ended question.” I would say I make decisions by talking a lot. That’s what people internally would truly say. I find the need to discuss things from different angles very important. I like to let other people talk. I like to get all the information together. I do believe there’s obviously a time where we need to stop and just move in a certain direction. The more important part for me, rather than how I make decisions, is I really like to explain decisions to people. At Atlassian we have a value: We’re an open company, no bullshit. I tell people all the time that does not mean transparency. We don’t owe you as an employee all of the details behind every decision. All the sausage making is not useful. It’s overwhelming to people. It doesn’t help us move fast. We do owe employees clarity on why we are making a decision and an explanation of said decision. That’s one of the most important things managers can do for training, for helping their talent grow is not, “We’re going left,” but, “We thought about it. We looked at the different options. We considered straight, right, and left. Here are the pros and cons. We’re going left. This is the reason why we’re going left.” Often people don’t do that because people don’t like the decision. I say that all the time. When you talked about our talent mix changes in March, we had to make a lot of hard decisions to run this business. I want to make sure that we are always explaining those decisions to people in the clearest and simplest way we can, why we’ve made a decision and what the decision is without any dressing. I really don’t like these corporate emails where I read the whole thing and think, “So what are you saying exactly? You’ve got a lot of words here.” I say, “Here’s what we’re doing. We need to change our skill mix. Here’s why. You should have the right amount of empathy to explain decisions, especially hard ones, but you also should not expect people to necessarily like it. We make a lot of hard decisions where I’m sure for 60% of the company, 7 percent of the company think it’s a great decision and 40 percent think it’s a terrible decision. We build trust as leaders by explaining our decisions and being clear about why we’ve made it and how we’ve made it. Hopefully we improve decision making. I always say decisions get to my desk when they’re presidential decisions. I think it was Obama that called it that, where it’s 51 to 49, that’s why it made it to my desk. If it was 90% certain what we should do, it should not have made it to my desk in the first place. I only get hard decisions. That’s a well-operating company. Any leader should be in that situation. If it’s an obvious decision, why did it make it to you? That then requires more thought and more explanation about those decisions. That’s probably the most important part than any framework we use to make them. Let me put all of this under some pressure. You’ve described Atlassian as a platform, and a lot of an organization’s decision-making and data and capabilities are in that platform. “What are we going to do next? How are we going to do it? Is it working?” The entire point of the SaaSpocalypse was that we don’t need your beautiful user interface. We don’t need your design surfaces or congruence of design across products. All we need is your data. If I unleash Codex upon your data store, it won’t even matter that it’s the same platform. You can have 15 different vendors and the AI will create itself some middleware and read and synthesize all the data and that’ll be great. Actually your business can leap ahead in meaningful ways. My favorite example of this from a previous era was I always had Daniel Dines from UiPath on the show. Instead of modernizing your healthcare billing system, he would just build robots that would use Windows 3.1 for you. This is a very early version of what I think every AI company is selling right now, which is, “Instead of modernizing your backend and spending all that money on a migration, our robots will just use your horrible old databases and present to you a new user interface.” That’s one version of SaaSpocalypse. It is. I see it coming for a lot of businesses. How do you expect to change all of Atlassian against that? Because a lot of what Atlassian sells is upgrade, upgrade the whole platform, use one product and use all the products, use the whole platform, modernize your entire business. The other answer is that it’s like $200 a month and you can just tokenmaxx your way into the future. There are a lot of different parts of that answer. You’re talking about the headless world, I guess. I say that headless is brainless. It is literally the same thing. It is submission in slow motion to do such a thing, if you’re a software vendor. I say that very carefully because again, it’s one of these answers that requires nuance. We have one of the most-used MCP servers in the world. We’ve just shipped a giant upgrade to it, rewritten it from scratch. It is massively used on a daily basis by millions of people and it’s amazing. [Along with] MCP servers, we’ve shipped three or four CLIs that are also used incredibly heavily. The ability to have inter-operation of tasks is very, very important for customers to get jobs done. I totally agree with that. However, people who use our MCP server and our CLI actually grow faster. They create more issues in Jira, they grow their seats faster. We’ve said this on public earnings calls, I think it’s like at least 5% faster. They grow their ARR at twice the rate of customers who don’t use our CLI and MCP servers. The customers using these things are growing their Atlassian spend faster than they have before. That world I totally believe in. When I hear the word “headless,” though, I think that is the only world that we’re going to exist in. We just sell an MCP server. That is the brainless part about it because then your software really has no value. If you truly just have a database, fine. If your interface sucks, and I think a lot of enterprise software has horrible, horrible, horrible interfaces— This is why I ask all the enterprise CEOs if they use their own products. It’s really asking, “How badly does your interface suck?” I’m a big fan of recording the Loom and saying, “Guys, this doesn’t make any sense.” If you do that, that yak-shaving over and over again, you end up with amazing interfaces and that’s why tens of millions of people wake up to our software. The investment in design and experience that I talked about is the antithesis of that. The world is very different than those views. We’ll see in a few years time who’s right. The way the Atlassian platform works is not to be a single platform. I am probably the only CEO that goes to customers, massive large-scale customers, and says, “Do not centralize everything on my platform.” I do not believe that is a viable strategy for a CIO. I don’t think it is. Our platform is highly connected. It always has been. One of the things that Jira, Confluence, Trello, and the Service Collection do is interact with many, many, many, many other applications. You look at any work item in Jira that’s been going for a while, it is full of links to Workday and Salesforce and ServiceNow and Figma, links to source code, links to Google Drive and Teams. Links are what build our context graph. The reason we have the best context graph in the world is that we have a lot of links to start with and we have had a 10-year history of having to understand how to be collaborative with other applications. Our applications are by nature highly connected to the rest of your IT world, the rest of your technology world. We don’t want to be a control tower. It’s a stupid analogy. Every CIO has been told, “Just trust us. We’ll deal with all the problems.” It’s never worked. I don’t think it ever will. The reality of technology is you’re going to have more— Once you add AI to the mix, maybe it’ll work this time. No. This is the thing that is being sold: You have now a powerful intelligent orchestration layer that will finally realize the dream. Three parts snake oil, two parts bullshit. The reality is it will help all the applications if built well to work better with each other. There’s no doubt. Look at Figma and Atlassian. We have a deep partnership. We use each other’s MCP servers a lot, and we can build better solutions for marketing teams and designers and technology teams between the two platforms than we ever have been able to before, because we can understand each other’s data in ways that we never have been able to before, and we can solve workflows on different ends. If you’re sitting in a deep design, you’re sitting in Figma. It’s a visual experience. If you need that to be tied to work and then understand how to find that, that moves back into our world and we want to make those journeys as seamless as possible. We want to be not a control tower. We want to be a subway station that is highly connected. We want to be Union Square. We want to be the town square. We want to have a lot of lines going through us and we want to make that experience for our customers the best that they can. I think 98% was our stat on the earnings call. More than 98% of people using our MCP server also use the user interface, and that is not changing. The ability to do work faster doesn’t mean I just stop using it. If you invest and have a good interface and experience and you’re solving higher level problems, if you look at the features we’re shipping around agentic uses, how do these harnesses — Claude, Codex and everything else — work inside the Atlassian world and vice versa? Now, one world is I open up Claude. I say, “Go work on this work item.” It uses our MCP server to go and get the work item. It uses our teamwork graph to go and get all the context around the work item that is necessary and related, and that gives it the best chance of doing that job the cheapest way possible. That is certainly a way that people are going to work and Claude’s going to move the work item along to the next stage. Awesome. There are a lot of other people who start in Jira and will say, “Send this work item to Claude.” On the board, it tells them the work item’s running and going. They’re in an interface that is built for the task, to go back to the saw and hammer and screwdriver analogy. It’s built for the task they’re doing. It needs to understand, “Claude needs your attention” inside of Jira. You can open Claude, say you need to do it, send it off. Most people don’t use Claude. Most people don’t use Codex. Every customer I go to has a whole bunch of harnesses running. They’re going to have internal ones they’ve built. They’re going to have two, three, four vendors. Most people aren’t making a choice. They’re all amazing. They’re all leaping over each other in progress. Our job is to bring all that together into a coordinated surface, to understand your agentic sessions, your human sessions, your work, and to connect all of your enterprise applications into a giant graph and give you phenomenal answers across the whole thing. That requires a lot of interface investment. That is hard. That takes time. That takes great design and experience folk. That takes taste-oriented product managers and engineers. That is a non-trivial challenge, and I do think a lot of people are avoiding that challenge, and I don’t think it’ll end well for them. I don’t think the world is going to live in a CLI as an end tool. You don’t think the chatbot is the final user interface? I don’t think chatbot is the final answer, no. We learned this in 1970 with the CLI. We don’t know. We’re all learning what the future interface is. It doesn’t mean that chat’s not going to be a part of it. It’s just, “Where does this world work?” You talked about Dia. Dia is a good example where we’re trying to build an AI-driven experience for knowledge workers that is based in the world they understand and still magical and different, not based in taking them out of the world they understand to a chatbot. So Dia is the browser. You acquired a company called The Browser Company . We spent a lot of time with The Browser Company when they were working on their products. At first they were building a consumer product. Then the notion that what they were building was more like an operating system came into focus as we covered Dia in their product strategy. Most of those conversations and most of that coverage — most of which was done by my colleague David Pierce — was happening simultaneously with Google search traffic dropping and the very notion of the browser as an information document viewer was changing in big ways. If you think the homepage of Chrome opens to Google and that’s going to take you to a bunch of webpages, Google actually has a lot of ideas in between there that is changing traffic on the web. And yet at the same time, the web as an application environment is a thing that can make AI agents vastly more capable because the applications all live somewhere that they can go address them. It was obviously explosive at that time. As you and I are talking, it’s like a week or two into Meta launching Muse. The power of Muse is that Mark Zuckerberg has given everybody 8GB of RAM and 8GB of storage in the cloud with the web browser. Now you have an agent, they can just go use a web browser for you. This is very powerful and Amazon is furious about it and has blocked this web browser. Do you have Dia because you need to control the application layer, or is it something bigger than that? No, no, no, no. Two things are