Everyone's Launching an AI Startup — Here's Who's Actually Winning

A developer recounts the AI startup gold rush, noting that out of 400 Twitter followers, 61 claimed to be building AI startups. The developer built a 'smart' email assistant in late 2024 that failed to gain traction, and observes that while many rush to launch AI products, the real winners are those selling the 'shovels'—the infrastructure and tools—rather than the applications themselves.

I counted once, out of pure curiosity. Scrolled through my Twitter following list and tallied how many people had "building an AI startup" or some version of it in their bio. Out of maybe 400 people I follow, 61 of them were building something with AI in the name, the tagline, or the pitch. Sixty-one. That's not a niche anymore. That's basically everyone I know who touches a keyboard for a living. And look, I get it, because I was one of them too. Back in late 2024 I spent a weekend building a "smart" email assistant that was, if I'm being completely honest with you, a system prompt and a nice UI wrapped around an API call. I put it on Product Hunt. It got 40 upvotes, three DMs from people asking if it did something it didn't do, and then it just... sat there. I still have the domain. I pay $12 a year to keep a graveyard. That's the part nobody puts in the launch tweet. Everybody shows you the "we hit 1 on Product Hunt" screenshot. Nobody shows you the folder on their laptop from six months later, the one they haven't opened since, with a README that still says "TODO: add payments." So here's the thing I actually want to talk about. Not another "AI is changing everything" post — you've read forty of those this month already, I don't need to be the forty-first. I want to talk about the boring, unglamorous, actually-useful pattern I've noticed after watching a lot of these companies launch, and a smaller number of them survive. Because a few of them are genuinely winning. Making real money. Growing without burning cash they don't have. And weirdly, almost none of them look like what you'd expect. Whenever this many people rush into one space at once, it's worth remembering we've seen this movie before. Not the AI part specifically — the rush part. Every gold rush works the same way. A resource becomes suddenly, dramatically easier to access than it used to be. Word spreads. Thousands of people show up with pans, convinced they'll be the one who strikes it. Most of them go home with nothing but blisters. And the people who actually made consistent, boring, reliable money weren't the ones panning at all — they were selling the shovels, the tents, the overpriced jeans. That's basically what happened with app stores in 2010. That's what happened with "get rich with dropshipping" in 2015. And it's what's happening right now with AI, except the barrier to entry dropped even lower than an app store ever did. You don't need to know how to code anymore, not really. You need an API key, forty-five minutes, and the willingness to call something "AI-powered" in your landing page copy. The problem isn't that this is bad. Honestly, it's kind of amazing that a solo person with no funding can build something that would've taken a team of engineers and a year of R&D back in 2018. The problem is that "easy to build" and "worth building" got confused for each other, and a lot of people are finding out the hard way that those are two completely different questions. Let me describe the exact thing I did with that email assistant, because I think a version of it is happening to at least a third of the people reading this right now. You take a general-purpose model. You write a decent system prompt for a specific use case — summarizing emails, writing product descriptions, whatever. You slap a clean interface on top. You launch. And for a few weeks, it genuinely feels like magic, because it kind of is magic, the model really is doing something impressive. Then a few things happen, usually in this order: First, someone points out you can basically get the same result by just pasting your email into ChatGPT directly. You didn't want to hear that, but they're not wrong. Second, the model provider ships an update — maybe a memory feature, maybe a native integration — and suddenly the thing you built for months is a checkbox in someone else's settings menu. Third, and this one stings the most, a competitor looks at your product for twenty minutes, reverse-engineers your prompt, and ships their version in a weekend. Because there was nothing to protect. No moat. No data you'd built up. No workflow anyone was locked into. Just a thin layer of UI on top of someone else's brain. I'm not saying this to be harsh about it — I built exactly this thing, remember. I'm saying it because the founders who are actually winning right now figured this out early, sometimes the hard way like I did, and changed the question they were asking. They stopped asking "what can this model do?" and started asking "what specific, annoying, expensive problem can I solve using this model that the model by itself absolutely cannot solve?" That's a small shift in wording. It's a massive shift in outcome. I've talked to a good number of founders over the last year — some through my agency work, some just from being around in dev communities, some from cold DMs after reading their launch threads and asking "okay but what's actually happening under the hood here." A pattern keeps showing up. It's not exciting. It's kind of the opposite of exciting. But it's consistent enough that I trust it. Nobody's going to lean in with interest when you say "I built AI tooling for construction permit compliance." People's eyes glaze over a little. But that founder is closing deals worth five figures a month while the "AI journaling companion" app with a beautiful landing page and 30,000 downloads has never made a dollar. Boring wins because the buyer already knows exactly what the problem costs them. Nobody has to be convinced that manually checking permit compliance is expensive and annoying — they live it every day. There's no education tax on the sale. And because it's not glamorous, there's a lot less competition crowding in, because most people building AI products right now want to build something they can show off, not something a compliance officer needs. This is the part that actually matters long-term, and it took me a while to really get it. The moat was never the model. It's whatever mess of data, context, and workflow sits around the model that took months of painful, unglamorous work to accumulate. Maybe it's a dataset of real customer interactions that gets the product smarter every week in a way a brand-new competitor can't replicate on day one. Maybe it's being so deeply wired into how a team already works that ripping the tool out means retraining an entire department on a new process. Maybe it's the hundred small edge cases you fixed after actual users broke your product in ways you never predicted, and a copycat has none of that scar tissue. None of that shows up in a demo video. All of it shows up in a renewal rate eighteen months later. Here's something I noticed rereading a bunch of "we hit $1M ARR" posts from founders actually doing well: almost none of them lead with "we use AI." They lead with the outcome. The invoice got processed in four minutes instead of two hours. The support ticket got triaged correctly on the first try. The contract got reviewed before the client's lawyer even finished their coffee. Nobody buys a car because they're excited about the torque specs on the engine. They buy it because it gets them to the thing they actually wanted to get to. The founders doing well seem to understand, almost instinctively, that customers don't care what's under the hood. They care whether the pain went away. This is the part almost nobody says out loud in public, because "AI made me 10x faster" gets way more engagement than the truth. I've used AI coding tools heavily in client work for over a year now, and if I'm being straight with you, the honest multiplier is closer to two — not ten. Some days it's barely 1.2x, when I'm untangling something the model got confidently wrong. Some days it really does feel like 5x, when it's boilerplate I'd have hated writing anyway. The founders building durable products seem to have internalized this same honest math for their own users. They're not promising a 10x miracle. They're promising something believable — save an hour a day, cut review time in half — and then they actually deliver it. Under-promising and consistently delivering turns out to be a much better growth strategy than a viral tweet that sets expectations nobody can meet. This one's less about the product and more about what happens the day you launch it. I've watched two nearly identical products launch within a month of each other. One founder had been writing in public for over a year — dev.to posts, a small newsletter, replying to people in his niche. The other had a genuinely better product, by most technical measures, but had never published anything before launch day. Guess which one had actual signups on day one, and which one posted "check out my new AI tool " into a void of twelve likes, nine of which were from other founders doing the exact same thing to him in return. Distribution isn't something you bolt on after the product's done. It's infrastructure you build in parallel, usually slower and less visibly than the coding part, which is probably why so many people skip it. Here's what I think is the most honest thing I can tell you about this whole space right now, and it's the part that doesn't make for a great headline. Most AI startups aren't dramatically failing. There's no crash, no dramatic "we're shutting down, thanks for the ride" post. What actually happens is quieter and, honestly, a little sadder to watch. A founder launches, gets a nice initial spike of users off a launch post, and then growth just... flattens. Not to zero. To a slow, thin trickle. A few hundred users. Some revenue, but not enough to quit the day job. A Discord server that used to have daily chatter and now gets a message every few days, mostly from the founder trying to keep it alive. They're not dead. They're also not winning. They're drifting on the momentum from launch day, running the numbers in their head every Sunday night, telling themselves the next feature will be the