Mike Cannon-Brookes

1. Key Themes

Headless Is a Brainless Bet: Design Gets More Valuable as Agents Proliferate

Cannon-Brookes rejects the idea that agents hollow out application interfaces. He argues that as technology creation becomes cheap and abundant, experience becomes the differentiator. He is simultaneously "big on MCP" and wants the interface to keep evolving: the graph is exposed to agents, but the product must still solve the customer's problem better through its own UI.

"Headless is brainless, by definition." 00:05:22

"I think design and experience is one of the keys to success in the AI era... If you argue that technology creation is increasingly cheap, technology creation is abundant and verdant, the differentiators, to me, clearly one of them will be design and experience." 00:06:31

Agents Won't Replace Interfaces: Horizontal, Vertical, and Embedded Agents All Coexist

He rejects the "one uber agent" thesis. Task type matters: functional tasks get absorbed by agents, while tasks involving taste, collaboration, and creativity keep interfaces. The Atlassian model is a two-way, A2A-style integration, with Figma and Databricks agents inside Atlassian and Atlassian inside their interfaces.

"I don't think agents will replace all the interfaces. I think all three of these can be true at the same time." 00:00:00

"I'm a big bear on the magic button thesis of AI that there'll be one button we press and it does it all for us. I don't think that's how it actually is going to work." 00:07:52

The Enterprise Context Graph Is the Core Moat (and It Must Be Fresh)

Atlassian's Teamwork Graph indexes corporate decisions, documents, people, and code ahead of any agent query, like a Google-style index tailored to each company. Cannon-Brookes stresses that the value lies in freshness plus historical learning, not in stockpiled raw data.

"The teamwork graph's job is to infer ahead of your agent. Before your agent has asked a question, to infer the best possible answer to that. Think about it like Google's index of the web." 00:36:21

"An organizational graph has to be very fresh and current with facts, and very historical in its learnings and memories to help the organizations best as possible." 00:37:53

Foundational Design Problems of the Agentic Era Are Unsolved

He says Atlassian is solving foundational design challenges for the first time since mobile: the blank text box, provenance of AI vs. human content, accountability, and how to supervise processes without drowning in detail.

"For the first time in a long time, we're solving foundational design challenges again, right? I would say the mobile era is the last time where we had to think about completely new design paradigms." 00:00:00

"I mean text boxes suck, firstly. Giving you a little text box and saying, 'Here is a one-line piece of text that you can type characters into, and it has unlimited power, and it's just a blank white line.'" 00:11:10

Specialized Models: "The Jury Is Out," and Most "Own Models" Are Fine-Tuned Chinese Open-Weight Models

He is skeptical of claims about "training our own model," arguing that most are fine-tunes of open-weight (often Chinese) models, that cheap fine-tunes of someone else's billion-dollar model aren't a from-scratch build, and that generalized models tend to erase specialized advantages over time.

"The jury is really out on the benefits of building specialized models at the moment." 00:00:00

"Almost every single time, that is, if it's not Nemotron, which it's usually not, it is a Chinese model that they've trained, or added their own fine tuning, specialization on top of, and then they call it their own model." 00:19:38

"The quality of the generic model has improved far faster than their ability to stay ahead with whatever specialization they were doing." 00:22:53

Open Models Yes, "Chinese Models" Are Sensitive: Atlassian Doesn't Use Them in Production

Customer pushback is on Chinese-origin models specifically, not on open models. Atlassian uses Nemotron (via its NVIDIA partnership), Mistral, and other open models, but no Chinese models in production. He argues the skepticism is partly inconsistent given Linux's global contributor base and the distillation loops between US and Chinese models.

"There is certainly some customer pushback in some places to quote unquote Chinese models. There is not customer pushback to open models." 00:00:00

"A lot of the US models are increasingly using Chinese models to train themselves, but they don't say that, and so you end up with all these weird loops." 00:17:19

Cost Discipline, Model Routing, and FinOps as a Core Technology Competency

Enterprise customers are becoming very sensitive to token spend. Atlassian runs an internal AI gateway team and does model routing and blending itself. His product guidance: build magic first, optimize cost only once adoption scales.