important to understand. When you get technological disruption, one thing that happens is that we all move up the interface stack a level. There are not a lot of people competing about DOS or operating systems anymore. In the modern internet world, one way to move up the interface stack is to move towards the browser. Features that were inside of applications move to being inside of the browser as we increasingly learn and operate that way. Second, the biggest thing that’s changed — Usually what happens is that people’s habits change. Knowledge workers, you, I, everybody spend 80% of our lives in a browser, but we’re not reading articles all the time. We are sometimes. We’re not reading webpages, we’re using applications. We’re using Gmail and Google Calendar and we have Slack open. We’re using the Atlassian platform, all of our applications, and Salesforce, and we’re watching videos like Loom and we’re interacting with those videos back and forth. That set of tabs is no longer a set of webpages. It is a set of documents and applications that have massive power underneath them. We’re trying to rebuild the web browser in a totally new form for knowledge workers who live in applications who have their browser open 80-90% of the day as their front surface area, and they want a better way of managing those applications and getting work done across those applications. The browser has some incredible power to be able to do that. It requires us to understand SaaS apps very deeply, which we do. It requires us to understand enterprise security very deeply, which we do. It requires us to not chase the consumer audience. Chrome and Safari are built the way they are because they have to cater to everyone from five-year-olds to 100-year-olds, who are doing fewer work-related knowledge tasks and are doing more browsing and entertainment and other tasks. That doesn’t mean that someone who’s a knowledge worker in a corporate job isn’t looking at YouTube or Facebook sometimes, but it’s not the purpose of their job. We believe we can build a better browsing experience, a better work experience. One of the pitches we had for the product is we’re not building a browser, we’re building a doer. It’s a totally different thing. We’re helping you do things. We’re not helping you browse passively, and hence why there are tabs and split screens and pin tabs and groups and AI and dynamic tab groups. Live tab groups as we have of the documents you have is a great example. Our documents live tab group in Dia shows you the documents. It will automatically open and close tabs if you want to think about it that way, for you across any content application, Google Drive, Confluence, SharePoint. Any application you use that has documents, if someone comments or shares a document with you, the tab automatically gets opened by your web browser. And if you’re done with it somewhere else, it automatically gets closed by your web browser because that is the set of documents you’re currently working on. Same for pull requests or anything like this. Then we add AI over the top. AI lets us change how we use these knowledge applications in insane ways. The browser has your entire history of what you’ve done. It has skills, it has an ability to run AI. You can use Dia to make Minesweeper, and you can use Dia to make Minesweeper out of the tasks that you did in the last week. It’ll go start a computer underneath the browser. It’ll run a bunch of code generation. It’ll create an application for you that plays Minesweeper based on the tasks you have completed in the last week. Now, I don’t think that’s a very useful application, but the fact that it can do it is amazing. The most popular feature of Dia is the Morning Brief. People are blown away and often suspicious about how we can understand what you need to do today with a fidelity and a design quality and an experience that makes it a joyful experience when it pops up every morning. Mine opened about 45 minutes ago. That is a design challenge. That is an amazing team of designers that are thinking about our most-used application surface area and reimagining it for knowledge workers and for the AI era simultaneously, with access to every SaaS application that you have open on your computer in a way that corporate security is going to say, “This is a good idea and I trust it.” That is the mission that we embarked on a year ago, that The Browser Company embarked on three years ago, and it’s usable today. It’s coming out on Windows any day now, it’s in beta on Windows, general availability is very soon, and it’s used on the Mac by a huge number of people on a daily basis and only increasing. Like I said, we’ve been covering that product for a long time. I’m glad it found a home with you all. Let me just push you a little bit on moving up the stack to the next level of user interface. You could argue that already happened with web browsers. Windows was disrupted by the very presence of the web browser. The iMac was only successful because the most interesting thing you could do with a computer at that time was go on the web. The presence of the web browser on the Mac allowed a bunch of applications to move away from Windows binaries to being web apps and then cloud apps and now productivity apps in the cloud. The web as an application environment is probably more vital and important than ever before. It is the thing you’re describing. It’s where all the applications live and a custom web browser designed to do work is very important. At the same time, I would just say, didn’t that already happen? Are you actually moving up the stack or are you just building a more custom web browser, or is AI changing that in some other way? AI is changing that. AI is giving us the ability to help you understand what work you have to do, not just your browsing history. One of the most powerful things about AI that people don’t broadly understand is… An LLM’s ability to generate content is well known. I can say, “Tell me a dad joke,” and it spits one out. It streams out tokens. What is more important for application vendors like us, people who are trying to make tools and products and solutions for customers, is the ability to read text, not write it, but read it, because we can understand meaning for the first time ever. We can understand what a page or a comment or a piece of text is about. Then we can make application-level choices, which we never could do beforehand. Word didn’t know what document you were writing. When it said, “It looks like you’re writing a letter,” it was really shithouse and it was a terrible experience. Now it can actually tell with super high fidelity, “Are you writing a letter, a PhD? Is it a high school essay or a university essay?” It can adapt and change its interface. The ability of Dia and AI is to understand what it is that you are browsing, what it is that you are doing, to be able to reach into applications to get content and to pull it back to give you tools to manage your day. The Morning Brief looks across your calendar and your messaging tools and your document tools to tell you what it is you have today in a very intelligent manner, the way it would be if you gave me all your content and said, “Spend the days.” Now it reads millions of words in order to generate that and it has a lot of smarts to give you a very simple experience of the most important things you need to do today that are totally different from anybody else on the planet. That requires a lot of AI to do and it requires connections to all of your SaaS apps. The browser’s logged into all those apps so it can do that. It happens locally, it doesn’t happen in the cloud, it happens on your machine. That is something that we could never do before to understand, “Of these 10 documents you looked at yesterday, here’s the one you need to do today because I know in your calendar you got to do this tomorrow, you got to do these things. This is the most important thing for us to do today, and let me help you get started on that task. Let me draft some notes for you. Let me help you with that document. Let me help you do things.” That is something that was never ever possible beforehand. Lastly, it understands the people you interact with. People are left out of so many of these conversations and equations. You can think of Dia as having a mini CRM that is constantly updated based on your browsing habits. It’s a weird analogy, but it’s effectively what it’s doing, one of the many things it’s doing inside the browser, which is why its answers are so good. It knows that you and I don’t know each other very well. It knows I know my head of engineering extremely well. If that person asks me for something, because it’s maintaining a constant context graph of people relationships across every single one of your SaaS applications into duplicating and normalizing and using AI, all of that helps to understand what tasks are actually important to me. It can look at LinkedIn and see who you are and say, “That interview prep you got to do today, you should put that up your stack. Nilay’s an important person.” It does all of that for you, and help you- Can you put that in the Dia system prompt so people know when they come to talk to me? Just hard code that in there, you got to prioritize. We’ll see if we know some people. But it’s hard, and the experience, none of that should need to be known by somebody using Dia. The experience should just be a beautiful, joyful way to start your day. We want to leave you with more sense of calm, control, comfort about your day, not stress about AI. Leave that to us to do really well and we just give you the answers that you need from your tools and applications. It also does not mean you use less of those tools and applications. I don’t use Dia and stop using Slack. I don’t use Dia and stop using Confluence. It’s helping me understand which thing I need to do next by understanding what’s inside those things. By cross-referencing with my LinkedIn and my Salesforce and my Microsoft Teams conversations and meeting transcripts, but it’s doing that all for me in the morning and during the day to be able to say, “Here’s what you have to do next. Here’s the most important thing,” and providing an interface to allow me to organize and group those tasks in ways that hopefully are bordering on joyful. They’re beautiful and simple experiences and there are so many lovely touches in Dia that just make it a joy to use every day. We’re pushing that mission really hard. Can I add a wild card? It seems very obvious to me that you have a very clear view of how AI should work in enterprise and how work should get done using your tools, your tools enhanced by AI, your tools expressed in your browser. There’s focus there, and I appreciate that. On the other side of the house, on the consumer side, my argument has been that reason people hate AI and they’re mad about it is a little bit that the doomers have said there’s a 10% chance we’re all going to die and all of our jobs are going away, but there are also no great products. There’s not as much coherence and focus as I’m hearing from you in the enterprise context. Do you think that there’s some great consumer AI product in the offing that has the kind of focus that you have in enterprise? Because that seems like the missing piece for me. Yes and no. Yes, if you look at the history of technological disruptions, the products that came through the disruption often aren’t linearly related to the disruption. If you think about the smartphone era — and we all pick up our smartphones 100 times a day, don’t tell me you don’t. When the iPhone came out, we were all excited by little bits and pieces. We did not think, man, that camera thing’s going to lead to this Instagram phenomenon. We did not think the GPS is going to lead to Uber and so I can take transportation. The products that came as a result of the technology shift. The iPhone is an amazing business, don’t get me wrong, but the services and changes to life didn’t come from the iPhone. They came from the apps that it enabled, like Uber and Instagram and whatever else you like on the mobile phone. It can be a game, it can be anything. Those take a while for people to play with the form factor to understand the bits. Uber didn’t come along until the iPhone 6, because all it needed was to get better and then someone said, “Wait, we can take a car with a phone and it’ll just show up?” Yeah, we can show you where the car is and we can build enough trust and we can build a whole service layer. I suspect one part of it is in AI, we need to keep playing, and I know the Muse team pretty well. That’s an amazing example of an application. There are many, many more of these things. There’s a company that one of my kids loves called Tolan which makes these Hollywood AI agent chatbot things with personality. Someone from technology would say, “Ah, that’s just OpenClaw with some skin on it.” Kids say, “This is a character that I can interact with.” It’s an amazing psychology Hollywood game character. These sort of applications we’re just starting to see. It’s going to be an incredibly exciting five years of those sets of applications. I suspect it takes way longer than we think to understand, get to GPT-6, get to understanding memory and harnesses and putting all these things together. It probably won’t be OpenClaw, it’ll be some version of how an agent surfaces in a certain area of life. So yes, the products are still coming in the consumer space. Secondly, I think the “no” part