Limitless Podcast - The Story of Cursor: The Fastest AI Company and Massive Comeback

Episode Date: May 25, 2026

Things people aren't talking about enough: Cursor’s rapid growth, its role as an AI coding harness, and the data and training advantages that may support its moat. With the reported SpaceX and xAI deal, there's a real possibility of combining Cursor with Grok and large-scale compute.------TIMESTAMPS0:00 The Cursor Story1:54 The Harness Moat4:51 Budgets and Data8:59 The Deals12:31 The Vision14:58 Concerns17:23 Final Takes------🌌 LIMITLESS HQ ⬇️NEWSLETTER:    https://limitlessft.substack.com/FOLLOW ON X:   https://x.com/LimitlessFTSPOTIFY:             https://open.spotify.com/show/5oV29YUL8AzzwXkxEXlRMQAPPLE:                 https://podcasts.apple.com/us/podcast/limitless-podcast/id1813210890RSS FEED:           https://limitlessft.substack.com/------RESOURCESJosh: https://x.com/JoshKaleEjaaz: https://x.com/cryptopunk7213------Not financial or tax advice. See our investment disclosures here:https://www.bankless.com/disclosures⁠

The story of Cursor is quite possibly the greatest comeback story in the history of AI. Just 90 days ago, some of the smartest people in tech had completely buried this company. They gave up on it. And Cursor was just an AI rapper that had raised too much money and Athropic was about to crush them. Then seemingly overnight, everything changed. If you look back to January of last year, they were running at about $100 million in ARR. Fast forward to today, that number is $3 billion. This is one of the fastest software ramps in history, way faster than Open AI.

So why? And on this episode, we're going to discover the moat that just about every single critic on the internet missed. To truly understand how impressive of a run-up Cursa's had, we have to start from the beginning. So Cursa was founded by four MIT graduates in 2022. Their founder, Michael Trull, is only 25 years old right now. He was 23 when it was founded. So it's an incredibly young team. All four of these founders, by the way, are on the Forbes 30 under 30 list. So massively successful. young company, but what were they building? Well, they discovered a problem. If all these new AI models like Claude, like ChatGPT could code for people, you would still need some sort of a system, some form of a harness, some form of a coding editor to be modernized around new AI coding models. Traditionally, you would have an IDE and integrated developer environment and you would use this to kind of like edit your software, see what the code looks like in reality, like right next to it side-by-side screen, but there was nothing quite for that when it came to coding AI models. Cursor wanted to be that. And so they forked VS code, which is a popular open source software

program from Microsoft, and they created what critics ended up calling it a wrapper, a harness around AI models. Now, initially, the criticism was, this is just an AI wrapper and the actual brain. The AI models are the smartest thing and the most valuable thing, and they'll eventually replace the wrapper. That actually didn't end up happening. If you look at this crazy growth curve that we have on the screen right now, they started off in Jan, 2025 with $100 million ARR. That then five-xed a few months later at $500 million. And then by the end of 2025, they were at $1 billion. And then we just got the news this week that they hit $3 billion.

Now, what's crazy about this is their projection for the end of this year is $6 billion. I actually think they might kind of blow the roof on this entirely. it sounds similar to Anthropic because Anthropics started with Claude Codod and they had this kind of like meteoric exponential rise. The same thing is happening with Cursor, but they have a very different mode.

67% of the Fortune 500 companies all use Cursor and they went from being the hottest commodity to the most criticized commodity to now potentially being acquired by one of the biggest frontier AI labs in the world. Yeah, it's amazing how quick the ties have turned. I remember back in the day, Cursor was actually the hottest piece of software in AI

because everyone was using it to write code. Then the AI models that came out like ChatGBTGBT and Anthropics Claude, they kind of crushed it. Everyone forgot about it. The comeback story is huge and it's a testament to where the AI industry is headed now. A big part of this we're going to talk about all the time is harness. This harness, if you think about like a human being and the LLM being the brain, a basic LLM is kind of like a brain that's been concussed a thousand times. It has no memory. It has no ability to kind of function outside of itself.