"I think that is a core strength and attribute of a great technology company will be its ability to manage technological cost." 00:25:15

"If five people are using a feature, I don't care how much it costs. If 5 million people use a feature, I'm going to care a lot about cost, and in the middle, I usually have a fair bit of time to keep optimizing." 00:29:23

Inference Prices Will Fall Further Than Expected

Asked what he changed his mind on, the answer was the future price of inference. He argues that the enormous volume of capital and talent targeting price-per-watt (chips, data center design, distillation) makes a precipitous decline likely, and that Jevons-style demand growth will be bounded by consumption limits.

"Right now, the production of technology is faster than the consumption of technology. It is ahead. We've reached a tipping point. We can build features faster than customers can consume them in a realistic way that changes their business." 00:00:00

"I think so, further than I thought." 00:49:40

From Information Asymmetry to Imagination Asymmetry

Borrowing a phrase from Culture Amp's Didier Elzinga, he frames the new competitive edge as creative application of AI rather than knowing something others don't.

"We're moving from a world of information asymmetry to imagination asymmetry... can I think about how to apply AI and these technologies to problems in unique ways?" 00:00:00

Provenance and Accountability Over AI-Detection

He predicts that in five years, AI-touched content stops carrying a stigma, much like CGI or color grading in film. What matters is how much human contribution is involved and where accountability sits.

"We don't attribute quality or subjectivity to AI affected content, AI improved content, AI of any form. Just like today, when you go to the movies, we don't say... 'I want to see films that aren't color graded.'" 00:15:42

2. Contrarian Perspectives

Agents Make Design More Valuable, Not Less

The consensus view says agents reduce human interaction with software UIs, turning apps into data providers. Cannon-Brookes argues the opposite: abundant technology makes experience the scarce differentiator. His backup is Atlassian's 20-year record of upleveling UX, and the fact that Jira and the teamwork collection posted their fastest growth ever at massive scale.

"I don't think agents will replace all the interfaces." 00:00:00

"Jira and the teamwork collection have had an amazing year of growth and accelerating and growing faster than they ever have at massive scale, and that is in the number of users and attention. And a large part of that is because of the design experience." 00:43:24

"We Trained Our Own Model" Claims Are Mostly Marketing

When Harry cites Harvey and Ramp touting their own models, Cannon-Brookes dismisses most such claims as relabeled fine-tunes, noting "You see they train them for like a million bucks and I'm like, 'There's no way you trained this for a million bucks.'" His stronger claim is that specialized models are structurally fragile because the generalized frontier keeps absorbing their niche.

"You took someone's billion dollar model and tailored it... That's totally fine, but that's not really training your own model, I would say, right? If you're not starting with raw data." 00:21:04

Enterprises Refusing Chinese Open Models Is Partly Incoherent

Against Jason Lemkin's claim that big companies won't touch Chinese open models, he argues the geopolitical question is separate from the technology question, and that enterprises already accept Linux despite heavy Chinese contribution. He also argues AI makes open code easier to audit than ever.

"I think there's a geopolitical issue around who has the sort of strategic control of the world. That is a different issue from, should we use a model that was trained in China." 00:18:58

"But for an enterprise to say, 'I'm not going to use a cheaper alternative that'll give me better results, when I know what's in it,' I think is going to be a very hard long-term choice." 00:19:38

Well-Funded Startups Ignoring Cost Are Building the Wrong Muscle

Harry notes that private founders often say they don't care about model cost. Cannon-Brookes replies that skipping cost discipline is a strategic weakness, and that using cheaper or older models well is a real advantage.

"I love all these private companies that have raised a bunch of money and don't have to care about cost, cuz they're not building the discipline to care about cost." 00:24:57

Inference Is Headed Toward Near-Free, Despite Jevons Paradox

Where most argue Jevons paradox means ever-rising spend, he argues the unit cost will fall "precipitously" and that consumption will saturate, as with phone CPUs, which are never run at 100% continuously.

"There's not going to be an infinite usage of all of this stuff, right? Your phone has amazing superpowers... Do you burn the CPU 100% all the time? No, you don't." 00:31:12

The "Software Factory" Analogy Is Flawed

A passing but sharp challenge to a fashionable metaphor: software is creative, not mechanistic.

"A factory is a very mechanistic, defined edge thing, and software is fundamentally very creative, right?" 00:12:22

3. Companies Identified

Atlassian

Enterprise collaboration and developer workflow software (Jira, Confluence, Rovo, Teamwork Graph). Mentioned as Cannon-Brookes' own company; key metrics and strategy were shared. Confluence is north of $1.5B ARR, growing north of 25%.

