Mike Cannon-Brookes and 20VC on AI, Software Markets and Leadership

Harry Stebbings welcomes Atlassian co-founder Mike Cannon-Brookes to a roundtable with investors Jason Lemkin and Rory O’Driscoll. They begin with a reported optimistic scenario in which Anthropic reaches $149 billion in annual recurring revenue in 2029. The figure is a projection attributed to the report, not achieved revenue or a forecast the panel independently verifies.

O’Driscoll compares that scenario with an estimated OpenAI projection and the size of the global software market. He says both firms achieving very large revenue would require either expanded spending or displacement elsewhere. His market-size numbers and combined forecast are rough illustrations, not verified budgets or established future outcomes.

Lemkin asks whether model spending competes with Atlassian for a finite enterprise budget. Cannon-Brookes says the answer requires understanding the revenue stack: his company uses multiple model providers and selects among them for cost, quality and speed. As Atlassian’s co-founder and executive, he has a direct interest in that assessment.

Cannon-Brookes gives a hypothetical example in which a customer pays an application vendor, which pays a cloud provider, which in turn pays a model supplier. Each layer may report revenue from the same end-customer spending. His amounts are explicitly illustrative and should not be added together as if each represented a separate purchase.

He says software companies can partner with model providers while also competing with some of their products, as large technology firms have done in other markets. He expects model providers to remain available to others because of competition. That is his strategic view, not a guarantee of future access, pricing or vendor conduct.

Cannon-Brookes questions the assumption that AI spending is zero-sum. If firms can build more products or improve operations, he argues, technology budgets may grow. O’Driscoll agrees that market expansion could occur but says the very large proposed revenues would require substantial additional spending beyond a simple continuation of past trends.

O’Driscoll notes that revenue stacking complicates the arithmetic: spending on a coding application and on the model behind it can appear in separate company figures. He suggests consulting and systems-integration budgets as possible sources of additional spending, but his market estimates are approximate and the extent of substitution is unknown.

Stebbings thinks consulting revenue could grow during deployment and integration, while O’Driscoll expects some routine systems-integration work to be automated. Lemkin argues that effective implementation may require more technical skill. These are competing forecasts about different kinds of work, not measured industry outcomes.

Lemkin questions whether enough specialists can implement complex agent workflows. Cannon-Brookes responds that vendors have an incentive to simplify tools so more people can adopt them. Their exchange highlights a tension between expert-led deployment and self-service, without establishing the correct staffing model for every enterprise.

Cannon-Brookes responds that mass adoption requires tools ordinary employees can use, not an unlimited supply of specialists. He describes customers beginning with small automations and learning gradually what agents can and cannot do. His comments reflect Atlassian’s product perspective and do not independently establish how widely or safely agents are deployed.

Asked whether AI means the end of software, Cannon-Brookes rejects the category-wide claim. He says some companies will fail and others will grow, as they did in earlier technology transitions. This is a prediction from the leader of a software vendor, not proof that any individual company will survive.

Lemkin distinguishes engineering and product tools, which he thinks could benefit from more software creation, from other categories that may face shrinking seats. He cites observations from companies in his network. His claims about hiring, tool demand and named companies are anecdotes and forecasts, not a representative labor study.

Lemkin describes product and engineering as a comparatively resilient category because teams may build more software rather than immediately reduce engineers. Cannon-Brookes notes that service products are also significant for Atlassian. Both interpretations reflect their own business contexts and do not establish industry-wide employment trends.

Cannon-Brookes says product and engineering represent a significant part of Atlassian’s business but service management is also important and growing. He describes a newer customer-service offering and argues vendors must make service cheaper or better. Those are company performance and product claims from an interested executive, not independently verified results.

He cites Atlassian cloud-revenue growth and remaining performance obligations, or RPO, as evidence of customer commitments. He uses a simplified multiyear-contract example to explain that future contracted amounts are not all current-period revenue. The company figures are his account, and RPO is not the same as cash collected or a guarantee of future recognized revenue.

O’Driscoll agrees that software as a whole will not disappear but suggests an architectural shift might increase the number of companies that fail. He explicitly presents his percentage examples as invented illustrations. He asks whether AI changes only development speed or the character of the software itself.