of the answer is that AI is not a product, it’s a technology. It’s a piece of technology. If it just makes everything about your life a bit better, you won’t even notice you’re using AI, and actually, that’s the beauty of it. You can use Dia all day long and not realize how much AI it’s doing. Just like the Morning Brief thing, it’s really good. I look at it every morning, it’s one of the most highly engaged features we’ve ever built. And it just tells you, “Here are the three things I think you need to do today. Here’s what your day looks like. You want me to help you prep for this thing?” And you say, “Yeah, help me prep for that thing.” You don’t need to know it’s AI. We don’t tell you it’s AI. We’re purposefully not trying to hide it, it’s not secret, but you need to have a great experience. You don’t care there’s an LLM and tokens and gateways and massive amount of technology going on behind the scenes. You just want to know what the three things you have to do this morning are. And if you build trust in it, that’s great. The hard part when you mention the data centers and stuff is that people want those experiences and they’re going to be very joyful and very contented and more calm and happy about the experiences they’re getting. They’re just not going to be using AI. I tell this to people here all the time, every time you take a photo on the iPhone, there is an insane amount of AI processing going on to make that photo look good. Insane amount. And photos look amazing on my phone compared to what they did 10 or 20 years ago. Do I need to know that? Nope. I want to point at my kid, snap a photo, and I want it to look good. That’s great technology, when it disappears. That’s what it should be in the enterprise or for consumers. There’s no reason people in the enterprise should put up with shitty technology. That’s consistent, but it’s not going to be necessarily knowable that I’m using AI. It’s just going to be an amazing product experience like Dia that says, “Here are the three things you got to do today.” And it gets it right. Even beginning to describe the conversations I’ve had with the Apple camera team about the use of AI on those photos is a full other hour of this show that we do not have time for. Mike, this was a great conversation. What’s next for Atlassian? What should people be looking for? We have a massive set of announcements in, I’m scared to say, a bit over a week, 10 days. We have a biannual conference, our Team 26 EU in Amsterdam in 10 days’ time, announcements in every collection. You will no doubt see lots of AI features that are magical and disappear. You don’t need to know that you’re using them. Dia launches, Teamwork Collection launches. A lot of fantastic Loom features. We haven’t talked about Loom, massive in the AI era. We have so much on Loom’s agentic use, some of which has been online. Kudos to the partnership with the Cursor team, we built some amazing Cursor and Loom features. Really exciting stuff coming up in 10 days’ time. We will have to have you back very soon to talk about what happens with enterprise software in just one year’s time, because I’m confident it’s going to change. Thank you so much for being on Decoder . Thank you for having me. Thanks, man. Questions or comments? Hit us up at [email protected]. We really do read every email!
Today, I’m talking with Matthew Prince, who is CEO of Cloudflare. This episode is part of a two-part series on the future of business. Matthew last joined us on the show about two and a half years ago, at what we thought then was a wild pivot point for the internet — and now it turns out things are even wilder because of AI, and Matthew and Cloudflare are right at the center of it. Cloudflare found in June that bots made up more than half of internet traffic. That number just keeps going up as more and more AI companies scrape more and more of the web, and now as more and more AI agents try and do things for people on the web. Cloudflare sits between websites and all those agents, and allows website owners some level of control: Owners can block all those AI tools, allow them, or, as you’ll hear, only allow those that might pay money for access. Verge subscribers, don’t forget you get exclusive access to ad-free Decoder wherever you get your podcasts. Head here . Not a subscriber? You can sign up here . So Matthew and I talked about how to control all these bots, what kind of mess they’re making of the web, and what kinds of information might be valuable in the future as some of the payment schemes come into focus. You’ll hear me ask pretty directly if some of the outcomes he’s describing are actually good — Matthew is a thoughtful guy, and his answer is something I’m still thinking about well after we had this conversation. AI is also upending everything inside of Cloudflare. Earlier this year, the company laid off more than a thousand people — 20 percent of the company — and then Matthew wrote an op-ed about that decision that ran in the Wall Street Journal under the headline, “How I choose which Cloudflare employees to replace with AI.” That’s pure Decoder bait in every single dimension: AI, big controversial decisions, and org charts, all in one. And one last thing, before we get going: You can subscribe to Decoder on YouTube , where we put out new episodes every Monday and Thursday. Okay: Matthew Prince, CEO of Cloudflare. Here we go. This interview has been lightly edited for length and clarity. Matthew Prince, you’re the cofounder and CEO of Cloudflare. Welcome back to Decoder . Thanks for having me. I’m really excited to talk to you again. It’s been about two years since you were on the show. I was just looking back over that episode and I said something to you like, “it’s a momentous time for the internet.” And I was not even close. It’s now an even more momentous time for the internet. You are at the forefront of rebooting how companies work with AI. That is the most Decoder bait of all time. I want to talk about bots and the internet and publishers and Google and all the things that you are in the middle of. But let’s start at the very beginning. Cloudflare is a complicated, important company. The last time you were on the show, you said, “Cloudflare is a service that makes the internet faster and protects it from bad guys.” Is that still how you would describe what Cloudflare does? That’s usually what I say when I’m trying to answer a question at a cocktail party and I don’t want to talk to the person any more. What I say if I’m interested in talking to the person more is that Cloudflare is trying to rebuild the internet the way it should have been built from the beginning if we knew how important it was going to be. We run a giant network. We’re in over 350 cities worldwide, literally thousands of data centers. We have equipment running in all those places, and then we do a handful of things. If you’re trying to put an application or content online, we make sure that it’s safe. If you’re a person or an employee online, we make sure that wherever you go online is safe. We help give people the ability to write applications that can scale to an entire internet audience and run on the infrastructure that we have. That’s the fastest-growing part of our business. We’ve recently started to think about how the business model of the internet is changing dramatically and we’re trying to help shape what that future business model of the internet looks like and make it as healthy as possible. Can I connect the dots between the business model of the internet, how the internet should have worked, and Cloudflare trying to build the internet as it should have been? That is different than, “Cloudflare protects from bad guys.” I will note that’s what you said to me last time, which now makes me feel like you don’t want to talk to me at the cocktail party, but that’s fine. That’s a shift. It’s just a notable shift. The business model of the internet has dramatically changed because of AI, because of bot traffic, because of AI scrapers, the whole thing. Describe what you think the business model of the internet is right now and what it should be. For at least the last at least 30 years, the business model of the internet has been advertising. It’s not the entire business model of the internet, but it’s really driven all of the growth of the web and built what we all enjoy today, which really is a miracle. The company that is responsible for that more than any other is Google, which for the last 28 years has really defined that. We all think about it for search, but they really built out all of the ecosystem around advertising online. They bought DoubleClick, they built things like Google Analytics that let you actually see who was coming to your property. It really is the hero of what I call the first generation of the internet and the first generation of the web. What’s changing though, very quickly, is who is actually using the web, or what is using the web. Last November I was at Web Summit, the big European tech event, and I was asked, “When do you think that non-human traffic — agents and all of those things — are going to pass human traffic online?” We have a lot of data because we see a huge percentage of the internet, so we pulled all of that data and looked at it and said that in the second half of 2027, automated traffic would become larger than human traffic. It felt like a big deal that we could see that for the first time in the internet’s history. I was asked again at South by Southwest in March of this year, 2026, and we pulled the data again and it had moved up, where it was going to then be the first half of 2027. We’re like, “Wow, this is growing. This AI thing is a big deal and people are using it like crazy and it’s driving a huge amount of traffic.” I was stunned when just a few months later in May, the team came to me and said, “You won’t believe it, but automated traffic has now passed human traffic online.” And that’s just expanding like crazy. If you extrapolate that out, with the giant caveat that I’ve gone wrong so far in every prediction that I’ve made on this, five years from now, we think that automated traffic will be 1,000 times human traffic online. Not because human traffic is going to decline, we think it’ll stay around the same, but because we’re just seeing such an explosion in all of the rest of the traffic. The challenge of that is, if you have 1,000 times more traffic, someone’s got to pay for the infrastructure to power that. That’s going to require bandwidth, that’s going to require servers, that’s going to require a lot of things in order to make that happen. And, the traditional model of how to pay for that, which was advertising, doesn’t work for bots. Bots don’t click on ads. They don’t respond to pretty swirls of paint and what we traditionally think of as a brand. We’ve got to come up with something else that’s going to power that incredible insatiable demand that’s going to be put on the internet going forward. I’m spending a lot of my time trying to think about what that’s going to look like. When you describe that as a business model, that implies that there’s going to be customers and revenue and profits. And I don’t know if any of that’s true. There’s certainly a lot of revenue. Well, maybe for the CloudFlares and the Googles of the world, but for the — Or for the Anthropics of the world! Anthropic added $10 billion of revenue in a single month. There is enormous demand for these AI services. And if you’re an AI company, what you really need are three things. You need great talent and researchers. And right now, that’s scarce. Right now we have a scarcity of people who really understand how to build these systems, but that’s going to change. Every single university in the world is standing back up their AI department. We’re training people like crazy. Labor markets are pretty efficient. We’re going to have more and more people coming into the space. The second thing that you need is chips. You need the silicon to run these things. Today, Nvidia is the best in the world at that, but you’ve got a whole bunch of others, whether that’s AMD or Qualcomm or all the hyperscalers, that are building their own chips in order to power these things. While we have a massive shortage in the availability of silicon and the ability to actually power it up and turn it on, that’s going to change as well. The third thing that you’ve always needed is the data to feed into all of these different models. And that’s been the one thing that I think is going to go the other direction. Historically we’ve just made the internet completely open and given bots access to all of the stuff that’s being created. I think that’s starting to change. We’re seeing a shift where more and more of the people who are creating content, really creating information, are saying, “Maybe we’ll give that away for free to humans, but if a bot is coming for it, then bots have to pay for it because they don’t have that traditional give-to-get. We can’t put an ad in front of them and have that be something that allows me to help pay for the content creation that’s there.” What we’re seeing actually goes back to some of the original protocols of the internet. The 403 protocol, which was written in the very first version of Netscape, was actually payment required, and you actually have to pay for the content that you’re receiving. Oh, excuse me, I said 403, it’s actually 402. We are working with other leading companies like Coinbase and Stripe in order to say, “How do we make