Varun Mohan (co-founder and CEO)20 April 2025

Building a magical AI code editor used by over 1 million developers in four months: The untold story of Windsurf

6Frameworks

14Insights

Episode overview

Varun Mohan explains how his company moved from profitable GPU infrastructure into Codeium and then built Windsurf after existing IDE interfaces constrained its AI capabilities. He describes Windsurf's differentiation through large-codebase understanding, custom retrieval and editing models, enterprise security, and support beyond its own VS Code fork. The conversation also covers high-agency hiring, deliberately understaffed teams, AI's effect on engineering work, and the need to cannibalize the product every six to twelve months.

Key ideas

Startups should concentrate on validating hypotheses, abandon assumptions when reality changes, and avoid remaining attached to their first idea.
Windsurf was created because conventional IDE APIs constrained agentic interfaces; moving to a custom surface tripled acceptance of the same refactoring models.
AI is shifting engineers from manually implementing solutions toward deciding which business problems matter and how they should be solved.
High agency becomes more valuable as AI lets product managers, salespeople, and other domain specialists build or modify software themselves.
The company hires only when a function is genuinely underwater, using constrained staffing as a forcing function for ruthless prioritization.
Windsurf combines frontier planning models with proprietary retrieval, editing, and autocomplete models informed by large volumes of user preference data.
Greater developer productivity can raise the return on software investment and cause ambitious organizations to hire more engineers rather than fewer.
Windsurf aims to make its existing product form obsolete every six to twelve months while still improving it through user feedback.

Transcript available · source text is retained privately and is not published

Frameworks in this episode

Mindset4 steps

Agency Over Title: The Force-Multiplier Mindset

Ignore role boundaries — use AI tools to execute your own ideas end to end

Read →

Strategy4 steps

Build Only Where You Can Be Best (Model Sourcing)

Build in-house only where your unique data or position lets you beat the frontier; otherwise buy

Read →

Innovation4 steps

Cannibalize Your Own Product Every 6-12 Months

Make your current product look silly on a fixed cadence instead of only shipping what users ask for

Read →

Productivity4 steps

Hills-and-Valleys: Getting Value From Probabilistic AI Tools

Be patient and explicit, start small, and learn where the model is strong vs weak

Read →

Entrepreneurship4 steps

Irrationally Optimistic, Uncompromisingly Realistic

Push hard toward a future you believe in while killing your own hypotheses as data arrives

Read →

Leadership4 steps

The Dehydrated Company Hiring Model

Only hire when a function is underwater, so headcount forces ruthless prioritization

Read →

People & resources mentioned

Attributed to the moment in the episode. Timestamps are approximate.

People · 7

Lenny RachitskyMentionstoday we've got another very special compilation episode something I've been pulling on more and more with the podcast and the newsletter
“Today, my guest is Verun Mohan.”
▶ Watch at ~0:30approx.
Varun MohanMentionsBaroon is the co-founder and CEO of Windservef
“Baroon is the co-founder and CEO of Windservef”
▶ Watch at ~0:30approx.
Uri LevineMentionsAuthor and startup founder associated with the principle of falling in love with the problem rather than the solution.
“Ori Lavine was on the podcast, Co-founder Ways”
▶ Watch at ~12:30approx.
Dario AmodeiMentionsI was sitting next to Dario yesterday and he's like I keep making these predictions and people keep laughing at me
“Daario from Thropic recently said the same thing.”
▶ Watch at ~17:00approx.
DouglasMentionsDouglas being my co-founder
“Douglas being my co-founder”
▶ Watch at ~27:00approx.
Carlos DelatorreMentionsCarlos Delator who used to be the CRO of MongoDB
“Carlos Delator who used to be the CRO of MongoDB”
▶ Watch at ~36:00approx.
Gaurav MisraMentionsgorov was an early employee at snap where he led the design engineering team
“We had Gar from captions on the podcast”
▶ Watch at ~1:08:30approx.

