Why an open model consortium is inevitable

To reliably sustain a frontier open model, a consortium where multiple companies pool funding will ultimately be needed. The training costs and opportunity costs of a single company continuing to push public models to the frontier alone are growing too fast.

A conversation with Percy Liang, and thinking about fully open model research like Marin, made this conclusion clearer. Nvidia's Nemotron Coalition is a case of testing this structure first within one wealthy company, but the argument is that in the long run, a broader coalition is the more stable solution.

Recent changes among senior figures at Qwen and Ai2, and Meta's retreat from its Llama-centered strategy, show this instability. Chinese startups like Moonshot AI, MiniMax, and Z.ai may also find it difficult to maintain their current open strategies if training and R&D costs keep rising.

Going forward, the following trends will strengthen.

• More companies will release smaller, more fine-tuning-friendly models rather than fully open frontier models.
• Revenue will tilt more heavily toward products and closed models.
• Companies will seek joint funding structures to secure future access to models.

Now that training costs have grown to the scale of billions of dollars, companies will prefer buying access at roughly one-tenth the cost, or as low as one-fiftieth if the consortium works well, rather than shouldering everything individually. In the end, demand for open intelligence combines with economic pressure, and the collaborative investment structures that support open models become a hard-to-avoid direction.