"Confluence is north of a $1.5 billion ARR business, that was our last disclosed number a couple of quarters ago. It would be a public company in and of itself. It is still growing well north of 25% per year." 00:41:37

"I think we have one of the top five, top three MCP servers in the world." 00:05:30

Figma

Design collaboration platform. Atlassian's partner with agent interop in both directions, cited as an example of A2A across multi-craft teams.

"It's one of the best examples of A2A across multi craft teams." 00:07:52

Databricks

Data and AI platform. Atlassian partner; users can query Databricks from inside Confluence via the teamwork graph's data context.

"If I'm reading a Confluence page and I want to just select a piece of text and say, 'Can you check this number for me?' and it goes off to Databricks or to Snowflake through the teamwork graph's data context and gets you back the answer or gives you a small chart, that's great." 00:07:52

Snowflake

Data cloud platform, named as another data source accessible through Atlassian's teamwork graph. Quote as above. 00:07:52

NVIDIA

Chip and AI company; deep Atlassian partner whose Nemotron open models are used in Atlassian's platform.

"We use a lot of Nemotron in our platform, for example. We have a deep partnership with NVIDIA." 00:17:19

Mistral

European open-model lab, used by Atlassian in parts of its platform.

"We use some Mistral. We use a bunch of other open models in different parts." 00:17:19

Anthropic (Claude)

Frontier lab. Atlassian partner whose agents access and contribute to the Atlassian platform; also cited for its watermarking features.

"Both of them are deep Atlassian partners and we really value our partnership with them." 00:41:37

OpenAI (Codex)

Frontier lab and Atlassian partner; Codex used as the example of agents creating pull requests on behalf of humans.

"If I'm using Codex and it does something for me, for example, if I'm just writing code and Codex creates a pull request, did Harry do that, or did Codex do that?" 00:11:10

Cursor

AI coding tool, named among the agent platforms that access and contribute to the Atlassian platform. 00:07:52

Microsoft (Teams) and Slack

Collaboration surfaces where agents run and connect to Atlassian. 00:07:52

Linux (open source project)

Cited as the example of enterprises happily using open code with significant Chinese contribution.

"I think it's like 40% of contributions or something like this is quite high nowadays come from China." 00:17:19

Culture Amp

Employee-experience software company; its founder is credited with the "imagination asymmetry" framing. 00:32:14

Fireworks

Open-model inference platform; Harry is an investor and believes in open models. 00:17:19

OpenRouter

Model routing gateway, recently acquired for "a lot of money" per Harry. Cannon-Brookes says large companies must build routing in-house, possibly layering OpenRouter among other things.

"For smaller businesses, or businesses where technology is maybe not their expertise, those are just very practical solutions." 00:28:10

Harvey and Ramp

Companies cited by Harry as touting benefits of building their own models; Cannon-Brookes is skeptical of the "own model" framing. 00:21:04

Mercor and Surge

Data marketplaces paying for large datasets; Cannon-Brookes explains they serve generic model-training needs, which differ from Atlassian's per-company context graph.

"Those type of companies are looking for large-scale, generated generic datasets to sell at scale to many other providers." 00:37:53

Spirit Airlines (data sale)

Cited as an example of corporate data (emails and other records) being sold for about $30 million to a data buyer. 00:36:21

Linear

Design-focused issue tracker; Harry cites it for design craft and perceived advantage over Jira. Cannon-Brookes says "I think it's unfair" and that Atlassian will "run our own race." 00:43:24

Instinct, Muse, and Jevons-style "decision models" (Jev)

Agent and model products mentioned in passing; Cannon-Brookes notes these small "decision models" and their clones are being applied to tasks previously too expensive for GPT-3/4-class models.

"They are often used for tasks that previously would have been too expensive to use even like a GPT-3 or GPT-4 level model for." 00:27:03

DX

Engineering-intelligence product within Atlassian's portfolio, used for visibility into token spend across engineering organizations.

"Either in DX in your engineering organization, right? Understanding where you use which model, where you're getting value." 00:25:15

Sponsors: Base44 and Laridin

Base44 (AI app builder) and Laridin (AI/human workforce impact measurement) appear as ad reads, not as subjects of the discussion.

4. People Identified

Co-founder and CEO of Atlassian. The guest; his candor on model strategy and design philosophy is the substance of the episode.