Resources · 47

CursorMentionssoftware · Anysphere
“is basically the main competitor to cursor with over 1 million users 4 months in.”
▶ Watch at ~0:30approx.
YouTubeMentionswebsite · Google
“don't forget to subscribe and follow it in your favorite podcasting app or YouTube.”
▶ Watch at ~1:30approx.
Lenny's NewsletterCoinednewsletter · Lenny Rachitsky
“Check it out at lennisnewsletter.com.”
▶ Watch at ~1:30approx.
SuperhumanRecommendssoftware · Superhuman
“you get a year free of Perplexity Pro, Notion Plus, Linear, Granola, and Superhum.”
▶ Watch at ~1:30approx.
GranolaRecommendssoftware · Granola
“you get a year free of Perplexity Pro, Notion Plus, Linear, Granola, and Superhum.”
▶ Watch at ~1:30approx.
LinearRecommendssoftware · Linear
“you get a year free of Perplexity Pro, Notion Plus, Linear, Granola, and Superhum.”
▶ Watch at ~1:30approx.
Notion PlusRecommendssoftware · Notion
“you get a year free of Perplexity Pro, Notion Plus, Linear, Granola, and Superhum.”
▶ Watch at ~1:30approx.
Perplexity ProRecommendssoftware · Perplexity AI
“you get a year free of Perplexity Pro, Notion Plus, Linear, Granola, and Superhum.”
▶ Watch at ~1:30approx.
TwitterUseswebsite · Twitter
“Thank you to everyone on LinkedIn and Twitter and my newsletter community for suggesting great questions”
▶ Watch at ~1:30approx.
LinkedInUseswebsite · LinkedIn
“Thank you to everyone on LinkedIn and Twitter and my newsletter community for suggesting great questions”
▶ Watch at ~1:30approx.
ChatGPTMentionssoftware · OpenAI
“Chat GPT was not out four years ago.”
▶ Watch at ~4:30approx.
MetaMentionscompany
“My co-founder who I've known since middle school uh worked on ARVR at Meta.”
▶ Watch at ~4:30approx.
GitHub CopilotUsessoftware · GitHub
“And at that time we were early adopters of GitHub copilot”
▶ Watch at ~6:00approx.
CodeiumCoinedcompany · Varun Mohan and Douglas
“we vertical integrated and actually took our infrastructure, our inference infrastructure to go out and build Kodium”
▶ Watch at ~6:00approx.
GoogleMentionscompany · Google
“build the next great apps like uh Google, like Amazon.”
▶ Watch at ~6:00approx.
AmazonMentionscompany · Jeff Bezos
“build the next great apps like uh Google, like Amazon.”
▶ Watch at ~6:00approx.
EmacsUsessoftware · Richard Stallman and Guy Steele
“That meant VS Code, Jet Brains, Eclipse, Visual Studio, Vim, Emacs.”
▶ Watch at ~6:30approx.
EclipseUsessoftware · Eclipse Foundation
“That meant VS Code, Jet Brains, Eclipse, Visual Studio, Vim, Emacs.”
▶ Watch at ~6:30approx.
Visual StudioUsessoftware · Microsoft
“That meant VS Code, Jet Brains, Eclipse, Visual Studio, Vim, Emacs.”
▶ Watch at ~6:30approx.
VimUsessoftware · Bram Moolenaar
“That meant VS Code, Jet Brains, Eclipse, Visual Studio, Vim, Emacs.”
▶ Watch at ~6:30approx.
JPMorgan ChaseMentionscompany
“we built out the kind of this enterprise motion to to work with these large companies like Dell, JP Morgan, Chase.”
▶ Watch at ~7:00approx.
DellMentionscompany
“we built out the kind of this enterprise motion to to work with these large companies like Dell, JP Morgan, Chase.”
▶ Watch at ~7:00approx.
Visual Studio CodeUsessoftware · Microsoft
