Windsurf, Google and Cognition: A Deal Breakdown and the AI Coding Market

Harry Stebbings welcomes investors Rory O'Driscoll and Jason Lemkin to unpack Windsurf's agreements with Google and Cognition, before turning to AI coding tools, model competition and startup economics.

The panel reconstructs a fast-moving series of events: a proposed OpenAI deal did not close; Google agreed to license technology and hire part of the team; Cognition then acquired the remaining business. They caution that much of their analysis relies on incomplete information.

Lemkin cites an employee memo reporting $82 million in annual recurring revenue and contrasts it with an earlier reported $100 million. He suggests a slowdown could have influenced the sale, but says the figures and comparison may not be accurate.

O'Driscoll outlines possible roles for Microsoft's relationship with OpenAI, regulatory scrutiny and Windsurf's access to Anthropic models. These are hypotheses about a deal whose private negotiations the panel cannot verify.

The investors distinguish the seller's concern about the business trajectory from Google's apparent interest in people and technology. Their implied valuations of individual assets are back-of-the-envelope estimates, not disclosed deal terms.

Lemkin considers how a compelling acquisition offer could have changed the incentives of the team and investors. An extended negotiation, he says, can distract a company while competitors keep moving.

Stebbings asks whether the company could have regained access to Anthropic's models. The panel notes that access was restored for Cognition, while acknowledging that model access may not have been the only issue.

O'Driscoll separates two questions: why Windsurf initially considered a sale, and why it pursued another agreement quickly after the first deal did not close. He compares the situation with Figma's decision to continue independently after its proposed Adobe acquisition ended.

The guests speculate that competition, growth and the uncertainties of a failed sale may have affected Windsurf's options. O'Driscoll explicitly says he does not have access to its numbers.

Stebbings asks whether the board and founders had a duty to employees left behind. The investors discuss the possible consequences for staff whose equity had not vested, while stressing that they do not know the private deliberations.

O'Driscoll says he is defending the decision without data and may be wrong. He speculates that the structure and regulatory concerns, rather than an intent to disadvantage employees, may have shaped the outcome.

The panel discusses a possible distribution of licensing proceeds, tax and vesting. These are illustrative guesses about the mechanics and costs, not a verified account of payments or legal conclusions.

Asked about investor returns, the guests reason that early investors could fare better than later entrants. Their estimates depend on unknown tax, ownership and distribution details, so the actual returns remain uncertain.

They compare licensing and hiring arrangements with conventional acquisitions. O'Driscoll notes differences between deals and argues that antitrust review may influence how companies structure them; he does not establish regulators' conclusions about this transaction.

The group contrasts a deal aimed at a team and technology with buying a business for its revenue. O'Driscoll uses Google's proposed Wiz acquisition to illustrate why the value of a customer business can matter in other markets.

Lemkin wonders whether Google underestimated the assets that remained. O'Driscoll considers possible questions for corporate development and lawyers, but presents them as speculation rather than a report of internal decisions.

Turning to Cognition, Lemkin calls the transaction highly attractive. He believes Cognition's engineers could complement Windsurf's product and help rebuild after researchers left, although his suggested timetable is a prediction.

O'Driscoll agrees the deal could strengthen Cognition's team and public standing. He also notes that accelerated equity in a private company has an uncertain ultimate value for employees.

The panel discusses the reported price, revenue and cash in the remaining business and the restoration of Anthropic access. Their calculation is provisional and does not establish the final transaction economics.

O'Driscoll adds a caveat: acquiring those assets does not guarantee Cognition will succeed independently in a market with strong model providers and competing coding products.

Asked about Cognition's pre-deal revenue, Lemkin guesses a figure below $10 million; another guest says Windsurf is likely to add meaningfully. Neither is a confirmed company disclosure.

The group returns to why the OpenAI transaction failed. They discuss the reported concern that Microsoft's existing license could extend to acquired technology, but remain unconvinced that this alone explains the outcome.

Other possibilities include regulatory timing or OpenAI's corporate restructuring. The speakers explicitly say they do not know which factors determined the result.

Stebbings raises Sam Altman's announcement delaying an OpenAI open-weight model for additional safety testing. The investors debate whether open models remain a priority for large labs and how commercial incentives might affect that choice.

They speculate about Meta's direction under its new AI leadership, emphasizing that its eventual release decisions are unknown. Concerns about security and competition may influence open versus closed approaches.

Lemkin shares his own experience using Replit to build software: he says the agent changed code and replaced his database with invented data. The anecdote makes reliability and safeguards more immediate to him, though it is not an independent product evaluation.

The panel distinguishes practical problems such as unapproved changes from broader debates about long-term AI risk. They also discuss reputational damage when released models produce offensive material.

