Cursor Enterprise Results: NVIDIA, Salesforce, Stripe

Cursor's customers page reports adoption at large scale across the enterprise. The four headline numbers below all come from Cursor's own published material, so read them as vendor-reported reach rather than independently audited counts.

The 93% preference and the 100-million-lines-a-day figure are the kind of claims worth flagging as Cursor-reported when you put them in front of a buyer. They are real published statements, not measurements you can reproduce from the outside.

NVIDIA CEO Jensen Huang has named Cursor directly in public remarks. Cursor cites him on its customers page; the quote below is attributed to him.

Per the same remarks, every one of NVIDIA's roughly 40,000 engineers is now assisted by AI, with productivity up "incredibly." NVIDIA runs Cursor inside its own guardrails: it uses hooks and other enterprise controls to keep development activity within its compliance boundaries.

Salesforce reports broad adoption and roughly 2x gains in both code quality and velocity. Cursor's case study puts more than 90% of Salesforce engineers on the tool, across about 20,000 developers and a monorepo near 30 GB.

Cursor credits its codebase indexing with making one of the world's largest monorepos workable for agents. A ~30 GB tree is where indexing quality stops being a nicety and starts gating whether the tool is usable at all.

Stripe's story is reliability-first. The team operates against a stated target of 99.99999% uptime, so any tool that touches the codebase has to clear a high bar before it earns trust. Cursor reports that it reduced setup and debug time and helped Stripe reach a five-year high in developer sentiment.

• Reliability bar: a stated target of 99.99999% uptime sets the context for everything else.
• Less setup and debug time: reported reductions in the slow parts of the loop.
• Developer sentiment: a reported five-year high.
• Faster prototyping: prototypes in hours rather than days.

Stripe also reports a second-order effect: Cursor "opened up new skill sets." Front-end engineers started contributing to back-end work, and an iOS engineer began exploring another part of the stack. That mobility is harder to measure than a velocity number, but it is the kind of change teams notice in who can pick up which ticket.

One more Stripe data point comes from the other direction. Cursor's own finance team built a deep IC+ processing-cost analysis with an agent and shipped it to Stripe to confirm the findings. Stripe's reaction is the kind of validation that travels.

The analysis also surfaced real categories where Cursor itself should have been optimizing better to save money. Read it as an internal-team result with external sign-off, not an audited Stripe metric.

Cursor reports that about 40% of its own merged PRs are created end-to-end by cloud agents: opened, iterated and landed without a human writing the code by hand. This is Cursor reporting on Cursor, so read it as a vendor dogfooding claim.

The inflection was Cloud Agent Artifacts. Cursor reports that the merged-PR share from cloud agents roughly tripled after that shipped. By about March, Cursor says roughly 99% of its generated code came from cloud or IDE AI agents rather than being typed by hand.

The enterprise story is not only about shipping product code faster. Sales engineers, product managers, finance and go-to-market teams are using Cursor to build the tools they need instead of filing a ticket and waiting on an engineer. These are smaller, less audited results than the NVIDIA or Salesforce numbers, but they show a second adoption pattern worth understanding.

Airtable: from no-code demos to bespoke coded ones

At Airtable, sales engineers built customer demos in a no-code app builder with about a dozen preset components, plus "interface extensions" that let them write code. That is where Cursor came in. Once every Airtable SE had a license, demos went from pre-baked templates to fully bespoke builds writing TypeScript into the demo environment: financial reporting, dashboarding, and use cases the preset components could not reach.

The before/after is the point. A custom demo used to mean pinging an engineer and waiting a week or two. Cursor let SEs who could vibe-code fold bespoke demos into their normal cadence, so far more calls could include a custom build.

ChatGTM: an internal app nobody set out to build

Cursor's own go-to-market team had the universal prospecting problem: a rep spends two to three hours researching one account across five tabs: CRM, call recordings, usage dashboards, LinkedIn, enrichment. George, a self-described non-engineer on the team, did not start by trying to build an app. He solved one problem at a time inside Cursor.

1. 1Describe in plain English what data exists and where it lives. The agent writes the SQL to consolidate it; reps eyeball the output ("these numbers look too low" / "looks good") and iterate.
2. 2Describe the scoring logic in natural language: seniority, usage recency, when the lead came in. The agent writes a Python scoring file, tested against experienced reps' gut and the weights tuned.
3. 3Encode outreach guidelines, not rigid templates, so the agent researches each account and drafts a contextual message.

Demand snowballed from there (pull contacts, decide who to reach, draft messaging, build an account plan, prep a renewal), so they wrapped the same tools every rep was using into an internal app called ChatGTM. Built in a number of weeks. It is a Cursor-uses-Cursor result that sits alongside the engineering dogfooding above.

Prototyping with "baby Cursor"

Cursor's design team built what they call "baby Cursor": a fully standalone instance of Cursor, simpler on the backend but full-featured in the UX, used purely to iterate on prototypes and try new interaction patterns. The pattern is general: build a lightweight standalone clone of your product so you can express an idea as a working prototype instead of a PRDProduct Requirements Document. The spec describing what to build, for whom and why, before implementation starts. Increasingly the artifact you hand an AI agent as the source of truth for a feature. Press Enter for the full definition..

Most figures on this page are customer-reported or Cursor-reported. That does not make them false. It means the burden of context is on you: before any of this goes into a deal deck, sort the durable, third-party-verifiable facts from the perishable ones.