sure that we can allow people who are creating content, people who are doing things online, anyone who’s putting a website up, to say that in exchange for that thousands and thousands and thousands of times more traffic that’s going to come to you, you’re not going to maybe get the advertising revenue that comes from it, but maybe you can get a fraction of a penny every time somebody actually accesses that information.” Where does that fraction of a penny come from? It comes from the fee that people are going to be paying for their AI agents and other systems that are out there, similar to how a Spotify or an Apple Music works today. I like that we’re less than 10 minutes in and we’re already at the 1996 dream of micropayments on the web. We’re going to stick with that for one more turn here. When I said business model, what I was really pushing at was the idea of incentives. So maybe there is some business model for information. You’re going to publish some new information and bots will come and find it and yep, the 402 protocol will come back to life and we’re going to do, I don’t know, crypto micropayments using Stripe powered by Cloudflare. That’s a version of the future that many people have talked about for a long time. The incentives to stand all that up on the web is the thing that I’m worried about, when I say there’s a business model and customers and revenue. If I stand up a website, and I think most of my customers are going to be bots and not people, I might not do that. I might just start a TikTok channel instead, and then monetize my TikTok audience in whatever way I might want to monetize a TikTok audience. That seems like a really big inflection point right now at this second. You mentioned Google, and I’ve asked you about this many times. The incentives to put new information on the web just seem to be in permanent decline. It just doesn’t seem like a good idea any more. On the flip side, the web as an application program, it’s like the absolute apex. Every new app that comes out is a web app, or if it’s a desktop app, it’s just Electron. There’s something happening with the web as an application platform that is incredible, and something happening to the web as an information platform that is devastating. And you’re trying to connect those dots, right? You’re trying to say we can change the incentives over here. What’s a version of the future — maybe it’s microtransactions, maybe it’s not — where the incentives to make the web an information platform are as good as YouTube? You’ve illustrated a tough to reconcile dichotomy that’s happening right now. If you look at things like the publicly accessible data on Wikipedia contributions, it’s down significantly, because people are like, “What’s the incentive to put information on Wikipedia?” Something that, again, people are doing for a lot of different reasons. Back in the day, if you were contributing things to Wikipedia, you knew people were at least reading Wikipedia. Now, the interface through which people consume that information isn’t Wikipedia itself. It’s just reading the answer through whatever answer engine you’re using, whether that’s OpenAI, ChatGPT, or Anthropic Claude, or Grok on X or whatever it is. And that means that the people who are the editors of Wikipedia are saying, “Maybe it doesn’t make as much sense for me to do that.” The flip side of that is, the web as a whole actually had been declining since about 2012. The web grew like crazy in the late ’90s and through the 2000s. Starting around 2012, you really saw it plateauing and actually decreasing, up until about 2025, where something flipped. Starting in about October of 2025, the various vibe-coding platforms made it easy for anyone to create a website, and so more and more people were creating those applications. And as you said, even your desktop application, your mobile application, increasingly is just a wrapper around what is fundamentally a web application. The growth of the web itself — with really high-quality stuff and a lot more people contributing to it — is faster today than it has been at any time since the early 2000s. That’s the tension between those various things. What’s unique about Cloudflare is that because we sit in front of more than 20 percent of the web, we have the ability to overcome some of the incentives problems. We can very quickly turn something on and say, “Okay, you have to pay a fraction of a penny in order to get access to this information.” And that can kickstart the beginning of what we need. If you don’t have something like that, we face a massive tragedy of the commons problem. Let’s say today I ask my favorite AI agent where I should go to lunch.If I were doing research, I might go look at a couple of different menus as an individual. My AI agent today goes and scans every single menu in the local area in order to figure out what’s going on. Only one of those places is going to actually get my lunch dollars. If that just expands infinitely, if there’s no cost to doing that, at some point my agent’s just going to look at literally every resource that’s available everywhere on the internet, and then come back and say, “You know, you should go to Wendy’s.” That’s an enormous waste. We’ve got to have something saying that there’s a cost every time you load a webpage. There’s bandwidth, there are servers, there are things that are behind that, and someone has to pay for that. If it’s not going to be advertising and it’s not going to be direct commerce, then there has to be something that actually puts some constraints there. We are in a good place to say, yeah, it can be a tiny amount of money, a thousandth of a penny or something like that. But that’s enough that we can actually start to say, okay, that will help pay for the infrastructure. If you have incredibly valuable content, if you’re a news publisher or you’re an academic, then maybe there’s a premium on top of that that you charge and say, “Hey, I’m not going to give you this content unless you pay even more for it, but I believe that it’s valuable and we can create a market for it.” In order for that to happen, you’ve got to have a player like us that’s in the market that can kickstart that. But once you kickstart it, we’ve seen from the side of the buyers, the big AI companies, that they’re all willing to do this. We’ve seen from the side of the sellers, the big content creators, that they’re all really excited about this. What’s really been lacking is the technology that links those things together, and that’s what we’re spending a lot of time building. It feels like the theoretical underpinning of this conversation is the very notion of scarcity itself. I’m an old copyright lawyer, and copyright law for years and years and years had a built-in mechanism to be important, which was that one copy of a CD was one copy of a CD. And if you wanted another one, it was pretty hard to make another copy of a CD. Even when it got easy, you still needed another physical CD-R, and that imposed some cost on how many copies you could make, and then it was hard to distribute them. All of that went away with the internet. We moved everything to digital files and the burden of making another copy fell to zero. And I think a bunch of consumers expected everything would be free. There’s that famous quote, “Information wants to be free.” There’s a second half of that quote everyone forgets, which is, “Information also wants to be expensive because it’s hard to generate.” And the internet just turned that upside down. The gating mechanism of physical media, which provided some scarcity, and thus some economic value that you could measure, went away and we decided information should be zero and maybe supported by advertising. There was actually a tension that became valuable, because that was scarce in its way. You’re talking about imposing scarcity with your technology. We always talk about markets needing supply and demand. That’s not exactly right. You need demand for sure, and you want infinite demand ideally, or as much demand as you can get. Infinite would probably be bad, because then it would be hard to discover price, but you want demand. And then, you actually want constrained supply. There’s no market for air where either of us are sitting right now, because there’s plenty of air. But if we go scuba diving, then all of a sudden there’s a market for air, because air is constrained underwater, and so you have to buy it in order to be able to do it. Music is the example that I look to when I think about what this could look like in the future. I flew up to Stockholm to meet with Daniel Ek, who started Spotify, and it was just a fascinating conversation. If you think about the history of music, once upon a time, the majority of music sales were from CDs or albums or whatever it was. And then, along came the internet and along with it, Napster and Grokster and Kazaa, and all of the things that essentially commodified music and made it available for free for everyone. Even if you’re an incredibly law-abiding human, the majority of people were just downloading music because they wanted access to music. If you go back 23 years, the music industry in total was valued at about $8 billion, which is a lot of money, but it’s not a lot of money for the entire music industry. That’s The Beatles and The Rolling Stones and everything else. But people were like, “We can’t make any money off of this.” And then, Steve Jobs steps on stage, and announces iTunes and that it’s going to be 99 cents a song, but they include cover art and they’re going to make sure it’s high-quality and all these things. But mostly it appealed to this emotion of, “you should be paying for music.” Now, that’s not the business model that won, but it was a flag in the ground that said that this information is actually worth paying for, and it’s really valuable. And it was the iTunes that then eventually begat the Spotifys of the world. And the incredible thing is, just last year, Spotify sent something like $12 billion back into the music creator ecosystem. We can debate whether the right people are getting it and whether it’s fairly allocated. This is my favorite thing to argue about because I think whatever happens to the music industry happens to everybody else five years later. So I spend a lot of time thinking about it. And the turn there, which I think is fascinating and is either good or bad, is that the amount paid or the music files themselves became very small. You can get a million streams on Spotify and you’re not making any money, but the amount generated by touring, commercial sponsorships, sync licensing, and advertising all skyrocketed. That’s just simply not true. Why is private equity buying all of the music catalogs for hundreds of millions of dollars? The answer is because actually making money off the streaming of the music is extremely lucrative. Spotify alone is sending $12 billion back to the music industry, back to the actual rights holders behind these various things. There’s way more money coming from Spotify into this than there is from touring or any of the other things that are there. Those are other ways to make money, but it’s the streaming that is really driving all of the real growth. I just want to draw a distinction here. You’re not here to argue with me about music, which is my favorite thing to do, so I apologize for just doing it. But private equity is going to make that money, not the musicians. They’re paying some of the catalog holders for some of the things, but that money flooding in — If The Beatles or The Rolling Stones or whoever’s the latest ones to sell their music catalogs had held onto the music catalogs, then the musicians would’ve made that. They are making the determination. And when they sell the rights, they get the check. They do get the check, but the comparison I’m making here — If you win the lottery, is it better for you to take the annuity of payments over the course of the next 50 years or to take the lump sum up front? That’s the trade-off that they’re making, but there’s more money going into music creation at this point in time than there ever has been in human history. Technology is not inherently a destroyer of value of information. In this case, it has been a massive enabler of value and allowed people to find audiences, and yes, sell more tickets to their concerts as well. That’s actually the model that we need to think about. How do we take what Spotify has done, which is to say we pool together the resources of a bunch of people that are paying for access to the entire catalog of music, and then give that back to musicians based on some sort of algorithm that hopefully rewards real value creation. What if you had the same thing where with a portion of what is being paid for for the various AI companies, you had to pay for the researchers that built the AI systems and you had to pay for the chips, but also to pay for the content, which is actually the real knowledge that is training these things? That’s exactly the type of model that you need to actually unlock what could be a real golden age of information creation. The comparison I’m making is not whether a bunch of aging rock stars are going to sell their catalogs and cash out because they’re old. That’s fine. They can do that. It’s more that the shift you’re describing, into gaining access, into wide open access on Spotify to every song ever made, and then we’re going to move some pennies around and some people get rich, and we’ll create some