O’Driscoll distinguishes normal turnover among software vendors from a potentially sharper architectural change. He stresses that his illustrative failure rates are invented for the discussion. The question is how quickly incumbent products and new competitors adapt, not a measured prediction that a particular percentage will fail.

Cannon-Brookes says Atlassian is using language models to build software and investing in context, search, chat and infrastructure. He says some inference costs have fallen substantially and cites improved gross margin across recent quarters. These are company claims without a product-level cost schedule or independent verification in the interview.

Stebbings asks whether falling software share prices mean investors should sell particular holdings. Cannon-Brookes believes fear is influencing valuations but says he cannot promise every company will be worth more in five years. His view of the sector and his comments about competitors are not personalized investment advice or an independently tested valuation model.

Lemkin credits Atlassian with accelerating metrics, while worrying that many peers have not regained growth. He argues that public companies should show how AI spending becomes useful products and revenue. His company figures and predictions about peers are his estimates and judgments, not a verified forecast for the sector.

The panel discusses whether the public software cohort may look weaker partly because fewer fast-growing companies have listed and private equity or larger firms have acquired others. Cannon-Brookes recalls a group of software CEOs that became smaller as companies were acquired. Their observations are possible composition effects, not a statistical decomposition of index performance.

Cannon-Brookes says some former public software companies have been acquired, while other fast-growing firms remain private longer. He argues those changes make simple comparisons of a public-company average difficult. The examples are his recollection and do not quantify the influence of each factor on valuations.

Cannon-Brookes adds that marketing-channel changes, access to private capital and different company sizes can affect the comparison. He believes customers will continue buying software, but acknowledges multiple forces are operating at once. His expectation reflects his role at Atlassian and does not resolve which suppliers will gain or lose.

Stebbings raises reported Harvey financing of $200 million at an $11 billion valuation and cites possible future revenue of $600 million. The panel also references roughly $190–200 million in ARR. These are reported or estimated figures, and the future number is a scenario; annualized recurring revenue is not a completed year of recognized revenue.

O’Driscoll notes the contrast between concern that a model provider’s legal plugin could disrupt incumbents and investor enthusiasm for a legal AI company. He thinks the important question is how long growth can compound and how large the addressable market becomes. His estimated market shares and future trajectories are hypothetical.

Lemkin argues that realized growth may reveal demand that earlier market estimates missed, especially if AI captures spending on services rather than only software. Stebbings and O’Driscoll question the price required to own that growth. Their disagreement is about uncertain investment assumptions, not whether the product has achieved a specified future revenue.

Cannon-Brookes says technology can change the size of a market and that a product customers like can be valuable even when it relies on another model. He believes customer relationships and distribution may provide an advantage. His favorable view of Harvey is an opinion, not proof of durable protection from competitors.

O’Driscoll clarifies that his use of the phrase “model wrapper” was meant to describe a market narrative, not dismiss the value of Harvey’s product. Cannon-Brookes says software often organizes an existing process into a useful interface. Both still acknowledge that durable customer value and competitive advantage must be earned.

Stebbings compares paying a high private valuation for Harvey with buying shares in Atlassian after a decline. Cannon-Brookes distinguishes company quality from the decision about where to invest and recounts Accel’s early secondary investment in Atlassian. His recollections of the price, financial model and eventual return are approximate and not independently checked.

O’Driscoll agrees a company can be impressive while still being a difficult investment at a high price. He uses a roughly 50-times run-rate revenue calculation to illustrate how much growth a buyer might need to justify its entry. The example depends on the speakers’ unverified valuation and ARR figures and makes no claim about realized returns.

Cannon-Brookes notes that an earlier investment model for Atlassian underestimated subsequent growth, but warns that the example is a successful outcome viewed in hindsight. O’Driscoll says the higher the entry multiple, the more future performance may already be reflected in the price. Neither example proves a particular return for a new investor.

Lemkin says his view of AI replacing some work has changed and points to legal labor spending as a possible source of demand. He stresses that not every company will prove the thesis. His estimates of spending and potential displacement are hypotheses, not evidence of actual jobs eliminated or revenue captured by Harvey.