"I think design and experience is one of the keys to success in the AI era." 00:06:31

Scott Farquhar

Atlassian co-founder, now retired but still highly engaged. Cannon-Brookes sees him roughly monthly and gets "near-daily text messages."

"He's very active for a retired person, which is great." 00:45:19

Didier Elzinga

Founder of Culture Amp, credited with the "information asymmetry to imagination asymmetry" phrase.

"I give uh Didier, the the founder of Culture Amp, the credit, cuz it's his his phrase." 00:32:14

Jason Lemkin

SaaStr founder and investor. Cited for the claim that enterprises won't use Chinese open models; Cannon-Brookes jokes he is "trying to run one of the smallest companies in the world."

"Jason's increasingly trying to run one of the smallest companies in the world, I think." 00:17:19

Karri Saarinen

Linear co-founder (named as "Karri" in the transcript), recognized for design craft. 00:43:24

Harry Stebbings

Host of 20VC and venture investor (including Fireworks); shares candid views on founder quality and VC-driven easy fundraising.

"We've made it too easy in some respects, and that's venture capitalists' fault." 00:48:32

Unnamed former Goldman Sachs CEO

Source of the quote about people who "grow into the job or swell into it," which Cannon-Brookes uses as a personal standard.

"You can tell a lot about people when they get promoted as to whether they grow into the job or swell into it." 00:50:15

5. Operating Insights

Don't Over-Optimize Cost Before Product-Market Fit; Tier Your Cost Effort by Usage

Cannon-Brookes' rule for product teams: build the magical version first, think about cost reduction ahead of time, and invest in optimization only once adoption proves out.

"Don't over-optimize for cost in the short term. Build amazing, magical features. If they're taken up by customers, and what we thought was magical, the customer thinks is magical and it scales, now I start working about the cost." 00:29:23

Build an Internal AI Gateway Team, and Optimize the Quality/Cost/Speed Triangle per Feature

At Atlassian's scale, routing and blending models is run by a dedicated, non-trivial team. The trade-off triangle is applied feature by feature, and cost savings are returned to customers as a product advantage.

"We have an entire AI gateway team. It's a not a small team that is doing that." 00:29:23

"The best quality answer for the cheapest quality price at the best speed. This triangle trade-off we're always looking at in any given feature." 00:28:10

Give Customers Token-Spend Visibility

Large enterprises care intensely about price-performance. Surfacing model usage, value, and optimization opportunities is becoming a product requirement.

"We have a whole raft of products now that will show you your token spend... because this is a bigger and bigger budget item, and they see it's only going to be a bigger budget item, and so they have to care about the spending." 00:25:15

Define "Own Model" Honestly and Use Fine-Tuned Open Weights Where They Win

Rather than claiming an own model, Atlassian fine-tunes open-weight models for specific purposes and presents the price-performance benefit, not the label.

"We have a bunch of very specialized open weight models that we fine tune for particular purposes that give great price performance for our customers, that I don't claim to be our own model." 00:21:04

Schedule Deep Thinking Time Deliberately

His explicit self-management practice: carve out even "a couple hours a week" for long thinking, since agents and fractured attention fill every gap.

"It's very easy to be busy, and I think in the AI era it's even easier to be busy." 00:34:04

6. Overlooked Insights

Customers Will Quietly Accept Chinese-Derived Models Embedded in Products They Already Buy

A throwaway remark suggests the real exposure is invisible: most "proprietary" enterprise models are fine-tuned Chinese open-weight bases, and US frontier models reportedly use Chinese models in training loops. That implies enterprise "no Chinese models" policies are already leaky through vendors, which creates both a compliance diligence opportunity and a differentiator for vendors who can credibly certify provenance (Atlassian itself says it uses none in production).

"A lot of the US models are increasingly using Chinese models to train themselves, but they don't say that, and so you end up with all these weird loops where I'm like I don't know if that disappears over time." 00:17:19

A Company's Context Graph Gets More Valuable, and More Defensible, as It Stays Live, Not as It Accumulates Data

He passed over this quickly: data marketplaces buying static corporate datasets (like the Spirit Airlines email trove) are a different business from a continuously refreshed organizational graph. A static dataset decays almost immediately (people leave and join, projects shift), so the enduring moat is the live, per-customer graph embedded in daily workflows, which can't be bought or sold in bulk.

"The social graph of a company is valid today, and only today. Tomorrow, there will be different people in a company than there were today." 00:37:53