winners and losers, fundamentally change the entire business model of music. The conversation with Dan was fascinating, but one of the things he said was he said, “Okay, listen, if you go onto Spotify and you search for Taylor Swift, ‘Shake It Off,’” they return a result and they’re pretty sure that they have given you what you are looking for.” On the other hand, if you go to Spotify and you search for, “I want a song to a disco beat about how much fun it is to dance with my cat,” there aren’t a lot of songs like that out there. They know that whatever they return is a pretty bad result. But the interesting thing is what happens next. They then take those searches for things that they don’t have good results for, and they publish that back to music creators. There’s something that’s pretty amazing about that. They’re saying, “Here’s an emotion that someone is searching for, which we don’t have a good answer for, which we’re then going to go and publish back to music creators. This is one of those moments where everyone’s going to be like, “I’m in the wrong profession.” There’s a guy in Denmark, who makes 40 million euros a year writing songs for unfulfilled Spotify queries, and he’s not alone. He’s the most successful, but there’s a whole bunch of people that are making literally millions of euros or millions of dollars a year doing this thing, writing songs for what people are searching for that isn’t out there. Now extrapolate that to the next level. For the first time in human history, we’ve built a mathematical model of human knowledge. That’s what the LLMs are. We know where they know things, but we also know where they’re missing things. I picture it like a giant block of Swiss cheese, and there’s a lot of cheese, but there’s a lot of holes in the cheese. The really interesting thing is, when you talk to the leaders at the big AI companies and you say, “What do you want to pay for?” They don’t want yet another story about what’s happening at 1500 Pennsylvania Avenue [ sic ], which is what the current media environment is feeding us like crazy. What they want is to fill in the holes in the cheese. They want new knowledge that no one ever knew about before. I have an example where this is actually working: My wife and I own a small local newspaper in our hometown of Park City, Utah. I think we will make more money off AI licensing deals this year than we do off digital advertising. Local media is exactly the sort of thing that is much more valuable in the media world that we’re going into, but was completely decimated in the media world that we are coming out of, where what mattered was volume and scale and dividing things. Whereas local media is all about, let’s tell you about the cool new restaurant that just opened down the street. If you’re an AI company, and you want to be able to be the best travel planner that’s out there, you want to have artificial general intelligence, you need to know what the hot new restaurant is in Park City, Utah. And if you don’t have access to the Park Record , our newspaper, you don’t know that. We’re not going to protect all media. In the same way that with the move to Spotify, there are a whole bunch of losers in the musician space, but there are going to be new winners, like that person in Denmark who’s creating things off unfulfilled Spotify queries. The winners in this new space might actually be the kind of things that people really want to come back to media: more local news, more unique things, more Reddits of the world, more of what the internet used to be when I was first on it in the ’90s. I think that’s actually what most internet users are craving. And I think if we get the incentives right, we actually have a way of incentivizing more of that unique original content as opposed to what we have today, which is a media ecosystem that is largely just rage-baiting people into clicking on things so that they can serve them an ad. One of the things you would need to build in order to make that work is a way to stop the AI crawlers, to stop the model companies from showing up. Just today, there’s yet another lawsuit. The music companies are going to sue Anthropic . The White House came out in support of OpenAI in its case against The New York Times today, saying that training should be fair use. I think I need to disclose that in some series of corporate mergers, The Verge ‘s parent company is now suing Google in some way. I have literally nothing to do with it. It’s just all that. It’s all the swirl, right? And that is a legal swirl. We’re going to use the law to say, “This is illegal and we’ll punish you if you do the bad thing.” That’s just up for grabs. Those are 50/50 and maybe existential problems. Maybe not. The law right now in the US is different than everywhere else, and so you’ve got this patchwork around it. But the best case on this is part of the Anthropic case, where the judge basically said that training is fair use against all the books. And by the way, you shouldn’t have stolen the books. That was bad. But if you hadn’t stolen the books, if you had actually bought the books, then it would be okay. That was a valuable enough result for Anthropic that they settled the rest of the case for $2 billion, which is a lot of money, even for Anthropic. So when you’re building technology, you’re keeping that in the back of your mind, right? The market is coming to some sort of understanding of what’s valuable and what’s not, and you need to stop them from showing up — Different regulations are going to be passed in different places around the world. Again, as an also recovering intellectual property attorney, I very quickly go to like, “Oh, let’s just use intellectual property.” But that’s such a kluge versus the much easier thing, which is, let’s just use technology. I remember sitting with a bunch of media execs and they’re like, “Oh, how are we going to stop these nerds in Palo Alto from scraping our stuff?” And I was like, “I go to war every day with North Korean and Iranian and Chinese hackers.” They’re really good at it and they hide. Whereas the nerds in Palo Alto have a Delaware-based C-corporation. It’s pretty easy to identify them versus the others. And it’s really difficult for them, at scale, to hide from us. What we’re seeing is actually much more of them being willing to say, “Listen, we will specifically identify when we are coming to a site that it is us, that you can rely on us, and we’ll tell you exactly what we’re doing, so that you can have the right to control how that information is being taken.” I think that was an interesting question, like, two years ago. But even the Googles of the world, who have been at times challenging through this — Google is sort of like a Marvel character, like the hero of yesterday becomes the villain of tomorrow. The challenge has been that they’re like, “We struck all these deals to get access to all the internet and now we can use it for whatever we want.” We were like, “Eh, in the past you were sending people traffic. Now you’re training on things and sending them no traffic. That’s a different give-to-get.” And what I’ve been really actually impressed by with the Google team is that they are much more willing to engage here. They understand the value of the ecosystem. There are people at Google who really do believe in making sure that there is a healthy, sustainable ecosystem going forward. Some of the things that they’ve already committed to around transparency of their crawler, and some of the things that I expect that they will do over the next little bit, I think that’s a good sign and it’s going to make it so that if even Google is doing it and is willing to say, “I’ll announce what I’m doing, and in many cases I’ll even be willing to pay for content,” that actually makes it even easier to convince all of the other AI companies to do the same. Do you think you need to turn the screws all the way and actually block the Google crawlers to get them to pay? As of September 15th, we’re going to set the defaults across all of our free customers for it to be that Google is going to be blocked for AI training, but if they won’t differentiate between AI training and search engine training, we’re just going to block them across the board. If you’re a site owner, you’re like, “I don’t want that to be for me,” we’ll make it easy for you to turn it off. It’s just about what we set the defaults to. But I’m really encouraged that the good forces at Google are realizing that this has to be a healthy ecosystem, and that they have to play by somewhat the same rules for this new AI market that everyone else is playing by, and that they can’t leverage the monopoly that they had in search yesterday to create a monopoly in AI tomorrow. I’ve talked to Sundar [Pichai, Google CEO] about this many times. I don’t think he loves the fact that I am the person who keeps calling it Google Zero , but so be it. Sundar is very thoughtful. He’s very kind. Every time I talk to him, I get the sense that what he wants to say is, “Well, you didn’t do anything about it. All you publishers are complaining. You got super addicted to my fire hose of traffic, and you have no leverage. You built no audience of your own. You didn’t do anything. Look at all of these other platforms that I have to compete with. TikTok showed up and YouTube had to compete with them. ChatGPT showed up and the search team had to compete with them, and you did nothing.” This feels like the something. We’re going to block the Google search traffic. We’re going to block the crawler. I think the CEO of People Inc. is talking about literally blocking Google. Steve [Huffman, CEO] at Reddit has really become more aggressive in this line. Google doesn’t work if it can’t search these things. And it’s actually more existential for them. The very nature of how Google’s search ranking works is it builds a tree. That’s what page rank was. It sort of says, “Okay, here’s a super reputable thing and then let’s see how it’s connected to everything else that’s online.” The problem is, with Cloudflare being 20-plus percent of the internet, if that just disappears, that’s in a giant hole in the middle of the tree. It doesn’t just break it for this, it breaks it for everything. That’s what we realized, that on behalf of and in conjunction with a lot of our customers, we could say, “Listen, this isn’t fair anymore. You are creating costs, you are taking content, you are getting value from that content, either in terms of the subscriptions that you’re selling to your AI tools, or to the ads that you’re running against that. And that’s just not a fair give-to-get anymore.” The deal has to change. I’m really proud of the fact that we’ve played a role in helping the publishing industry go from what was two years ago when I had dinner with Neil [Vogel] from People. He’s like, “Woe is me. What are we ever going to do?” I’m like, “I think we can fix this.” Now, last time I saw him, we gave each other high-fives and said, “I think we’re making progress.” And the deals that large publishers are doing are significantly better. Now the question is, how do we bring that to the rest of the internet? Because if we don’t, we have this massive tragedy of the commons problem, where it’s just going to be, take, take, take, impose cost, impose cost, impose costs, and there are no incentives for actually creating things, for contributing to the Wikipedias of the world, for putting up a new website, for writing about what the new local restaurant is, for creating local news and local media, for being an academic. If there’s not some way that you can do those things and still make enough to eat, then people aren’t going to do it. And that harms us all. Again, I’ve been very critical of Google over the years. I will say that their tune appears to be changing over the last few months. I believe that, at their core, they really do understand that this is an ecosystem, and that they play an important role in that ecosystem, and that they need to give back to that ecosystem and play by rules that allow the ecosystem to flourish. You said Google’s going to make some changes soon. What changes would those be? I think the thing that they’ve already committed to is just a lot more transparency on what their crawler is doing. So what we had pushed them to do is split their crawler apart and say, “We’re going to crawl for AI separate from crawling for the web.” They, for a lot of technical reasons, pushed back and said, “That actually is incredibly inefficient. We now have to crawl twice. It’s going to put twice as much load on everything that’s out there. What if instead, we just said when our crawler comes to a page, we’ll announce what it’s doing, and then give you the ability that if you don’t like it doing something to say, ‘no, no, that’s not allowed, but this is,’ and put that together.” They have signaled that, largely in response to what was a ruling out of the United Kingdom , that they were going to put those procedures in place to allow publishers to make that choice, and that they wouldn’t just do it in the United Kingdom, but they’d actually do it on a broader basis. I think we come out of this on the other side with a stronger web, and frankly, with even a better Google, which would be great. Again, I could argue with you about the music industry all day and all night. I think you know that, but it is true that changing access and copyright law and all that stuff around music industry changed the business. The incentives to put on shows and do