“we decided to go out and fork VS Code and build our own ID”
▶ Watch at ~7:30approx.
WindsurfCoinedsoftware · Codeium
“because of that, we decided to go out and fork VS Code and build our own ID with some of these new agentic capabilities.”
▶ Watch at ~7:30approx.
Fall in Love with the Problem, Not the SolutionMentionsbook · Uri Levine
“His book is called this. Fall in love with the problem, not the solution.”
▶ Watch at ~12:30approx.
IntelliJ IDEAMentionssoftware · JetBrains
“they might use the Jetrains family of ids that like Intelligj”
▶ Watch at ~16:00approx.
MITMentionscompany
“So I I went to MIT, a lot of us at the at the company went to MIT together”
▶ Watch at ~19:30approx.
JuliaUsessoftware
“I took a parallel computing class in Julia”
▶ Watch at ~21:00approx.
CodaUsessoftware · Coda
“I personally use KOD every single day to manage my podcast and also to manage my community.”
▶ Watch at ~34:00approx.
MongoDBMentionssoftware · MongoDB, Inc.
“Carlos Delator who used to be the CRO of MongoDB”
▶ Watch at ~36:00approx.
UberMentionscompany · Uber
“say at Airbnb or Uber like it understanding what the hell you have and how it works”
▶ Watch at ~38:30approx.
BoltMentionssoftware · StackBlitz
“we had the founders of bolt and lovable on the podcast.”
▶ Watch at ~38:30approx.
LovableMentionssoftware · Lovable
“we had the founders of bolt and lovable on the podcast.”
▶ Watch at ~38:30approx.
FedRAMPMentionsother
“we have fed RAM compliance, right? We have which means we can sell to like very large government entities.”
▶ Watch at ~40:30approx.
AirbnbMentionscompany · Brian Chesky, Joe Gebbia, and Nathan Blecharczyk
“what I would like an Airbnb for dogs website to kind of look like.”
▶ Watch at ~41:30approx.
ReactUsessoftware · Meta
“This is a very basic React project.”
▶ Watch at ~41:30approx.
ReplitMentionssoftware · Replit
“having just looked at lovable and bolt and replin and apps like that”
▶ Watch at ~48:30approx.
GPT-4oUsessoftware · OpenAI
“OpenAI's GPT40 is also good.”
▶ Watch at ~54:00approx.
Claude 3.5 SonnetUsessoftware · Anthropic
“I would say right now sonnet is a really really good planning model”
▶ Watch at ~54:00approx.
OpenAIUsescompany
“OpenAI's GPT40 is also good.”
▶ Watch at ~54:00approx.
AnthropicUsescompany
“And then what we try to do internally is run all the models necessary to do high quality retrieval for the agent.”
▶ Watch at ~54:00approx.
GitHubMentionssoftware · GitHub
“It's not like it is on GitHub.”
▶ Watch at ~57:00approx.
CaptionsMentionscompany · Gaurav Misra and his co-founder
“We had Gar from captions on the podcast”
▶ Watch at ~1:08:30approx.
MicrosoftMentionscompany · Bill Gates and Paul Allen
“The answer is Microsoft is going to be the be the winner in any software category”
▶ Watch at ~1:09:30approx.
WindsurfRecommendssoftware · Codeium
“download with surf, start coding, ask it to build things for you.”
▶ Watch at ~1:11:00approx.
GitHubUsessoftware · GitHub
“And then can you push straight to GitHub”
▶ Watch at ~1:12:30approx.
Lenny's PodcastCoinedpodcast · Lenny Rachitsky
“You can find all past episodes or learn more about the show at lennispodcast.com.”
▶ Watch at ~1:14:00approx.

