GTC 2026 Highlights: Blackwell Ultra, DGX Updates, RTX PRO Workstations, and What's Next

GTC 2026 (GPU Technology Conference) was NVIDIA’s biggest event yet, with announcements spanning every segment of the AI hardware and software ecosystem. From the Blackwell Ultra architecture to new personal AI supercomputers, here are the highlights that matter for AI infrastructure buyers and developers.

Blackwell Ultra: The Next Compute Tier

The headline announcement was the NVIDIA Blackwell Ultra architecture, the successor to Blackwell that powers the new B300 GPU and GB300 systems. Key improvements over standard Blackwell:

50% more FP4 compute per GPU compared to B200
50% more HBM3e capacity, up to 288GB per GPU
2x attention performance, critical for transformer-based models and AI reasoning
Enhanced NVLink for tighter multi-GPU coupling

Blackwell Ultra is specifically optimized for the emerging class of AI reasoning workloads, multi-step inference where the model “thinks” before answering, exemplified by test-time scaling techniques.

DGX GB300 NVL72: Rack-Scale AI Factory

The DGX GB300 NVL72 integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs into a single liquid-cooled rack. At 1.1 exaFLOPS of dense FP4 compute with 38TB of fast memory, it delivers up to 50x the AI factory output of Hopper-based systems.

NVIDIA positions the DGX GB300 as the building block for “AI factories”, purpose-built infrastructure for continuous AI training, post-training, and inference at enterprise scale.

DGX Spark and DGX Station: AI for Every Desk

On the personal computing side, NVIDIA announced partnerships with ASUS, Dell, Gigabyte, HP, and Lenovo to bring DGX-branded personal AI computers to market:

DGX Spark: 1 PFLOPS, 128GB, $249-class form factor, AI development for every developer
DGX Station: 20 PFLOPS, 748GB, desktop AI supercomputer for research teams

RTX PRO Blackwell for Workstations

The professional visualization team announced the full RTX PRO Blackwell lineup:

RTX PRO 6000: 24,064 CUDA cores, 96GB GDDR7, 4,000 AI TOPS at 600W
RTX PRO 5000: 14,080 CUDA cores, up to 72GB GDDR7 at 300W
RTX PRO 4500/4000/4000 SFF/2000: Completing the lineup from 200W towers down to 70W SFF systems

Lenovo, Dell, and HP all announced workstations featuring RTX PRO Blackwell GPUs for AI-ready design and simulation workflows.

Jetson Thor: Blackwell at the Edge

The Jetson AGX Thor developer kit became generally available, bringing 2,070 TFLOPS of Blackwell AI to edge deployments. Key applications highlighted at GTC include humanoid robots, autonomous vehicles, and AI-powered medical devices.

IGX Thor: Industrial and Medical AI

The IGX Thor platform was announced for industrial and medical edge AI, featuring functional safety certification (ISO 26262 ASIL D) and 10-year lifecycle support. The T7000 production board and T5000 SOM enable OEMs to build certified AI products for regulated industries.

Software Announcements

NVIDIA NIM: Expanded catalog of inference microservices for faster model deployment
NVIDIA AI Enterprise 6: Updated enterprise software platform with Blackwell optimization
NVIDIA Omniverse: Enhanced digital twin and simulation capabilities
NVIDIA Isaac: Updated robotics platform with Thor support

What This Means for Buyers

GTC 2026 signals a clear direction: AI is moving from experimental to industrial. The product announcements reflect this, from personal development tools (DGX Spark) through production edge platforms (Jetson Thor, IGX Thor) to enterprise-scale AI factories (DGX GB300). Every tier of the stack got a significant upgrade.

For organizations planning AI hardware purchases, the message is clear: the Blackwell generation is here, with products available across every deployment scenario.

Want to understand how these announcements affect your AI infrastructure plans? Contact us for a consultation on product selection and upgrade timing.