Part IThe Story
The photograph hangs on the wall behind his desk, so on a video call it hangs just over his shoulder. A Fortune reporter noticed it during a Zoom interview for a profile published on March 21, 2026. The subject is Robert Caro, sweatered, bespectacled, bent over a page and writing intently, and he watches the frame the way a saint watches a side chapel. Below him sits Michael Truell, twenty-five years old, soft-spoken and, in the reporter's phrase, "with a spine." Truell is the chief executive of Cursor, a four-year-old company then valued at $29.3 billion. By its own estimate it was used by 67 percent of the Fortune 500 and was writing roughly 150 million lines of enterprise code every day. Caro is the man who spent seven years on The Power Broker and most of the following half-century on Lyndon Johnson. Four volumes are published. The fifth is still unfinished.
The reporter admitted she had expected someone else on the wall: Wozniak, perhaps, or Jensen Huang, some patron of speed. Instead the founder of what is often called the fastest-growing software company in history had chosen the slowest serious writer in American letters to look over him. Asked why, Truell answered with the flat precision that people who have worked with him describe, and that his interviews confirm.
— Michael Truell, to Fortune, March 2026He's a motivational example of someone who's done useful, consequential work that's taken a long time. I actually most enjoyed his memoir about his writing process. There are few examples of people who set off to do multiple decades-long projects consistently.
The memoir is Working, a thin book about the labor of making thick ones. In it Caro describes typing every draft, interviewing the same sources over and over, and moving to the Texas Hill Country so he could understand what the land had done to the people who lived on it. Truell said he liked it best of all of Caro's books. That is a telling preference. He was not drawn to the finished monuments so much as to the method behind them.
The irony is close to unbearable, and Fortune said so. Truell admires work measured in decades. He runs a company measured in weeks, in an industry where a quarter is a geological era and slowing down for seven days can leave you behind. The same profile carried the headline Cursor's crossroads. It noted that a fair amount of social media had settled on a three-word verdict: "Cursor is dead." Truell told the magazine he wanted to build a long-lasting, independent, generational company focused entirely on professional developers.
Thirty-one days later, SpaceX announced it held the right to buy him.
By the Numbers
Four Years, Compressed
$60BImplied equity value of SpaceX's all-stock acquisition of Anysphere, closed August 14, 2026
389,289,254Shares of SpaceX Class A common stock exchanged for every outstanding Anysphere share
$3BAnnualized revenue as of May 2026, up from $4M in April 2024
67%Share of the Fortune 500 using Cursor, per the company
150MLines of enterprise code generated daily, per company estimate
~60Employees when Cursor was already past $100M in ARR
$8MSeed round, led by the OpenAI Startup Fund, announced October 2023
A Game Whose Players Are Programs
The word cursor has a long history that most of its users never think about. In Latin a cursor is a runner. Lucius Papirius Cursor, a Roman senator of the fourth century BC, was consul five times and dictator twice, and the cognomen is thought to have honored his speed on foot. Many centuries later the word attached itself to the slide rule's sliding indicator and then to the blinking vertical bar on a screen. That bar does one modest, essential thing. It tells you where you are. It marks the position at which the next human decision will land.
Keep that bar in mind. Most of what follows is an argument about where it should sit.
Truell was born in New York City in September 2000 and grew up there. He attended the Horace Mann School, a private preparatory school in the Bronx, where one of his classmates was Aman Sanger, later his co-founder. He started programming at eleven for an ordinary reason. He and his brother had tried to make a mobile game and failed, and the failure sent him to learn Objective-C. The first project of his that strangers noticed was not especially high-minded. It was an app that spoofed high scores, and it went viral. A child who learns to program by rigging a leaderboard has already understood something useful about systems: the scoreboard is code too, and code can be argued with.
Then came the robots. As a teenager, with a collaborator named Josh Gruenstein, Truell built what their research summary called "a novel control system that learns from humans." Robotic arms learned, through reinforcement learning, to play ping pong from very few data points. The paper's title, A Universal Robot Control System using Reinforcement Learning with Limited Feedback, has the earnest overreach of a science fair entry. It was one, and a successful one. The work collected honors from the New York City Science and Engineering Fair, ISEF, NASA and CERN. The resulting library, Fido, a lightweight C++ machine-learning library for embedded electronics, still sits pinned to his GitHub profile. As of this writing it has 464 stars, and its last meaningful commit predates the existence of ChatGPT.
The defining project of his adolescence, though, was a game in which humans did not play.
In 2016, with a collaborator named Benjamin Spector, Truell co-created Halite, Two Sigma's first artificial intelligence programming challenge. Halite's players did not move pieces. They wrote bots, in whatever language they liked, and the bots fought it out on a grid. Across three seasons more than ten thousand contestants took part. Somewhere in that span a teenage Truell was working at a quantitative hedge fund. One commenter on an Andreessen Horowitz post, years later, put it plainly: "Impressive working at Two Sigma at 15-16 years old." In October 2017, Truell and Spector posted a paper to arXiv titled The Design and Implementation of Modern Online Programming Competitions. It treats ranking systems, sandboxing and the engineering of a fair arena with the seriousness of people who had watched thousands of strangers try to break theirs.
That July, The Riverdale Press ran a short item beneath a headline about kindergarten books donated by KPMG. Michael Truell of the Horace Mann School had been one of eighty students selected for the Research Science Institute, a free six-week summer program at MIT where high schoolers do original research and defend it before professors.
Look at the shape of these early projects. A boy fails to build a game, then builds the tool that fakes the game's scores. He builds a machine that learns to play a game from a human's sparse corrections. Then he builds the arena where programs play games while their human authors watch, tweak and resubmit. In each case the interesting object is not the game. It is the relationship between the human, whose intent is underspecified and corrections are few, and the machine, which does the playing. The human sets the goal. The program executes. The human looks at the result and adjusts.
Years later Truell would describe the future of software engineering in nearly those terms, as "specifying your intent for how exactly you want everything to work." He would call the job of the future engineer that of a "logic designer." He was describing the Halite contestant.