Spot an error or want something removed? Request a correction or removal.

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 6

Hot Take08:30

Why Value in AI Accrues at the Application Layer, Not Infrastructure

Varun explains that their original bet on heterogeneous model architectures collapsed once everything converged on the transformer by mid-2023, invalidating the infrastructure thesis. He argues the application layer is the deep place to differentiate because there is effectively no ceiling on how much better developer workflows and UX can get.

The first thing a founder believes they should work on is unlikely to be the right thing.
They initially assumed many model architectures (CNNs, RNNs, LSTMs) would need an abstraction layer.
By mid-2023 everything looked like a transformer, so the infrastructure-value hypothesis was wrong.
Value accrues at the application layer, where there is effectively no ceiling on improving developer workflows.

“It's very unlikely the first thing that you believe you should go work on uh is going to be the right thing which is like…”
Varun Mohan · 08:30

#ai-strategy#value-capture#startups

Hot Take21:30

Agency Is the Most Undervalued Skill in the AI Era

Beyond problem-solving and systems knowledge, Varun singles out agency as the skill society and schooling most undervalue. School gives people well-defined paths and tells them exactly what to do; startups instead need people who want to build something without being told, and Windsurf explicitly screens for high agency.

School conditions people to follow well-defined paths and do exactly what they're told.
Agency, the drive to build something without instruction, is undervalued in the last decade.
Windsurf explicitly hires for high agency because a startup dies if it doesn't innovate.
Most large-company software engineering interviews don't select for agency.

“we look for people that are really high agency because we just recognize that by default if we don't innovate and do crazy things, we're…”
Varun Mohan · 22:00

#hiring#skills#agency#career

Hot Take23:00

The 'Dehydrated Company': Only Hire When You're Underwater

Varun explains their hiring philosophy: only hire for a function when the team is genuinely underwater, because adding people you don't need creates politics as they manufacture low-value work to justify themselves. He wants the company to stay 'dehydrated,' with each hire a little water added only when people are raising their hands saying they're drowning.

The goal is to be the smallest company that can satisfy their ambitions, not lean for its own sake.
Only hire for a role when that function is actually underwater.
Unneeded hires manufacture unimportant work and lobby the org that it matters, creating politics.
Staying understaffed is a forcing function for ruthless prioritization.

“I wanted the company to almost be like this dehydrated entity right and every hire is is like a little bit of water, right? And…”
Varun Mohan · 25:30

#hiring#org-design#startups

Hot Take31:00

'I Work Smart' Is a Hiring Dog Whistle

Varun says when he asks candidates how hard they're willing to work and they answer that they 'work smart,' it's a red flag. In a company that already has many smart people who also work hard, someone who only works smart will pull the team down, since a company is a massive group project where the cost of a coasting teammate is felt by everyone around them.

Windsurf tells candidates up front it's a very competitive, hard-working environment, not a chill one.
Answering 'I work smart' when asked how hard you'll work is a dog whistle.
The real cost of a coasting person is the signal it sends to teammates about the internal bar.
If four of five teammates don't care, the fifth stops caring too.

“one of the biggest dog whistles I sort of hear is when I ask people like how hard are you willing to work? Some people…”
Varun Mohan · 31:30

#hiring#culture#work-ethic

Hot Take50:30

Verticalized Niche Software Will Get Competed Down

Varun predicts that highly verticalized niche products, like sales, legal, or finance software, will be competed down because domain specialists can now build the exact tools they want. Those vertical vendors historically survived by bundling a kitchen sink of features when each customer only needed 10%, and companies couldn't build or maintain their own; that constraint has now disappeared.

Very verticalized niche products (sales, legal, finance) are most at risk of being competed down.
Vertical vendors won because customers couldn't build or maintain custom software for their needed 10% of features.
Domain specialists can now build the tools they always wanted in minutes, often more customized.
Off-the-shelf software often frustrates buyers with missing integrations; custom builds fix that.

“these very very verticalized niche products I think are going to get they're going to get competed down a ton.”
Varun Mohan · 51:00

“the domain specialists now have access to build the tools that they ultimately wanted”
Varun Mohan · 52:00

#software-industry#ai-disruption#saas

Hot Take1:06:30

Cannibalize Your Own Product Every 6 to 12 Months

Varun's most counterintuitive lesson is the tension between visible short-term wins and the longer, riskier bets that actually determine survival. He tells the company they should cannibalize their own product every 6-12 months so it makes the existing version look silly, warning that being first with the agentic IDE will depreciate quickly unless they keep re-proving themselves in ways users aren't even asking for.

The company works on 3-9 month bets internally even as the public celebrates weekly incremental wins.
The goal: cannibalize the existing product every 6-12 months so it looks silly by comparison.
Being first with the agentic IDE depreciates fast without continuous reinvention.
Incremental feature requests feel safe but are table stakes, not why they'll win.
You still can't ignore users, since they're the reason you exist.

“we should be cannibalizing the existing state of our product every 6 to 12 months. Every 6 to 12 months it should make our existing…”
Varun Mohan · 1:07:00

#product-strategy#innovation#roadmap

Explainer· 2

Explainer17:30

How Engineering Changes When AI Writes 90% of the Code

Varun breaks an engineer's work into three buckets: what to solve, how to solve it, and solving it. He argues AI will handle nearly all of the 'solving it,' and increasingly the 'how' as codebase understanding improves, pushing engineers toward the highest-value work of choosing which business problems and capabilities actually matter.

