According to NVIDIA's SIGGRAPH 2026 blog post, the SIGGRAPH 2026 conference is running through Thursday, July 23, in Los Angeles, where attendees can explore graphics research, neural rendering, simulation and AI (E1). NVIDIA said its own keynote took place on July 20 at 3:45 p.m. PT, featuring NVIDIA AI research and engineering leaders Neil Ashton, Edward Liu and Ming-Yu Liu, who discussed neural rendering techniques, world models and simulation methods for AI built by AI (E2). In an introductory video for the keynote, NVIDIA founder and CEO Jensen Huang said: "Creators and designers need tools powerful enough to expand their imagination, malleable enough to give them the freedom to shape ideas and precise enough to realize their vision exactly as intended" (E3).
NVIDIA said Neil Ashton described how the latest model architectures developed by NVIDIA Research and its partners compress Earth-2 model checkpoints down to a couple hundred megabytes — a million-times compression that the company said enables physically accurate visualizations within a second (E4). NVIDIA also announced Cosmos 3 Edge, a 4-billion-parameter model built to run in real time on-device, released the same day as the keynote (E5). Ming-Yu Liu announced Cosmos-Dreams, a collection of closed-loop simulators; a colleague demonstrated an autonomous-vehicle simulator that generates an entire world from a single frame while running on a single NVIDIA RTX PRO 6000 GPU (E6).
NVIDIA said Cosmos 3 Edge ranked No. 1 on the VANTAGE-Bench vision-analytics benchmark within its parameter class (E15), and that partners including Agile Robots, Doosan Robotics, Siemens and Skild AI are evaluating Cosmos 3 Edge for robotics workflows (E16). Cosmos 3 spans three sizes:
| Cosmos 3 tier | Parameters |
|---|---|
| Edge | 4 billion |
| Nano | 16 billion |
| Super | 64 billion |
(Source: NVIDIA, E17)
NVIDIA announced Nemotron 3 Ultra, a 550-billion-parameter open model optimized to run on DGX Station GB300 systems as a customizable model layer (E18). The company said the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip delivers up to 20 petaflops of FP4 AI compute and 748GB of coherent memory to run large models such as Nemotron Ultra from a desktop-class DGX Station (E19). NVIDIA also said its ConnectX-8 SuperNIC delivers up to 800GB/s of bandwidth within DGX Station and supports linking up to two DGX Stations to scale model capacity and performance further (E20). According to NVIDIA, DGX Station is built and available to order from ASUS, Dell Technologies, Exxact, GIGABYTE, HP, MSI and Supermicro (E21).
| DGX Station spec | Value |
|---|---|
| FP4 AI compute | up to 20 petaflops |
| Coherent memory | 748GB |
| ConnectX-8 SuperNIC bandwidth | up to 800GB/s |
| Max linked DGX Stations | 2 |
| Nemotron 3 Ultra parameters | 550 billion |
(Source: NVIDIA, E18–E20)
NVIDIA reported that, in its own testing, the Synthetic Video Detector NIM microservice's accuracy reached up to 92% on uncompressed video, 87% at 15% compression and 82% at 50% compression (E12). The company said the microservice can process 1080p video in as little as 22 milliseconds on NVIDIA RTX systems and approximately 30 milliseconds on NVIDIA L40 GPUs (E13).
| Compression level | Detection accuracy |
|---|---|
| Uncompressed | 92% |
| 15% compression | 87% |
| 50% compression | 82% |
| Hardware | 1080p processing time |
|---|---|
| NVIDIA RTX systems | as little as 22 ms |
| NVIDIA L40 GPUs | approximately 30 ms |
(Source: NVIDIA, E12–E13)
NVIDIA said Wowza is embedding the microservice through its Wowza Video Intelligence Framework, bringing real-time synthetic video detection into livestreaming workflows used across more than 35,000 deployments in over 170 countries (E14).
According to NVIDIA's blog post, several creative-software makers are adopting the Model Context Protocol (MCP) to connect AI agents to their tools. NVIDIA said Adobe is expanding its creative agent across Firefly, Express and Creative Cloud, and that Adobe now provides the Adobe Express Developer MCP Server so AI coding assistants can build Adobe Express add-ons using official documentation and APIs (E7). Affinity by Canva has introduced an AI Connector for Claude that uses MCP to bring natural-language automation directly into Affinity, NVIDIA said (E8). Boris FX Silhouette now includes an MCP server that lets AI assistants work directly inside projects (E9), while SideFX is bringing MCP support to Houdini 22 through its new APEX Script workflow (E10). NVIDIA also noted that Unreal Engine recently announced the ability to connect AI clients to Unreal Editor through MCP, enabling AI workflows to interact with editor capabilities through a standardized protocol (E11).
NVIDIA said it had 21 technical papers accepted at SIGGRAPH 2026, describing them as the foundation of real-time systems that generate virtual worlds and drive machine training in the real world (E22). Among the results NVIDIA highlighted is MotionBricks, a real-time motion model trained on more than 350,000 motion clips and running at game-engine speeds, which lets creators direct and connect character movements. NVIDIA said the same model that drives an animated character on screen also drives a Unitree G1 humanoid robot in the room, using computer graphics and simulation to accelerate physical AI development (E23).
Taken together, the figures NVIDIA disclosed at SIGGRAPH 2026 trace a consistent pattern: models and hardware are being scaled along parallel tiers, from Cosmos 3's 4-billion, 16-billion and 64-billion-parameter options (E17) to DGX Station configurations that scale from a single desktop unit with 748GB of coherent memory up to two linked stations over an 800GB/s ConnectX-8 link (E19, E20). The Synthetic Video Detector figures show accuracy declining as compression increases — 92% uncompressed versus 82% at 50% compression — a tradeoff NVIDIA disclosed within the same set of tests (E12). And two separate demonstrations, Cosmos-Dreams running on a single RTX PRO 6000 GPU (E6) and MotionBricks driving both a screen character and a Unitree G1 humanoid robot (E23), both link simulation output directly to physical or physically modeled systems, consistent with the "physical AI" framing NVIDIA used across its keynote and its 21 accepted papers (E22).
According to NVIDIA, the keynote took place July 20 at 3:45 p.m. PT, featuring Neil Ashton, Edward Liu and Ming-Yu Liu.
NVIDIA said it had 21 technical papers accepted at the conference.
In NVIDIA's own testing, accuracy reached up to 92% on uncompressed video, 87% at 15% compression and 82% at 50% compression.
According to Intel's newsroom announcement, Intel's Fab 9 advanced packaging site in Rio Rancho, New Mexico has scaled its 'reticle limit' to 8 times the current industry standard, with a target of over 12 times by 2028. Intel says the facility grew from 25 employees in 1980 to 2,700 employees and 500 suppliers today, using Foveros and EMIB/EMIB-T technologies to package AI chips.
According to NVIDIA's July 21, 2026 blog post, Vera Rubin NVL72 delivers up to 10x more tokens per megawatt and one-tenth the cost per million tokens than GB200 NVL72. CoreWeave's DeepSeek-R1 benchmark and Google Cloud's new A5X instance each independently report the same 10x per-megawatt token multiplier on live hardware.
According to a NVIDIA blog post, Bristol Myers Squibb (BMS) is deploying its second NVIDIA DGX SuperPOD, built on eight DGX Vera Rubin NVL72 systems delivering up to 10x the performance per megawatt of the infrastructure it replaces. The move follows roughly three years running its first SuperPOD, which BMS says had become saturated.