The AI Boom: When Building a Startup Isn’t About Coding Anymore

By some estimates, our planet is home to more than 10,000 artificial intelligence startups. That’s more than the total number of wild cheetahs left on Earth, and far greater than the global population of ancient dawn redwoods. Of course, that number is little more than an educated guess—startups rise and fall faster than market trends can track them. But even if the total is off, we know that more than 2,000 of these new ventures secured their first round of venture funding last year alone. As investors pour billions of dollars into the AI boom, one question is impossible to ignore: What exactly are all these thousands of new companies actually building?

To answer that question, I set out to speak with as many recent AI startup founders as I could. My goal wasn’t to spot the next billion-dollar winner—it was to get a ground-level view of what building an AI product actually feels like right now: how new AI tools have reshaped the very work of starting a company, and how terrifying it is to compete in such a crowded, fast-shifting field. From the outside, the whole pursuit looks a lot like tap-dancing on the roiling surface of the sun. One new update from OpenAI drops, and X is instantly flooded with hot takes predicting the demise of a hundred young startups. It’s brutal.

Will this revolution leave most founders burned? Almost certainly—there’s no way every one of these startups survives. After all, every startup is an experiment, and most experiments fail. But when thousands of these experiments are running across every corner of the global economy, they’re already giving us a clear preview of what the near future will look like.

The Teenagers Building AI Pesticides

Take Navvye Anand, cofounder of agricultural AI startup Bindwell. When we connected over video call, he spoke with a half-smile and easy confidence as he walked me through his work building custom pesticides designed by proprietary AI models. Bindwell’s old website billed these models as “insanely fast,” claiming they can predict the results of experiments that once took days in just a matter of seconds. Listening to him explain how he’s adapted AI drug discovery principles for crop protection, it’s easy to forget he’s only 19 years old.

Anand grew up in India, reading Hacker News alongside his dad from a young age, and was building his own large language models halfway through high school. Before he even graduated, he, his 18-year-old cofounder, and two friends they met at summer camp published a preprint paper on bioRxiv about an LLM they built to predict a key aspect of protein behavior. The paper got scientists talking on X, earned a citation in a top peer-reviewed journal, and convinced the group they should try turning their work into a company. After brainstorming use cases, they settled on protein-based pesticides. Then, like something out of a startup fairy tale, a venture capitalist reached out to them on LinkedIn, offering $750,000 to drop out of high school and college to work on the company full-time. They said yes, and got to work in December of last year—despite knowing almost nothing about the agribusiness industry at the time.

Five months later, Anand and his cofounder opened their first biological testing lab in the San Francisco Bay Area, then moved to a larger space where they personally test promising new molecules by hand, dropping tiny drops of compound into small vials. The core idea is that protein-based pesticides can target specific pests like locusts or aphids far more precisely, without harming humans, earthworms, or bees. When I asked how he picked up the wet lab skills he needed to run tests, he laughed and told me “I hired a friend.” That friend coached him through the basics over the summer before heading back to college in the fall. “Now I can do some biochemical assays,” Anand told me. “Not the full range, but enough to do basic wet-lab validation for our models.”

That story stopped me in my tracks. That a handful of teenagers could build their own custom LLMs, learn the basics of pest control biochemistry, use their models to identify promising new molecules, and set up their own working lab in just a few months is already impressive. When I really stopped to add it all up, it felt almost absurd. I’d gone into these interviews expecting to hear that AI tools speed up small parts of building a company, but I had no idea just how transformative the speedup has been. So when I spoke next to the cofounders of Roundabout Technologies, a 14-month-old startup, I got straight to the point: break down exactly what’s changed, and by how much.

Speed Is the New Superpower

“We’ve shipped an insane amount of work,” Collin Barnwell told me. His four-person team (two of whom joined only last April) is building a real-time computer vision system for traffic lights that adjusts red and green light timings based on actual traffic flow. If you’ve ever sat waiting at a red light for an empty cross street, you know how much wasted time this could fix—making an AI overhaul of intersections a no-brainer.

Barnwell rattled off a long list of what the team has pulled off since April, all enabled by jumping between AI tools: training custom vision networks on their own data, using LLMs to deep-dive into city regulations and infrastructure, writing code for their on-site GPU, building data dashboards, and engineering custom hardware components. “You really feel like these tools are pushing you straight to the cutting edge,” he said. Barnwell describes himself as an average coder, and he’s still giddy about what he’s able to build now. “I’d say we’ve got ‘Collin AGI’ working for us,” he joked. “We’re not quite at ‘Sabeek AGI’ yet.”