Engineering work splits into: what to solve, how to solve it, and solving it.
AI will handle the vast majority, if not all, of 'solving it.'
Deep codebase and organizational understanding will let AI increasingly handle 'how to solve it' too.
Engineering shifts toward prioritizing the most important business problems and technical decisions.

“what should I solve for how should I solve it and then solving it”
Varun Mohan · 17:30

#future-of-work#engineering#ai

Explainer1:04:00

Why 90% AI-Written Code Means Hiring More Engineers, Not Fewer

Varun uses Amdahl's law to explain why AI writing 90% of code doesn't make engineers 10x productive: writing code is only one part of the job alongside reviewing, testing, debugging, and deploying, so even eliminating coding time yields maybe 30-40% gains. More importantly, when technology gets cheaper to build, its ROI rises, so high-ceiling companies like JP Morgan should invest more and hire more engineers.

AI writing 90% of code doesn't make engineers 10x productive; they also review, test, debug, design, and deploy.
Amdahl's law: parallelizing one task to near-zero only shifts the total by that task's share (e.g. 100 to 73 units).
They see roughly 30-40% productivity improvements, not 10x.
Cheaper technology raises the ROI of building it, so high-ceiling companies invest and hire more.
JP Morgan Chase has a ~$17B software budget and 50,000+ engineers, each now more productive.

“if AI is writing over 90% of the code, that doesn't mean engineers are 10x as productive, right? Engineers spend more time than just writing…”
Varun Mohan · 1:04:00

“the ROI of building technology has actually gone up.”
Varun Mohan · 1:05:30

#productivity#economics#engineering#ai

Story· 3

Story04:30

From GPU Infra to Kodium to Windsurf: The Two Pivots

Varun recounts starting the company nearly four years ago building GPU virtualization and compiler software, before ChatGPT existed. When generative models got good in mid-2022, the team realized their bespoke-model infrastructure was losing value, so they vertically integrated into an AI coding product (Kodium), then forked VS Code six months before the interview to build the Windsurf IDE.

Started ~4 years ago in GPU virtualization; Varun came from autonomous vehicles, co-founder from AR/VR at Meta.
By mid-2022 they had a couple million in revenue, ~10,000 GPUs managed, 8 people, free-cash-flow positive.
Realized that if everyone runs the same generative models, an infrastructure company can't differentiate.
Vertically integrated their inference infra into Kodium, an AI coding product offered free across all IDEs.
Forked VS Code ~6 months ago because existing IDEs capped the agentic capabilities they could ship.

“And in a world in which everyone was going to run generative AI models, why would an infrastructure company be a differentiator?”
Varun Mohan · 06:00

#startups#pivots#ai-coding#company-history

Story13:30

Forking VS Code Tripled Their Acceptance Rate With the Same Models

Varun describes why they built a custom IDE instead of staying inside VS Code: they believed AI would write over 90% of software, shifting the developer's job toward reviewing code. When they ported their autocomplete/refactor UI to Windsurf, acceptance rate tripled using the identical ML models, proving the surface matters as much as the technology.

They believed AI would write well over 90% of software, making code review the developer's main job.
VS Code's limited APIs forced hacks like dynamically generating images next to the cursor.
Porting the same models to Windsurf tripled the acceptance rate.
Lesson: great technology delivers little value if the surface and interface can't showcase it.

“immediately by porting over to Windsorf our acceptance rate tripled same ML models it just tripled”
Varun Mohan · 15:00

#product#ide#ux#ai-coding

Story49:30

Their Sales Team Built Their Own Software and Saved $500K in SaaS

When Windsurf launched, they tasked everyone in the company, including the go-to-market and sales teams, with building an app using it. The result: over half a million dollars of SaaS products they were going to buy were replaced by internally built tools, including a partner portal built by a head of partnerships who had never written software.

Everyone at the company, including sales, was tasked with building an app with Windsurf.
They saved over half a million dollars in SaaS products they would have bought.
The head of partnerships built his own partner portal instead of buying one, despite never coding before.
The company built ways to deploy these internal apps securely.