The Starred Circle
He arrived at MIT to study computer science and mathematics. The record is uncertain on a small point and certain on the large ones. Wikipedia credits him with a Bachelor of Science. Fortune, Forbes-derived coverage and a long tail of founder hagiographies call him a dropout. A venture newsletter from 2024 calls the whole founding team "class of 2022 grads." A careful profile would note the disagreement and move on. That he spent the right years at the right institute is not in dispute, and those years were busy.
After his first year, at eighteen, he took a summer internship at Google working on what his LinkedIn describes as "language models for feed ranking," or, in another account, models for news recommendations. Feed ranking is an unglamorous corner of machine learning. It is also where a model's guess about what a human wants next turns directly into behavior at enormous scale. He later worked at Octant, a drug-discovery company, on computational chemistry, and built internal tools there. He would remember that job long after leaving it. "The off-the-shelf tools that existed were horrible," he told Lenny Rachitsky in 2025, "and did not fit their use case at all." The demand for internal software, he said, "far outstripped just the things that I could build in the time that I was with them." At MIT he also did statistics research. With Jan-Christian Hutter he co-wrote a paper titled Maximum Likelihood Estimation for Brownian Motion Tree Models Based on One Sample. It is the sort of title that tells you its author is comfortable being the only person in the room who understands the problem.
During the Google summer he met Ali Partovi, an early investor in Facebook and Airbnb, who was recruiting for Neo Scholars, a program that picks about thirty exceptional technical students a year and mentors, connects and eventually invests in them. Partovi gave Truell a written coding test, handwritten in some accounts. Truell finished it in what Forbes called "record time." After the meeting, Partovi drew a star with a circle around it next to Truell's name on his list of candidates. In his private notation, as Forbes reported it, that mark meant he would invest in anything Truell chose to do.
The starred circle deserves a moment, because it is the purest form of the credential and Truell would later distrust credentials in hiring. It is a judgment rendered before any product exists, on the evidence of speed and mind alone. Partovi kept his word. When Anysphere was formed, he was among its first investors.
The three people Truell built the company with fit the same mold.
Aman Sanger grew up in New York like Truell and went to Horace Mann like Truell. His father, Arvind, is an IIT Bombay graduate who runs Geosphere Capital, a long-short hedge fund focused on natural resources. His mother, Shilpa, is an orthodontist who also invests in startups and sits on the board of a nonprofit for education in India. Sanger worked at You.com and at Bridgewater. He was also a Neo Scholar. He would eventually become the founder who had to publicly own the company's most awkward disclosure, as a later section describes.
Sualeh Asif came from Karachi, Pakistan, and competed in the International Mathematical Olympiad. He interned at IBM and worked as a research assistant at MIT. He became Cursor's chief product officer, which is a slightly odd title for the mathematician of the group and an apt one for a company whose product is a theory of how people should think.
Arvid Lunnemark worked in engineering at Stripe and at Jane Street, the quantitative trading firm whose culture of puzzles, humility and enormous quiet competence is close to legendary among people who care about such things. He was the founder most associated with security and cryptography. In October 2025, at the height of the company's ascent, he would leave to start a safety-focused AI research lab called Integrous Research.
Two quantitative trading firms, a hedge fund father, an Olympiad, a Silicon Valley talent scout's private symbol. Read together, these are the biographies of people trained to price uncertainty and act before the market agrees with them. None of them had ever run a company. None of them had ever been a mechanical engineer either, which would matter shortly.
Sleepy and Uncompetitive
In 2021, a year before ChatGPT, the four of them were asking a question Truell later recalled at Y Combinator's AI Startup School: "Do we go and work on AI in academia? Or … do we go join, you know, a big existing AI effort? Or do we start our own thing?"
Two things pushed them toward the third choice. Truell has told the story often enough that its structure is fixed. The first was GitHub Copilot, which they used in its early beta. "This was the first time we had used an AI product that was really, really, really useful," he said, "and was actually just useful at all, and wasn't just a vaporware kind of demo thing." Copilot was, in his estimation, also "one of the most useful, if not the most useful dev tool we'd ever adopted." The second was a run of scaling-law papers from OpenAI and elsewhere, which argued that "even if we had no new ideas, AI was going to get better and better just by pulling on simple levers."
What they did with those two convictions made sense on paper and turned out to be wrong.
They ran what Truell calls "this whole grand exercise." For each area of knowledge work, they asked how it would change as models improved, what the work would look like at the end, what the tools would need to become, and what would happen once pre-training ran out of steam. Coding was obviously the answer, and for exactly that reason they ruled it out. "We thought, 'Oh, coding's great, coding's totally going to change with this AI, but people are already doing that,'" he said. At the Compile keynote in 2026 he put it more starkly: "We thought, there's just, there's no room, there's not much to do, people have got that covered." So they chose mechanical engineering, which in his words was a field that seemed "relatively uncompetitive, and sleepy, and boring, and no one would be looking at it." The plan was a copilot for people who design parts in SolidWorks and Fusion 360.
It was a textbook move, and a textbook failure.
"There was a little bit of a blind man and the elephant problem from the get-go," Truell said. Nobody on the team was a mechanical engineer. They had friends who were, which is not the same thing. The data didn't exist in usable form. There is not much on the internet in the way of three-dimensional CAD models together with the sequence of steps that produced them, and getting that data from the companies that have it was hard. The market was fragmented, and the training infrastructure was primitive. Elsewhere, two of the founders were working on an end-to-end encrypted messaging system meant to solve metadata leakage, a Signal-like problem that stalled on scalability and found no business buyers. There were prototypes for security vulnerability detection and for tools aimed at quantitative researchers. An Andreessen Horowitz account adds an AI email client to the list.
Truell's accounts of how long this lasted do not quite agree. On Lenny Rachitsky's podcast the mechanical engineering detour was "a period of four months." To a Notion interviewer it was "almost a year wandering in the desert working on mechanical engineering tools something we weren't passionate about. It had really bad founder market fit." Both can be true, depending on where you start the clock and how honest you are about the stretches spent pretending to be excited.
