Analyzing the Fastest-Growing Software Category I’ve Ever Seen

Was this newsletter forwarded to you? Sign up to get it in your inbox.

---

The fundamental promise of AI is that it will dramatically reduce the cost of making any product that has a digital component. 

Writers such as myself love to publicly wring our hands about how LLMs’ rapidly improving writing ability may soon put us out of a job. In a satisfyingly ironic twist of fate, though, it’s the field of software engineering that has so far been hit hardest. The very people inventing the technology that may put many knowledge workers out of a job have been among the first to experience disruption, with software engineering job postings at a five-year low. This change is partially due to the tens of billions of dollars that have been invested in building AI coding tools, which have proven incredibly effective.

These startups have grown faster than I have ever seen in my career in the tech industry. There are individual companies in tech history that have grown at somewhat similar rates, but I’ve never seen it happen to so many organizations in the same category: 

• Cursor went from $1 million to $100 million in annualized recurring revenue (ARR) in less than 12 months.
• Coding agent startup Lovable grew to $10 million in ARR two months after launch.
• Anthropic’s Claude series of models are best in class at coding and have seen explosive growth—growing 40 percent this year alone by hitting $1.4 billion in annualized revenue. 

As such, automated coding tools are a useful litmus test for what happens to the rest of us as the models improve. What does work automation look like? What kind of companies are winning? Even if you don’t code, you should care about our brothers and sisters of the coding faith—they are the over-compensated sacrificial lambs whose blood sacrifice will help us divine the future.

These new tools teach us three things:

1. Software wins on the basis of what work it replaces and where in the tech stack it resides.
2. There are only two choices that matter: whether to build a model, and how to position your product relative to your peers.
3. Beware AI revenue (but not for the reasons you think).

Let’s unpack them.

What makes these companies so successful?

AI can either integrate into an existing developer’s workflow (i.e., help me write the exact code I want faster) or replace the need for the developer (i.e., make an application that does something I want).

You’ll note that both of these jobs-to-be-done are tool agnostic. “Write code gud” or “me no want to do spreadsheet automation” are things that many other existing solutions have attempted to tackle. This matters far more than you might think.

Typically, when people buy software to use at their company, they aren’t looking just at what that software can do. Not only does it need to do a particular thing, but it also has to work within the context of an existing software stack and how the employees use them. Just as you would never buy budgeting software that didn’t integrate with your credit card, neither will an enterprise purchaser buy an expense management tool that doesn't talk to their accounting system.

It is remarkable how many founders I meet that get this wrong, so it bears repeating: Users need to understand exactly how your new boondoggle fits into their existing workflows.

This is why AI coding tools are so compelling. Whether they create entire applications or just help engineers code faster, these tools sit independent of the standard software stack. They’re extremely low-friction to set up and, even if you’ve used one for months, require little effort to switch away from. One engineer I talked to has changed his AI code editor every month for the last four months, just because he can. From the perspective of the companies behind these tools, that makes acquiring new customers extremely easy.

When you add in the fact that high-quality software engineers are exceedingly rare and functioning code is one of the most expensive things in the world to generate, you have a perfect recipe for explosive growth. However, explosive growth also brings explosive competition.

Context is king

When a founder is thinking about selling their product, they have a crucial decision to make: Do I build something no else can make, or do I focus on one specific customer type more than anyone else possibly could?

In the AI coding world, there are three buckets of companies that matter:

1. Selling the model: The bet by these companies is that if you make the models generating the code good enough, many of the problems around memory or integrations will sort themselves out. Competitors include the big research labs of OpenAI and Anthropic, but also relatively smaller competitors like Poolside and Magic. Customers define what “good” means—ranging from having memory to intelligence to security compliance—but these companies are all gunning to build and sell the very best proprietary model they can.
2. Selling to developers: Tools like Cursor, Windsurf, and Devin aim to replace development environments like VScode by wrapping applications around LLM’s capabilities. The AI acts as a pair programmer, helping the developer write code faster. In the hands of talented engineers, they obviate the need for junior staff and allow the production of code at a significantly faster rate.
3. Selling to everyone else: These tools attempt to bring the full capabilities of a developer to someone who isn’t familiar with code. Tools like Lovable, Replit Agent, and v0 aim to be an engineer in a box. Generally speaking, the applications or webpages these tools make aren’t currently ready for the enterprise use cases, but they are good enough for personal applications. Because code is such a flexible digital building material, these tools can be used to produce web pages, workflow applications, and even video games.

The trillion-dollar question for these companies is whether models are ever going to be good enough to not require significant technical scaffolding to help them accomplish their task. For categories two and three, the models cannot do what customers need them to do out of the box. So the companies make a highly complex string of automations, prompts, and tools to guide the AI. The LLM “brain” is amazing! It is crazy that this even works. But for now, it still requires lots of direction for specific applications.

