NVIDIA AI Updates: August 12, 2026

1. NVIDIA Releases Nemotron 3.5 Lightning, an Open MoE Model for Long-Running Agents

NVIDIA. NVIDIA released Nemotron 3.5 Lightning, an open-source 30-billion-parameter Mixture-of-Experts model with 3 billion active parameters per token, optimized for the high-volume, low-latency execution layer of always-on agents handling tool calls and subagent delegation. NVIDIA reports up to 4x output speed versus similar-sized models and 86% accuracy on PinchBench while completing 10,000 tasks 30% faster than comparable solutions, using speculative decoding and NVFP4/BF16 checkpoints. The model ships under OpenMDW-1.1 with weights, training data, and recipes on Hugging Face and ModelScope, alongside NeMo Switchyard for routing agent workloads across models. Source

2. NVIDIA Frames AI Factory Compute as an Investable Asset Class

NVIDIA. NVIDIA said it is working with six major financial institutions, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, to build financing platforms that treat AI factory compute as a financeable productive asset, with the goal of mobilizing more than $500 billion of third-party capital over time. The company argues its GPUs behave like durable infrastructure that generates revenue and can be redeployed, citing that its 2020-era A100 remains in active commercial use, and noted rising rental rates such as H100 hours near $2.35 and Blackwell B200 commanding $5.30 to $7.05 per GPU-hour. NVIDIA may provide residual-value support of up to 25% per project while partners conduct independent underwriting. Source

3. NVIDIA Details an 800 VDC Power Architecture for AI Factories

NVIDIA. NVIDIA introduced an 800-volt direct-current power distribution architecture designed to remove electrical bottlenecks in scaling AI compute, delivering power at higher voltage with fewer conversion stages than traditional AC systems. Components include a Power Rack that fits existing AC infrastructure in the second half of 2026, a Row Power Center supporting up to 2 megawatts per row in 2027, and facility-scale DC Power Blocks for new builds. Developed with Google and Microsoft through the Open Compute Project, the specification already has more than 80 equipment makers building compatible products. Source