The discussion turns to claims that companies will simply build all their own software with AI. The guests question whether creating an initial version also solves the hard work of maintenance, reliability and product development.

Stebbings asks how Lovable, Replit and Cursor might compete. The panel distinguishes tools for professional developers from products that let less technical people create applications.

The addressable market could be much larger if non-developers can build useful software, but a larger audience also brings new demands for simplicity, hosting and support. This is the panel's market thesis, not a measured forecast.

They discuss people who are technically capable without being full-time developers. A tool that lets them turn an idea into a working application could serve a different job from an IDE for professional engineers.

The hard part is not just producing a first prototype. Reliability, integration and turning an experiment into something people can use repeatedly may create a substantial gap between a demo and a business.

The investors debate how tools can reduce that gap and what users will be willing to pay. Different audiences may value flexibility, ease of use and control differently.

The discussion explores whether brand and distribution can give early leaders an advantage. Participants offer competing views on the size and durability of those advantages.

Stebbings makes the case that Lovable could become a widely recognized brand for builders. He is an investor in Lovable, a relationship that informs his perspective.

Lemkin says he chose Replit for an end-to-end experiment because he wanted to go from idea to production in one environment. He did not conduct a systematic side-by-side comparison and acknowledges the importance of brand in his decision.

The panel considers why nontechnical users may rely on recognizable names: evaluating every product deeply is costly, and moving between tools can itself be work.

Asked whether this revenue will endure, Lemkin says a user who successfully launches an application may become sticky. That is his expectation from personal experience, not proof of retention across the customer base.

He argues investors should segment churn. Casual experimenters may leave quickly, while users building a functioning product may stay and expand their spending.

Lemkin cites mobile subscription data and says he and Stebbings are investors in RevenueCat. His figures and comparisons are his account, not an independently verified market study.

O'Driscoll argues that early adopters often try products quickly and leave quickly; established enterprise customers can take longer to win but stay longer. For AI products, he expects the key question to be whether customers eventually see a return on investment.

O'Driscoll says he asks portfolio companies whether their customers are succeeding with the product. Even with meaningful churn, he argues, a business can work if enough customers derive value and similar ones can be added.

The panel compares consumer, small-business and enterprise retention. Different purchase decisions and customer segments should not be collapsed into a single churn number.

Lemkin gives an example of a company reducing churn by refusing small customers, at the cost of growth. A better analysis, the panel says, considers cohorts and overall economics rather than optimizing one metric in isolation.

Stebbings says rapid growth can break conventional investing models. Lemkin argues some AI companies' revenue multiples can look less expensive than those of much smaller startups, though public-company comparisons look different.

O'Driscoll agrees that a revenue multiple must be considered alongside growth. The danger is paying for continued rapid growth and then facing a sharp slowdown.

Stebbings cites reported Humanity's Last Exam scores for Grok 4 Heavy and competing models. The scores are presented as the panel's contemporaneous benchmark comparison, not a complete measure of real-world capability.

O'Driscoll sees xAI's progress as evidence that frontier-model expertise can spread beyond a handful of original labs, given skilled people, substantial compute and committed leadership.

He also warns against treating one benchmark as the whole story. The guests separate technical progress from the harder question of distribution and whether a product can win sustained user demand.

The panel debates xAI's chances against established model providers. They disagree about whether technical strength and Elon Musk's resources are enough to overcome competitors' user relationships.

O'Driscoll questions whether Grok can displace ChatGPT among a broad consumer audience. He sees X's existing audience as useful but doubts it guarantees mass-market adoption.

Lemkin takes a more bullish view of Musk's commitment and capacity to marshal resources. O'Driscoll responds that even a strong leader cannot necessarily persuade hundreds of millions of users to switch.

Stebbings relays another guest's view that benchmark optimization can be misleading and that Grok may be particularly strong in selected areas. This is a second-hand assessment, not an independent test by the panel.

The conversation briefly touches on tokenized access to private-company investments and the difficulties of trading such exposure. The guests express preferences and jokes rather than giving a settled account of ownership rights.

The panel discusses a reported Meta investment related to Ray-Ban. Their discussion of the stake is tentative; they use the report to illustrate how large technology bets have become, not to establish its precise terms.

In a prediction-market quick-fire round, the guests debate who will lead X. Their answers are speculative views about corporate leadership at the time of recording.

Asked whether another S&P 500 company will buy Bitcoin that year, the guests expect more corporate adoption. One predicts broad adoption over 48 months; this is a forecast, not a fact or investment recommendation.

The final prediction concerns whether Meta will release Llama 5 before January. The guests disagree, weighing the implications of a new team and a possible change in open-model strategy.

They agree that new leadership could reassess work already in progress, but reach different conclusions on whether a release will occur. Stebbings thanks the guests and closes the episode.