residencies and all that changed because of the distribution changing. That I think we can generally grant. You’re describing an information market where the incentives will shift again, where it might be more economically lucrative to make information for the bots than for people. And that might shape the very nature of the information. You run a local paper, maybe your people are publishing restaurant reviews that are best ingested by an LLM and spit out. Or instead of reviewing a hotel, review every hotel room, right? For every word that appears in a story in The New York Times , that reporter has written down probably a hundred words somewhere else. That’s all that other metadata, which traditionally has been constrained by how many column inches you had in the physical paper, or how much attention a human would actually spend on that. Imagine if you could say to the LLMs, “Hey, listen, we’re gonna sell — “ and again, there are all kinds of things around protecting sources you have to get right, but if you get it right, there’s an enormous rich catalog of additional information. For every picture you see in a magazine or a newspaper, there are probably 50 or 60 that were taken of that same thing. And that’s all valuable to these AI systems that are out there. There’s just a bunch of content that’s literally being thrown on the floor today. That can be incredibly valuable. This is my universe. I feel confident about this one. I was in a packaging meeting today for one of our big stories, and we had a pretty fulsome debate about the lead image in a story we’re going to run in a few weeks. Then we disagreed, and eventually we picked one, or we picked a direction, and that was an editorial choice that was designed to elicit some reaction in humans. We could publish all the rest of the photos. They’re all really, really good. We decided one would be the winner and the rest wouldn’t. But if I publish all the photos and I give them to an LLM, it will change the thing that we made, because its distribution will necessarily change, and its intended audience will change. Maybe the biggest Decoder trope of all is that your distribution inevitably changes the thing you make, and at the end of the day, maybe my future is just making YouTube face slinging AG1 and that is the future of all podcasting. There’s just some force of distribution that changes the thing that you make. Is the outcome you’re describing good, where we’re just making an infinite flood of information GEO-optimized for some human to consume, digested by a chatbot in the middle? First of all, is the current model good? I’ve talked with lots of people at your parent corporation. I think you guys do a good job. There’s a lot of media that today just tries to create content as inexpensively as possible, and then A/B tests headlines in order to either stimulate a dopamine or cortisol response to that. If only the media was that scientific. But a lot of it is. I’ve sat with some of the folks from Huffington Post , some of the folks from BuzzFeed , where they were like, “Yeah, that’s the game and that’s the game that we play.” I don’t know that that’s a game you play everywhere, but you can track The New York Times , you can track The Wall Street Journal , you can track the FT and just watch how much more inflammatory and almost tabloid-ish the headlines have gotten over the last 20 years. Every one of those is an amazing media organization, but in order to win in this space, it’s been, “how do I provoke really a deep —” Well, what you’re describing is they’re playing to their distribution, which is largely social media. It’s algorithmic, social media algorithms, and that is the distribution, and it’s shaping the content. So I’m asking you about the new distribution that you’re describing — The sentence that I say that gets everyone at Google to yell at me is, “everything wrong with the world today is Google’s fault.” That is not fair to Google. Google has by and large been a great company, but Google begets Facebook, which begets TikTok, which has us in this attention economy hole, where we’re trying to get as much attention as possible. And the best way to do that is to actually stimulate, again, a deep mammalian response. That’s the media landscape that we live in today. And I think it’s deeply broken and I think it’s what’s divided the world. It’s what’s led to the rise of very destructive populism around the world, rather than real intellectual debate. So what could we move to? There’s a lot that can go wrong here, but if you talk to Sam [Altman] at OpenAI, or Dario [Amodei] at Anthropic, or you talk to the teams at Google, that are building the AI systems at DeepMind and others, what do they really want? They want new true knowledge. They want the thing that nobody knows about yet. They want the story about the interesting thing that no one else is covering. They want that thing which is actually advancing human knowledge forward. That’s what I want to read too. I don’t actually want to read yet another take of what happened in the Trump White House today. There’s plenty of that. And they don’t want it either. The best evidence of this is actually how some of these distribution deals have been done. The New York Times , amazing media organization, and Reddit, also an amazing media organization, have about the same amount of tokens. The New York Times has been publishing for a lot longer. Reddit is a lot higher volume. They do deals. Who gets more money for their tokens? The answer is Reddit, by at least seven, or by some measures, 14 — Just be clear for the audience, you’re using tokens as a measure of amount of new content. Of content, yeah. Content that’s in there. It’s not just new content, but it’s the legacy content as well. The question is, why? This is deeply unfair to The New York Times , but if you don’t have The New York Times , then you can just license The Wall Street Journal and ask AI to rewrite it as if it’s a New York liberal and you get The New York Times . Deeply unfair, except that all of these major media publications have made their business of telling the exact same story to their individual tribes. Whereas if you don’t have Reddit, there’s no substitute for it. Reddit is this unique thing that’s out there. I think a media of the future that looks more like truly unique storytelling around, again, local communities, unique stories that no one else has told. Real knowledge creation is what is valuable and is proven valuable by the market that exists today. But the other side of that market — just to be reductive here, just so I understand the model — the market, the buyer in this market is a bunch of model companies. It’s not people. But then ultimately, it’s the customers of those model companies, which is all of us. What’s been amazing is it doesn’t appear like any of these model companies, on their own, is going to run away. People are like, “Oh, well, what if OpenAI gets AGI?” I’m like, “Two days later, Anthropic will have AGI, and then Grok will have it a couple of days after that and Google will have a couple of days after that. And then AGI won’t be enough. It’ll be AGI Plus or AGI Plus Plus or whatever.” What is going to be interesting is, what’s going to turn these things from commodities into actually sticky products that people sell? The answer is going to be who has access to the most true knowledge that gets you the right answers. It seems like all of the incentives from the end consumers say, “I want to get as much back to the creators of real knowledge as I possibly can.” I think we can solve the financial piece. The piece that I’m actually much more worried about — and again, you should have Daniel Ek on your show, because he’s fascinating to talk about these things. But he’s like, “Listen, you’ve been thinking a lot about how to get content creators paid.” He said, “If you think about musicians, there are two reasons that people become musicians: to get rich and to get famous. And if you, at the end of the day, had to cause musicians to rank which one is more important, fame probably beats rich.” By the way, it’s a stupid strategy. Being rich and anonymous is the best. But that’s true. I think that a lot of people really do want recognition. This is part of why Wikipedia is seeing a drop in the number of contributors and information that’s contributed, because people don’t get recognized as much because you’re not going to the original sources. I think we’ll solve the financial problem, but the next problem is, how do we actually recognize the creators? What I’ve been pitching to the big AI labs is we should create, I don’t know if it’s the Nobel Prize or the Academy Awards, but some recognition where we use math to measure who contributed the most to, I don’t know, mammalian biological research in some specific field, and measure it in the last year, and then have a big ceremony and give them an award and celebrate them, and talk about how wonderful they were at advancing this. I think figuring out how we take the people who are actually out there creating the knowledge, and even though the medium through which that knowledge is going to get disseminated might be removed from the original research, it might be the AI bot or the chatbot that’s out there, we still have to say, “But you’re doing really important work, and we’re going to recognize you for that important work.” The last time you were on, we spent a lot of time talking about content moderation, and Cloudflare using its power on the internet to essentially make it harder for sites like Stormfront, the neo-Nazi website, to exist. And you were in torment about it. You had a lot of power, you were going to use it. You published some op-eds about it. We talked about it for a while, and eventually you came to decide you didn’t want to be in business with these folks, and the consequences are the consequences, but there should be some frameworks to regulate your power. Many things have changed, but we’ve kind of had a long conversation about Cloudflare using its power to block AI scrapers and create economic opportunity. It’s the same power in many ways, right? It’s a power of a layer of the internet that is pretty dominant to stop things from happening. Have you changed your mind? Is it just the economic rationale that’s made that more comfortable for you, or is there something else happening? It’s been interesting inside of Cloudflare. When we see these bad things that are using us, there’s a very human reaction. And it’s often my cofounder, Michelle [Zatlyn], who’s like, “Listen, I realize all the hard, thorny moral questions, but this is just per se bad. We should do something about it.” We have largely come out to say, “It’s tough.” We don’t want to be the ones that are deciding what’s good and bad. And there actually aren’t that many things that are just per se bad that are out there. But wouldn’t it be great if we could change the incentives, so that there’s less incentive to create these various things? And what I’ve found interesting around the question of what the business model of the future of the internet is going to be is… Trying to deal with a bad site here or there is basically just cutting down dandelions, more are going to spot up. You’re not solving the root problem. I think if we can get the incentives around the next business model of the internet to be less about “how do we create rage and emotion” — being a neo-Nazi is one of the most effective ways to just really piss people off, and they tend to click on your things, and then buy nutritional supplements or whatever. All these things have very weird business models that are behind them. But if we can get back to something where we’re actually rewarding not rage creation, but information creation and knowledge creation, maybe that’s actually getting at the fundamental roots that are there. And maybe I’m too much of an optimist, but I think the way that you stop a lot of horrible things we see online, is by solving the business model that is really broken, which has led to the online communities that we have today, and by getting away from what has been a very much an attention-based economy to one that is much more of an information knowledge-based economy. And so, as we are thinking about what the future of the internet looks like, that’s what we’re playing for. So it’s the same power, right? Cloudflare has a lot of power on the internet. It’s a platform layer, it has a lot of influence. You’re able to stop things. In the case of the neo-Nazi websites, you declined to use your power, right? No, no. We kicked them off our systems. Right. You said you can’t have Cloudflare protect you from DDoS attacks, and then maybe they’re just going to DDoS into oblivion or whatever. Yeah, they’re all still around, so at some level it shows that it’s limited. But in this case, you’re going to use your power. You’re going to say we will use Cloudflare to stop things from happening and that will create a market. I guess the question is power. So it takes five minutes to sign up for Cloudflare. We have a free version of service. It takes 30 seconds to leave. So if we’re ever doing something that’s not in our customers’ interest, we’ll lose our customers. And so, we are very much at the service of our customers. I remember the first time that a media company called me and said, “We have this new threat, you have to stop it.” And I was like, “What is the threat?” They’re like, “It’s the AI companies.” I rolled my eyes. I was like, “That’s the