The same CTO who watched Firefox lose the browser wars now wants to win the AI infrastructure race by open-sourcing the whole stack.
The Summary
- Mozilla's CTO Raffi Krikorian argues AI should be built on open-source infrastructure, mirroring how the early internet scaled without proprietary gatekeepers
- While consumer attention focuses on ChatGPT and Claude, enterprises are quietly choosing open models they can modify and control
- The bet: companies care more about customization and data sovereignty than they do about marginal performance gains from closed models
The Signal
Raffi Krikorian sees the pattern because he lived through it once already. At Mozilla, he watched Chrome eat Firefox's lunch not because Google's browser was technically superior, but because Google controlled the entire stack and made strategic bets Mozilla couldn't match. Now he's watching the same movie play out in AI, except this time the open-source side might actually win.
The consumer narrative around AI is a head fake. Everyone talks about ChatGPT's user numbers and Claude's context windows. Meanwhile, enterprises are building on Llama, Mistral, and other open models because they can't stomach the idea of piping their entire business through someone else's API. When you're a bank or a healthcare provider or a logistics company, "trust us with your data" isn't a product feature. It's a dealbreaker.
"Companies are choosing models they can customize and control over closed systems, even if it means accepting slightly lower performance."
Krikorian's thesis is that AI infrastructure will follow the same path as internet infrastructure. The web didn't scale on CompuServe or AOL's walled gardens. It scaled on open protocols: TCP/IP, HTTP, HTML. No single company owned the stack, so anyone could build on it. The same dynamic is starting to play out in AI:
- Open model performance is closing the gap with frontier models faster than most people expected
- Inference costs are dropping so quickly that the "moat" of expensive training runs matters less every quarter
- Developers who learn on open tools build products for open platforms, creating a flywheel effect
The Mozilla angle matters because they have credibility here. They're not a VC-backed startup pivoting to AI. They've been building open-source infrastructure for 25 years. When Krikorian says AI should be built like the internet, he's not speaking metaphorically. He's describing an architectural choice: distributed, interoperable, impossible to own.
The counterargument is obvious. OpenAI and Anthropic will point to their safety work, their Constitutional AI frameworks, their red-teaming efforts. They'll say open models are dangerous because anyone can fine-tune them for harm. Krikorian's response, implied in his framing, is that security through obscurity has never worked. The internet didn't get safer by keeping the code closed. It got safer because millions of eyes could audit it, fork it, and improve it.
The Implication
Watch where enterprises spend their inference budgets over the next 18 months. If Krikorian is right, we'll see a quiet migration away from proprietary APIs and toward self-hosted or hybrid deployments. The companies that win won't be the ones with the best chatbots. They'll be the ones building the picks and shovels for the open model economy: fine-tuning platforms, inference optimization tools, model evaluation frameworks.
For individual builders, the lesson is simple. Bet on what you can own and modify, not what you can rent. The agent economy doesn't run on permission. It runs on forkable code.