Now that Cursor is part of SpaceX, should it build its own frontier coding models or keep relying on OpenAI, Anthropic and Google?

Introduction

Cursor started as an editor that let developers use the best models from OpenAI and Anthropic. It then built its own models: the Tab model (strengthened by the Supermaven acquisition), Composer in Cursor 2.0 in October 2025, and Composer 2 in March 2026, which was built on Moonshot AI's open-weight Kimi base and post-trained with reinforcement learning in Cursor's agent environment. In June 2026, SpaceX agreed to acquire Anysphere for a reported $60 billion in stock, and the deal was reported to have closed in August, placing Cursor inside SpaceX's AI organisation with access to very large compute.

That raises a strategic question: should Cursor now try to train its own frontier coding models from scratch, or remain mainly a product company on top of models from OpenAI, Anthropic and Google, which are also its competitors?

Key objectives for the strategy:

1. Keep Cursor the best place to write and ship code with AI.
2. Protect gross margin as usage grows.
3. Reduce dependence on suppliers who compete in the same market.
4. Use the new parent's resources without losing focus.

I'll clarify goals, analyse the market, define a vision and roadmap, prioritise initiatives, plan execution, set KPIs and manage risks.

Step 1

Clarify the Strategic Goals (3-4 minutes)

• Is the primary goal product quality, cost, or independence? I'll assume all three matter, but product quality wins any tie, because developers switch tools quickly when a rival is better.

Why it matters: a pure cost goal would push heavy routing to in-house models; a quality goal keeps frontier models central. Impact on approach: I'll favour a layered strategy rather than all-or-nothing.

• What compute and talent does the SpaceX relationship actually provide? I'll assume meaningful GPU capacity and a research team from the parent's AI division, but on a shared roadmap.

Why it matters: training frontier models costs billions; without dedicated compute it is not realistic. Impact on approach: phase investments behind proven compute access.

• How have OpenAI, Anthropic and Google behaved as suppliers? I'll assume they still sell API access to Cursor, but each is pushing its own coding agent (Codex, Claude Code, Gemini CLI and Antigravity), so terms and early access could tighten.

Strategic assumptions:

1. Users care about results, not model provenance, except enterprises that need transparency.
2. Cursor's unique asset is real usage data from millions of developer sessions and its agent environment.
3. Supplier risk is rising, not falling.