For Varun Mohan, Insights Are Always Depreciating

“Every single insight we have is a depreciating insight,” said Varun Mohan, CEO of Windsurf, in the Lightcone Podcast hosted by Y Combinator. “Technology moves fast. If we’re not continuing to have insights—and executing on them—we’re just slowly dying.”

That mindset helps explain how a team of eight engineers pivoted from GPU infrastructure to building one of the most widely used AI coding platforms. It also sheds light on why OpenAI is now reportedly acquiring Windsurf for $3 billion, in what would be its largest acquisition to date.

Windsurf began as Exofunction, a GPU virtualization company. In 2021, Mohan and his co-founder were helping enterprises manage over 10,000 GPUs. Revenue was solid, surpassing a couple of million dollars, and they had raised $28 million. The bet was that many companies would soon need to train their own deep learning pipelines.

But by mid-2022, that assumption began to unravel. “Our belief was that a wide range of companies would train their own deep learning models,” Mohan said. “But even an early version of GPT-3 could outperform them on tasks like sentiment classification. It became clear people wouldn’t need to train models like BERT anymore.”

The team responded by making a sharp and immediate pivot. Over one weekend, they abandoned their existing business and started building a developer tool. By Monday, they were working on what would become Codium, an autocomplete extension for VS Code. It shipped in two months.

The first version wasn’t great. “It was materially worse than GitHub Copilot. The only advantage was that it was free,” Mohan admitted. But Windsurf had one asset Copilot didn’t: its own inference stack. That allowed the team to quickly improve model performance, focusing on gaps Copilot couldn’t yet handle like completing the middle of code lines instead of just the end.

The bet worked. Developers started adopting the product via Hacker News, and enterprises soon followed—asking for secure, private deployments.

From the beginning, Windsurf’s approach to AI tooling was technical, infrastructure-heavy, and opinionated. While others centered around chat interfaces and basic autocomplete, Windsurf prioritized codebase comprehension, fast edits, and multi-language support across IDEs. “If we were going to work with companies like JP Morgan Chase, where half of developers use Java, we couldn’t just be a VS Code plugin,” Mohan said.

This pushed them to build support for JetBrains, Eclipse, Vim, and eventually fork VS Code itself. Within three months, they shipped their own IDE, Windsurf optimized for AI agents that could read, interpret, and make changes to massive codebases.

Today, Windsurf is used by over a million developers. Hundreds of thousands use it daily. Inside enterprises, it powers onboarding, internal tool development, and even production deployments. “Deploying software to our servers is now done via workflows built entirely in Windsurf,” said Mohan.

Their stack also deviates from prevailing AI trends. While many rely on vector search for retrieval-augmented generation (RAG), Windsurf uses a layered approach: abstract syntax tree (AST) parsing, keyword-based search, and real-time GPU ranking of code snippets. That system is built to ensure high precision and recall—essential when navigating codebases with over 100 million lines.

This technical depth extends to their evaluation strategy. Windsurf uses open-source commits, historical tests, and user telemetry to benchmark agent performance across retrieval, intent interpretation, and test pass rates. “We don’t strive for complexity,” Mohan said. “We strive for what works and we measure whether it’s actually working.”

One of the more surprising outcomes is how accessible the tool has become. At Windsurf, even non-technical staff are using it. A partnerships lead without formal coding experience has built internal apps that replaced multiple sales tools. “Before, that would’ve required looping in product and engineering teams. Now he builds it directly.”

Traditional developers may no longer be the only ones building software, Mohan noted. “I think everyone is going to be a builder.”

Windsurf, originally launched as Codeium in 2021, has become a leading player in the emerging field of vibe coding—where developers use AI prompts to generate code. According to PitchBook, the company has raised $243 million in venture funding. Bloomberg reported in April that OpenAI is in talks to acquire Windsurf for around $3 billion.

On the podcast, Mohan described how AI is set to democratize software creation. “Instead of downloading an app, people might just ask their AI assistant to build a custom tool tailored to their needs—one they can keep tweaking over time,” he said. “I can imagine a future like that where effectively everyone is building but people don’t know what they’re building is software.”

Still, the agent experience isn’t perfect. “If you give vague instructions, the agent can over-modify the code,” Mohan noted. “But if you structure the workflow well and commit often, the experience is fast and surgical.”

Internally, the company treats engineering as a research function. “Boilerplate has gone away,” said Mohan. “So we’re spending more time testing hypotheses, building evals, and shipping systems no one’s tried before.”

That R&D orientation, grounded in real product feedback and architectural control, is also what differentiates Windsurf in an increasingly commoditized AI tools market. “We had a product generating millions in revenue and had no churn,” Mohan said. “We still decided to kill it in a weekend, not because it wasn’t working, but because we didn’t know how to scale it.”

Where most teams bolt AI onto existing workflows, Windsurf treats AI as a native, evolving medium. Its core question isn’t just “how do we integrate models”—it’s “how should software development itself change when AI is in the loop?”

That’s the story OpenAI appears to be buying into: a team willing to rebuild the developer experience from scratch again and again, if necessary.