dumbest thing I’ve ever heard. Why are all media companies such Luddites?” But then we pulled the data, and we saw that really, again, that there was an existential threat to how the internet was working. I certainly became convinced that this was something that was worth us spending our time and our resources fighting. We have a really privileged position, because we have provided so much value to so many companies that are out there that they trust us. But if we ever screw that up, they’ll leave us in a second. Is that power, or is that just being a good steward to the internet? It’s 2026. The valence of speech on the internet has changed. If you had to make the Kiwi Farms, Daily Stormer, decisions again today, would you make them the same way? Each of those things is a moment in time, and I think there are puts and takes under all those things. At some level, at the time, when we made the decision around the Daily Stormer, no one knew what Cloudflare was, and we were seeing a bunch of regulation that was happening online that really could have been a real threat to the underlying way the internet worked, regulating key protocols like DNS and TLS. And we were nervous about that. And so, the question was, how did you make that point? And so, one of the real rationales of kicking Daily Stormer off, and then writing about it, talking about it, going on the news about it, writing another Wall Street Journal editorial about it, was because it helped us then frame the right policy decision that was out there. We talked about Rawls last time. I’ll talk about Kant this time. We very much offended Kant in that. We were using this site as a means to an end that was not into itself. But if you’re going to use something as a means to an end, to make a point, neo-Nazis are pretty fun to use for that point. I think we were able to make the point. We were able to change the policy discussion. We were able to talk about what the challenges were, and we could do it. Today, if the exact same set of facts came up, it just wouldn’t be the exact same set of facts, because today obviously Cloudflare is much more known. We’ve had those policy conversations. It feels like every five years another one of these things pops up, so we’re probably due for one sometime soon. But the situation behind each of them is going to be very, very different. I usually do the Decoder questions first, but we just got into it. So I want to ask the first one first really quickly, and then I want to spend a lot of time on structure. The last time you were on the show, I asked you how you made decisions, and we had a long conversation about values and mission and how you kind of came to figure out what your values and mission were and you came back to that. It’s been two years. Are you still there? Is that how you make all your decisions? Totally. Yeah. So, the big decision you recently made: In May, you laid off 1,100 people, which you said was about 20 percent of the company. How many employees is Cloudflare today? Somewhere between 4,500 and 5,000. So you’re growing compared to two years ago. Yeah. You very openly attributed the layoffs to AI usage inside the company. And I know that’s true because you literally published an op-ed in the Wall Street Journal . The title was, “How I Choose Which Employees to Replace with AI.” I didn’t get to choose the title, because again, the way media works today is they create the titles, but yes, but I wrote the rest of it. The op-ed supports said title. I don’t know that that’s how you tell that something is true, that somebody wrote a Wall Street Journal editorial, but that’s the editorial editor. That was the headline. And the opening is, two weeks ago you laid off more than 20 percent of your workforce. You wrote that, and then you said you didn’t do it because Cloudflare is struggling. You did it because in the future, Cloudflare needs to change. And then I’m just going to run through your rubric and I just want to ask you about that rubric, specifically. You said you broke people into builders, sellers, and measurers, and you’re basically going to cut all the people who did measurement, all the audit functions. Not all, but a lot. The majority of the people that we laid off came from that category. This all comes back to some old-school basic business research, which is that there are three functions with any firm. There are people that build things, so the engineers, the product managers, the folks that actually create new products. Then people that sell things, the people that are out there actually doing the deals and selling those things. And then the third category is what a lot of every organization is, which is the people that actually measure things. Each of those is going to be impacted by AI in very different ways. I’ve had a lot of software CEOs on the show recently, and we’ve talked about how AI is scrambling every software company. And I usually ask about product managers, designers, and engineers, the builders. All of those roles are totally scrambled. They’re all kind of doing one another’s jobs. That’s true. I do think that there’s an increasingly jack-of-all-trades aspect to this, where because you can have these tools, someone who’s a product manager can do a lot of what engineers do, an engineer can do a lot of what product managers do. The people that are winning in that space are the people that are ambidextrous. They can do multiple different things out there. But where I get lost is with some of the AI maximalists who are like, “We’re all going to lose our jobs.” If I can hire a builder, and they are now 10 times as productive — which they are, it is wild to watch how much more productive the builders on our team are today — I’m going to hire as many of them as I can, because I’ve got lots of stuff to do. That hasn’t decreased the incentive for hiring engineers and product managers and everyone else. It’s actually increased the incentive, because the return that I get for the salary dollars that I spend on one of these people is now essentially 10 times as much. Of course I’m going to hire as many as I possibly can, which is exactly what we’re doing. So, just walk me through this decision. You woke up one day and said, “I’ve got a…” It’s Peter Drucker, I think is who you quoted. Yeah, the builder / seller / measurer framework comes out of Drucker. For the Decoder heads out there, this is old-school management philosophy. How’d you make this decision? You were like, “I’ve got to do this. I’m going to sort these people into these categories, and we’re going to start making cuts.” Did you see any evidence in the data? Walk me through it. One thing that we saw was that the world was dividing into two camps. One camp tended to be two different demographics within the organization. It was either people who were very, very senior, or people who were very, very junior. And those folks were adopting AI like crazy. The junior folks, because they were just native to it. The senior folks, because they had the confidence to bet their career, that these paradigm shifts would be a way for them to learn new things and take on new challenges, and they were confident enough in their jobs that they could do that. The other camp was the folks who were earlier in their careers. They might not have been the most senior folks, but they weren’t just the brand-new folks who had come in. So they’d come through and been trained and been taught that the way you succeed at a business was by playing by a certain set of rules. And then, they watched around them as their colleagues were all of a sudden using these new tools that were out there to be able to deliver these things. There’s a wildly imperfect analogy that I use. If you imagine our job was to screw screws into wood, we hired the best people at using manual screwdrivers. And then all of a sudden we invented an electric screwdriver, or it came along, we were able to buy it. And for not every job, but for most jobs, using the electric screwdriver is just better and you can get a lot more done with it. The first thing that we started to do back in the middle of 2025 was say, “Hey guys, this is what’s going on. I understand all the incentives, if you’re in that second camp, are to fight against the electric screwdriver, but let’s give you the resources to train you. Let’s give you the confidence that you’re going to have a job and you’re going to do these things, but let’s make sure that everybody across every role is learning how they can do that.” That’s just a really important step that you have to do in the first part. What we then learned, though, is that as we got everyone, not just the engineers, but people on finance and legal and everything else, to start using these tools, we found that one place where AI just shined was in measuring things. One of the places where we don’t talk about the advantage of AI enough is that AI has bias, but the biases are uncorrelated to the rest of the organization. Humans have biases like crazy, but a group working together has biases that are massively correlated together. Even if you’re working on internal audit and you’re supposed to be the bad guy who’s looking over everyone’s shoulder, you still go to lunch in the same cafeteria, you talk to the same people, you participate in the same all-hands, you end up developing the same biases, whereas AI doesn’t. It has a very different set of things. We found that we could use these tools to do things significantly more efficiently. There’s a woman on our team named Heather. Heather was on our investor relations team. Every time we would close the books before we’d have earnings, because we’re a public company, her team of about 20 people would spend about two weeks working like crazy to generate all of the information. They basically took the measurements and then generated documents that we would then distribute during earnings to all of our investors. And Heather was like, “I think we can use tools to do this better.” We took what used to take two weeks and we reduced it down to three minutes and our investors are like, “Wow, these documents are much better. There are fewer errors. There are less mistakes, and our tools are auditing all of those different things.” So, for those 20 people, we looked for other places for them, but a lot of them, what they liked doing was that sort of work. And they would be great at doing that sort of work as the Heather at some startup or somewhere else, but we just didn’t need some of the functions that were there. And it tended to be all of those functions that were largely measurement. There were things like internal audit, which we were able to just get actually much more efficient at doing. There are also things like middle management. Traditionally the Harvard Business School number is that on average every manager should have six direct reports. We found with tooling, we could actually be much more efficient with managers. When we enabled managers to have more tools to better participate in surface issues early and see how their team was doing, we could get up to 12, on average, direct reports and everyone was actually happier. That matters because as you increase the number of direct reports, what you actually do is decrease the amount of hierarchy in any organization. That’s the way of measuring how flat versus how hierarchical an organization is. Again, we could use these tools to say, “Hey, let’s flatten the organization,” which has made the organization much faster and more nimble. But in the process, there are a whole bunch of middle managers that just weren’t the right folks. I think we sat there and we were like, “Gosh, we know we’ve got to get rid of these jobs, so the question is, do we do it now or do we do it later?” The problem with doing it now is we feel very exposed, we feel very alone. The number of death threats that I got, not even from our employees, but from just random people who are anti-AI, was really pretty scary. But at the same time, we’re like, “Is it kinder to say we’re going to make these changes now, and then do the work to not only give great severance and everything else, but actually go place these people? Because they’re great people, and we’ve been very successful at doing that across the board. Or is it better to cover our own ass and wait until everybody else in the industry is coming to the same conclusion?” By the way, that’s coming. Uber just did a layoff , that same rationale. You’re going to see across the industry, not just in tech, but everyone, start to get to this realization. I think we were early, but was it going to be easier to get a job back in April when we did this, or next April when everyone is doing it? Once we realized that, we were like: If we’re real leaders, we should be the ones taking the arrows. We should be protecting our team. We should be working to make sure that even the people that we said, “listen, we don’t have a role for you any more,” we can help them find great roles at other places because they’re great people. That is what real leadership is. When I talk to a lot of my peers, they’re sitting around saying, “Yeah, I know we need to do this, but gosh, I saw your stock went way down, and you got all these really scary death threats and other things, and I don’t want to have to go through that.” I think that’s chickenshit. If you’re a real leader, get out there and actually do the work. There are a lot of things about the economy that are going to change. It doesn’t mean jobs are over. We’re going