Varun Mohan

Varun Mohan is a cofounder of Exafunction, the GPU virtualization company that became the foundation for Codeium and later the Windsurf AI editor. His consequential decisions included using the team's infrastructure to make code completion free and building an editor when extension APIs constrained the experience. He joined Google DeepMind in July 2025; Google's May 19, 2026 developer announcement, which he coauthored with Logan Kilpatrick, identified him as Director, Software Engineering.[3][1][2][4][5]

Before those agentic coding tools, Mohan worked on infrastructure for deep learning at Nuro. That earlier work provides concrete context for the serving technology his team subsequently used to build Codeium.[1]

Mohan studied computer science at MIT and worked at Nuro, where he ultimately became tech lead manager for autonomy. He described managing large-scale offline deep-learning infrastructure there and finding that even sophisticated companies struggled to build and maintain it. He and Douglas Chen founded Exafunction in 2021 to make serving deep-learning workloads easier.[1][3]

At Exafunction, the team decoupled GPU resources from workloads to improve hardware utilization. Mohan explained that assigning a GPU to a single container could leave much of its compute idle. Alongside keeping models running, the team built compiler technology to execute them more efficiently, including operator fusion: combining operations so their joint execution took less time than running each separately.[1]

Mohan recalled that the team initially doubted how much GitHub Copilot could help with its CUDA and C++ work. The turning point was watching it generate difficult template code in a roughly 500-line file that then compiled. In his March 2023 interview, he described the realization in simple terms: “This is not a toy anymore.” According to investor Leigh Marie Braswell's account, the founders used Exafunction's infrastructure to build Codeium directly after their Series A.[1][3]

Codeium's team already had serving infrastructure when it began building the coding assistant. It bootstrapped internally with open models, then shipped its own model with fill-in-the-middle capabilities. Mohan explained the serving tradeoff: autoregressive generation produces tokens sequentially, so longer outputs take longer; batching requests improves GPU utilization, but must preserve the low latency an assistant needs when responding to keystrokes.[1]

Mohan also distinguished training-compute optimality from serving practicality. A larger model might have improved Codeium's quality, he said, but increased latency and made fitting the available hardware harder. His argument was to consider deployment constraints before training, rather than discover afterward that a model was difficult to serve.[1]

Efficient serving made free code completion feasible. In 2023, Mohan explained that the aim was to build a user base that would stay as the team added more products, with paid and enterprise offerings anticipated. He qualified the origin of that decision: giving every piece of software away had not been the initial goal; the team realized it could offer completion free by cutting costs and optimizing heavily.[1]

Mohan framed AI product design around latency, quality and correctability—how easily users can recover useful work from imperfect output. His example was a ten-file change in which two files were modified consistently and eight were not: reviewing and repairing that output could become harder than its impressive demonstration suggested. He recommended that most founders begin with an existing API and concentrate on these product constraints, unless their task required substantial custom training.[1]

By the December 2024 interview, Mohan described a further decision: building Windsurf. Codeium had first supported developers in their existing IDEs, including environments important to enterprise customers. The team then found that extension constraints limited its agentic experience. Mohan gave a specific example: Supercomplete could suggest refactors, but VS Code lacked the desired display API, so the team dynamically generated PNGs to show its suggestions.[2]

Anshul Ramachandran explained the combination behind the editor: codebase retrieval, access to tools and awareness of the developer's trajectory—files opened, edits made and terminal commands attempted. Mohan described Cascade, Windsurf's agentic component, as using both human and AI trajectories to propose changes and execute code. He also emphasized making generated changes easy to review, since an unusable review interface would undermine the underlying capability.[2]

Mohan described evaluating Cascade with open-source commits and associated tests. The team removed changes to leave code incomplete and withheld the commit message, then tested whether retrieval, planning and iterative execution could produce code that passed. This approximated a developer who had not fully stated their intent. Ramachandran separately described retrieval evaluations that sought many relevant snippets rather than one isolated answer, using files changed together in historical commits to construct reference sets.[2]

The model strategy had also evolved. Mohan said keystroke-level autocomplete and Supercomplete used the team's own models, while proprietary systems handled retrieval and applying changes. High-level planning used third-party models. In discussions with Chen, he argued for buying undifferentiated tools while retaining competencies such as inference infrastructure where they could provide lasting differentiation. That preference was conditional: if the team could not keep up, it should buy rather than maintain the pretense of a useful internal capability.[2]

Mohan and Ramachandran also learned enterprise sales directly. Mohan recalled that the team had sold to approximately 30–40 customers before hiring a sales leader. Ramachandran described their continued involvement in deal cycles and deployment engineers' work alongside account executives. Mohan argued that a sales organization needed to understand the rapidly changing technology—including customers' questions about retrieval—so it could remain a useful partner.[2]

Mohan joined Google DeepMind in July 2025 alongside Chen and some Windsurf researchers. The following May, he and Kilpatrick coauthored Google's I/O developer announcement covering Antigravity and other agentic development tools. They presented the goal as making the shift from single-turn prompts to collaborative, persistent agents easier by reducing infrastructure friction.[4][5]