It was supposed to take four months. NAB, Australia's largest bank, needed a new payment app that wasn't locked into specific hardware.
It took three weeks.
Trace that speed back far enough and you land on one company. Failed as an email client, failed as a CAD tool, rebuilt itself five times in four years, and just got bought by SpaceX for $6 billion. That's Cursor.
It wasn't the smartest team, and it wasn't the biggest compute
a16z's investors kept asking themselves one question while writing up Cursor: in AI coding, who actually wins in the end? They figured it had to be one of three answers.
You could bet on the smartest team. Except in this industry, talent switches companies every other week, so "the smartest team" is a title that doesn't hold for long. You could bet on the biggest compute. That's mattered historically, sure, but it's never been enough on its own. a16z's actual answer was "the fastest iterating team" — the bet that, over the long run, holds up the most reliably.
Cursor's own history is the proof they point to. In 2022, four MIT dropouts first tried an AI email client. Next was an AI CAD tool for mechanical engineers — they killed that one because nobody on the team actually had domain expertise. Frustrated with how slowly GitHub Copilot improved as a plugin, they forked VS Code entirely and built what's now Cursor. a16z counts tab-completion IDE, token reseller, AI lab, and agent platform as later identities, adding up to five reinventions in four years. When Claude Code and Codex showed up in 2025 and the internet decided Cursor was old news, the founder called an all-hands and cranked up the intensity. What followed was a run of releases — the cheap, high-performance Composer 1.5 and 2.5, the multi-agent Cursor 2.0, and Cursor 3, which rewrote what an IDE even means.
| What to bet on | The problem | |
|---|---|---|
| The smartest team | Leadership changes with every talent move | |
| The biggest compute | Matters, but never enough alone | |
| The fastest iteration | Wins eventually, even after failures | ★ |
Easy to say. Here's what it looks like in numbers
Anyone can say iteration speed matters. What's different about Cursor's enterprise customers is that the claim shows up in production data.
Over 30,000 NVIDIA engineers use Cursor daily. Committed code volume tripled after adoption, and the bug rate stayed flat. That's not "faster but sloppier" — it's faster without quality collapsing.
NAB standardized more than 6,000 developers on Cursor. A mainframe migration moving an Assembly-based core banking system to the cloud is running three times faster, and a project originally scoped to spend two of its six months purely on prep work — documenting legacy code, writing requirements — got that entire phase done by one engineer in a single week. One payment app went from a four-month plan to three weeks. The engineer who shipped it put the speedup at "5 to 8x".
What's notable is that Cursor didn't slow down after the acquisition news broke. Composer 2.5, released in May, matches Claude Opus 4.7 and GPT-5.5 on coding benchmarks at roughly a tenth of the cost per task. The iteration didn't stop once the deal was signed.
How do you actually apply this to your own team
This reads like a coding story, but pull apart the NAB and NVIDIA cases and the pattern underneath applies regardless of industry.
- Compress the prep work first
NAB turned two months of documentation and requirements-gathering into one week. Look at your own repetitive prep — research, first drafts, briefs — before you touch the "real" work. That's usually the easiest place to cut time. - Match the tool to the difficulty
NAB used cheap models for routine work and reserved the expensive, high-reasoning models for judgment calls like architecture. Not every task deserves the same level of resource. - Build small failure into the pipeline
Cursor kept going after the email client and the CAD tool both failed, with a stated attitude of pushing forward through "the first mistake, the second, even the tenth". Iteration only survives if each individual failure is cheap. - Compounding iteration becomes your reference
Cursor's revenue mix flipped from over 80% consumer in early 2025 to nearly 80% enterprise by the time of the acquisition. Early enterprise customers grew their usage more than 10x in year one. Fast iteration builds the case study that wins the next contract.
The core idea
Neither NAB nor NVIDIA started with a flawless plan. They cut prep time, split tools by task difficulty, and let small failures happen fast. The starting point isn't the direction — it's shortening the loop.
Want to go deeper
Cursor + SpaceXAI: the fastest iterating team wins a16z's own writeup — the full pivot history and the original source of the "iteration speed" thesis. a16z.com
NAB's Cursor case study How a 6,000-developer banking org applied Cursor to legacy migration, in detail. cursor.com
NVIDIA's Cursor case study How 30,000 engineers hit 3x code commits with a flat bug rate. cursor.com
Composer 2.5 announcement Benchmarks and pricing for Cursor's own frontier-tier model at a fraction of the cost. cursor.com
Cursor (Anysphere) company history Founders, early pivots, and valuation timeline. en.wikipedia.org
Korean coverage of the SpaceX-Cursor deal Deal structure and IPO timing, reported locally. digitaltoday.co.kr
