Cursor, the developer of a widely used artificial intelligence code editor, has closed a $2.3 billion later-stage investment at a $29.3 billion valuation.

Accel and Coatue led the Series D round. They were joined by Google LLC, Nvidia Corp., Thrive, Andreessen Horowitz and DST. Cursor disclosed in its funding announcement today that its annualized revenue recently topped $1 billion.

The company, which is incorporated as Anysphere Inc., has built a code editor that uses AI to help developers write software. It’s a modified version of Microsoft Corp.’s open-source VS Code editor. Cursor has extended it with large language models that can generate code and fix bugs in existing software.

One of the LLMs embedded in the editor is a custom model dubbed Composer. According to Cursor, it’s a mixture-of-experts algorithm that runs four times faster than LLMs with comparable output quality. It can complete many coding tasks in under 30 seconds.

An LLM is comprised of kernels, parallelized snippets of code that can run across a large number of graphics card cores at once. Developers usually write kernels with the help of CUDA libraries that abstract away some of the associated complexity. They contain prepackaged code that removes the need to write everything from scratch. 

Cursor says that it didn’t use any CUDA libraries while building Composer. The company implemented the model’s kernels using PTX, the low-level machine language in which Nvidia chips express computations. That approach helped Cursor achieve a more than threefold performance increase across some Composer’s components.

The company trained the model using a custom AI cluster. The cluster is powered partly by Ray, an open-source tool for training and running Python-based AI workloads. The technology was co-developed by Anysphere co-founder and Chief Technology Officer Philipp Moritz a few years before the company launched.

One of the contributors to Composer’s speed is that it stores information in a data format called MXFP8. The format takes up less space than many alternatives, which speeds up processing. Cursor customized its MXFP8 implementation for the Blackwell B200 chips it used to power Composer’s training workflow. 

Cursor rolled out Composer to its code editor last month as part of a release called Cursor 2.0. The upgrade also introduced an embedded browser that the editor can use to test the web applications it generates. According to Cursor, users can run multiple AI agents in parallel to speed up code generation.

The company disclosed today that its installed base includes several million developers. Cursor said Composer and its other custom AI model, a code autocompletion algorithm called Cursor Tabs, generate more code than “almost any other” LLM. Its editor also uses models from external providers such as Google, one of the contributors to today’s funding round.

Cursor will use the new capital to finance AI research initiatives.

Image: Cursor