Open models spread to the device while capacity keeps moving up the stack

August 12, 2026

NVIDIA released Nemotron 3.5 Lightning as an open model with weights, training data and recipes, and also released the NeMo Switchyard routing library for directing requests across open, proprietary and NVIDIA models on PCs, edge devices, workstations, data centres and the cloud. Meta launched the open-weight Muse Glimmer model under an Apache 2.0 licence, said it will open the weights for Muse Spark 1.2, and said the models are designed for direct download, local use and integration into runtimes such as llama.cpp and Ollama.

This appears to move local AI another step from hobbyist option to normal delivery form, because both companies paired model release with the practical pieces needed to run and route models outside a single hosted service. For Microsoft, that is likely to narrow the distance between Copilot-style distribution through the cloud and software that customers can run on their own devices or inside their own stack.

At the same time, TSMC's board approved about US$29.44 billion in capital appropriations for process technology, packaging, capacity and new fab construction tied to AI and high-performance chips. Intel also increased its stock offering to $20 billion, priced the shares at $95 and said the proceeds are for general corporate purposes including capital expenditures and factory tooling to meet rising data centre and AI compute demand.

This suggests the market is still behaving as if AI demand will be constrained less by model ideas than by the physical ability to build and package enough compute. It also likely means the shift to local and open models does not reduce Microsoft's exposure to infrastructure competition, because cheaper distribution at the edge still rests on a supply chain that remains capital-heavy upstream.