NVIDIA AI Updates: July 16, 2026

1. NVIDIA Launches Jetson Thor T3000 and T2000 Modules for Mainstream Robotics

NVIDIA. NVIDIA introduced two new Jetson Thor modules, the T3000 and T2000, to bring higher AI compute to mainstream robotics and edge systems. The T3000 delivers 865 FP4 teraflops of AI compute with 32GB of memory, while the T2000 provides 400 teraflops with 16GB, giving developers scalable options across performance tiers. Both modules are slated to become available in Q1 2027, with developer emulation support arriving later in July 2026 through JetPack 7.2.1. Source

2. NVIDIA Introduces Nemotron Labs to Push Open, Customizable Enterprise Models

NVIDIA. NVIDIA introduced Nemotron Labs, framing open models as a way for enterprises and nations to build AI systems they can trust, control, and customize. The company argues that competitive advantage comes from how organizations adapt available models rather than which model they pick, and it cited deployments spanning clinical documentation and legal work. NVIDIA said the approach can deliver significant cost reductions compared with closed alternatives. Source

3. NVIDIA CUDA 13.3 Adds Carryless Multiply Instruction to Speed Cryptography

NVIDIA. CUDA 13.3 introduces clmad, a hardware-accelerated carryless multiply-accumulate instruction for binary extension field arithmetic on Ampere and newer GPUs. NVIDIA reported that the instruction speeds GHASH operations in AES-GCM encryption by up to 18.8x over the prior bitsliced state of the art, reaching roughly 6.3 TB/s on a B200. The same primitive accelerates zero-knowledge proof protocols by 3x to 13x, positioning it for cryptography, error-correcting codes, and modern verification systems. Source

4. NVIDIA DeepStream 9.1 Adds Multi-View 3D Tracking and Agentic Skills

NVIDIA. NVIDIA released DeepStream 9.1 with Multi-View 3D Tracking (MV3DT) and AutoMagicCalib to keep object identities consistent across multiple camera feeds. MV3DT projects detections from calibrated cameras into a shared 3D coordinate system and maintains globally consistent object IDs, addressing limitations of single-camera 2D tracking. The release ships 13 agentic skills that let developers build multi-camera pipelines using natural language prompts with coding agents. Source