The chip vendor becomes the model vendor
NVIDIA is training a trillion-parameter open model and shipping a router designed to keep work away from frontier models. Both moves sell accelerators.
Two NVIDIA releases in one week. A trillion-parameter open model in training, on a $7bn capped budget. And NeMo Switchyard, an open router built to route work to the cheapest model that can do it. Those look like opposite strategies. They are the same one.
Weights are a complement, not a product
NVIDIA does not need model revenue and is not pretending to seek it. Nemotron 4's flagship is meant to exceed a trillion parameters — roughly twice Nemotron 3 Ultra's 550B total, 55B active — and to be free to download and modify. The stated goal is parity with the best open models anywhere.
The commercial logic is that every capable open model tuned for NVIDIA hardware enlarges the market for NVIDIA hardware, and every hour inside CUDA is an hour not spent evaluating alternatives. A $7bn cap through fiscal 2028 is what a strategic complement costs, not what a moonshot costs.
The router is the sharper move
Switchyard argues that the frontier model should be a planner invoked rarely. It ships three tuning-free strategies — an LLM classifier with session affinity, a stage router reading recent tool activity, an escalation router that starts cheap and promotes on sustained difficulty. Nemotron 3.5 Lightning, 30B total with 3B active and a 1M context, exists to be the cheap worker underneath it.
NVIDIA sells the picks. It does not care which side of the argument wins.
Who this hurts
Not NVIDIA. If frontier models win, they run on NVIDIA. If routed swarms of small models win — and Cursor's swarm data points that way — those run on NVIDIA too, in greater numbers.
The exposed party is the lab whose revenue assumes the frontier model is called once per step rather than once per task. Routing does not reduce the work. It reduces who gets paid for it.
Reuters via The Star — Nvidia is developing Nemotron 4 open-source models → · MarkTechPost — NVIDIA AI Releases Nemotron 3.5 Lightning and NeMo Switchyard → · NVIDIA Developer — Nemotron 3.5 Lightning for long-running agents →