O’Driscoll offers a counterexample using estimated revenue per active lawyer: if a product remains a relatively low-cost tool, its addressable spending could be smaller than a model that replaces labor assumes. The inputs and customer count are rough and unverified. He says increasing value per customer would change the calculation.

Lemkin predicts some agents could command much higher prices if they perform work previously done by people. O’Driscoll asks whether productivity could instead produce more output with the same staff. Their competing scenarios do not establish pricing, staffing or legal-service outcomes for a particular customer.

The guests disagree about the effect of automation on headcount. Lemkin sees substantial risk to seats outside engineering and product; O’Driscoll asks whether more output might absorb some productivity gains. These are forecasts about possible responses, not demonstrated outcomes for a profession.

Cannon-Brookes says legal firms’ billing practices complicate the economics of efficiency and cautions that he is not a legal-market expert. He proposes distinguishing functions constrained by incoming requests from activities such as software creation where more output can be made. The categories are a framework, not a universal rule about employment.

He illustrates the distinction with in-house legal questions and customer support, where demand may depend on the number of requests, versus engineering, where a roadmap can expand. O’Driscoll applies the idea to systems integration: reducing the cost of installing one system does not necessarily mean a customer buys two.

Lemkin counters that cheaper assistance can also create demand for work previously not worth doing. He describes a SaaStr experiment using Agentforce on lower-priority leads and says it worked. The example comes from his own organization and has no controlled results in the interview, so it cannot establish a general return on investment.

Stebbings asks whether reported Sierra growth represents new support demand, replacement of existing software or work previously done by people. He cites a quarter above $50 million and more than $150 million in ARR. Those are reported figures with no independently verified revenue recognition or customer-level breakdown in the discussion.

Cannon-Brookes says Atlassian already has a customer-service product and that its service collection is substantial and growing, without disclosing detailed results. He sees support as a large conversational market with both internal and external use cases. These are claims by an executive whose company competes in the category.

He says a buyer can compare the cost of a service product with the savings it creates, but that vendors must keep delivering value in later years after initial savings are embedded. His suggested support-spending percentage and return framework are rough estimates, not verified outcomes for every business.

Lemkin cites reported AI-support revenue at an incumbent and claims hiring in support has fallen while new products attract spending. He also argues better experiences can create an immediate “aha” moment. These claims about revenue, hiring and causation are unverified; reduced hiring alone would not prove how much budget moved to AI.

Lemkin argues that an effective AI support experience can be markedly different from a traditional tool and could eventually expand into adjacent tasks. He describes what he has observed and expects, not a controlled comparison of support products or proof of customer satisfaction.

O’Driscoll argues that support is not a single uniform market. Internal help desks, external business-to-business support and high-volume consumer queries have different workflows, customer sizes and communication channels. That segmentation may explain why several companies can exist, though it does not establish how many will ultimately succeed.

Cannon-Brookes suggests that teams using support agents may need to write more documentation, because answers depend on available knowledge. He reports seeing this pattern internally and among customers, and speculates that work could shift toward documentation. Those examples are his observations, not a measured employment forecast.

He says agents may help people answer questions or take actions such as filing a request or resetting a password. In his view, completing an authorized task can add value beyond reducing support costs. His examples are product aspirations and company observations; they do not establish that any system acts accurately or safely in every case.

Stebbings remains concerned that support is crowded by established firms, well-funded newcomers and customers building their own tools. O’Driscoll agrees that investing directly against well-capitalized leaders can be difficult and that a neighboring provider might bundle a smaller product. The outcome of any individual competition remains unknown.

Stebbings emphasizes the opportunity cost of investing in a category with many funded entrants and strong incumbents. O’Driscoll agrees that even in a large market, a company can struggle if an adjacent competitor bundles its functionality. Their comments are investment judgments rather than a ranking of companies.

O’Driscoll suggests investors may pay a high price for a perceived consensus leader rather than identify an earlier challenger. He uses a one-times preference as part of the example; its protection depends on actual terms and available proceeds and is not a guaranteed return. His financing statistics are estimates without a source in the conversation.

