In 2014 we gave our best product away.
All of it. The code. The docs. Free, forever.
It was the smartest business decision we ever made.
Developers found our gateway on GitHub. They argued about it on Hacker News. They snuck it into banks and telcos through the side door. A company grew around it.
That was the open source playbook. Give it away. Get discovered. Get trusted. Get paid.
Every infrastructure founder knows it. Half of them are betting their company on it right now.
Here’s the problem.
It doesn’t work any more.
It was never about the licence
The open source go-to-market was never really about freedom, or community, or the licence.
It was two hacks.
A discovery hack: a developer with a problem went hunting. Google. GitHub trending. Comparison threads. An unknown vendor with good code could be found, and finding us cost an evening.
A trust hack: the code was the proof. No sales call. No marketing claims to decode. Read it, run it, judge for yourself.
Both hacks depended on one assumption so obvious nobody ever said it.
A human does the looking.
Nobody’s looking
That assumption died quietly, sometime in the last two years.
84% of developers now use AI tools. More than 90% use an assistant every month.
Watch a developer solve a problem in 2026. They don’t hunt. They ask.
And the machine answers. Three options. Pre-filtered. Pre-ranked. Working config generated on the spot.
The time spent doing an evaluation is gone.
The trending page is gone.
The scrappy repo with the brilliant README is gone.
The shortlist arrives fully formed, out of a model. And the model recommends what it knows: the established, the documented, the endlessly discussed.
Ask an assistant to suggest an API gateway. Same handful of names every time. Ours among them, thankfully.
The fifteenth result on GitHub used to have a chance. Merit could move you up the page.
The fifteenth option in a model’s head doesn’t exist.
You cannot out-open-source a training corpus.
The trust hack died too
A decade of relicensing dramas taught buyers that “open source” describes Tuesday. Not Thursday.
And the newest consumers of software don’t care at all.
An AI agent choosing tools doesn’t read your licence. It doesn’t admire your commit history. It cares whether your docs make sense and your errors are honest.
Source availability. The credential we built everything on.
Machines don’t even look at it.
What I’d do starting today
The licence was never the magic.
The magic was what the licence forced on us: everything public. Code. Docs. Issues. Arguments in the open.
The licence was the container. The legibility was the product.
Today you’d chase the legibility directly. Docs so good the models absorb them. Examples everywhere. Teach your problem space so thoroughly that every model learns your framing of it.
Then the machines do the distributing for you.
Yes, I know the rebuttal: the models were trained on open source, so open code is in the corpus by default. True. That’s exactly why the old winners are so entrenched.
It does nothing for the new repo starting from zero. One among a hundred million.
Easy for me to say
Obvious objection: this is precisely what someone at the top of a ladder would say.
Fair.
We benefit from a decade of open source presence baked into every model on earth. And our gateway stays open source. Not as marketing. For the banks and telcos we serve, it answers a real question about risk.
But I won’t pretend the repo does the job in 2026 that it did in 2014.
It was the engine. Now it’s the paperwork.
If you’re starting an infrastructure company today: open source is still a fine licensing decision. A fine community decision. A fine values decision.
It is no longer a distribution decision.
When a developer, or their agent, asks the machine your category’s question, why would your name come out?
Answer that first.
Distribution used to be a licence.
Now it’s a corpus.
Further reading