There is a more cinematic version. In 2026 a widely shared LinkedIn post described a Truell who was "losing hope" in December 2022, who late one night pasted his own buggy code into an early GPT-4 and watched it fix the bug and explain why. "His hands started shaking," the post says. He called his co-founders, and they threw problems at the model until sunrise. It is a good story, perhaps too good. Its author attributes it to Truell's Y Combinator interview, and that interview's published summary contains no shaking hands. Truell's own accounts are flatter and less flattering to himself. "We came to our senses," he said. "We realized we're not super excited about mechanical engineering, it's not the thing we want to dedicate our lives to." At Y Combinator he was blunter still. They were "desperate," he said, after their first ideas failed to take off.
What changed was not the market. It was their reading of the market's ambition. Copilot was reportedly earning around $100 million a year by 2022, and the incumbents looked unbeatable on paper. Looked at again, the people building in the space "weren't being sufficiently ambitious about where everything was going to go in the future," as Truell said, "and how all of software creation was going to blow through these models." Nobody, he said at Y Combinator, was seriously pursuing a future in which all of coding as we know it would be automated.
The a16z version of the moment is the most human one. "We caught the bug," Truell said. "We couldn't stay away from it."
In a Cave, in Our Underwear
At the end of 2022, by Truell's account at Compile, the founders went "in a cave" and coded "in our underwear" for about two weeks to get a prototype together. Here the prudent move would have been a plugin. Copilot was a VS Code extension. VS Code dominated the editor market. Any sensible reading of that market said you should live inside the incumbent's house and pay rent.
They built their own house, and at first they built it from nothing. The first Cursor was, in Truell's word, "hand-rolled": built on CodeMirror, with its own SSH layer and its own implementations of everything a modern editor must do. That means support for many languages, navigation, error tracking, an integrated terminal, and connections to remote servers. "We kind of just went on this blitz of building things incredibly quickly," he said. After about five weeks the founders were "living on the editor full-time, and had thrown away our previous editor." About three months after the first line of code, they released it.
"The three-month version wasn't very good," Truell said, and he has never softened it.
The first user did not like the product. Neither did the second. Some early testers, he told the Compile audience, "ran away from us kicking and screaming." A few stayed and began what developers call "daily driving" it, meaning they used it as their real editor for real work, in the way you discover whether a car is reliable by commuting in it rather than test-driving it. Those few were enough. Their feedback came quickly, and much of it pointed in one direction: rebuilding the whole editor was costing too much. Cursor switched to a fork of VS Code. Truell compares that to "many browsers use Chromium as a base" and then go their own way.
It is worth being exact about what the fork meant, because it was the company's first and most important strategic decision and it is often described backwards. Cursor did not give up on owning the editor. It gave up on rewriting the parts of the editor that didn't matter, so it could own the parts that did. A plugin, Truell argued, operates under someone else's limits, and "the extensibility that existing coding environments have is so, so, so limited." If you believe "the form factor programming is going to change a lot," you "necessarily need to have control over the entire application." The fork kept the whole application, including the cursor, in their hands. It let them inherit everything else.
They had expected to build for "a couple hundred people for a long time." Instead there was an immediate rush of interest. In 2023 Cursor reached a million users with a team of fewer than ten people. Most of the marketing was the founders posting on Twitter, often about technical details as obscure as open-source models like Flan-T5.
The users gathered in a Discord server. That server began as a place to report bugs, swap workflows and complain. Over time it turned into the company's best recruiting channel. "We actually hired many members of [the] Discord server," Truell said at Compile in 2026. The developers who most loudly told Cursor what was wrong with it were offered the chance to fix it. "Through all of this craziness," he said, "there's a group of people whose fundamental kind of core ethos is this group of developers building for developers."
Halite worked the same way. The people who argued hardest about the arena's rules were the ones who understood it best.
Letting the Fires Burn
Ask Truell about Cursor's growth and he describes not a rocket but a curve that looked like nothing for a long time. "To be honest, it felt fairly slow to begin with," he said, "and maybe it comes from some impatience on our part." The growth was "fairly just consistent on an exponential, of just consistent month-over-month growth." He then made an observation that sounds obvious and is not: "an exponential to begin with feels fairly slow and the numbers are really low."
The numbers bear this out. The OpenAI Startup Fund led an $8 million seed round, announced in October 2023. Nat Friedman, former CEO of GitHub, which made Copilot, and Arash Ferdowsi, Dropbox's co-founder, joined as angels. That the former head of the company that built the product Cursor most needed to beat chose to back Cursor is the sort of detail novelists are told to cut for being too on the nose. In April 2024, Cursor's annualized revenue was about $4 million. That August it raised a $60 million Series A led by Andreessen Horowitz at a $400 million valuation. By November 2024, revenue was around $50 million, and TechCrunch was reporting that Benchmark, Index Ventures and others were in what it called a "wild unsolicited bidding war" that pushed the valuation to about $2.5 billion, more than six times higher in four months. That same month Cursor bought Supermaven, an AI code-completion startup founded by Jacob Jackson, and folded its team in.
In January 2025, Cursor crossed $100 million in annualized revenue and raised a $105 million Series B. Usage tracked through Y Combinator tells the story from the inside. In 2023, Cursor users were a single-digit share of YC batches. In 2024, by the account summarized from Truell's YC talk, they were about 80 percent.
What Cursor did not do in those years matters as much. "Some of the normal things that people would maybe reach for in building the company early on," Truell said, "we really let those fires burn for a long time, especially when it came to things like sales and marketing." The founders worked on the product. They staffed support rotations themselves. They hired at a pace that, by his own later account, was a mistake.
— Michael Truell, on Lenny's Podcast, May 2025Many people you hear hired too fast, I think we actually hired too slow to begin with.
