Nvidia is spending $26 billion to subsidize open-weight AI models, betting that a world of cheap, distributed inference creates more GPU demand than the investment costs. The logic is structural: more companies training and deploying models means more chips sold. This is not altruism. It is vertical integration dressed as ecosystem building.
The Linux analogy breaks here. True open-source AI, meaning full training recipe, data, and code, requires capital that individual contributors cannot supply. What most developers actually use are open-weight models: weights and inference code only, closer to a compiled binary than a living codebase. The author, who helped build Ai2's OLMo models, draws this distinction carefully, and it matters because it defines who can actually sustain the ecosystem long-term. Companies like Databricks and 01.ai have already exited training. The ones still in are increasingly dependent on Nvidia's balance sheet or revenue-share licenses appearing on Hugging Face repos from Moonshot and Qwen.
The author lays out two futures: open models stay competitive and generate returns proportional to Anthropic and OpenAI's API profits, or they fork into a long-tail ecosystem focused on efficiency and specialization, running enterprise agents on-prem while closed models dominate drug discovery, knowledge work, and software engineering. He calls the second outcome his most likely scenario. The deeper argument, about how reasoning training is becoming as opaque as pretraining was three years ago, and what that does to the open-source lexicon itself, is worth reading in full.
[READ ORIGINAL →]