The harness is the body. It's the containing mechanism that provides the system with. memory. It provides the system with custom tool sets and inputs and outputs, and it gives it everything it needs to think critically. So that's why he'll notice that on some of these harness companies, they're actually just making a living off of building harnesses for AI models in the hopes that they could do so better than companies like Anthropic, better than companies like Open AI. And funny enough, even though they made some progress in creating these harnesses, people were still kind of writing them off as trivial and non-necessary. Because, I mean, if you think about it,

Anthropic has the intelligence, opening eye has the intelligence, why it's just a rapper going to make a difference on that? And I think that was the single point that a lot of people got confused on and tripped up on. And that's not necessarily true. Yeah, I think the best way to analogize around this is if an AI model is the brain, you still need a bunch of appendages. You still need the eyes to take in data. You still need ears to hear things. You still need a body, a medium, to kind of like translate that intelligence into the real life, right? That is basically what cursor is to all these different AI models, and that's what they have proven.

So what is this exact product? What is this exact moat that cursor has? It's known as the agent harness, as you just mentioned. And the way to think about it is it's kind of like a scaffolding that turns like a raw model into an autonomous coding agent that can manage its time well, learn when to use a prompt for what particular intent or purpose, and most importantly, do it efficiently.

There was news that broke this week that Microsoft is now relinquishing all their employees' access to ClaudeCode because they're just using way too much money of the budget that Microsoft has for it. Uber, they had an annual budget at the start of this year for all their employees. All their employees burned through it in four months.

It's completely gone. So everyone's trying to reassess how the best way is to use the different coding models. It's kind of like they're using it as kind of like a blunt tool. they're kind of like throwing kind of like stuff at it and like hoping it would like come up with an answer. It's very inefficient. Curser is the number one product to kind of like formalize all of this and make you kind of like an additional genius on top of like what you're already prompting with these

coding models. So the way that the harness works is you kind of build up different aspects of it. You kind of like teach it how to use muscle memory in the correct way when it comes to specific flows or coding programs that you're working on. It teaches you how to manage your work. It spins up different types of agents. It gives the agent's specific personalities. It augustrates these agents across effectively all the workforce. So if you're a company, this is hugely important because you have many different subteams that are doing this particular thing. And then of course, this is translated into a distribution for cursor themselves. So I mentioned companies, 67% of Fortune 500 companies use cursor. And they may say like, hey, we used Claude Code or we

use GPT 5.5, but it is through cursus. So they're the surrounding environment that has been able to build this. Yeah, when we think about moats, there are some pretty unexpected ones that I found out in researching about cursor. Did you know that three quarters of cursor's final compute was their own reinforcement learning training? A lot of it was done internally and proprietary. This isn't stuff that you can find on GitHub. This isn't open source data. And they also have a tremendous amount of data on how millions of developers actually accept, reject, and edit their suggestions. And that feedback loop is a really big data set that's also proprietary. So they take all these data sets, they loop them together, and they build in this really

refined reinforcement learning stack on top of the harness that creates a bit of a moat that other companies don't have, and that data is incredibly valuable because a lot of other companies, they'll create a wrapper, it's just an API to another model, they don't actually get the full training loop there. Curser has figured this out and they've done so. And their new model proves this thesis because now Curser isn't just a harness company. Curser is actually creating very highly intelligent models most recently with Curser 2.5, which I believe is just as good as Claude Opus 4.7 and GPT 5.5 high, at least on some benchmarks. Like this is incredibly impressive to go from just a wrapper company to an incredibly powerful harness with a leading edge frontier model. I mean, we just

