Atlassian CEO, Mike Cannon-Brookes on Why Everything is Overvalued & Are We in an AI Bubble

Summary

• “Most of the things are vastly overvalued” — but some are undervalued and will be worth far more, and “the far more is probably overweight”: the dot-com pattern where Amazon emerged from the rubble. The circular-revenue adage — an AI company gives $100M to a model company, which gives $150M to a cloud, which gives $200M to Nvidia, “and you’re all like losing 50 million bucks on the way through” — is why “we aren’t in an era where we have a durable business model,” and anyone claiming to know where value settles “is probably not telling you the truth.”
• Of at least ten CEOs Harry has asked, Cannon-Brookes is the only one who wouldn’t pay 10x for AI coding tools: “the competitive vector between them is not going to let you do that” — charge 10x and engineers switch — and coding is only 10-30% of a developer’s week anyway. Search, debugging and ops fill much of the rest, and more services — hypothetically, 100 becoming 1,000 — means more gremlins: a “Gordian loop of AI solving the problems AI created with AI.”
• The moat logic investors should keep: “You cannot build a business fast that has high switching costs.” LLMs were “dropped on the whole world at the same time” — unlike PCs or mobile, which took years to permeate — so switching costs for almost everything built right now start small, and they will have to build up from value delivered, data, workflows and familiarity. The race remains: “can the startup acquire distribution before the incumbent acquires innovation?”
• Per-seat pricing isn’t dying. Value-based pricing fails twice — buyer and seller must agree on measuring the outcome, and the shared-savings math works “for one loop, one lap of the world” (once the $100 support ticket costs $90, you can’t keep charging for the discount). Customers dislike consumption models, which tend to flip into another form of payment, as with unlimited storage. He expects a blended model: “the answer is probably a nuance and subtlety like always.”
• More engineers in five years, not fewer — Mike spoke right before Matt from AWS, and the two “violently agreed.” Atlassian is hiring more grads than last year, some 10x engineers may become 100x with the new tools, and the roadmap is unlimited: “find me a software company that gets to the end of its roadmap.” As creation gets cheap, design becomes the scarce, hard-to-copy differentiator.
• On what may take longer than expected: the “delta between magical demos and actual value delivered is quite high” — enterprise AI deployment may take much longer than hoped, requiring workflow changes, data cleanliness, security, education and more. And the future interface isn’t necessarily a chatbot (“otherwise we’d all just use the terminal”) — it’s stable UI plus prompt layers, with a better click-to-value ratio.
• 23 years in, the operating creed: “the founder’s job is to fight the entropy of ambition” (Tobes’s line, which Mike wrote down), and Atlassian’s defense through technology transitions is creativity — “we’re not going to be able to defend our way through it. We’re going to have to create it.”

Deep dive

1. “Most of the things are vastly overvalued” — and many lack a durable business model

• Asked how an investor should assess durability amid the chaos, Cannon-Brookes gives it straight: “Most of the things are vastly overvalued. Probably true. Some of the things will be worth far more and are undervalued. And the far more is probably overweight.” The classical rhyme: “for all the dotcoms that went bust, Amazon turned out to be quite a good business” — which is precisely what makes this era hard to invest in.
• His sharpest structural point — many of these things lack a durable business model. The circular-revenue adage: an AI company gives $100M to a model company, which gives $150M to a cloud provider, which gives $200M to Nvidia, “and everyone’s like look how much revenue we’ve all got and you’re like yeah but you’re all like losing 50 million bucks on the way through.” Maybe scale fixes it — “I don’t think anybody really really knows,” and “anyone who tells you that they know it is probably not telling you the truth.”
• When Harry offers his own changed mind — he never saw OpenAI and Anthropic as $3-4 trillion companies “in any world,” and now “if you squint you can” — Mike agrees they’ll be “big and important and impactful,” and adds that the speed with which AI can be understood and deployed may take longer to play through: “the delta between magical demos and actual value delivered is quite high.” Asked why — adoption? data? security? — “Every single one of those and more… Like most technical disruptions, we’ve got a hype cycle going on.”

2. Atlassian’s two bets: multi-model on purpose, design as the scarce asset

• Bet one, made very early: multiple foundational models would compete, so Atlassian built multi-model “in almost every way” and decided not to make foundational models. The expertise it chose instead: “there’s going to be a new model about every 3 months and there’s going to be four to six companies competing” — so the skill is picking up a model, testing it, and shipping it to customers fast. “That’s proven to be I think a really good bet.”
• Bet two: it’s really all about design — lowercase d, not colors. Every major technology transition forces fundamental design; his trite example is pull-to-refresh, which didn’t exist until phones did, then “one app did it and everyone copied it.” In a world where software gets cheaper and more abundant, “what differentiates it is probably how it feels and how it works. And that’s very hard to copy.”
• The corollary for cheap creation: “good design, scarce resources probably get more valuable” — and cheapness buys iteration. Plenty of technology is built once, badly, because “some team somewhere was like, man, I got to move on. I’m out of budget.” At lower cost, “maybe I can build them two or three times to get it right.”

