Cursor

Michael Truhl and the team at AnySphere started with a grand exercise in thinking about how AI would transform different knowledge work domains over the next decade. In late 2021, they were inspired by two moments: the first beta of GitHub Copilot, which felt like the first truly useful AI product they'd encountered, and a series of scaling papers from OpenAI showing that AI capabilities would improve dramatically through simple lever-pulling on model size and data. Rather than chase the obvious play—AI for coding—they initially pursued mechanical engineering tooling, thinking the space would be less competitive and overlooked.

That experiment lasted four months before they "came to their senses." They weren't mechanical engineers, the data challenges were immense, and crucially, they realized they weren't excited about it. Looking back at the coding space, they noticed something critical: despite the hype around GitHub Copilot and other entrants, they felt the ambition in the space was insufficient. "The people that were working on the space maybe had a disconnect with us and it felt like they weren't being sufficiently ambitious about where everything was gonna go in the future," Michael recalled. This realization—that an overcrowded space could still have room for a more ambitious player—set them on the path to building Cursor.

The team moved with extraordinary speed. They built the first version of Cursor entirely from scratch, hand-rolling their own editor rather than extending existing tools. This meant building language support, navigation, error tracking, integrated terminals, and remote server connectivity—all the scaffolding that modern code editors require. "After maybe five weeks we were living on the editor full time," Michael said. Within three months of the first line of code, they launched to the world.

What surprised them was the form factor question. Rather than just a plugin or wrapper, they bet heavily on building a full IDE, reasoning that programming itself was going to change so dramatically that the extensibility constraints of existing editors would become fatal. "If you think that the UIs may change a lot, if you think that the form factor program is going to change a lot, you necessarily need to have control over the entire application." That architectural bet proved prescient. User feedback in the early days was so strong that they decided to switch from their hand-rolled editor to a VS Code base, but retained full control over the application layer—a choice that enabled them to iterate on the product surface as their vision evolved.

Cursor's growth followed an unusual pattern: not a hockey stick, but consistent exponential growth from day one. "The growth has been fairly just consistent on an exponential," Michael noted. "An exponential to begin with feels fairly slow when the numbers are really low." This consistency—month-over-month acceleration rather than sudden inflection points—suggests that product-market fit arrived almost immediately upon launch, driven entirely by organic adoption and word-of-mouth.

The team's strategy was radically focused on product quality over marketing. "The customers we've seen have most success with AI I think are still fairly conservative about some of the ways in which they use this stuff," Michael explained. Users gravitated toward features like next-action prediction and edit inference—high-fidelity AI assistance for discrete, well-scoped tasks—rather than fully agentic programming. This insight guided the product roadmap: rather than chase the dream of fully autonomous AI agents, Cursor doubled down on making the AI invisible and reliable at specific tasks, keeping humans firmly in the driver's seat.

The most counterintuitive decision: building custom models. When Cursor started, the team calculated that training models from scratch would be prohibitively expensive and concluded that using off-the-shelf foundation models would be sufficient. They were wrong. "Every magic moment in cursor involves a custom model in some way," Michael revealed. The team discovered specific weaknesses in foundation models—especially around multi-file code completion, context retrieval, and diff generation—where custom, task-specific models could dramatically improve both quality and speed.

They approached model development pragmatically. Rather than chase pre-training (a path dominated by well-funded labs), they focused on post-training: starting with the best open-source and closed-source models and fine-tuning them for specific use cases. Their auto-complete model, for instance, is trained specifically to predict sequences of diffs across a codebase—a task where foundation models struggle because they optimize for generic next-token prediction rather than multi-file coherence. "We found a ton of success in training models specifically for that task. So that's a place where no foundation models are involved. It's kind of our own thing."

The team also refused to over-invest in sales and marketing early on. "Some of the normal things that people would maybe reach for in building the company early on, we really let those fires burn for a long time, especially when it came to things like sales and marketing." Instead, they rotated through support, obsessed over product quality, and let the tool's utility do the selling.

Cursor hit $100M ARR in 20 months and crossed $300M ARR within two years—among the fastest growth trajectories in startup history. The team attributes this to sustained paranoia about what could be better, not to any single breakthrough moment. "It's been kind of, a lot of it hasn't been over-rotated on kind of that initial push, but instead is like the continued evolution of the tool and just making the tool consistently better," Michael said.

Michael's vision remains ambitious: a world "after code" where software builders move away from writing formal programming languages and toward specifying intent in something closer to pseudocode or English. This transition, he believes, will unfold over years, with the IDE—or whatever form the tooling takes—evolving to support it. The ceiling remains high. There's room for niche players, but Michael believes one company will eventually build "the general tool that builds almost all the world's software. And that will be a very, very generationally big business." Whether that's Cursor, only time and relentless execution will tell.