Key Highlights
· NVIDIA confirmed $1 trillion in orders for Blackwell and Vera Rubin systems through 2027, up from $500 billion previously
· Vera Rubin is a fully integrated AI factory platform comprising seven chips and five rack-scale systems, with all chips confirmed in full-scale production
· The Feynman architecture, NVIDIA's next generation platform after Vera Rubin, was revealed at GTC including the Rosa CPU and LP40 Liquid Processing Unit
· Rack densities confirmed on a path from 120kW today toward 600kW under the Kyber architecture expected in the second half of 2027
· OpenClaw and NemoClaw introduced as an open-source agentic AI operating system, described by Huang as the Android for AI agents
· Delta Electronics presented 800 VDC power architecture co-developed with NVIDIA as the baseline for next-generation AI data centre builds
· NVIDIA confirmed Australia as an active sovereign AI deployment market within the NVIDIA Cloud Partner program
· DGX Station delivers 20 petaflops at desk scale, capable of running trillion-parameter models locally, confirming distributed AI as a parallel infrastructure trend
Announcement | Significance for Data Centres |
$1 trillion orders confirmed | Signals scale of committed global infrastructure investment |
Vera Rubin NVL144 | First fully integrated AI factory platform, H2 2026 |
Feynman architecture revealed | Roadmap beyond Vera Rubin, Rosa CPU + LP40 LPU |
OpenClaw and NemoClaw | Agentic AI OS, raises sovereign infrastructure demand |
800 VDC power architecture | New baseline for AI facility electrical design |
DGX Station 20 petaflops | Confirms distributed AI as parallel infrastructure trend |
Vera Rubin Space-1 | Orbital AI compute signals long-term workload distribution thinking |
NVIDIA GTC has served as a forward indicator for data centre investment cycles for several years. GTC 2026, held in San Jose from March 16 to 19, confirmed that the infrastructure shift already underway in Australia is not a local phenomenon.
The Inference Inflection Is the New Demand Driver
Jensen Huang opened his keynote by confirming that the inference inflection has arrived. Training workloads made the current generation of AI infrastructure necessary. Inference at scale, serving hundreds of millions of users in real time, is what defines the next phase of demand. Huang confirmed Vera Rubin as the platform for prefill inference workloads, with the Groq 3 LPX integration addressing the decode stage. The Vera Rubin NVL144 is confirmed for the second half of 2026, with Microsoft Azure confirmed as the first hyperscale provider to deploy Vera Rubin NVL72 systems.
Inference is less bursty than training but more continuously active and more latency-sensitive. Facilities being planned in Australia today will run inference workloads at scale for the majority of their operational lives. The design implications of that reality need to be in the brief now.
NVIDIA Is No Longer a Chip Company
The most consequential framing from GTC 2026 was Jensen Huang's explicit statement that NVIDIA has completed its transition from a chip company to a full-stack AI infrastructure company. NVIDIA now positions itself as the architect of the complete physical plant of the AI economy: compute, networking, storage, software, and the infrastructure systems that connect them at facility scale.
The commercial weight behind that statement is significant. NVIDIA confirmed $1 trillion in orders for Blackwell and Vera Rubin systems through 2027, up from $500 billion previously. AWS confirmed deployment of more than one million NVIDIA GPUs across its global cloud regions this year, spanning Blackwell, Rubin, and Groq LPU architectures. These are not pipeline figures. They are committed infrastructure programs that will drive facility construction, power procurement, and cooling investment across every major market including Australia.
The AI Factory Is the New Unit of Compute
The Vera Rubin platform, confirmed at GTC as the fundamental unit of the new compute paradigm, is the first vertically integrated AI system built from the ground up for agentic workloads. It comprises seven chips, five rack-scale systems, the Vera CPU designed for long-term memory and planning, BlueField-4 STX storage, Spectrum X co-packaged optics, and the Groq 3 LPX rack tightly integrated for token acceleration. Vertiv presented at GTC on how factory-assembled modular blocks combining power, cooling, and IT infrastructure reduce on-site construction time and enable future-ready capacity through digital-first design using NVIDIA Omniverse.
NVIDIA also revealed the Feynman architecture, the platform that follows Vera Rubin, including the Rosa CPU and LP40 Liquid Processing Unit. The reveal of the architecture beyond the current roadmap reinforces the argument made at the Data Centre Leaders Summit in Sydney this week: Australian facilities must be designed with sufficient flexibility to absorb compute generations that do not yet exist commercially.
Agentic AI Raises the Sovereign Infrastructure Stakes
A theme running through GTC 2026 with specific weight for Australian enterprise and government buyers was the agentic AI framing. Huang introduced what he called the Android for AI agents: OpenClaw as the open-source foundation and NemoClaw as NVIDIA's production-ready deployment stack built on top of it. Live workshops at GTC demonstrated deployment of working AI agents in under an hour. Agentic systems, which involve AI taking autonomous actions across enterprise environments around the clock, carry data handling and security implications that make sovereign infrastructure a requirement rather than a preference for Australian government and regulated industry deployments. Facilities with certified architecture and Australian data residency are directly in the path of that demand.
Power and Density Require Planning Decisions Now
Delta Electronics used GTC 2026 to present an 800 VDC power architecture co-developed with NVIDIA, positioned as the baseline for next-generation AI data centre builds. The configuration includes in-row power racks with battery backup, megawatt-scale liquid cooling distribution units, and a microgrid combining solid-state transformers, fuel cells, and energy storage, with a confirmed migration path from existing 415 VAC infrastructure. NVIDIA's confirmed roadmap points toward 600kW per rack under the Feynman generation, expected in the second half of 2027. For Australian operators making facility design decisions today, GTC 2026 confirmed that liquid cooling is a baseline requirement and that power infrastructure must carry headroom for density levels that do not yet exist at commercial scale.
Distributed AI Changes the Capacity Equation
One of the most commercially significant announcements at GTC 2026 for enterprise infrastructure buyers was the introduction of DGX Spark and DGX Station. DGX Station delivers 20 petaflops at desk scale and is capable of running trillion-parameter models locally.
The Australian Angle
Australia’s data centre sector enters this period with structural advantages that GTC 2026 reinforced. NVIDIA confirmed Australia as an active market within its Cloud Partner sovereign AI deployment program, aligning directly with Australia's Five Eyes positioning, data residency requirements, and the demand from government and enterprise for infrastructure operating under Australian law.
The $1 trillion orders figure from GTC is the global capital base from which Australia's $142 billion annual opportunity is being drawn. GTC confirmed the direction that capital is flowing: toward certified, AI-ready, liquid-cooled facilities with integrated power and thermal management and sovereign credentials. Huang also announced plans to deploy AI data centres into Earth's orbit under the Vera Rubin Space-1 program. The practical implications for Australian ground infrastructure remain distant, but the question of where AI compute physically lives is one NVIDIA is thinking about at every layer of the stack.