covered Google last week, they didn't even get to the frontier and now suddenly Cursor is. This is a pretty big deal. Okay, so some fun facts from their recent model that I really, really enjoy. So number one, it is trained on a completely open source model, which is significantly smaller than the current frontier models that come out of Anthropic or Open AI. It's called Kimi K2.5. Now, if you are reading that or listening to that and thinking, isn't there like a Kimi K2.6, you'd be right, but they used an older model and still built a harness around it that competes directly with Opus 4.7 Max and GBT 5.5 extra high. I'm showing you on the screen here. If you look at it, the discrepancy is literally like 0.5 or like a one percentage point. But a few things that I want to

suggest here is look at how cheap Composer 2.5, which is Curse's Frontier model, is compared to GPD 505 extra high and Opus 4.7, despite being as good. It is four times, sorry, it is eight times cheaper than GPD 5.5 and 11 to 12 times cheaper than 4.7. But I also want to point out that another Frontier AI Lab released their model this week. It is Google Gemini 3.5 Flash, and it not only is dumber technically than Composer 2.5, this harness that is wrapped around just a small, careless open source model, but it is also four times more. expensive. So the point I'm trying to make here is this proves the thesis that the harness is the moat. And actually Sam Altman recently commented on this saying, I don't get all the criticism

around agent harnesses. It is quite obvious that you have foundational pre-training and post-training of a foundational model, but then you also need to wrap all this orchestration layer around that. And Curser is the only one or the number one company right now to have built that mode that Sam Altman himself tried to acquire and they reject it. That's funny. Yes. I got to of Sam Credit, he saw this coming, he knew what was happening, he tried to acquire them. I believe the number was $3 billion at the time. Now their ARR is double that. So clearly it was a good deal that he would have had. They rejected it for a good reason. But there is a new entrant into this deal space.

We've mentioned it before, if you've been listening to the show, it's Elon and SpaceX. They have acquired the option to purchase cursor. Now, that option was $10 billion guaranteed to cursor. No matter what happens, cursor gets $10 billion, SpaceX gets paid out, zero for that. That gives them the right, the option to buy cursor outright for $60 billion within the first 30 days of their IPO. Now, we have new information on this IPO and that it is likely going to launch June 12th. So SpaceX is now on the clock. They have until the middle of July to decide if they want to spend $60 billion or not

to acquire this company. It seems as if the decision has already been made. it seems that Elon is going to indeed buy SpaceX, and now the new number is 20 times what Sam Altman offered at $60 billion. So clearly the value is being realized, and now this is being bundled into a pretty impressive frontier AI company. I mean, SpaceX has the data centers in space. They have all the infrastructure. They've been leasing out their Colossus data centers to Anthropic, but all of those resources are going to be available to the cursor team. And I have to imagine that combining XAI with Cursor is going to create a pretty meaningful run on

the new frontier. And Anthropic and Open High are going to have some serious competition in coming. Yeah, I think it's important to point out why this deal is so symbiotic between both of these companies. So Cursor runs into a problem, right? They have this amazing agent harness, but they're always reliant on waiting for Frontier Air Labs to reach a new model and then build a new harness around them, right? So they're like, okay, well, I wish we could build our own models, but like we don't have enough compute. So I guess we'll take this open source model and wrap our harness around it. Ideally, you get access to the foundational frontier American model first. So Elon saw this and said, hey, not only will I give you the compute, I'll give you access to my frontier model,

Croc, which gets to train on a million plus bleeding edge GPUs that I have in Colossus 1, 2, and in the future, three. So Michael Trull saw this and was like, hell yeah, we'll take the compute and you give me an optional buyout of $60 billion if this goes well, or $10 billion for the compute if this goes like a pear-shaped. So that's completely fine for me. Now on Elon's side, he wants nothing more than to beat Sam Altman, Open AI specifically. Currently, Sam Oatman is sitting somewhere near the top. So he did two things. He partnered with Dario, Anthropic recently, to give them compute and make more money off of that. But most importantly, he intends to acquire Cursor specifically

so that he can get access to this harness and wrap it round GROC. So GROC overnight ends up becoming probably somewhere near the number one coding AI model and that instantly puts Elon in the mix for the most important AI race in the world, which is coding.