Notion's interview series described the early hiring as "just one person every few months." Truell said he wanted a particular blend of person: "intellectual curiosity and experimentation," together with "an intellectual honesty, and maybe micro-pessimism, bluntness." Micro-pessimism is a phrase worth noting. It describes a temperament that is optimistic about the decade and suspicious of the week. It is close to the Caro temperament. He also admits that the early search began with the wrong profile: "we actually biased a little bit too much towards looking for people who fit the archetype of well-known school, very young, had done the things that were high credential in those well-known school environments." In effect, they went looking for starred circles. They did better when they found "fantastic people who are willing to do this with us who were later careered," and very young people who looked "slightly different from being straight out of central casting."
The method they settled on is now well known. Candidates spend two days on site working on a self-contained project in Cursor's actual codebase. It is not work the company ships. It is "a canned list of projects," but it is real, and the candidate eats meals with the team. Truell expected it would not scale beyond a handful of hires. It has lasted, partly because it serves as what he calls a "do you want to be around this person type test," and partly because it sells the company. In the early days, he said, "when the product is comparatively not very good, really, the only thing you have going for you is a team of people that some people find special and want to be around." Later, Business Insider reported that Cursor bans AI tools in the first round of its coding interviews. A company that sells AI coding wants to see you code without it first.
In spring 2025, Cursor was past $300 million in annualized revenue and Truell told Rachitsky the company was "going on 60 people." Many software companies with forty engineers, he said, "would be over 100 people," because they are sales-led and "that's just quite labor-intensive." By June 2025, revenue was over $500 million, and Thrive Capital led a $900 million Series C at $9.9 billion. By August, the company had about three hundred employees.
He described the culture he wanted in terms that would fit a Caro research assistant: people "less focused on external validation, more focused on building something really great," people for whom "the highs aren't very high, the lows aren't very low."
Every Magic Moment
Truell's answer to "the most counterintuitive thing you've learned" is a confession that the founders misjudged their own business.
"We definitely didn't expect to be doing any of our own model development," he said. Early on they had tried. The original Codex paper suggested a model could be trained for around $100,000, and they made some attempts. Then they backed off and used API models for pragmatic reasons in 2023. Training foundation models from scratch "felt a bit like focusing one's attention in the wrong area." Why re-teach a neural network the entire internet when others already had?
They came back to it, one narrow use case at a time. "At this point," Truell said in 2025, "every magic moment in Cursor involves a custom model in some way."
The clearest example is the autocomplete Cursor calls Tab. Truell explains it with a comparison to writing. Gmail's autocomplete helps only a little, he points out, because "it's just really not clear what you're going to be writing just by looking at what you've written before." Code is different. "Sometimes, really, the next 5, 10, 20, 30 minutes of your work is entirely predictable from looking over your shoulder." Change a function signature in one file and the edits needed in three other files follow mechanically. A model that predicts those edits, not the next token of text but the next set of diffs across a codebase, has to be extremely fast: "they need to give you a completion within 300 milliseconds." It also has to be cheap, because it is re-run on every keystroke. No frontier lab was going to build that model. Cursor did.
Other custom models sit around the large models from Anthropic, OpenAI and Google. On the input side, Truell describes "a mini Google search" that finds which parts of a codebase to show the large model. On the output side, smaller specialized models take a frontier model's sketch of a change and expand it into full code edits. "The high level thinking is done by the smartest models," he explained, and "these smaller specialty incredibly fast models, coupled with some inference tricks, then take those high level changes and turn them actually into full code diffs." The design is a division of labor between an expensive mind and inexpensive hands. It is also, not by accident, the same planner-and-worker hierarchy Cursor's research team would later use for its long-running agents.
This is the point where the "GPT wrapper" label, applied to Cursor for years by people who had not looked closely, stops being accurate. It is also where the company's capacity for embarrassment grew. A company that owns more of its stack has more ways to get things wrong.
In April 2025, an AI support agent named "Sam" made up a login policy that did not exist. Users who believed it canceled their subscriptions before staff stepped in, apologized and issued refunds. Ars Technica's headline did the work of a satire: Company apologizes after AI support agent invents policy that causes user uproar. In July 2025, Cursor moved its $20 Pro plan from 500 requests to usage-metered limits. Developers were hit with surprise charges, and the company apologized, rolled back the limits and promised refunds. In March 2026, Cursor released Composer 2, a coding model it presented as its own. Within a week, users on X had worked out that it was built on Kimi 2.5, an open-weights model released in January by the Chinese startup Moonshot AI. Lee Robinson, a Cursor vice president, said that "only ~1/4 of the compute spent on the final model came from the base, the rest is from our training." Sanger called it "a miss to not mention the Kimi base in our blog from the start." Moonshot, for its part, congratulated Cursor and noted an authorized commercial partnership through Fireworks AI.
Each episode is the same failure in a different form. In each, the company let the user's view of what was happening fall out of step with what was actually happening, whether about policy, price or provenance. For a company whose core promise is that humans stay in control, these were not minor lapses. They were small breaches of the thesis itself.
An Innovator's Dilemma at Four Years Old
From the start Truell has said the moat in his market is thin. Most founders who say this are playing at modesty. He appears to mean it.
"I truly just think that the ceiling is so high that no matter what entrenchment you build, you can be leapfrogged," he told Rachitsky. He compared the market to search engines in the late 1990s and to personal computers in the 1970s, 1980s and 1990s, eras in which "you could keep getting value for the incremental hour of a smart person's time, the incremental R&D dollar for a really long time." Ask him about defensibility and he gives an answer most chief executives would never say out loud:
— Michael Truell, on Lenny's Podcast, May 2025I think maybe the sad truth for people like us, but then the amazing truth for the world is, I think that there are many leapfrogs that exist, there's more useful things to build.
He also explained why GitHub Copilot, the product that had inspired Cursor, fell behind. Part of it was structural. This is "a market that's not super friendly to incumbents," where switching costs are low, the product never becomes a commodity, and bundling doesn't help. Part of it was history: "the group of people that worked on the first version of Copilot have, by and large, gone on to do other things at other places." It was a clean diagnosis of a dominant player losing to a quicker one. It would describe Cursor itself within a year.
The leapfrog arrived under the name Claude Code. Anthropic launched it barely a year before Fortune's March 2026 profile. It was a terminal-based coding agent from the company whose models powered much of Cursor's product. That is the uncomfortable fact of Cursor's position: one of its main suppliers had become one of its main competitors, with a reported $380 billion of financial strength behind it. Fortune reported that customers were leaving. One Cursor investor said several startups in his portfolio were "decoupling" from it. The head of engineering had left. "The thing about this market is that things change so quickly," a venture capitalist backing a competitor told the magazine. "If we were sitting here a year ago, you'd be writing the opposite article, 'Cursor is on top of the world,' right?"
The numbers, meanwhile, did not show a dying company. Cursor's Series D closed on November 13, 2025: $2.3 billion, co-led by Accel and Coatue, with Google and Nvidia joining as strategic investors, at a $29.3 billion valuation. Annualized revenue was above $1 billion. Forbes estimated each founder's net worth at roughly $1.3 billion. Sanger explained the financial approach to The Economic Times in the language of a hedge fund manager's son: "If you look at the amount that other foundational model companies are spending relative to revenue, it's significant. We are planning to build a more sustainable business." By February 2026, two investors told Fortune, annualized revenue had passed $2 billion. By May, according to Bloomberg, it had passed $3 billion.
The product moved quickly as the ground shifted under it. Cursor 2.0, released October 29, 2025, ran multiple agents in parallel. There were agents for the web, for mobile, for the command line and for the cloud. Bugbot, launched in July 2025, reviewed pull requests. In December, Cursor agreed to buy Graphite, a New York code-review startup, for what was reported as well above Graphite's last $290 million valuation. A January 2026 research post described pointing hundreds of agents at "an ambitious goal: building a web browser from scratch." The agents "ran for close to a week, writing over 1 million lines of code across 1,000 files." A three-week internal migration from Solid to React added 266,000 lines and deleted 193,000. After trying flat self-coordination, which failed, the team arrived at a hierarchy of planners, workers and judges. Its finding read like a line from one of the old Halite papers: "The best system is often simpler than you'd expect." In February 2026, Truell published an essay about a "third era" of software development, in which the unit of work shifts from editing code to directing and reviewing agents.
All of this complicated the story Cursor had told about itself. Truell had always said humans should be "in the driver's seat," with "a fast iteration loop and not just having something in the background that's super slow and takes weeks." In 2025 he warned that an approach of "let's write a giant thing telling the model exactly what to do" was "a little bit of a recipe for disaster right now." He had told a Notion interviewer, without qualification: "We're not about vibe coding. Developers still need control." Now his research team was running agents for a week without supervision. Whether this was a betrayal of the thesis or its next phase depends on where you think the cursor sits when a thousand agents are typing.
The fairest reading is that Truell had always set the boundary in time rather than in principle. Right now, he kept saying. Today. "We're very far from this right now." The human stays in the driver's seat because the car still drifts. It was the realism of someone who had trained a robot arm to play ping pong from limited feedback and knew exactly how much feedback that was.
Sixty Billion Dollars of Someone Else's Stock
On April 21, 2026, SpaceX said it had reached a deal with Cursor giving it the right to buy the company later that year for $60 billion, or else to pay $10 billion for what CNBC's headline, quoting SpaceX, called "our work together." That afternoon Truell posted on X: "Excited to partner with the SpaceX team to scale up Composer. A meaningful step on our path to build the best place to code with AI."
The tweet is precise. It does not call SpaceX an owner. It names one thing, Composer, Cursor's in-house model, and one verb, scale up. Read alongside the April 2026 headline in The Information, "Behind Cursor's Deal With SpaceX, Anthropic and Compute Costs Loomed Large," the logic is plain enough. A company whose main supplier had become its main competitor needed a different source of intelligence and the computing power to build it. When a16z's partners later wrote about the deal, they made the same point with a wink: "consider that only one gigawatt-scale data center is actually operational today, and who built it." The data center is called Colossus. Cursor's model menu now listed Grok first.
On June 16, a few days after what the New York Times described as SpaceX's record initial public offering, SpaceX said it would exercise the option in an all-stock deal. It was called the largest acquisition of a venture-backed startup in history. The deal closed on August 14, 2026. Under the merger agreement, every outstanding Anysphere share, common and preferred, was converted into a combined 389,289,254 shares of SpaceX Class A common stock. The $60 billion figure was not a price. It was an implied equity value, calculated from the volume-weighted average closing price of SpaceX shares over the seven trading days before closing. Unvested options became SpaceX equity awards. Cursor became a wholly owned subsidiary, folded into a unit called SpaceXAI. Wikipedia now lists two headquarters: 33 New Montgomery Street in San Francisco, and Bastrop, Texas.
There was some theater. Within weeks, according to a stream of headlines, OpenAI, whose Startup Fund had led Cursor's seed round three years earlier, had walked away from a reported $1 billion partnership and cut Cursor off from its models. One headline told the rest: "Sam Altman doesn't trust Elon Musk's cursor, Musk says he couldn't care less." The first institution to bet on the company became, by the end of that summer, the first to treat it as a rival's asset.
Truell's GitHub profile changed too. The line under his name, where developers say what they are working on, now reads "Working on @xai-org." His personal site says he is building SpaceXAI. In late August, Time named him one of the 100 most influential people in AI.
There is one more small detail, for a reader who enjoys these things. In July 2026, Wikipedia's article on Cursor was flagged with a warning that it "may incorporate text from a large language model," and "may include hallucinated information." So the encyclopedia entry about the company that taught machines to write code for people now carries a notice that a machine may have written it.
The tension at the center of the story can't be resolved neatly, and it would be dishonest to try. In March 2026, Truell said he wanted a long-lasting, independent, generational company. By August, the company was neither independent nor, in any legal sense, his. The cynical reading is that the photograph of Caro was decoration. A kinder and probably more accurate reading is that Truell did what an "intense realist" (his phrase for his own approach to technology) would do once the input he could not produce himself, computing power at gigawatt scale, became the input that decided the game. Caro had the patience to work for decades because the material would wait for him. Lyndon Johnson was dead and the archives were not going anywhere. Truell's material moves. The models he relies on are owned by companies that now compete with him, and every week the frontier shifts.
