NVIDIA builds 1-trillion-parameter Nemotron 4, assembles $500B compute coalition

NVIDIA is pushing on two fronts at once. On models, it is building Nemotron 4, an open-model family topping out at a 1-trillion-parameter variant that could be ready by late fall, plus Nemotron 3.5 Lightning targeted at code review and NeMo Switchyard for agent routing. Moving into frontier-scale open weights puts NVIDIA in more direct competition with the labs it sells GPUs to — a notable strategic wrinkle.

The bigger structural story is financial. NVIDIA assembled a coalition with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize more than $500 billion for AI compute infrastructure, explicitly framing compute as a new investable asset class. This is Wall Street being enveloped in the AI capex frenzy, per Bloomberg, and it fits this week's dominant theme of compute-financing megadeals — alongside Amazon lifting 2026 capex to $220B and Mistral's European Compute Units push.

The technical detail on the smaller Nemotron tier is concrete: a 30B MoE with 3B active parameters using a Mamba-2 hybrid attention architecture and up to 1M-token context, optimized for autonomous agents and local inference. The skeptical framing writes itself: turning compute into a securitized asset class raises questions about circular financing (NVIDIA financing customers who buy NVIDIA chips) and bubble dynamics. Watch whether the 1T Nemotron actually ships on schedule and how the labs react to their hardware supplier competing on models.