For all three categories, context is king. The more an AI model can know about your codebase, your taste in code, and what problems you need to solve, the better the output will be. That context may come from enhanced memory or fine tuning (the bet that model providers are making), or from careful prompting and tooling (the bet the application layer is making).

What matters is that the context is legible to the models. The model needs to have data about your circumstance, either given in its training or in instructions fed to it, to be successful. So the more easily explainable a task or job function is, the less it relies on secrets, taste, and intuition.

Since it’s near and dear to my heart, let’s use the field of writing as an example of how the impact of these tools is likely to be felt. Writing at the highest levels of skill will see very little change from AI writing models. Great writing has tens of thousands of subtle unexplainable cultural touches, things that even the writer would struggle to tell you why they did. However, good enough writing, like SEO or explainers, will see large changes in workflow.

These changes for software engineers are not all bad. Yes, job postings are down, but really talented engineers can create so much more than they used to. SEO writers can produce far more high-quality output than before. Knowledge work will likely continue—it’ll just be concentrated among people at top-tier skill levels. For them, AI will be a superpower allowing them to accomplish volumes of work at a high quality.

The law of publicly released numbers

Whenever you see a data point offered by a CEO in a TechCrunch article, you should think to yourself, “In what way is that number suspect?” A well-trained founder will try to remake public perception with a drip of data that warps how we think. Take, for example, the growth numbers from Lovable and Cursor I mentioned earlier—they specifically call out huge growth in little time, giving the perception that they are unstoppable juggernauts.

This is technically true. However, they will usually say “ARR,” but I spelled it out earlier for a reason: Most readers are likely to assume that ARR refers to “annual recurring revenue,” but sometimes it’s actually “annualized recurring revenue.” Companies multiply their current monthly revenue by 12 and hope we don’t notice the difference. In the case of products that have long sales cycles and complex integrations, this practice is justifiable—churn is very low in traditional enterprise software, for example. However, AI coding tools often cost 20 bucks a month, and as I mentioned before, switching costs are near zero. They’re easy as pie to churn. I talked with venture capital investors about the AI coding startups they were seeing and all of them expressed some form of hesitation about churn. To some, it was a deal killer, to others, no big deal, but the sector’s investors are keeping a close eye on this metric.

Even though these tools are currently focused on differentiating on capabilities or customers, what matters for long-term success in software is power. A software provider either has to hold the key data, generate unique sources of demand, or offer an experience that no one else on the market can beat.

This is why I have been cautious about participating in any of these AI-first software deals. Because the capability race is so fast, relatively few founders are paying attention to the stuff that actually matters in software: building long-term defensible moats and compounding with positive net dollar retention. You want to earn more money every year from each cohort of customers, not be reliant on industry tailwinds. My sense is that few AI coding tools are building that moat right now and instead are betting on improvements in models to make the winning strategies more obvious.

Hold me, I’m scared

It’s ironic that software developers are the first to face disruption. Why? Because coding is an expensive, digitally native, and relatively context-light profession. In other words, it's the perfect candidate for AI takeover.

But not all AI takeovers are created equal. Despite the explosive growth and eye-popping revenue numbers, most of these tools are on thin ice. Their switching costs are near zero. There is a big risk of churn. And beneath the carefully manicured headlines that founders drip-feed to TechCrunch lies a market built less on durable advantage and more on temporary convenience. This is software's version of musical chairs—everyone looks good when the music’s playing, but you better have a seat when it stops.

What really matters—and what's true across software—isn't raw capability. It's how neatly your product fits into the existing workflows, integrations, and habits of your customers. AI that merely generates better code is replaceable. But AI that understands exactly who it's serving and why—that weaves itself deeply and inseparably into a developer's life—is invaluable.

The broader lesson, and the reason everyone should care about software engineers becoming capitalism’s latest sacrificial lambs, is simple: What's happening to developers today will happen to knowledge workers everywhere tomorrow. Any role that’s context-light, digitally native, and easy to switch out is up for grabs. The middle will thin out, with value flowing to either end: to those who build irreplaceable systems and integrations, and to workers whose subtlety and intuition defy easy automation. AI doesn’t destroy jobs evenly—it concentrates them, reshapes them, and makes the winners disproportionately valuable.

---

Evan Armstrong is the lead writer for Every, where he writes the Napkin Math column. You can follow him on X at @itsurboyevan and on LinkedIn, and Every on X at @every and on LinkedIn.

We build AI tools for readers like you. Automate repeat writing with Spiral. Organize files automatically with Sparkle. Write something great with Lex. Deliver yourself from email with Cora.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.