to hire as many builders and as many sellers as we possibly can. And there are still some measurers on our team. Some people have to build the tools and do those things, but it’s going to look very different going forward. Once you realize that as an organization, the sooner that you can do that is the kindest way that you can be to your own team. Another Decoder trope is that structure is a rough proxy for culture. I’m always saying, if you tell me the structure of your company, I can tell you about 80 percent of your problems. The 20 percent is usually where the actual magic is. How has your structure changed and how has that changed your culture? The first thing is we’ve flattened the organization a lot. There’s a lot of middle management that went away, and not just because we laid people off. There’s a whole bunch of managers that are raising their hand and saying, “I don’t want to be a manager any more. I want to go back to being an individual contributor.” It’s wild to see all these really senior managers that are leaving big jobs to go to Anthropic to be an individual contributor. And that’s because there’s so much more leverage that you can have. What is a manager versus what’s an individual contributor is going to get really fuzzy. And that’s one of the things that we’re spending a lot of time thinking about. Does it even make sense to have those as two separate tracks? Maybe that all merges together into one thing, because even individual contributors now are managing a fleet of agents that are working on their behalf. The biggest thing is that we’ve flattened the organization quite a bit. And then the other thing is that we’ve centralized through something we call Cloudflare OS, which we’ve now open-sourced because enough of our customers were like, “That’s really cool. We want to use that too.” And so, we’ve got a whole bunch of customers that are taking it and turning it into XYZ company OS. It’s a set of tools that, essentially takes the knowledge that’s inherent to Cloudflare, the things that have to be true, the things that inform our decisions, and then allows anyone on our team to have access to a wide range of the different tools that are out there, but it’s continuously pulling and updating the information about what has to be true. And having it so that when you want to figure out what information is pulling out of Salesforce or Workday or any of the systems of record that we have at Cloudflare, being able to pull that into one consistent interface has allowed our teams to be significantly more productive and much better at cross-collaboration. There’s something here which maybe is just the nature of a software company, and that’s the way it is. It’s a worldview that I understand, because of how a software company or a tech company works, where everything that is happening in your company happens in a digital system and then that thing can be measured. You’ve got a line in your piece, “As CEO, I’ve never had better tools to measure exactly how the business is performing, including identifying our rising stars.” This implies that your rising stars are producing data that can be measured by an AI, and then the AI can tell you these people are rising stars. There’s something about that that is fascinating to me. Not every company works that way where everyone’s job is to tell Slack everything that’s going on or tell Cloudflare OS everything that’s going on. They’re out in the world doing stuff. How do you connect those dots? Because that seems to be the challenge of how Silicon Valley might see the world and how every other business might see itself. I think we’ve all at various times worked in jobs where we’re like, “Gosh, I’m just not being recognized for the hard work that I’m doing.” Or other times where you’re like, “I’m kind of over-recognized and maybe I don’t —” I’ve never felt over-recognized in my life. If I could get some more, that’d be great. You’re doing a great job.[ Laughs ] [ Laughs ] Now, can you get your chatbot to tell me I’m doing a good job? Because that’d be really good. It is absolutely the case that there have been great people who’ve worked at Cloudflare that we just missed in the past. And for a million different reasons. Maybe they’re part of the organization that we just didn’t recognize enough, maybe they didn’t have a great manager. One of the things that’s great about these tools is you can find different ways to say, “What do we as an organization value?” And then look more broadly across the people who are really performing incredibly well. But that has to be legible to the system, right? So it’s like the AI is going to watch every code commit and say that person’s doing a lot or that person is the nicest to their agents in Slack. You have to watch them in some way to get that data back out. Yeah. But there’s all kinds of signal inside of every organization. I’m just stuck on this and I’m running out time, so I’m sorry to interrupt. But give me an example of signal that helped you through AI identify rising stars. We knew people who were high performers. We trained models based on what their high performers were. We ingested a ton of things across that. Obviously code commits and all kinds of things, but those are gameable in various ways. And so, you really want to look at “this person did this thing” and then trace it all the way through the organization and see whether it resulted in either higher revenue or lower cost or better collaboration across the team. Because we are a very digital company, we’re able to pull that. But I think every company has signal in various ways. If you’re a supermarket, you’ve got cameras that are seeing things, you’re going to see the person who sees the spill on the floor and cleans it up. A manager may never see that, but the camera did. And if the AI can say, “Hey, that person just cleaned up a spill without having to be asked,” of course we should be rewarding that stuff. Now there’s plenty of Black Mirror horrible ways that this stuff can go wrong. We’re very privacy-respecting. It’s not a gotcha thing. But it was remarkable how much better I found it at being able to say, “Wow, there are some people across the team who might be very junior, but are just way outperforming.” And then what I can do as a leader is go to those people and say, “Hey, you’re doing a great job. Keep doing it. By the way, here’s my cell phone number, call me if you ever need anything.” And we’re just watching a bunch of those folks turn into the next leaders at the company. It goes back to one of the real values that these systems have, which is that AI has bias, but it’s uncorrelated to all the rest of the bias in the organization. That provides an independent outside lens that helps you then better run your organization in order to make smarter decisions. Whether that’s around internal audit functions or identifying great talent, these tools are various ways that you can do pretty amazing things. In our case, we didn’t go out and buy some widget. We just said, “Okay, let’s take all the things we’ve learned about who’s performing well or what products do well or whatever it is, train models on that, and then run it across the system and see what it shows up.” And yeah, there were some mistakes. There were some people where it said, “this person’s a huge star.” And then you actually look down and you’re like, “no, they’re not.” But for a lot of times there were people who really were incredible stars and it was great to be able to recognize them. Do you think being managed in an automated or quantified way will dehumanize or depersonalize your workforce? You can see how management by robot does get pretty blown up. If every promotion decision is made by some totally unaccountable AI system, that seems wrong. But on the other hand, if an AI system is better able to say, “Hey, Matthew, CEO, here’s this junior customer support person who’s just been giving amazing answers to customers. You may not have ever seen them before, but you should give them a call.” I think that’s actually incredibly humanizing. That means that you can be seen for the contribution that you’re doing. There are lots of ways that bad organizations will use these technologies to do bad things, but that comes back to the leadership. Don’t be a bad organization. Be organized around trying to do the right thing. Celebrate your employees, make them rich. That’s exactly what you want. I feel like we’ve become a better organization, better at recognizing where talent is, better at then being able to reward that talent and invest behind that because of these various tools. Let me connect that all the way. At the beginning, I asked how you make decisions, and you said you start with mission and values. And a thing that really struck me at our last conversation was that you came to that realization, that was emergent as you began leading Cloudflare. I think you had a joke that was like your mission was to just impress your mom. And then that became this much bigger mission about security and all these other things that you’re doing. How has your relationship to leadership changed as you’ve automated the management function? Because there’s something big in there. I think it is incorrect to say that we’ve automated the management function. I think that what we have done — We’re collecting a bunch of data on what everyone is doing, we’re measuring it perfectly, and then we’re calling the customers — I don’t think we’re measuring it perfectly. I think we’re measuring it. We’re giving managers more tools to help both recognize the people who are overperforming and to help the people who are struggling get the resources that they need. We’re surfacing data that allows managers to be better managers. I don’t think we’re replacing managers. I think we’re making managers better at their jobs. And that scales management, right? Actually, it’s funny you mentioned that you’re up to 12 direct reports. The last time, you said — By the way, we’re not there yet. You’re not there yet? We’re headed there. Two years ago you said your number was about eight, which you called high. So you’re going from eight to 12. Mark Zuckerberg is at like, “We should have 50.” And I keep saying Decoder has a long life ahead of it because it’s a show about org charts. We’re on the cusp of the weirdest org charts in history. Are you there? Are you at 50? Oh no, no, no. I really do think that humans are social creatures and we want to be able to know people. And the Dunbar number is real, which is the number of direct social relationships you can keep in your head, which is supposed to be something like 120, but people should have friends outside of work. So you can’t just occupy those all with work colleagues. If I said eight before, that was wishful thinking. We were probably closer to six, but we are making our way up. We’re probably closer to eight-ish now, but we’re headed more towards 12. You have to be very, very specific because you also don’t want people playing games with that, where they’re like, “I’m going to build out a big team of people” where we don’t need a big team. I think that can be aspirational and directional, but I don’t think you can be religious. I don’t think you can just say, “Oh, everyone has to have at least 12.” You’ve got to figure out where that makes sense and where it doesn’t make sense. As that evolves, you have to also have some foundation that aligns people, and that’s where mission comes in. And it’s absolutely true. Our mission in the beginning was to take advantage of this interesting market opportunity, hopefully make some money, and impress our parents so that they’d get off our backs about why we didn’t go work at a bank or whatever. It was only as we started to serve our customers, and we saw just how important the internet was, and how it really lacked defenders, that we realized that our mission was really to help build a better internet. If you talk to anyone at Cloudflare, anywhere through the organization, any country that we’re in, any office that we’re in, I think that what I find amazing is, time and time and time again, they come back to saying, “The reason I work here is because I believe that the mission is one of the most important things we can do.” There are a lot of days that people are like, “Why do you still work at Cloudflare?” And I’m like, “Because I can’t imagine anything more important right now than helping build a better internet, and I can’t imagine anything that’s more exciting to be working on.” Well, Matthew, that’s a great place to leave it. We’re going to have to have you back very soon because I feel like the internet’s going to change even faster than I did last time. Thank you so much for being on Decoder . Thanks for having me. Questions or comments? Hit us up at [email protected]. We really do read every email!
We seem to be entering into an AI agent renaissance. Meta's new consumer-facing AI agent, Muse, topped the App Store charts soon after its release and has 600,000 daily active users in the US, by an Apptopia estimate. And AI agent platform Instinct, whose eponymous creator is fundraising at a $2.5 billion valuation, has been making the rounds in the tech industry. Under the hood, though, there are some obvious similarities to the platform that started it all: OpenClaw. For the past week or so, some social media users have alleged that Muse is directly built on OpenClaw. The two platforms use the same names for core files (SOUL.md, memory, … Read the full story at The Verge.