Sabeek Pradhan, Barnwell’s cofounder, put the speed difference in stark terms. “What would have taken us a few weeks to build now just takes five minutes of waiting for a model to run,” he told me. By far the most time-consuming part of their work so far has been landing their first real customer. Last July, as the company approached its one-year anniversary, the pair worked with the city of San Anselmo, just north of San Francisco, to install their system at the city’s busiest intersection. They went live at a second crossing in October, with 11 more traffic light installations already planned.

Nearly every founder I spoke with echoed that same observation: a week or two of traditional coding work can now be knocked out in a single day. One founder told me building software has gotten so simple that it’s “not even fun” anymore. Another said that when his internet goes out, having to write code from scratch is “excruciating” and “really painful.” Dependence on large language models is very real for these young teams. And no group is more reliant on big foundation models than the founders building AI agents—tools that run entirely on top of existing large models.

When Taste Matters More Than Code

Justin Lee and Linus Talacko were working at a medical startup in Australia when they felt the pull of the AI boom. They knew it was time to jump. “We just could tell this was the once-in-20-years big shift,” Lee, 21, told me. The two software engineers looked at their own daily routines, saw how much of their work was repetitive and unautomated, and decided to build an AI agent: essentially a bot intern you can give orders to over Slack. They named their company Den.

They hit a wall almost immediately. They’d designed their agent to be conversational, but users didn’t have patience for back-and-forth chat. Instead, Lee said, users wanted to use it “like Python scripts or Zapier workflows” — less like chatting with a coworker, more like a button you click to get a repetitive job done. They scrapped their entire original build and started over. The biggest challenge, Lee explained, is that no one really knows how people want to interact with AI yet, so engineers are basically guessing and hoping for the best. “This is one of the fastest-moving, most dynamic periods of change in modern history,” he said. “Everything we know gets rewritten every single month.” As Talacko puts it, the only way to survive is to stay endlessly flexible. You have to be ready to throw out your entire codebase and start over from scratch. “I’m constantly updating my assumptions about what the world will look like in two weeks,” Lee said. “Most of the time, I still undersell how much things will change.”

One quick tool Talacko built was an agent that monitors new product sign-ups on Den’s website, then automatically googles each new user to see if they’re a notable public figure. “That’s how we noticed Ivan Zhao, the CEO of Notion, had signed up to try our product,” Lee said. “The agent basically acted like a sentry, patrolling our user database. That’s obviously not a job you’d ever pay a human to do.”

This points to a bigger shift in how work works, especially for early-stage founders. This isn’t just about eliminating jobs or avoiding hiring. Instead, AI lets individual founders stretch further than they ever could before. As Barnwell from Roundabout put it, AI lets him write code he never would have bothered to build two years ago, when coding was far more labor-intensive. Back then, he spent most of his time mired in the details of functions and algorithms. Now he can dedicate more time to research, exploration, and what really matters: thinking. “You’re building disposable code all the time now, just to help you get things done,” he said.

Code isn’t precious anymore, and that shift can feel destabilizing. Lee told me that a year ago, he spent three months struggling to carefully code an agent from scratch. Today, that same entire project “could probably be written in three days, without us even touching most of it.” Lee admits he sometimes feels a little discouraged, and finds himself wondering what skills are even worth learning these days.

“What’s crazy right now,” Talacko added, “is if a customer emails you with a bug report or a feature request, you can literally copy and paste what they wrote straight into Claude Code, and you’ll have a working new feature in 15 minutes.”

“So,” I asked him, “does that mean your brain isn’t even involved anymore?”

He huffed a laugh and paused to think. “I guess the judgment of whether you should even build that feature is what matters most now. It’s far less about being able to crank out code and throw something out into the world, and far more about being tactical and strategic about what you choose to build.”

When building anything becomes this easy, founders can run countless experiments, and wander down unproven paths that would have been too costly to try before. The penalty for testing a new approach has dropped so low that a single startup can test dozens of different versions of itself before settling into one smart, focused shape. When, as many people now note, you can write all the code you need for a startup in a single weekend, coding isn’t the core work of building a tech company anymore. So what is?

“The result of all this,” Lee told me, “is that taste becomes the most important thing. Everything else is just an expression or implementation of that taste.”

I paused, suddenly caught on an impossible question, then blurted it out: How do these founders define taste? I expected an awkward silence.

Lee just smiled and said, “Actually, we have a whole group chat where we’re trying to figure out what taste is.”

We talked about it for a few minutes, but we couldn’t land on a crisp definition. After our call ended, I got an email from Lee with a list of links to his favorite writings on taste. Some were posts on X, one was a thoughtful write-up on LinkedIn. But his favorite—and mine, too—was an old essay from startup godfather Paul Graham, written back in 2002, called “Taste for Makers.”

