Will OpenAI and Anthropic just build whatever AI product I make?

The frontier labs aren't going to build a newsletter builder plus a form system for local media publishers. The market's too small for them and the workflow is too specific. That's exactly why I build there.

"Frontier labs like Anthropic will not build hyper-specific, multi-step workflows for a narrow niche like local media, because it is not worth their effort. That's where an independent operator can build a moat."

How should I think about the AI labs?

As infrastructure. ChatGPT, Claude and Gemini are building the models. Yes, they build their own products too, but most products people use with AI won't be ChatGPT or Gemini directly. They'll be software like lightbreak that sits on top of that infrastructure.

General models can in theory do a lot. That doesn't mean they can get you to something publishable for a narrow job. It would make no sense for Claude to be super good at placing an ad into a newsletter in one specific style. That's too small for them to care about, which is why people build on top of these tools.

My view is that the whole economy is going to be people in every niche building specialized tools on top of the big ones. You get deep into a niche, learn its exact problems, then stitch several AI tools together to solve them. That combination is what's valuable.

Which AI products are safe from the labs?

• Narrow. One industry whose problems you know.
• Multi-step. Workflows with several stages that each need industry knowledge.
• Integrated. Tools that work together for one kind of buyer, like a newsletter builder plus a form system. If you have eight different things in one that all work together, Claude's not going to build that.

That's also why I picked local media over another general AI startup. It's small enough that nobody mainstream cares, and big enough that you can crush it. When a traditional local media company wrote an article asking who all these weird entrepreneurs doing local newsletters are, I took it as a good sign. When the legacy people hear about you but don't get it yet, they're ripe to be disrupted.

Which AI products are not safe?

Simple ones. A one-click AI receptionist is commoditizing quickly. Vibe coding is real and it's going to keep happening, and it makes sense for it to get rid of super simple software that does one thing.

The bigger danger is whoever already has distribution adding your product as a feature. That's what people talk about with OpenAI. Granola, the note-taking app, was the big example: ChatGPT made what Granola does a feature. In a niche, the version of this risk is somebody already in the industry putting $100,000 behind your idea instead of $10,000.

Still, being so scared that you never build anything is bad too.

Even with all that, I'd guess about 90% of business owners would rather pay $99 for a tool than build their own, because building it means signing up to constantly fix it.

What if a big platform copies my product?

Niche down. If two or three people come out with killer local newsletter automation systems, or a big platform launches its own version, move to a narrower segment: a government version, a version a business uses for its own clients, or one built for agencies to sell to their clients.

Lots of competition also means lots of buyers. If you have no competition at all, that's probably a bad sign. And don't obsess over competitors. There may be other people building newsletter AI tools, but nobody's going to build exactly the way I do things, and a lot of people will want to do things my way.

Can't people just build it themselves with Claude?

Some can. Fewer will than Twitter suggests. lightbreak took my team eight months to get out the door. It was 98% ready for a long time, and that last 2% was brutally hard. If a tool does one thing, maybe you can replace it in a week. If it has multiple moving parts, it's not that easy, even for people building with these tools every day.

Email sending is the clearest case. Sending email isn't difficult. Sending a large volume without landing in spam is. Once you reach a certain scale, if you aren't doing things correctly, the whole thing can blow up. That's what you're paying a platform like Beehiiv for. I expect Gmail and the other providers to get even stricter as more agents try to abuse email, so you can't half-vibe-code a sender and expect it to work.

Possible also isn't the same as easy. A big email platform can do much of what my survey tool does in theory. It isn't talking about it. The value often comes from making something easy enough that people actually do it, and then promoting it.

If building is easy, what's the real advantage?

Distribution. Ideas used to be kind of pointless. Now ideas matter, because people can build things so quickly. The real answer, though, is audience and marketing. Since building is easy, more people are building, and it's crowded trying to get in front of buyers.

The other answer is outcomes. Claude can draft a newsletter now. That's not the same thing as running local media, which is still mostly operational pain: content, social, events, directories, job boards and sales. Buyers increasingly want to pay for outcomes, not software where it's on them to do the work. So can the software still exist, knowing what Claude can do? I think the answer is 100% yes, even if it's a little harder.

Customers will keep paying well for software that does six, seven or eight things easily and at high quality. What they don't want is a product that doesn't change for a year, or one that charges more for every little new feature. I want people to feel like members who keep getting more at no extra cost, the way Costco works.

More in what the real moat is when anyone can build software.

Does a polished interface protect me?

Less every year. Agents like Claude will increasingly operate B2B software on a user's behalf, so visual polish matters less over time than workflow reliability and data integration. When I planned lightbreak's second version, I dropped drag-and-drop for an automation-first system.

A peer showed me how disposable simple software has become: he built a grant-review app in Lovable, used it with four teammates, then deleted it.

Should I build or consult?

For most people, the highest-probability way to make money from AI right now is consulting or implementation, because faster change increases demand for help. If you do build, build narrow. See build AI software or sell AI services.