3. Business apps may not collapse into a godlike Siri — that’s someone talking their book

• On a business-applications claim likely from Satya Nadella that applications collapse in the agent era: “it’s very easy to make big statements in the current era because you can’t be proven wrong.” His counter is the old enterprise adage that every piece of enterprise software can be replicated with email and Excel — true, and yet we have HR, CRM and project systems because specific tools do the task much better. “I don’t think we know in the AI era if all those apps just disappear into some godlike Siri agent… I’m not a believer in that world view.” AI will change every application “in massively fundamental ways” — some die, some profit — “but I don’t think that they’re all just simple CRUD apps. I think that’s usually someone talking their book.”
• Nor is the chatbot the endgame interface: “I don’t believe the world will just end up as a chat box. Otherwise, we’d all just use the terminal on our computers.” AI could in theory regenerate the UI for every screen, but users want stability — “it’ll be too discombobulating.”
• His actual prediction: the interface will look more like today than we think, but smarter, with far less “click-to-value ratio” — somewhere between buttons and the prompt that grows so long “I’m like just give me a damn button.” The Word example: tell it you’re not a lawyer, and it redesigns the toolbar for who you are.

4. More engineers in five years, not fewer — and finance people vibe coding Python

• Mike spoke right before Matt from AWS, and the two “violently agreed that 5 years from now we’ll have more engineers working for our company than we do today.” His reasoning: technology creation isn’t output-bound — “find me a software company that gets to the end of its road map” — and the R&D bandwidth constraint “will be I believe forever.”
• Vibe coding at Atlassian is live — “I’m in Barcelona and we’re launching one at the moment” — but his frame dissolves the category: the finance professional writing Python “to do crazy complex analysis,” who previously would have used Excel or filed a ticket “or they just wouldn’t have bothered,” isn’t a software developer. “That’s a financial professional who’s trying to get their job done and has a new set of tools.” Customers will probably make more apps — perhaps more apps per customer — many thrown away, “and that’s okay. That’s fine. That’s great.”
• On entry-level engineers, a contrarian call: “I think we’ll have far more of them” — Atlassian is hiring more grads this year than last, more than the year before. The 10x engineer of a decade ago “is probably still a 10x engineer today. Maybe they’re a 100x engineer.”
• The Loom parallel is the argument as told: one way Loom gets into businesses is through graduates, because the TikTok/Snapchat generation natively communicates by video, and “the business goes, man, those people are more productive” — the business or a majority of it adopts it. Today’s hyperproductive AI-assisted grads will shake up incumbent engineering talent in a similar way: “I better copy that or maybe I should go do something else.”

5. The only CEO of ten who wouldn’t pay 10x for coding tools

• Atlassian runs its own Revo Dev plus Cursor, Copilot and more — “we’ve been very liberal. They’re quite cheap, these tools.” Asked if he’d pay 10x: “No, probably not.” Harry: “I’ve asked 10 CEOs this at least, and you’re the only one who said no.”
• His two reasons: “the competitive vector between them is not going to let you do that” — charge 10x and you switch — and they’re all good at different things: most handle greenfield well, some are “very bad at long form… manipulate this piece of code across 500 repositories,” some can’t ingest a large codebase.
• The deeper one: coding is 10-30% of a developer’s week — search takes about as much time as writing code, plus meetings, debugging, production alerts. If a hypothetical medium-sized startup’s 100 services became 1,000, the permutations would multiply gremlins. “The answer to all these questions can’t just be throw more AI at it. You end up in this weird Gordian loop of AI solving the problems AI created with AI.” Harry’s reply, worth keeping: “that’s what venture investors are telling our LPs at the moment. So don’t ruin the money drain for us.”

6. Switching costs can’t be built fast — and margins are unknowable today

• Harry invokes Hamilton Helmer’s Seven Powers and Salesforce’s lock-in; Mike’s response is the episode’s most quotable moat logic: these AI businesses were built very fast, “which almost by virtue of that means the switching costs are quite low… You cannot build a business fast that has high switching costs. It logically makes sense.” Whether they’ll have them in future “is a very different question.”
• Why this era is unique: “Large language models have been dropped on the whole world at the same time… which is something I don’t believe we’ve seen in technology before.” Mobile and PCs took years to permeate; here a billion technically literate people got access at the same time — hence the verdant creativity, and hence low switching costs throughout. Future switching costs will come from “value delivered to customers, data to some extent, workflows, familiarity” — and the incumbent race: “can the startup acquire distribution before the incumbent acquires innovation?”
• On the [ __ ] margins critique: “we’re too early to really understand what those margin profiles are” — pricing schemes change every three months, and there’s the time degradation of model value: “I’ve trained a model. It cost me a billion dollars. Am I going to get that billion back in the 9-month period before the next model’s better?” These “will be sorted out in some form” — he just can’t tell you how.
• Where value settles: some investors think chips capture it all, or the power producers upstream — “if you’re selling electricity right now, it’s kind of a good business” — but he doubts it all flows upstream, because design and feel build willingness to pay at the other end. “If you’re just an OpenAI wrapper as a startup, I do think you’re in trouble” — though that might prove a fine way to start; “plenty of businesses made money on mobile without being iOS and Android.”

