Meta publishes Llama's weights. DeepSeek publishes its own. Both are downloadable by anyone with enough storage and patience, and commentary treats this as a democratizing moment: power that used to sit exclusively with a handful of labs is now, in principle, available to a student, a startup, a government with modest resources. In principle is doing a lot of work in that sentence.
Downloading Llama is not the same as running it well. Serving inference at meaningful scale — low latency, high throughput, acceptable cost per query — requires specialized hardware and an optimized serving stack. Most organizations that download an open-weight model end up renting inference from Together AI, Fireworks AI, or Groq, three of the providers that already solved that problem, which quietly recreates the concentration the open release was supposed to dissolve.
The same pattern holds for data and fine-tuning. A base model like Llama is a starting point, not a finished product. Making it genuinely useful for a specific domain requires proprietary data, and the organizations with the richest proprietary data — large enterprises, established platforms, the labs themselves — are best positioned to extract value from an open checkpoint. Open weights lower the floor for who can experiment. They do not lower the floor for who can compete.
Distribution compounds the effect. A model is only useful embedded somewhere a user already is. Meta can wire Llama into WhatsApp and Instagram without needing to win users from scratch. A superior open-weight model with no distribution channel loses to a mediocre proprietary model already sitting inside software people use every day.
None of this makes open weights meaningless. DeepSeek's releases mattered enormously for researchers who need to inspect internals rather than query a black box, and for organizations with data-residency rules that cannot send information to a third-party API regardless of quality. The openness is real. It just does not translate into the competitive leveling the word implies to people outside the infrastructure layer.
What is actually happening is a reallocation of where the moat sits, not its disappearance. When weights were closed, the moat was the model. With Llama and DeepSeek open, the moat migrates to inference infrastructure at Together and Groq, proprietary data, and distribution — three things concentrated before the first open-weight release, showing no sign of becoming less so now that the release has happened.