Yeah, like if you believe that AI writes most software over the next five years, then Cursor becomes a toll booth. And this is something that GROC didn't have previously. They've been trying to build a coding model. it hasn't been going super great. They're training some now, but Cursor gives them a shortcut right to the front and owning that toll booth is far more valuable than owning the cars that run by it. And that's basically what GROC was prior to the acquisition of Cursor and the harness. So now Elon is assembling this full

vertical stack. He has rockets to deploy the satellites, satellites to deliver bandwidth, X to capture attention, Colossus to train the models and Cursor to write the software as the toll booth. Then there's GROC on top of all of this to talk to the users. It's this really unbelievably compelling stack. And I think SpaceX has this super unique ability to kind of change the AI industry at their will. Like currently they're empowering Anthropic by giving them the data centers. But we both know there's a three-month option on that that they're not beholden to. So in the case that cursor does really come up with some breakthrough training, in the case that cursor comes up with some breakthrough models, it needs all the training data. They can claw that back. And it sounds like

they're kind of testing that now with these larger models. I mean, this post is teasing Elon with a half trillion GROC model and then a 1.5 trillion. trillion parameter model. So they're increasing the size fairly quickly. Yeah, Elon has publicly stated recently that he has got a 15 trillion parameter model and in the future, very near future, a 20 trillion parameter model. Now, the 15 trillion parameter model plans to go live in July. If he's using Cursors, Agent Harness and Datamote to feed GROC into becoming a frontier coding air model, we may get this number one frontier coding model to compete with Anthropic and Open AI as soon as July, which is insane, given that the IPO is happening very soon,

and they're looking for different new revenue incomes, and they need a reason to kind of propel them even further. And then, like, there's the other part of this, which you kind of touched upon just now, which is you obviously need so much more compute to kind of, like, feed these different models. And SpaceX is the number one company to scale that massively. They're doing this through space, through AI data centers,

which people still think is a fairy tale, but ends up probably becoming the most feasible one, given all the power constraints on Earth. So cursor not only gets acquired by a frontier AI lab, by a guy that is obsessed with scaling things. And so, like, if you can scale the harness, it's a win-win for them, but also the tollway booth to infinite compute that could make Composer 2.5, 3, 4.5, 5, the number one leading model in the entire world. It's pretty cool. It seems cool. But there are a few risks.

There's a few things that we need to look out for, a few landmines that still could be stepped on, things to take note of one of those being that Composer 2.5, it does still run on Kimi K2.5 weights, which is the model from the Chinese lab, Moonshot AI, and Cursor runs it on Cursor's own infrastructure. So user code routes to Cursor servers, not moonshots. So the data is defensible, but the actual model is not. We'll see if they could continue that with the GROC models. I'm assuming that's what's going to happen is cursor is going to shift their harness to fully support GROC, and we'll see. I mean, XAI has not historically been at the frontier, been able to create the most capable

models. I will say they're the most price performant models. And that's really interesting for the case for a cursor. But it is something to note that the current underlying technology is still that Chinese lab from Moonshot. It's not an actual AI model that they own. So that is one wildcard that I think is noteworthy. I see the other criticism is also the commodity thing, right? They're saying, okay, well, if Curse has been able to build this agent harness, what's stopping a really well-funded competitive from doing the same. The same reason why you would give that to Anthropia, you wouldn't make that argument to Anthropica Open Air.

They have the lead right now. And I would actually argue that they have a more distinct lead because the harness isn't just one thing. It is multiple singular things that you need to kind of orchestrate and combine into one thing. You've got the memory side of things. You've got the agent creation. You've got MD file management.