7. Per-seat pricing survives — value-based pricing works for one lap

• Only two logical replacements exist: per-unit-of-value and consumption. On the latter: “go talk to customers about how much they love consumption based models. Generally they don’t” — they can’t predict usage and don’t want to police employees. Consumption models tend to flip into another form of payment, as with vendors giving away “unlimited” storage while pricing overall to cover it. “I do wonder if AI ends up that way around.”
• Value-based pricing has two failure modes: both sides must agree on measuring X — “a big problem in a lot of these scenarios” — and the savings-share math works “for one loop, one lap of the world.” His worked example: cut a $100 support ticket to $80 and I’ll pay you $90 in year one; in year two “it only cost me $90 now… I can’t keep charging you for the discount because we’re going to end up at negative.”
• His landing spot: “I don’t believe per seat or singular models like that are going to go away, but we’ll probably have to get used to a little bit more of consumption based pricing… the answer is probably a nuance and subtlety like always.”

8. The co-CEO playbook: 60-80% overlap and the convince-Scott test

• Two decades of co-CEO with Scott worked because of equality — not just equity but “age, life stage… naivety and experience and wisdom” — and mutual awe: “it really helps to think the other person is just far better than you are at pretty much everything because you’re both stretching.”
• The geometry he’d hand Spotify’s new co-CEOs (Alex Nordstrom and Gustaf Sodestrom): “Scott and I are probably somewhere between 60 and 80% overlapped. If we’re 100% overlapped, there’s not much point in having two of us. If we’re 0%, you’re going to have a lot of conflict.” Swim lanes rotated — Mike ran sales and marketing for 13 years, then Scott for four.
• Conflict resolution as told: the first shareholder agreement literally specified scissors-paper-rock, best of three — never invoked, and no longer in the bylaws. The real mechanism: “If I can’t convince him that something’s a good idea, it’s probably not. That is a really good marker.” No epic disagreements in 20+ years — though a security incident once required radioing Scott’s bush plane on honeymoon in Africa, “hundreds of kilometers from anywhere.”

9. “Founder mode” denied — but this is the era for making bets

• Harry’s framing — that the last 18 months (Williams F1, the Browser Company and DX acquisitions) are “the most aggressive but also obvious example of founder mode in action” — gets flatly rejected: “there’s literally zero of that.” It’s like a new prime minister getting credit for interest rates; many moves “were planned many months, years, quarters in advance,” and Loom three years ago was just as bold. “I ring him all the time.”
• What he does concede: this is a “verdant, creative, Cambrian explosion period” demanding decisions. “One of the only predictions I think I would make accurately is 10 years from now, we’ll look back and say, man, you remember that era? That was crazy.”
• The operating discipline: make hunches about the next three years, “but about every quarter we reassess” — and hold the willingness to walk away. The stabilizer amid the noise: customers as “a rock in a storm” — “talk to enough of them that you get internal conviction about what is or isn’t happening.”

10. Fight the entropy of ambition — creativity is the defense

• Asked what he and Scott must protect fireside (drinking Southern Comfort, apparently), the answer isn’t a product — it’s creativity. An early maxim was to be a multi-decade technology company like Microsoft, Adobe and Intuit, which means surviving transitions: Atlassian has crossed SaaS, mobile, and now AI. “In the long arc of time, it’s going to be whatever’s next after AI… we’re not going to be able to defend our way through it. We’re going to have to create it.”
• The Tobes line he wrote down: “the founder’s job is to fight the entropy of ambition.” The moment ambition stops, “we are freaking tiny as a business… on the scale of the global economy we’re like tiny — which means we got a long way to go.”
• He rejects the unconstrained-resources hypothetical outright — “constraints are what makes everything happen… a football game with no 90 minutes would just be boring” — and reframes: 10,000 people in R&D, billions in the bank, “now let’s go build something. You’d be like, oh my god, we can’t complain about anything.” In the unconstrained hypothetical, the constraint would be talent acquisition.
• Quick-fire keepers: biggest flop was adding Twitter-style status updates to Confluence 20 years ago (“nobody understood it”); unlike Jensen’s “if he knew how hard it was, he would never have done it,” his average of 23 years says “no, I do it again every single day”; and for Atlassian’s future — “you don’t want your company to live on as some sort of weird zombie. You want it to live on as a vibrant competitive place.”