The Unfinished Fifth Volume
In 2025, Truell was asked how he keeps a team focused in the most frenzied corner of the most frenzied industry on earth. He gave the answer of someone who reads history in order to discount the present. Rachitsky noticed that the shelves behind him held a history of "some old computer company" the host had never heard of. Truell said the founders had worked on AI since "2021, 2022." Since then they had seen InstructGPT, DALL·E, Stable Diffusion, ChatGPT, GPT-4, and a long run of new modalities arrive and recede, "and only a very small number of these things really kind of affects the business."
"So I think we've kind of just built up a little bit of an immune system," he said.
He compared this to the history of his own field. There have been "so many papers on deep learning in academia," he said, yet "a lot the progress of AI can be attributed to some very simple elegant ideas that have stayed around, and the vast majority of ideas that have been put out there haven't had staying power."
This is a historian's way of thinking, and specifically Caro's. The Power Broker explains how a city was remade by a few decisions that compounded over forty years. Robert Moses did not win because he was fast. He won because he understood which few things mattered and held on to them while elected officials came and went. When Truell describes the market he is in, the comparisons are historical too: the search engine wars at the turn of the millennium, the personal computer through three decades. When he describes what he wants, he describes duration: "Companies that have continued to really ride the wave of many leapfrogs," he said, "but they're kind of rare too, it's a hard thing to do." He has sat across from Ed Catmull of Pixar and Patrick Collison of Stripe, founders whose institutions outlasted several technological eras. He has said, more than once, that the transformation he is working on will take "a couple of decades."
He has also said it will be more consequential than the internet. "More consequential than any shift in tech that we've seen since the advent of computers." Then, almost immediately: "And I think it's going to take a while."
So a twenty-five-year-old with a portrait of a slow writer above his desk sold his company four years after starting it, because the timeline of the work was longer than the timeline on which he could pay for it himself. Whether that counts as defeat or as patience depends on whether you think the project was the company or the work. Caro has never owned the Johnson archives. He has spent his life in other people's buildings, inside their institutions, writing the book.
On Truell's GitHub profile, the line about xAI is new. Below it, the pinned repositories have not changed. One is the Halite arena from season one, built when he was a teenager, where ten thousand strangers sent bots to play a game none of them would touch with their own hands. Next to it is Halite II, then the Brownian motion trees. At the bottom of the page is Fido, the C++ library for teaching a robot arm from a human's sparse corrections. It was written in high school and has not been updated in years. It still lists its honors in the description: NYCSEF, ISEF, NASA, CERN. Below the last line of text, an empty field is waiting for input.
Part IIThe Playbook
What follows is a set of operating principles drawn from Truell's decisions, and from his mistakes, which he has been unusually willing to name in public. They are not rules for every company. They are rules for building a product inside a technological shift whose ceiling is still rising: a market where incumbents are fragile, suppliers can become competitors, and the founder's job is to keep the human cursor in the right place while the machine underneath gets more capable every quarter.
The Truell Playbook
- 1.Enter the crowded room if the ceiling is high.
- 2.Choose the domain you would suffer for.
- 3.Own the whole pane of glass.
- 4.Be an intense realist about the technology.
- 5.Let the fires burn.
- 6.Hire slowly, then admit when slow became too slow.
- 7.Recruit from the room where your users already argue.
- 8.Build your own models only at the seams.
- 9.Grow an immune system.
- 10.Assume the leapfrog, including your own.
- 11.Treat legibility as part of the product.
- 12.Know which input you cannot manufacture.
Principle 1
Enter the crowded room if the ceiling is high
The standard advice is to look for a sleepy, uncompetitive market where no one is paying attention. Cursor's founders followed it, chose mechanical engineering, and lost months, or most of a year, depending on which account you trust. The lesson is not that the advice is wrong. It is that the advice answers the wrong question. Truell's later framework asks two things: how high the ceiling is, and how ambitious the people already there are. In 2022, Copilot was making around $100 million a year and looked unassailable. Truell saw that "they weren't being sufficiently ambitious about where everything was going to go in the future."
A crowded market with a low ceiling is a commodity fight. A crowded market with a high ceiling is a sequence of leapfrogs waiting to happen, and the incumbents' existing customers are proof that demand is real. "You need there to be leapfrogs that can happen," Truell said. "You need there to be things that you can do." Competition tells you the market exists. The ambition gap tells you whether there is room for you in it.
Tactic: Before you rule out a market for being crowded, write down what its best product will look like in ten years. If the leading players aren't building toward that, the crowd isn't a barrier.
Principle 2
Choose the domain you would suffer for
Truell's diagnosis of the CAD detour was not about the market. It was about the founders. "It had really bad founder market fit. We weren't mechanical engineers, and the space didn't click for us." The "blind man and the elephant problem" meant every product decision passed through friends' secondhand descriptions. The data problem, with no public corpus of 3D models and the steps behind them, would have been solvable for a team that cared enough to grind through it. This team didn't care enough. "It's not the thing we want to dedicate our lives to."
The phrase dedicate our lives matters. In a fast market, endurance is what decides things, and endurance comes from caring about the work. Cursor's product process depended on dogfooding: "We never wanted to ship anything that wasn't useful to us." That only works if you are the user. A founder who doesn't feel the pain can't tell when the painkiller is working.
Tactic: Pick the domain where you can personally tell, within an hour of using a prototype, whether it got better or worse. If you need a customer interview to know, you are too far from the work.
Principle 3
Own the whole pane of glass
The consensus in 2022 said: build a plugin and ride the incumbent's distribution. Cursor forked VS Code instead. The reasoning followed from a thesis about the future: if "the form factor programming is going to change a lot," then a plugin's "so, so, so limited" extensibility becomes a hard ceiling on what you can build. Owning the application meant Cursor could add multi-file editing, diff-level predictions and agent workspaces without asking permission.
