Cursor's new Acquisition Means a lot, a Vector Database Joins the 10×ARR Club
AI debugging is becoming a critical bottleneck in the AI coding workflow
Cursor’s acquisition of Graphite, which uses AI to review and debug code
Yesterday on X, I came across an interesting post from Dimi Nikolaou, the co-founder of Wondercraft. He posted a simple question: ElevenLabs and Lovable are both valued at $6.6 billion today. If you could invest $1 million in one of them now and only exit after 2030, which would you choose?
Despite the same valuation, their fundamentals differ. ElevenLabs is reportedly doing around $300 million in ARR, while Lovable just hit $200 million but is growing quickly. In the replies, 88% chose ElevenLabs, while only 12% picked Lovable.
The reasoning was straightforward. ElevenLabs owns its proprietary models, faces fewer direct competitors, and has stronger moats and margins. Lovable, by contrast, operates in a much more competitive environment, relies heavily on third-party models, and—despite strong revenue growth—some believe it may not yet have reached durable product–market fit.
This highlights just how intense competition has become in the AI coding and “vibe coding” space. Even Lovable’s own growth lead has publicly said that PMF now has a very short half-life—something that needs to be re-earned every few months.
To strengthen its moat, Cursor recently acquired Graphite, an AI-powered debug and code review startup. The deal combined cash and stock, with Graphite continuing to operate independently. According to reports, the acquisition price significantly exceeded Graphite’s previous $290 million valuation.
As AI coding tools have taken off, AI bug detection has emerged as one of the hottest subcategories which I summarized in a previous post. Graphite was an early player, initially backed by Anthropic, and raised a $52 million Series B led by Accel in March.
Two other companies in this space recently raised nearly $90 million and $30 million respectively, with top valuations reaching $550 million and ARR exceeding $15 million.
Interestingly, Graphite didn’t start as a code review product. The team originally built a mobile app development tool, and code review was just an internal feature. Once released publicly, demand exploded. The team quickly abandoned the original direction and focused entirely on AI-powered code review.
In 2024, Graphite’s revenue grew 20×. By September, its customers already included over 500 companies—such as Shopify, Snowflake, Figma, and Perplexity—used by tens of thousands of engineers. The product uses AI to flag bugs and oversights, suggest changes based on repository context, summarize diffs, and propose fixes.
Before acquiring Graphite, Cursor had built a similar internal tool called Bugbot, which reportedly generated $10 million in ARR within a month.
The acquisition signals that AI debugging is becoming a critical bottleneck in the AI coding workflow. While AI has dramatically accelerated code generation, review, collaboration, and merging have emerged as the new constraints.
Graphite’s founders described the goal as unifying the “inner loop” (coding and generation) with the “outer loop” (review and merging). This would allow Cursor to cover the full lifecycle from code generation to deployment, strengthening retention and improving its enterprise appeal.
Industry-wide, more acquisitions and partnerships are expected as companies race to fill gaps in their stacks. I think other AI debug and review startups may soon become acquisition targets as well.
A New Vector Database Joins the 10×ARR Club
Meanwhile, the vector database space—once one of the hottest infrastructure categories—has grown quieter. Many startups raised large VC rounds over the past few years, but recently, few have shared meaningful ARR growth numbers.
But a new vector database startup just hit a big step forward. With a team of just over 20 people, it grew revenue more than 10× in a single year, reaching tens of millions of dollars in ARR, while landing customers like Anthropic, Cursor, and Notion at an early stage.
Its approach runs counter to the previous industry consensus. For years, the belief was that vectors had to live in memory to achieve low latency—at a cost of several dollars per GB. But it took a different view:
