Open weights just became the new venture capital
Every major model release now comes with open weights because distribution trumps differentiation in the foundation model game.
The pattern is everywhere now. NVIDIA drops Nemotron with open weights. Google ships Gemma under Apache 2.0. Everyone’s giving away the crown jewels. This isn’t altruism. It’s the smartest business move in AI.
Distribution beats differentiation
Open weights solve the same problem venture capital does: getting your product into as many hands as possible, as quickly as possible. When you’re building foundation models, network effects matter more than margins. The company that gets developers building on their architecture wins the platform game.
Closed models optimise for per-query revenue. Open models optimise for ecosystem capture. NVIDIA doesn’t make money selling Nemotron inference calls. They make money when every AI startup fine-tunes Nemotron and buys H100s to run it.
The real moat isn’t the model
The uncomfortable truth is that model weights have become commoditised faster than anyone expected. If your business model depends on keeping a transformer architecture secret, you’ve already lost. The value moved upstream to training infrastructure and downstream to deployment tooling.
Open weights let you skip the “trust us, our API is reliable” conversation entirely. Developers can inspect, modify, and guarantee uptime themselves. That’s worth more than any benchmark score.
We’re watching the foundation model layer become truly foundational. Like Linux, like TCP/IP, like HTTP. The money isn’t in owning the protocol. It’s in building everything else.