The subtler point is that they first built the editor from scratch, found that doing so was costing them focus, and switched to the fork. Owning the surface does not mean writing everything yourself. Truell's own comparison is browsers built on Chromium. You inherit the commodity layers and control the layer where the change is happening.
⌘
Four Years, Compressed
The pace at which Cursor moved from a hand-rolled prototype to a subsidiary of SpaceX.
2016
Truell, still in high school, co-creates Halite, Two Sigma's first AI programming challenge.
2022
Anysphere is incorporated. Mechanical engineering, encrypted messaging and other ideas fail. At year's end, the founders start building Cursor.
2023
Launch about three months after the first line of code. Switch to a VS Code fork. One million users with fewer than ten employees. $8M seed led by the OpenAI Startup Fund.
2024
$60M Series A at $400M. Revenue grows from $4M to $50M. Unsolicited bidding pushes the valuation to about $2.5B. Supermaven acquired.
2025
$100M ARR in January. $9.9B valuation in June. Pricing backlash in July. Lunnemark departs in October. $29.3B Series D in November. Graphite deal in December.
2026
Claude Code pressure. Composer 2 and the Kimi disclosure in March. SpaceX option on April 21. Exercised June 16. Closed August 14.
Tactic: If your thesis says the interface will change, make sure you control the interface. Fork the commodity layers, and do not build on a platform whose owner sets the limits of your product.
Principle 4
Be an intense realist about the technology
"We've tried to be intense realists about where the technology is today," Truell said, "very, very, very excited about how AI is going to mature over the course of many decades." Holding both positions at once, long-term conviction and short-term skepticism, explains Cursor's product choices better than any roadmap. He warned specifically against the habit of seeing a model do "magical things in one area" and assuming it is smart everywhere. It isn't, and daily dogfooding, he said, "instills a realism in you around where the tech is right now."
In practice this produced Cursor's long commitment to keeping the human in control. It also produced its advice to users, which runs in two directions. In daily work, break tasks into small pieces. On a side project, "explicitly try to fall on their face" so you find the model's real limits. Junior engineers, he observed, trust the AI too much. Senior engineers "underrate what AI can do for them." The realist position sits between those two errors. It has to be re-checked every time a new model ships, which is why Cursor's agents began running for days once the models could handle it.
Tactic: Schedule regular, deliberate attempts to break the newest model on low-stakes work, so your sense of what it can do is based on evidence instead of last quarter's demo.
Principle 5
Let the fires burn
Every early-stage company has a list of urgent things that are not the product: sales, marketing, process, partnerships, conferences. Cursor ignored most of the list. "We really let those fires burn for a long time, especially when it came to things like sales and marketing." The founders did support rotations themselves and otherwise worked on the editor. As late as spring 2025, the company had about sixty employees and more than $300 million in annualized revenue. Truell noted that many software companies with forty engineers "would be over 100 people."
◧
Conventional Wisdom vs. Cursor's Approach
Where the early company went against standard startup advice.
| Conventional wisdom | Cursor's approach | Outcome |
|---|---|---|
| Find a sleepy, uncompetitive market. | Returned to the crowded market they actually cared about. | Validated |
| Ship as a plugin to the dominant editor. | Forked VS Code to own the whole application. | Validated |
| Build sales and marketing early. | Product only. "Let those fires burn." | Validated |
| Hire fast to capture the market. | Hired one person every few months. | Overcorrected |
| Hire from elite pedigrees. | Moved toward later-career and unconventional hires. | Learned |
| Stay independent at all costs. | Sold to the party that controlled compute. | Contested |
The principle has a precondition that founders often skip. Letting fires burn only works if the product can carry distribution on its own. Cursor's growth came from word of mouth among professional developers, spreading through YC batches until usage approached 80 percent. A product that can't spread that way doesn't earn the right to neglect sales. It just neglects them.
Tactic: Make a written list of the fires you are choosing not to fight this quarter, and check it monthly against one question: is the product still growing without them?
Principle 6
Hire slowly, then admit when slow became too slow
Truell's most quoted lesson is a correction of his own earlier view: "Many people you hear hired too fast, I think we actually hired too slow to begin with." The slowness had reasons. He wanted "a world-class group of engineers and researchers" with curiosity, intellectual honesty and "micro-pessimism." But the bar was set partly in the wrong place. The early search focused on "well-known school, very young" candidates who had collected high credentials. The company improved when it widened the search to later-career people and unconventional young ones.
The method that lasted was the two-day on-site project. It turned interviewing into a sample of real work, a test of whether the team wanted to be around the person, and a sales pitch all at once. When the product was "comparatively not very good," Truell said, the team was the only thing the company had to offer. The two days let candidates find out whether they wanted to "throw in with us." The a16z description of the process states its purpose: it tests "for agency over credentials."
Tactic: Replace at least one interview round with a paid, multi-day work sample in conditions close to the real job, and treat it as much as a pitch to the candidate as an evaluation of them.
Principle 7
Recruit from the room where your users already argue
Cursor's Discord began as a bug tracker with a social life. It turned into a talent pipeline. "We actually hired many members of [the] Discord server," Truell said. The people most active in shaping the product from outside had already shown the traits the two-day trial looks for: initiative, technical judgment and stubbornness about quality. They had shown it publicly, for free, over months.
This is Halite again. The people who care most about a system are the ones trying hardest to break it, and the ones complaining most precisely are the ones who understand it. A user community is a live, ongoing interview where candidates show their work without knowing they're being evaluated.
Tactic: Assign someone to watch your most active community members for people who give unusually precise feedback, and approach them as recruits rather than as power users.
Principle 8
Build your own models only at the seams
Cursor did not set out to train models. Truell said explicitly that pre-training from scratch "felt a bit like focusing one's attention in the wrong area." It ended up training many, by finding places where frontier models were too slow, too expensive or simply not designed for the job. Tab had to respond within 300 milliseconds and run on every keystroke, and it had to predict diffs rather than text. Retrieval needed a "mini Google search" for codebases. Applying edits required small, fast models to turn a frontier model's sketch into real code.