You've got all these tool orchestrations. There's so many different things that cursor does well, which is exactly why Winsurf and a bunch of other competitors hasn't really been able to translate what cursor has been able to do. So I don't think that's a major issue. I also think, like, to be honest, like, the Chinese open source model, whether it is that or not, isn't really a point of contention here, because I know when they get access to a Western model, specifically GROC, they'll be able to do an even better job. The proof in the pudding is they took Opus 4.7 and they took GPT 5.5 and they wrapped Cursor's harness around it

and it ended up giving them a 10 to 20% percentage point increase across all key benchmarks. So the point is, the agent is definitely very valuable. And I think if Elon and Groc and Cursor work closely together right from the foundational model training, you know, they see what the architectures being used on, a Colossus one right from the ground up, they'll be able to create something pretty special, in my opinion. Yeah, I think I am going to remain pretty bullish on SpaceX, bullish on Cursor. I mean, we saw what happened when OpenAI acquired WinSurf.

codex very quickly got very good. And I'm sure they took a lot of that data and rolled it into it. And now codex is probably the top coding harness, if not very close right neck and neck with Claude Code. So these things do make a meaningful difference. It's clear that SpaceX has been trying to get into the world of coding. Curser is going to be that company. One thing to note is when this company sells for $60 billion. The CEO is 25 years old. We mentioned this on the roundup last week, if you were listening. But he's probably going to become the youngest billionaire in the world. That's pretty exciting. That's got to feel good. good. And I wonder what the integration between those companies looks like. A lot of these acquisitions

that we find are aqua hires. They are to acquire talent. So we'll see how the talent gets distributed. But a huge exit nonetheless. I mean, I remember, what was it last year? Maybe it feels like an eternity ago, but like cursor actually was the hottest software on the block. Everyone was using cursor. Everyone was really excited to write agentic code. That was when vibe coding didn't even exist. It could just kind of help you complete sentences. Now the whole world has shifted. And cursor has gone through the entire narrative cycle of being the hottest thing in the world to being dead and now being right back at that frontier helping XAI push themselves to the frontier. So it's a huge comeback story and just like it's a good time to be bullish on cursor. You know, I just

realized something. There's a really hot new job role that is becoming more pervasive in AI and it is called the forward deployed engineer. Anthropic and open AI are like obsessed with this role. And what it is is basically an engineer slash consultant, which, goes into enterprise companies and teaches them how to use Claude or GPT in an effective manner. And the number one problem they're trying to solve is how does all this tool use actually make sense? Like we need to create custom tool use. We need to create a custom agent. We need to create custom orchestration management.

Sounds pretty much like cursor. I wonder if there's a world in the future where you don't need to get a forward-deployed employee to come in and like kind of build it from the ground up, custom systems for your company. you just have Curse, which has an orchestrated tool, which does all of this for you. And that becomes the kind of like new SaaS version of this forward-deployed engineer. Either way, I'm impressed with Curses kind of like bouncing around, but like undeniable assent as of recent. At 25 years old, Forbes 30 under 30, he's not in jail. So he's hopefully breaking the stereotype.

We still have time. We have around probably 60 days until the IPO and potential acquisition. But that is the story of Curse. was a really important story to tell, mainly because it was burst off of my, I'll speak for myself, my own kind of like incorrectness around calling cursor a rabbi. Sometimes last year I was like, they're going to get replaced by Anthropic. But we had a learning recently that I was like, that's definitely not the case. And they have a moat that we haven't spoken about on this show so far. So we hope you found this informative. And we're going to keep tracking all of these processes

going forward. Josh, do you have any final words? Well, that's the moat. That's the change of heart. In fact, you wrote about it this week on our newsletter. So if you're not subscribed to the newsletter, you can do and go read that piece on where this change of heart came from. And kind of a complimentary version to this podcast episode. If you enjoyed, please don't forget to share it with your friend. Leave us a nice comment, letting us know what you think about, maybe about cursors, SpaceX IDEO, you investing. You think this is bullish or not. Share it with your friends.

Don't forget to give us a five-star rating wherever you get your podcast. And like always, we will see you guys in the next episode. Thank you so much for watching. See you guys.