What happens when three companies own the thinking

I was debugging a side project last week and realized every AI call I make goes through one of three companies.

Not metaphorically. Literally. I checked. My code editor uses Claude. My search uses Gemini. My chatbot prototype calls GPT-4o. Three companies. Three API keys. Three single points of failure between me and the AI capabilities I’ve started depending on without really deciding to.

I didn’t choose this. Nobody sat me down and said “hey, would you like to build your entire workflow on top of three corporations?” It just happened. The way most dependencies happen. Gradually, then completely.

The Foundation problem

Asimov wrote about this in the Foundation series. Not about AI specifically, but about the idea that civilizations become dependent on technologies they no longer understand. The people of the Galactic Empire used nuclear power for everything but had forgotten how to build a reactor. The knowledge concentrated in fewer and fewer hands until it was functionally gone.

We’re not there. Not yet. But the shape rhymes.

Right now, roughly 90% of paid frontier AI inference goes through OpenAI, Anthropic, and Google. Frontier means the models good enough for the hard stuff: reasoning, coding, analysis, the tasks where you actually need the best available model and not just a fast autocomplete.

Three companies. For the thinking layer that thousands of products are built on top of.

The feeling

There’s a specific feeling that comes with this kind of dependency. It’s not fear exactly. It’s more like the sensation you get when you’re on a plane and you realize you have absolutely zero control over the next four hours of your life. You trust the pilot. You trust the engineering. But you can feel the absence of agency in your chest.

That’s what building on these APIs feels like if you think about it long enough. Which, honestly, most days I don’t. The APIs work. The latency is good. The models keep getting better. So you just… keep building. Like everyone else.

The math

I don’t think this is evil. I think it’s just physics.

Building a frontier model costs somewhere between $100 million and $1 billion. The next generation will cost more. The talent pool is maybe a few thousand people worldwide who know how to train models at this scale. The compute requirements are measured in tens of thousands of GPUs running for months.

Three companies can afford this. Maybe five. Not fifty. Not five hundred.

That’s the math. It doesn’t have a villain. It’s just expensive, and expensive things concentrate.

The questions I can’t stop asking

What happens if Anthropic has a bad quarter and raises prices 3x? Every product built on Claude recalculates overnight. Some of them die.

What happens if Google decides Gemini is a loss leader that isn’t worth subsidizing anymore and kills the API? It’s happened before with Google products. The graveyard is long.

What happens if OpenAI’s next terms of service update restricts a use case that 10,000 startups depend on? Those startups didn’t vote on the terms. They just woke up to an email.

What happens if two of the three have outages on the same day? It happened in cloud computing (AWS us-east-1 and Azure had overlapping issues in late 2024). Half the internet wobbled.

I keep asking these questions and not arriving at good answers.

The indie developer’s honest assessment

You can self-host. Llama 3.1 405B is genuinely good. Meta gives away the weights for free.

But honestly? Who does? Who actually self-hosts a 400-billion-parameter model in production? You need specialized hardware. You need operational knowledge most teams don’t have. You need to handle the scaling, the monitoring, the failovers. You need to do all the things that OpenAI and Anthropic and Google do, except you’re three engineers and a credit card.

The open-weight path exists. It’s real. It’s important. And for most developers, it’s theoretical. The practical reality is: you pick one of the three, you paste in the API key, and you move on.

I do the same thing. Every day.

What Asimov would say

I think Asimov would look at this and say: of course. Of course the thinking concentrated. Technology always concentrates before it distributes. The printing press was controlled by a handful of printers before it wasn’t. Electricity was controlled by a handful of utilities before it wasn’t. Computing was controlled by IBM before it wasn’t.

Maybe AI follows the same pattern. Concentrate first, distribute later. Open models get better. Hardware gets cheaper. Self-hosting gets easier. The oligopoly weakens.

Or maybe it doesn’t. Maybe frontier AI is different because the cost curve never flattens enough. Maybe the three become two. Maybe the two become one.

I don’t know. And the not-knowing is the part that keeps me on the roof at night, looking at the sky, wondering what kind of future we’re building on top of three API endpoints that we don’t control.

Related thinking: