Cursor raises $2 billion to scale AI coding infrastructure

Cursor raises $2 billion in a Series E round at a $50 billion valuation. Thrive Capital and Andreessen Horowitz are reported co-leads, and Nvidia participates. This valuation follows a period where the company value sat near $29.3 billion. As of February 2026, the company reported $2 billion in annual recurring revenue. The founders, Michael Truell, Sualeh Asif, Aman Sanger, and Arvid Lunnemark, all graduated from MIT in their mid-20s. The tool attracts users ranging from independent developers to executives like Patrick Collison of Stripe. Jensen Huang, the CEO of Nvidia, has praised the company in public. The amount signals what institutional capital believes about enterprise AI value. This capital concentration shows that institutional investors view coding assistants as infrastructure. The scale of the round places Cursor among a tiny list of private AI companies. Software engineering teams evaluate these tools as line-item budget decisions. The company’s revenue grew from $100 million to $200 million as the developer community embraced the product.

Composer 2.5 and Agentic Intelligence

Composer 2.5 uses Moonshot AI’s Kimi K2.5 checkpoint as its base. Cursor dedicated 85% of its training compute to continued pretraining and reinforcement learning. The model targets the agent loop, specifically long-horizon tasks and multi-step instructions. It reaches 79.8% on SWE-Bench Multilingual and 63.2% on CursorBench v3.1. Cursor uses targeted textual feedback during reinforcement learning to correct model missteps at the exact point in a trajectory. They also used synthetic tasks like feature deletion to improve performance. One training instance saw the model reverse-engineer a Python type-checking cache to recover a deleted function signature. Another instance saw it decompile Java bytecode to reconstruct a third-party API. The system employs a distributed variant of the Muon optimizer running Newton-Schulz orthogonalisation asynchronously across shards. I find the model most useful for refactors that span dozens of files or dependency upgrades where the agent needs to take hundreds of actions. For engineers running parallel agents via git worktrees, the cost-per-task delta between Composer 2.5 and frontier alternatives like Claude Opus 4.7 provides the single most compelling argument for switching to this model. Cursor is also training a much larger successor model from scratch alongside SpaceXAI on Colossus 2 using 10x more compute. This follows the SpaceX-Cursor option deal signed in April.

Managing Enterprise Migration and Pricing

Organizations migrating to Cursor must treat the tool as infrastructure spend. The June 2025 pricing transition replaced fixed fast request allotments with usage-based credit pools. This change caused significant community backlash. Developers reported unexpected charges when the auto-mode selected expensive models for complex tasks. In April 2026, vulnerabilities in the Cursor environment allowed attackers to execute arbitrary code and access sensitive credentials. I suggest that teams forecast usage per engineer to avoid budget surprises. You already know that switching IDEs usually involves a learning curve.

The company has experienced several pricing incidents. In June 2025, the Pro plan changed, and users found their usage was depleted in a single day. In late 2025, rate limit changes broke workflows for power users. These incidents led to a developer exodus toward alternatives like Claude Code and Windsurf. Claude Code provides terminal-based agentic development, while GitHub Copilot remains the leader for broad IDE integration. GitHub Copilot Pro costs $10 per month and includes 1,500 AI credits. Cursor’s Tab feature provides multi-line edits and smart rewrites that predict the next cursor move.

Enterprise plans include pooled usage, invoice or PO billing, and SCIM seat management.

Will the current pricing instability drive more developers to local models like Cline?