Codeium AI

Codeium AI built AI coding assistants for existing development environments before launching Windsurf Editor in November 2024 and renaming the company Windsurf in April 2025. Its work expanded from autocomplete into Cascade, a system combining codebase retrieval, developer activity, planning, and code execution to make changes across files. Cognition signed an agreement to acquire Windsurf in July 2025; in June 2026, it introduced Devin Desktop as the next generation of Windsurf.[2][1][10][11]

The coding products grew out of an earlier GPU infrastructure business. That history supplies context for Codeium’s existing ML infrastructure at launch and its later decision to retain inference and distributed systems expertise internally.[4][3][1]

The company began as Exafunction in 2021, optimizing GPU workloads for deep learning. Its July 2024 retrospective described inefficient GPU use in workloads such as autonomous vehicle simulation and a business that took a share of customers’ savings. The company reported reducing customers’ GPU bills by almost 97%, raising a Series A, and exceeding $1 million in annual recurring revenue within a year. These figures concerned the infrastructure business.[2][4]

In the same retrospective, Codeium said the emergence of generative AI in summer 2022 changed its assessment of the infrastructure market: providers running a handful of standard model architectures might optimize inference themselves, while competition could commoditize that work. It recalled pivoting entirely over a weekend to a vertically integrated generative AI application using its GPU software expertise. The company informed investors that its thesis was changing and customers that its existing products would be sunset.[4]

Codeium launched its beta on October 28, 2022: a free code-generation tool running on its existing ML infrastructure, accompanied by a Discord community for developer feedback. The announcement joined two priorities—models paired with scalable infrastructure and widespread developer testing—and explicitly proposed offering tools free so developers could try them and give feedback.[3]

By January 2024, Codeium offered autocomplete, chat, and context awareness using proprietary code-biased language models. It supported more than 70 languages and over 40 IDEs, including Visual Studio Code, JetBrains, Eclipse, and Jupyter Notebooks. Enterprise options included SaaS, self-hosting in a VPC or on premises, integration with existing source code management systems, and optional local model fine-tuning. The company also said it was working with Atlassian to enhance the developer experience.[6]

In the interview published December 13, 2024, Varun Mohan explained that enterprise developers used varied IDEs and source control platforms, making compatibility important. Anshul Ramachandran said Windsurf’s launch did not mean abandoning extensions: enterprise customers continued to use JetBrains and Eclipse. Mohan described shared infrastructure serving individuals and enterprises, with indexing and custom-model serving maintained internally. Ramachandran pointed to the containerized system as an advantage for enterprise deployment requirements.[1]

Mohan, Ramachandran, and other engineers participated directly in sales before hiring a sales leader. Mohan recalled having sold to roughly 30–40 customers by that point; Ramachandran described hundreds of deal cycles that taught the team which messaging and deployment work mattered. Codeium later employed deployed engineers who partnered with account executives, helped customers use the product, and brought customer learning back into product development.[1]

The move into an owned editor gave Codeium more control over developer activity and the interface for reviewing AI changes. Mohan recalled that displaying Supercomplete’s suggested refactors inside VS Code had required dynamically generating PNGs because the extension API did not expose the needed presentation capability. Ramachandran described a broader requirement: observing navigation, edits, and terminal activity together could help the system infer what a developer was trying to accomplish.[1]

In his November 13, 2024 launch account, Kevin Hou credited the team with a September breakthrough combining codebase understanding, tool use, and coordination with the user, which it called Cascade. Hou described using typing and navigation to infer intent, suggest multi-cursor edits through Supercomplete, identify the next file to edit, and continue a task. In the December interview, Mohan explained that Cascade used both human and AI action trajectories to propose changes and execute code.[5][1]

The December 2024 architecture divided responsibilities across models and systems. Codeium’s own models powered autocomplete and Supercomplete, which ran on every keystroke. Proprietary systems handled retrieval and applying plans to the codebase, while third-party models supplied high-level planning. Ramachandran acknowledged that the team’s earlier preference for first-party models needed qualification: improvements from external model providers had helped make Cascade and Windsurf possible.[1]

Codeium evaluated Cascade against incomplete code rather than requiring a fully stated task. Mohan described taking open-source commits with tests, stripping some changes, withholding the commit message, and checking whether retrieval, planning, and iterative execution could bring the code back to a passing state. Ramachandran described a complementary retrieval evaluation: assessing approximately 50 retrieved snippets and using files edited together in historical commits to construct relevant-file sets. This tested whether retrieval captured distributed knowledge across a codebase.[1]

Mohan and Ramachandran also described improving Supercomplete with synthetic data and user-preference signals. Acceptance alone was an incomplete signal: subsequent edits to an accepted suggestion could reveal what the developer actually wanted, and this information was available through extensions as well as Windsurf. Other product work included Cortex, announced in August 2024 as a reasoning engine, and Forge, an AI-assisted code review tool. Windsurf’s SWE-1 announcement later described its full-size model as built for advanced reasoning and tool use.[1][7][8]

On January 30, 2024, Codeium announced it had raised a $65 million Series B, led by Kleiner Perkins with Greenoaks and General Catalyst participating. On August 29, it announced secured $150 million Series C funding led by General Catalyst, with continued support from Kleiner Perkins and Greenoaks. The latter announcement put total funding at $243 million and the valuation at $1.25 billion.[6][7]

Codeium reported more than 300,000 individual users in January 2024. Its August announcement reported over 700,000 active developers, a team of 80, eight-figure enterprise ARR, enterprise ARR growth above 500% since early 2024, and more than 100 billion tokens processed daily. The December Latent Space introduction reported over one million developers using its products; in the accompanying interview, Mohan put extension usage above 800,000 developers. These counts cover different product scopes and activity measures.[6][7][1]

On April 4, 2025, Codeium announced its renaming to Windsurf, making the editor’s name the company identity. It also announced plans to rename the original Codeium Extensions “Windsurf Plugins” and continue developing the editor alongside new products.[2]

On July 11, 2025, TechCrunch reported that Google DeepMind was hiring CEO Varun Mohan, co-founder Douglas Chen, and some researchers; Google confirmed the hiring of Windsurf’s leaders. Google obtained a nonexclusive license to certain technology without taking a stake or control. TechCrunch attributed a $2.4 billion licensing-and-hiring price to Bloomberg and the collapse of a proposed $3 billion OpenAI acquisition to The Verge. It also reported Jeff Wang becoming interim CEO, with most of Windsurf’s 250-person team remaining.[9]

Three days later, Scott Wu announced that Cognition had signed a definitive agreement to acquire Windsurf’s IP, product, trademark, brand, and business, and welcome its staff. Wu reported $82 million ARR, more than 350 enterprise customers, hundreds of thousands of daily active users, and enterprise ARR doubling quarter over quarter. He said he and Jeff Wang had structured terms giving all Windsurf employees financial participation, waived vesting cliffs, and fully accelerated vesting for their work to date.[10]

On June 2, 2026, Scott Wu and Jeff Wang introduced Devin Desktop as the next generation of Windsurf, following Windsurf 2.0 earlier that year. Devin Desktop made Agent Command Center the default IDE surface for managing local and cloud agents, pull requests, and context. Spaces let related agents share context, while support for the Agent Client Protocol allowed compatible agents to run alongside Devin. The product retained a full IDE compatible with Windsurf, including editor workflows for inspecting and editing code.[11]