“we've saved over half a million dollars of of SAS products we were going to buy because our go to market team has now built…”
Varun Mohan · 50:00

#ai-coding#productivity#internal-tools

Takeaway· 3

Takeaway10:30

Never Fall in Love With Your Ideas

Reflecting on pivoting away from a profitable business after raising $28M, Varun argues founders should assume many of their hypotheses are wrong and build a truth-seeking organization where morale doesn't collapse after a wrong bet. He frames each year's growing engineering output as a fresh chance to test new hypotheses.

They pivoted overnight from a profitable infra business, having already raised $28M, because focus is everything.
Assume a lot of your hypotheses will be wrong, then do concentrated work to validate them.
A truth-seeking org keeps morale up even after a wrong decision, preserving optionality.
Each year the company's total engineering output exceeds all prior years combined, a new lease on life to test hypotheses.

“it's awesome when you have a great idea, but you should never be too in love with your ideas”
Varun Mohan · 11:30

#decision-making#startups#culture

Takeaway26:30

Get an A+ on the One Class That Matters, an F on the Rest

Varun contrasts school with company-building: school optimizes total GPA, but companies win by doing one thing exceptionally and failing at nine others. Keeping the team underwater forces this ruthless prioritization, since a team will never ask to work on something unimportant when they can only take on one thing.

You don't win by doing ten things well; you win by doing one thing really well.
School optimizes total GPA; a company just needs an A+ on the one class that matters.
An F in other classes means deprioritizing what doesn't matter, not doing anything illegal.
Fewer people is a forcing function for organizational prioritization.

“In school, you optimize for your total GPA, but for companies, I just need to get an A+ on the one class that matters and…”
Varun Mohan · 27:00

#prioritization#focus#startups

Takeaway35:30

Why They Bet on Enterprise Sales Early (When Most Technical Founders Won't)

Varun pushes back on the technical-founder instinct that sales is negative and everything should be product-led growth. As a GPU-infra company he couldn't scale a sales function because he couldn't sell it himself, but once large enterprises reached out to Kodium and they ran tens of concurrent pilots, they hired a VP of Sales by end of 2023 and scaled an 80+ person go-to-market team.

Technical founders often wrongly see enterprise sales as negative and insist on pure PLG.
As an infra company they never hired sales because the founder couldn't yet scale the motion himself.
Large enterprises reached out to Kodium, driving tens of concurrent enterprise pilots by mid-2023.
They hired a VP of Sales by end of 2023; go-to-market is now 80+ people.
You can't sell to the Fortune 500 purely by swiping a credit card.

“I think this is like a interesting thing that technical founders sometimes don't really like. They think sales is like a very negative sort of…”
Varun Mohan · 36:00

“if you want to sell to the Fortune 500, it is very hard to do that purely by swiping a credit card.”
Varun Mohan · 37:30

#sales#enterprise#go-to-market

Key ideas

Frameworks in this episode

Agency Over Title: The Force-Multiplier Mindset

Build Only Where You Can Be Best (Model Sourcing)

Cannibalize Your Own Product Every 6-12 Months

Hills-and-Valleys: Getting Value From Probabilistic AI Tools

Irrationally Optimistic, Uncompromisingly Realistic

The Dehydrated Company Hiring Model

People & resources mentioned

People · 7

Resources · 47

Insights & moments

◆Hot Take· 6

Why Value in AI Accrues at the Application Layer, Not Infrastructure

Agency Is the Most Undervalued Skill in the AI Era

The 'Dehydrated Company': Only Hire When You're Underwater

'I Work Smart' Is a Hiring Dog Whistle

Verticalized Niche Software Will Get Competed Down

Cannibalize Your Own Product Every 6 to 12 Months

✶Explainer· 2

How Engineering Changes When AI Writes 90% of the Code

Why 90% AI-Written Code Means Hiring More Engineers, Not Fewer

❝Story· 3

From GPU Infra to Kodium to Windsurf: The Two Pivots

Forking VS Code Tripled Their Acceptance Rate With the Same Models

Their Sales Team Built Their Own Software and Saved $500K in SaaS

▲Takeaway· 3

Never Fall in Love With Your Ideas

Get an A+ on the One Class That Matters, an F on the Rest

Why They Bet on Enterprise Sales Early (When Most Technical Founders Won't)

Hot Take· 6

Explainer· 2

Story· 3

Takeaway· 3