NVIDIA is using Hot Chips to make its next-generation Vera Rubin platform about more than raw AI compute. The company’s latest announcements span inference accelerators, CPUs and networking, reflecting an industry increasingly focused on the infrastructure required to run AI agents efficiently at large scale.
Among the more notable disclosures is the first on-silicon performance data for the Vera Rubin NVL72 system. NVIDIA says its measurements using SemiAnalysis’ AgentX workload show Vera Rubin NVL72 delivering up to 30 times more work per watt for AI agents. That is a substantial claim, although it is based on NVIDIA’s own testing and will need broader independent benchmarking as Rubin hardware reaches customers.
Efficiency has become a particularly important battleground for AI infrastructure. Training enormous models remains expensive, but inference is increasingly the recurring cost that operators need to control. Agentic AI can amplify that problem because agents may execute multiple model calls, use tools and process long contexts while completing a single task. Performance per watt therefore matters almost as much as headline throughput for companies contemplating deployments involving thousands or millions of agent interactions.
NVIDIA is also putting more emphasis on specialized inference hardware. Groq 3 LPX has entered full production and is being positioned alongside Vera Rubin for workloads where low latency and long-context responsiveness are priorities. The move gives NVIDIA another architectural option as competition around AI inference shifts beyond simply adding more general-purpose accelerator capacity.
The company’s Hot Chips announcements also extend into the CPU side of the stack. SpaceXAI is adopting NVIDIA’s Vera CPU for large-scale agentic AI infrastructure, providing an early high-profile deployment for a processor intended to work closely with NVIDIA’s wider accelerated computing platform.
Networking is the other major piece of the strategy. NVIDIA detailed Spectrum-X Multiplane, an approach intended to support larger AI clusters with flatter network designs and improved resilience. It also introduced what it calls “scale-in” infrastructure, built around BlueField-4 and DOCA, as it looks for ways to move data efficiently between compute, storage and networking resources inside increasingly complex AI systems.
Taken together, the announcements show where NVIDIA expects the next phase of AI infrastructure competition to develop. GPUs remain central, but processors alone are no longer enough to differentiate an AI platform. CPUs, networking, data processing, interconnects and specialized inference silicon are becoming parts of the same system-level argument.
Vera Rubin NVL72’s early efficiency numbers will attract attention, particularly as power availability becomes a constraint on new data center capacity. The more meaningful test will come when customers and independent researchers can compare production Rubin systems against Blackwell and competing architectures under realistic agent workloads. For AI operators, the question is increasingly not just how fast a model can run, but how much useful work an entire data center can deliver for every watt it consumes.