Cannon-Brookes says major technology shifts can attract large amounts of capital, create a few winners and produce losses along the way. He is skeptical of treating the current cycle as immune to those patterns. He also suggests some investors value recognizable portfolio logos; this is an opinion about incentives, not evidence about a particular transaction.

The group turns to Anthropic’s Super Bowl advertising and OpenAI’s response. They summarize their understanding of the competing messages and debate whom the ads were intended to reach. Their descriptions of private motivations, employee messaging and audience reaction are speculation, not independently measured campaign results.

O’Driscoll suggests a consumer AI product may need advertising if relatively few users pay for subscriptions, while an enterprise product may choose differently. He presents this as an interpretation of business models, not verified conversion data or a statement of either company’s long-term policy.

Lemkin questions the strategic value of an ad focused on a product area where he considers Anthropic weaker. Cannon-Brookes compares highly visible advertising in technology cycles. Their opinions about effectiveness and executive motives are not supported by campaign attribution, spending records or an audience survey.

Cannon-Brookes distinguishes private firms with abundant funding from public firms asked to show profitability and growth. He suggests marketing decisions can look different under those constraints, but his description of spending at model companies is an estimate. It does not establish their budgets or the effectiveness of their campaigns.

Stebbings pushes back, citing a Wix marketer who regarded Super Bowl placements as effective. O’Driscoll questions whether tech companies would purchase them as consistently if budgets were tighter; Cannon-Brookes says the answer depends on the business and the expected return. The interview does not establish the return for Wix or either AI company.

The panel discusses whether public-company reporting pressures make it harder to compete with heavily funded private firms. Cannon-Brookes says his job is to meet near-term obligations while investing in future products and research. His view reflects his responsibility for Atlassian’s capital allocation, and private and public companies can face different constraints.

He argues that leaders must explain why research spending matters and make difficult choices, citing long-established public technology companies as examples. He believes public status has made Atlassian better at forecasting and execution, while warning that financial planning must not replace strategy. This is a self-assessment, not a controlled comparison.

He says it is difficult for investors to see exactly how research spending will affect a company’s competitive position in future years. In his account, a public-company CEO must explain the investment and execute on it. His claim that Atlassian can do both is a management view, not an independently verified forecast.

Stebbings asks whether Cannon-Brookes has communicated the company’s story effectively given a lower share price. Cannon-Brookes says Atlassian delivered against many goals but acknowledges that the market’s judgment about future value is weaker. He says the team must keep adapting and building; his assessment of company performance is not independent financial analysis.

Cannon-Brookes urges leaders to accept changed conditions and create new customer value rather than wait for disruption to pass. He lists Atlassian’s customers, capital and research staff as resources, using approximate figures, and says faster building can be an opportunity. Those company numbers and future prospects are his claims.

Lemkin wonders whether all software CEOs want to lead through another demanding transition. Cannon-Brookes and O’Driscoll say a leader may reasonably decide to step aside and make room for someone else. They frame that as a personal and organizational decision rather than a moral judgment about the leader.

Asked whether he is working harder, Cannon-Brookes says his co-founder’s retirement and the pace of change have increased demands, and that he currently starts early. He stresses that a leader should also consider whether the work is enjoyable and whether family, exercise or time away remain part of life. This is personal advice, not a prescribed schedule.

Lemkin jokes that many founders might quit if they assessed balance too honestly. Cannon-Brookes clarifies that he does not expect every day to be enjoyable; over a longer period, he believes leaders should value creating products, building teams and serving customers. His reflection acknowledges difficult periods without prescribing one response for everyone.

The group discusses the tension between intense ambition and life outside work. Cannon-Brookes says time with his children and friends helps him and that spending every hour at work can lead to worse decisions. His experience is personal, not clinical evidence that a particular schedule works for all founders.

O’Driscoll agrees that time away from work can help prevent exhaustion and poor decisions, drawing on personal experience. The panel also notes the trade-offs of family, friendship and demanding work. These are individual reflections, not medical or universal productivity advice.

Stebbings says working with friends creates regular time together. Cannon-Brookes closes by valuing trusted colleagues with whom he has experienced successes and setbacks. He says a leadership team needs mutual help and shared purpose when facing hard problems; the conversation ends with thanks.