"Picking your spots carefully, not trying to reinvent the wheel," he said, "focusing on their weaknesses, and how you can complement them." Specialization is the strategy. Use the big models for reasoning. Own the latency, the context and the last mile, where product data gives you an advantage the labs can't easily match. "Every magic moment in Cursor involves a custom model in some way." The magic was at the seams.
Tactic: Map every user-facing interaction to the model that serves it, and train your own only where speed, cost or a task-specific objective makes general models a poor fit.
Principle 9
Grow an immune system
In four years Truell watched InstructGPT, DALL·E, Stable Diffusion, ChatGPT, GPT-4 and the shift to video come and go, "and only a very small number of these things really kind of affects the business." His defense was partly about hiring, choosing people who are "level-headed," and partly about historical perspective. In AI research, he noted, "the vast majority of ideas that have been put out there haven't had staying power," and progress came from "some very simple elegant ideas that have stayed around."
The Caro photograph is part of this immune system, as are the books on forgotten computer companies. Studying long careers and failed institutions trains a founder to ask of each announcement whether it is one of the few things that compound. The question costs nothing. It is still rarely asked by people running fast companies, who are rewarded for reacting quickly.
Tactic: For every major industry announcement, write one sentence on whether it changes your next twelve months, and review those sentences each quarter to measure how good your filter is.
Principle 10
Assume the leapfrog, including your own
Truell predicted Cursor's vulnerability in 2025. He described his market as one where "no matter what entrenchment you build, you can be leapfrogged," and where the moat is closer to a consumer's loyalty than an enterprise contract. You have to be the best product, continuously. He diagnosed Copilot's stall as the result of a market unfriendly to incumbents and the departure of its original team. Within a year, Cursor was dealing with both. Claude Code was the fast newcomer, and Cursor's own head of engineering had left.
His answer was not a defensive moat but R&D at scale: "If you're in a place where big investments, and having more and more great people working on the right path can keep giving you value, then you can get these economies of scale of R&D." The parallel agents, Graphite, the planner-worker research and the "third era" essay all applied that answer to his own company. If the ceiling is high enough for others to jump over you, it is high enough for you to jump over your earlier self, provided you are willing to make your own current product less central.
Tactic: Keep a small team whose explicit job is to build what would make your flagship product obsolete, and evaluate them on whether they succeed.
Principle 11
Treat legibility as part of the product
Cursor's three main public setbacks all came from a gap between what users believed and what was true. In each case the cost was trust, and in each case the fix was a public correction.
!
Three Breaches of Legibility
Each episode undercut the company's core promise of human control.
| Date | Episode | What the user couldn't see | Response |
|---|---|---|---|
| April 2025 | "Sam," an AI support agent, invents a login policy | Policy | Apology, refunds |
| July 2025 | Pro plan shifts from 500 requests to metered usage | Price | Rollback, refunds |
| March 2026 | Composer 2 ships without disclosing its Kimi 2.5 base | Provenance | "A miss," per Sanger |
The analytical point is that for Cursor, legibility is not a matter of good manners. It is the product. Truell's whole account of why Cursor exists, that humans should keep "a ton of control over all aspects of the software" and should be able to "drill down when needed," is a promise to make what's under the surface visible. A company that sells control can't hide its pricing, can't let a bot make up rules, and can't leave out where its model came from. Each of these failures was a small contradiction of the company's own reason for being.
Tactic: Before shipping any change to pricing, policy or model provenance, ask whether a user who relies on your product to stay in control would feel misled. If the answer is maybe, disclose first.
Principle 12
Know which input you cannot manufacture
Cursor could build editors, retrieval systems, fast models and agent hierarchies. It could not produce the thing those depended on: frontier intelligence at scale, and the compute needed to train it. When its largest model supplier became its biggest competitor, that dependency became the strategic problem. Truell's April 21 tweet named the goal: "scale up Composer." The Information named the pressure: Anthropic and compute costs. The a16z partners named the asset: the only operational gigawatt-scale data center and the company that built it.
This principle is the most uncomfortable one, because it ran directly against Truell's stated wish for a "long-lasting, independent, generational company." Independence is worth something only if you control your critical inputs. A founder who admires decades-long projects has to decide whether the long project is the company or the work. Caro, for what it's worth, has always done his life's work inside institutions he did not own.
Tactic: Identify the one input your product can't function without and can't build yourself, and decide in advance under what conditions you would give up independence to secure it.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Michael Truell, on Lenny's Podcast, May 2025More and more being an engineer will start to feel like being a logic designer, and really, it will be about specifying your intent for how exactly you want everything to work.
— Michael Truell, on Lenny's Podcast, May 2025At this point, every magic moment in Cursor involves a custom model in some way.
— Michael Truell, on Lenny's Podcast, May 2025The three-month version wasn't very good, and so I think it's been a sustained paranoia about, there are all of these ways in which this thing could get better.
— Michael Truell, to Andreessen HorowitzWe caught the bug. We couldn't stay away from it. We started working on Cursor at the end of the year; fundamentally, because we wanted to build a tool that we really liked, and a tool that we wanted to use.
— Michael Truell, Compile keynote, 2026Through all of this craziness, there's a group of people whose fundamental kind of core ethos is this group of developers building for developers.
Maxims
- Measure ambition, not crowding. A crowded market where the incumbents aren't aiming high is an open market.
- Feel the pain yourself. If you aren't the user, you can't tell whether the product is getting better.
- Own the surface where change happens. Inherit the commodity layers and control the interface your thesis says will change.
- Be a micro-pessimist. Be optimistic about the decade and skeptical about this week's demo.
- Choose which fires burn. Neglecting everything but the product works only if the product spreads on its own.
- Interview by working together. Two days of real work tell you more than a day of questions, and they sell the job.
- Hire your loudest users. The people who argue hardest about your product already understand it.
- Build at the seams. Train your own models where general ones are too slow, too costly or aimed at the wrong task.
- Expect to be leapfrogged. The ceiling that lets you overtake the leader will let someone overtake you.
- Never hide how it works. When the product promises control, opacity about price, policy or provenance breaks that promise.
