Engineers at @paritytech ran an experiment to see how much of their AI coding stack they could run themselves instead of relying entirely on external providers like OpenAI or Anthropic. They set up a system where engineers could keep using their usual coding tools while freely swapping the AI model behind them. They tested an open-weight model, DeepSeek V4 Flash, on rented GPUs. The trial was meant to last one afternoon, but engineers kept using it for a month. 25 engineers, about 13 billion tokens, and roughly 1,200 USD in GPU costs. Seven of the eight engineers surveyed said the model could potentially be their main one, but with some considerations. The goal wasn't to completely replace hosted providers but to keep the freedom to choose. That's the same principle Polkadot is built on. Builders shouldn't be locked into a single provider's rules, prices or roadmap. On Polkadot, teams choose who runs the infrastructure and how it's governed, and the open-source code means they can always inspect, adapt, and implement the tech according to their unique needs.