• Releasing your best model openly competes directly with selling it, so labs drift closed once real revenue appears. The count of firms shipping near-frontier open weights shrinks even as the count shipping small open models grows.
• Nvidia's Nemotron is a one-company version of the consortium, bankrolling open models to sell GPUs. That support is fragile: it breaks if Nemotron starts competing with Nvidia's own customers, or if the cash cushion funding it disappears.
• The consortium pitch is arithmetic. Training a relevant model is heading toward billions, and a company that depends on open weights would rather pay a tenth of that, or a fiftieth if enough members join, than train one alone.
• Members would buy more than weights: a say in decisions like model sizes, plus early access to build internal and open-source tooling before the release lands.
• Chinese open-weight labs are expected to crack first. A tightening capital market punishes giving away your only durable advantage, and Moonshot, MiniMax, and Zhipu all still need funding to keep pace.

• open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.

If your stack depends on open-weight models, this maps which sources of near-frontier open models are likely to survive and where the ecosystem is heading.

• articleInterconnects: Open and closed models are on different exponentials
• articleMy bets on open models, mid-2026
• articleKimi K3: The open-weights escalation
• articleGLM-5.2

• articleMy bets on open models, mid-2026Nathan Lambert
• articleThe Open-Model Race has Split Four WaysAlphaSignalAI
• articleInterconnects: Open and closed models are on different exponentialsNathan Lambert

Checking sign-in…

Loading comments…