NVIDIA GTC 2026: Physical AI, Edge Computing, and the GB300 Era
NVIDIA's annual GTC conference in San Jose ran 16–20 March 2026, and Jensen Huang's keynote set the tone for what the company is calling the Physical AI era — intelligence that operates in the real world, not just in browsers and chat windows. The conference drew over 25,000 in-person attendees and more than a million virtual participants, cementing GTC as the defining event of the AI infrastructure calendar.
DGX Station GB300: Frontier AI at Your Desk
The first NVIDIA DGX Station systems, powered by GB300 superchips, began shipping to pioneering developers in early March. These deskside workstations bring frontier-scale model training within reach of individual researchers and small teams without a data centre contract.
| Spec | DGX Station GB300 |
|---|---|
| GPU | NVIDIA GB300 NVL72 |
| Memory | 1.4 TB unified GPU memory |
| Interconnect | NVLink 6, 1.8 TB/s bandwidth |
| Form factor | Deskside workstation |
| Target user | Researchers, university labs, AI startups |
IGX Thor: Real-Time Physical AI at the Edge
NVIDIA announced general availability of IGX Thor, an industrial-grade edge computing platform engineered for safety-critical machines operating in environments where cloud latency is unacceptable:
-
Caterpillar is deploying an in-cabin conversational AI assistant powered by IGX Thor to enhance worker productivity and safety on heavy machinery in remote mining and construction sites.
-
Hitachi Rail is using IGX Thor for real-time rail condition monitoring and predictive maintenance, reducing unplanned downtime on high-speed rail networks.
-
The platform targets: construction, manufacturing, logistics, healthcare (surgical robotics), and space exploration.
Nemotron 3 Super: Open-Weight Coding Champion
Released at GTC on 11 March, Nemotron 3 Super achieves 60.47% on SWE-Bench Verified — the highest open-weight score currently published — making it the go-to foundation for enterprise coding agents that need to stay on-premises for IP protection or regulatory reasons.
| Model | SWE-Bench Verified | Context | Licence |
|---|---|---|---|
| Nemotron 3 Super | 60.47% | 128K | Open-weight |
| Claude 3.7 Sonnet | ~62% | 200K | Proprietary |
| GPT-5.4 Standard | ~65% | 1.05M | Proprietary |
| Qwen 3.5 9B | ~44% | 32K | Apache 2.0 |
Key Takeaway for Enterprise Architects
GTC 2026 confirmed that the AI infrastructure stack is bifurcating into two distinct tiers — and organisations that architect for both will have a decisive operational advantage through 2027:
Tier 1 — Hyperscale GPU clusters for training, fine-tuning, and batch inference workloads where latency tolerance is high and throughput is paramount.
Tier 2 — Edge platforms (IGX Thor, Jetson) for real-time inference at the point of action, where sub-50ms latency, air-gap operation, and functional safety certification are non-negotiable.
The companies that win the Physical AI era will not be the ones with the biggest cloud contracts — they will be the ones that successfully extend AI reasoning from the cloud to the machine.