Nvidia declared the official opening of the 'AI PC era' by unveiling a technology upgrade at this year's CES that signals a major turning point for PC-based generative AI.

NVIDIA announced that at CES 2026, it applied RTX acceleration across core generative AI tools such as CompiUI, LTX-2, Rama.cpp, and Olama to significantly boost video, image, and text generation performance, and also unveiled the next-generation DGX system capable of running up to 1 trillion parameter models in a desktop environment.

NVIDIA defined 2025 as the year when PC-based AI development took a full leap forward.

The accuracy of small language models (SLM) doubled compared to the previous year, and the number of PC-class model downloads increased tenfold.

To continue this trend, NVIDIA announced major performance improvements this year centered on RTX AI PCs.

The most notable change is the dramatic improvement in 4K video generation performance.

NVIDIA has PyTorch–CUDA optimization and NVFP4 in CompiUI.By adding FP8 precision support, the video generation speed was increased by up to 3 times and VRAM usage was reduced by up to 60%.

In addition, RTX Video Super Resolution has been integrated into CompiUI, enabling high-quality 4K upscaling to be processed in real time even on a local PC.

With the release of Lightrix's latest audio-video generation model, LTX-2, as an open weight, the landscape of local video production is changing.

LTX-2 generates up to 20 seconds of 4K video and features multi-keyframe and advanced conditioning capabilities, providing quality comparable to cloud models.

Based on this, NVIDIA also unveiled a new 4K video generation pipeline leading from storyboard to keyframe to final video.

Text and image-based AI performance has also improved significantly.

NVIDIA collaborated with the open source community to improve the SLM inference performance of Rama.cpp and Olama by 35% and 30%, respectively.

The same optimizations have been applied to DGX Spark and DGX Station, significantly increasing the execution speed of large-scale models.

In particular, the DGX Station is notable for being equipped with the GB300 Grace Blackwell Ultra Superchip, which allows it to run up to 1 trillion parameter models locally.

Local search technology has also evolved.

Nexa AI's 'Hyperlink' provides the ability to index documents, images, and PDFs to transform a PC into a natural language-based knowledge base.

With RTX acceleration applied, text and image indexing speeds become more than 100 times faster than CPUs, and a beta feature was also unveiled that allows searching for objects, actions, and voice within videos.

NVIDIA AI infrastructureMajor announcements continued in the field as well.

The next-generation Rubin platform and the DGX SuperPOD based on it have set a new standard for enterprise AI factories.

The Rubin platform is an integrated structure of six chips, including the Vera CPU, Rubin GPU, and NVLink 6 switch, which reduces inference token costs by up to 10 times.

The DGX SuperPOD features 576 Rubin GPUs, delivering 28.8 exaflops of FP4 performance, and has established itself as a next-generation AI infrastructure for large-scale MoE models and long context inference.

NVIDIA Broadcast 2.1 Update was also announced.

The virtual key light feature has been enhanced to provide more natural lighting effects on RTX 3060 or higher GPUs, and is expected to be more useful for streamers and creators.

NVIDIA CEO Jensen Huang stated, “With the demand for AI computing exploding, a new platform is needed that encompasses both local AI and data center AI,” adding that “the Rubin platform and RTX AI PCs will be the foundation for next-generation AI innovation.”