You probably know who Graham is: cofounder of Y Combinator, prolific blogger and X poster, the most influential voice in modern startup culture. Graham even made a surprise appearance in my conversation with 19-year-old Anand, the pesticide startup founder. Back in February, Graham posted on X that he doesn’t approve of venture capitalists luring teens to drop out of school. As Anand tells it, Graham had met him and his cofounder—both already dropouts by that point—for tea that same day. Graham shared his disapproval, then, in a classic plot twist, decided to invest in their startup anyway.

Graham writes that taste is simply the ability to make beautiful things. A person with good taste has enough experience to recognize what’s good, and the skill to build it. What counts as good design? For starters, it’s simple, timeless, and bold. Then Graham gets even more specific: good design is also a little funny. Good design is often strange.

I thought about those lines as I talked to founder after founder. Several teams I spoke with are building home robots, which are having a moment right now—all because AI lets them make real progress with far less money and far fewer people than ever before, plus access to cheap components from China. One of those robots, Isaac from Weave Robotics, is essentially a wheeled floor lamp with thick, crab-claw grippers. It carries a small wicker basket in one claw, and rolls from room to room picking up dirty cups and scattered toys, like an 18th-century farmhand gathering vegetables at the market. Right now in San Francisco, one Isaac lives in a building full of washing machines, folding clothes for a local laundry startup.

A more imposing home helper, K-bot from K-Scale Labs, has a familiar sleek, all-black humanoid shape. When it drops a slice of bread into a toaster, the movement carries a faint hint of menace—there’s something inherently weird about a big hunk of metal making your breakfast. If you’ve been jaded by a decade of flashy robot demos that went nowhere, this might sound ordinary. But open your eyes: it’s still deeply strange to see a humanoid robot serving you toast, just like it’s strange to see a wheeled “farmhand” with crab claws tidying your living room. This, I think, is where Graham’s “funny” and “strange” come in. They add a jolt of shock to the everyday, and catapult you into thinking about what a new future could look like.

Reaching For The Sun

Benjamin Bolte is the founder behind K-bot and K-Scale Labs. When he was searching for a startup idea, he concluded the only smart move was to go after the hardest problem he could find. “All the ideas that made sense a couple of years ago don’t work anymore,” he told me, referring to the business software that has dominated tech for decades. “You have to push your own horizons just to have a chance to not get completely crowded out by all the other companies doing the same thing.” That’s how he landed on building affordable, open-source humanoid robots: imagine putting six of them to work running a small construction company.

Bolte’s choice of name is no accident. “K-Scale” refers to the Kardashev scale, created by Russian astrophysicist Nikolai Kardashev. In a 1964 paper about searching for extraterrestrial civilizations, Kardashev proposed a scale to measure a civilization’s technological advancement. At the bottom is Type 1: a civilization that can harness all of the energy available on its home planet. Current estimates put humanity at around 0.7. Bolte’s X bio reads, “Moving humanity up the Kardashev scale.” He believes a global network of affordable, capable robots can lift us all the way to Type 1.

It sounds heady, but it’s far from the most ambitious idea I heard. Just a few days after speaking with Bolte, I had a second conversation about the Kardashev scale with the founders of Starcloud, a startup working to put data centers in orbit to power all of the world’s growing AI demand. Starcloud was founded in summer 2024, and plans to launch its first GPU into low Earth orbit this November. Earth-based data centers drain millions of gallons of water and already face local pushback—so why not take them literally up into the sky? The company’s founder has his sights set on getting humanity to Type 2: a civilization that can harness all of the energy of its host star. There’s no shortage of ambition in this new wave of startups.

So things are moving incredibly fast. Precisely how fast? It’s hard to pin down. Our perception of progress can outpace actual results, after all. A recent paper from Cornell University found that developers who rely heavily on AI actually end up working slower than developers who write code by hand, because time once spent on active coding is now consumed by other, unglamorous tasks like prompting and debugging AI output. And let’s be honest: the odds are stacked against most of these startups. It’s unlikely that more than a tiny fraction will still be around to see 2027, no matter how much AI speeds up their work.

Even so, there’s something undeniably exponential about the current moment. When you add up even the most grounded, down-to-earth ambitions of these thousands of startups—bringing AI to Rust Belt factories, independent grocery stores, and local county governments—the scale of change is staggering. AI is often called Promethean, for its unique mix of raw power and existential danger. So it feels only fitting that this new wave of driven, optimistic founders are reaching straight for the fire itself, just to see what happens next.