one that changes things. If any of that sounds familiar and it's your product — I want to be clear, that's not a verdict on you as a founder or a developer. It's information. Every founder I've talked to who actually broke out of that drift did the same unglamorous thing: they stopped adding features and started going back to actual users, on actual calls, asking what specifically was still painful. Almost always, the answer wasn't "we need more AI." It was something much smaller and more human — a confusing onboarding step, a price that didn't match the value, a workflow that almost fit but not quite. If you actually go looking — not at who's trending on Twitter, but at who's hiring, who's raising follow-on rounds without a press release, who's quietly doubling headcount — a pattern shows up that's almost the exact opposite of what gets attention online. Vertical tools built for boring, regulated, high-stakes industries are doing well. Legal document review. Medical coding and billing. Insurance claims processing. Compliance monitoring for financial services. These aren't the products getting viral demo threads, but the buyers understand the cost of the problem in exact dollar terms, and that makes the sale a lot less about hype and a lot more about math. Internal tools built for large companies are doing well too, and for a reason that's easy to miss — they don't need viral growth at all. They need one internal champion with budget authority who's tired of a specific process being slow. That's a completely different, much more forgiving growth motion than trying to win over the general public on Product Hunt. And then there's the group that gets almost no attention but is quietly one of the most durable — agencies and small studios that use AI to deliver client work faster and better, without ever trying to sell "AI" as the product itself. They're not pitching investors. They're just doing good work more efficiently and reinvesting the margin into building their own tools on the side, at a pace that doesn't require anyone's permission. None of these make for an exciting screenshot to post. All of them tend to still be around in two years, which honestly is the only metric that matters. If you're reading this while your own AI product is somewhere between "launched" and "not sure it's working," I'm not going to pretend there's a magic checklist that fixes it. But there are a few things worth doing before you write another line of code or ship another feature. Go talk to five real people who have the problem you're solving. Not a poll on Twitter — an actual call, screen-shared, watching them do the painful task with their own hands. You will learn more in thirty minutes of that than in a month of guessing. Build the smallest thing that removes real pain, not the most impressive thing for a demo video. Those are almost never the same product, and chasing the demo version is how a lot of people end up with something flashy and unused. Charge money early, even a small amount, even before it feels ready. Free users will tell you your product is cool. Paying users will tell you, very quickly and sometimes bluntly, whether it's actually valuable. Build in public while you're building, not after you've already launched and need people to notice. Your audience isn't marketing bolted on at the end — it's part of the actual foundation. And plan for the model to keep getting cheaper and better, because it will. If your whole value proposition evaporates the day a bigger company ships a native feature that does the same thing, you never had a product. You had a temporary head start, and head starts run out. Everyone's launching an AI startup right now because, for the first time, almost anyone genuinely can. That part is real, and it's not something to be cynical about — it's probably the most accessible moment to start something that this industry has ever had. But being able to launch something was never the hard part. It never was, even before AI made it easier. The founders actually winning aren't the ones with the flashiest feature or the loudest launch thread. They're the ones who found a problem specific enough, painful enough, and well understood enough by the person paying for it, that the AI itself basically disappears into the background. Nobody's buying the model. They're buying relief. The rush is going to keep bringing thousands of people in with their pans, myself included at various points, if I'm honest. Most of them will go home with nothing. The ones who actually make it will mostly be the ones who stopped chasing the shiny version of the idea and built something boring enough, and useful enough, that it's hard to kill. If you're building something right now, that's the real question worth sitting with. Not "is this impressive." Just — is this actually solving something someone would pay, gladly, to never have to deal with again. That's the whole game. It always has been. AI just changed who gets to play it. I write about developer tools, real startup lessons, and shipping actual products not just demos over at TheBitForge on dev.to. If any of this hit close to home, I'd genuinely love to hear what you're building — drop a comment.