Why I’m Betting On Owned AI, Not Rented AI

Two stories from this week made me rethink how much of my stack I actually control

Photo by Cash Macanaya on Unsplash

I run multiple products. Some of them run on API calls I don’t own, models I don’t control, and infrastructure that can change its rules on me overnight.

That’s the deal every indie maker or developer signs up for.

Free to read for non members!

Most days, I don’t think about it.

This week I did, because of two stories that landed almost back to back.

First, Palantir’s CEO went on CNBC and said enterprises paying frontier AI labs are also handing those labs the exact usage data that makes the labs better at replacing them.

Second, a Reddit user reverse-engineered Claude Code and found it was silently checking timezones and proxy domains to flag users connected through Chinese networks, encoding the result into invisible Unicode swaps in the system prompt.

Anthropic didn’t deny it. An engineer confirmed it was an anti-distillation experiment from March, said it wasn’t disclosed anywhere, and said it’d be pulled in the next release.

I don’t think Anthropic ran a spy operation.

I think a company under real pressure from distillation attacks (they’d separately accused an Alibaba-linked lab of running 25,000 fake accounts to scrape 28 million conversations) shipped a hacky detection layer and didn’t think hard enough about disclosure.

That’s a normal, unglamorous failure.

But normal failures are exactly the ones worth paying attention to, because they tell you what a company does under pressure when nobody’s watching, not what they say in the trust and safety blog post.

Anyways, I’ll come to

The part that actually matters for builders like me

I don’t build on Claude Code’s internals, so the specific bug doesn’t touch me.

But, what it reminded me of is something I already knew and kept ignoring:

every layer of my stack that I don’t own is a layer somebody else can quietly change.

My Gemini API bill can double with a pricing update I didn’t vote on.

My Dodopayments integration works until it doesn’t.

My Neon database is a phone call away from a policy I have to read twice.

None of this is a conspiracy.

It’s just what renting means.

You get speed and you give up control, and most of the time that trade is completely worth it, which is exactly why nobody talks about the downside until a story like this makes it visible for a week.

Karp’s actual argument, stripped of the enterprise sales pitch, is simple:

If a capability is core to what makes your business defensible, you shouldn’t be renting it from someone who has a commercial incentive to eventually compete with you.

For a solo builder that doesn’t mean buying a GPU rig, it means being honest about which parts of my stack are genuinely swappable and which parts I’ve quietly let become load-bearing without a backup plan.

[embed]Karpathy's 4 CLAUDE.md Rules Cut Mistakes by 30%. I added 4 more to further cut it down to 5% techaiguild.aibucket.org

Where I’m actually drawing the line

For GritGlean, the AI calls are a feature, not the whole product.

If Gemini’s pricing or policies shift tomorrow, I can route around it in an afternoon because the value is in the data pipelines and the UX, not the model call itself.

That’s the version of “owning your stack” that’s realistic for someone doing this in the hours before and after a day job: not self-hosting everything, just making sure no single vendor is a single point of failure for the thing you’re actually selling.

**BoutPredict is the one place where I already did the harder version of this without framing it that way at the time. The fight predictions don’t run through a rented API. It’s a six-layer neural network I built and own outright, trained on our own fight data, sitting on our own infra.**

Nobody can change its pricing on me.

Nobody can quietly retrain on my traffic and ship a competing feature.

Nobody can decide tomorrow that predicting MMA outcomes falls outside their usage policy and cut me off.

It’s slower to improve than pinging a frontier model would be, and I don’t get to coast on someone else’s billions in R&D.

But the core thing customers are paying BoutPredict for isn’t rented from anybody, and that wasn’t an accident: the model

Wrapper products stay on rented APIs because speed matters more than ownership there. But wherever the model is the product, wherever it’s the specific thing people are paying for, I want to own it the way I own BoutPredict’s prediction engine.

So,

The uncomfortable version of this question is:

how much of what you’re building would survive if the model you built it on changed its terms next month?

If the honest answer is “not much,” that’s not a reason to panic.

It’s a reason to spend a weekend decoupling before you have to do it under pressure instead of on your own schedule.

I’m not switching off closed APIs.

I’m just done pretending the convenience is free.

In case we are meeting for the first time, come over *here, it’ll be worth the roller coaster of articles that are gonna come up in the next few weeks.*