Codeium

Varun Mohan and his co-founder (known since middle school) started Codeium nearly four years ago with a different vision. Coming from backgrounds in autonomous vehicles and AR/VR at Meta, they believed deep learning would transform industries. They built GPU virtualization and compiler software that allowed complex machine learning applications to run efficiently without expensive GPUs. By mid-2022, the business was generating "a couple million in revenue," they had 8 people on the team, and were managing over 10,000 GPUs while remaining cash flow positive. By every conventional measure, it was working.

But Varun saw the ground shifting beneath them. "When these generative models started to get very good, we felt a lot of what we built was not as valuable," he recalls. The insight was brutal: Why would companies train custom sentiment classifiers when they could ask GPT? Why would an infrastructure company differentiate when everyone would eventually run the same generative AI infrastructure? "In a world in which everyone was going to run generative AI models, why would an infrastructure company be a differentiator?"

The team made a bet-the-company pivot. They'd raised $28 million, but instead of optimizing the business that was working, they announced a complete strategic shift: "We're going to pivot entirely from this... overnight." Early adopters of GitHub Copilot, they recognized the coding space was about to be "tremendously disrupted." They took their infrastructure expertise and built Codium, initially just an autocomplete product—as users typed, it would complete the next few lines of code. Crucially, they offered it free across every IDE developers used: VS Code, JetBrains, Eclipse, Visual Studio, Vim, and Emacs.

Their infrastructure background gave them an edge—they could optimize these workloads in ways competitors couldn't. But within months, large enterprises came calling. Dell, JPMorgan Chase—companies with massive, proprietary codebases—wanted more than autocomplete. They wanted secure, personalized offerings trained on their private code. Varun and a handful of founders began selling, running "tens of pilots concurrently with large enterprises." By the end of 2023, they'd hired a VP of Sales (Carlos Delatory, former CRO of MongoDB) and began scaling an enterprise motion.

Six months before this interview, Varun realized they'd hit a ceiling. VS Code's API limitations prevented them from building the UI experiences their AI capabilities deserved. When they built an inline refactoring feature in Windsurf with custom UI, "our acceptance rate tripled. Same ML models, it just tripled." The lesson was clear: technology matters, but user interface matters more.

They made another contrarian bet—fork VS Code and build their own IDE, Windsurf, with agentic capabilities baked in. In a market where engineers fiercely guard their tools, asking them to switch seemed insane. But Windsurf launched "a bit over four months ago" and reached "over a million developers" with "hundreds of thousands of monthly active users." The product worked because it solved real problems: reviewing AI-generated code (not writing it) became the primary developer task, requiring new interaction patterns.

Their enterprise sales bet also paid off. While many AI startups dismiss sales as uncool, Varun saw it differently. They now have "over 80 people" in go-to-market—nearly half their 160-person engineering organization. They invested in FedRAMP compliance to sell to government, built hybrid deployment modes to keep enterprise code on-premise, and engineered codebase understanding for companies like Dell with 100+ million line codebases. They can understand massive codebases in parallel across thousands of GPUs, rank relevant code snippets, and suggest changes without breaking existing systems—a capability Cursor and Copilot didn't have.

Windsurf is now competing head-to-head with Cursor as the leading AI IDE. Varun is explicit about what they won't do: they won't abandon support for JetBrains and other IDEs (which host 70-80% of Java developers) just to grow Windsurf's user base. Their philosophy is to "meet developers where they are." They're investing heavily in what comes next—Varun believes engineers will spend 90%+ of their time reviewing AI-generated code, not writing it, and coding will shift from "how do I implement this?" to "what should we build?" The company that wins will be the one that deeply understands codebases, supports multiple IDEs, and handles enterprise security better than anyone else.