Technical Context
I dug into what NVIDIA actually showcased, and this story isn't just about another 'even faster GPU.' They are moving an entire AI factory stack: Rubin GPU, Vera CPU, NVLink 6, ConnectX-9, BlueField-4, and the network plumbing designed for massive inference workloads. For those implementing AI in production, this matters far more than beautiful slide decks.
The most practical element here, in my view, is the DGX Vera Rubin NVL72. Combining 72 Rubin GPUs and 36 Vera CPUs in a single rack, it promises fewer data transfer bottlenecks and a better cost per token compared to Blackwell. I'd take these vendor claims with a grain of salt for now, but the direction is clear.
What caught my attention is the Vera CPU. NVIDIA didn't just slap a CPU on 'for show'; they clearly engineered it for data movement, reasoning workloads, and tight integration with accelerators. When a single company controls the GPU, CPU, network, and DPU, you no longer have just a server—you get a unified AI architecture.
Another strong signal: they continue to push the idea of the data center as a simulatable object. Running everything through an Omniverse digital twin before physical construction sounds less like marketing and more like standard engineering practice. If this is truly battle-tested with clients, then designing AI infrastructure is becoming closer to a software workflow than the classic 'build and pray' approach.
The robotics story is similar. There are fewer public details than I'd like, but NVIDIA is once again weaving hardware, simulation, and local inference into a single loop. And this is the actual foundation for physical AI, not just exhibition demos.
What This Changes for Business and Automation
I see three major implications. First, large-scale AI automation is becoming less about 'which model to choose' and much more about cost per token, network topology, and memory. An architecture mistake will cost far more than a prompt error.
Second, those who build long pipelines, agentic systems, and robotic processes will win. Those who buy hardware without a clear use case and workload estimation will lose.
Third, digital twins for infrastructure will become the norm. I wouldn't be surprised if, in a couple of cycles, nobody builds serious AI capacity without running simulations first.
It is precisely at these crossroads that you don't need generic 'AI advice' but proper AI solution development: what to automate, where to run inference, when to go on-prem, and when to avoid it. If your business is already hitting a wall with AI process costs or implementation chaos, let's look at it pragmatically. At Nahornyi AI Lab, we build AI automation around real workflows so that it actually pays off, rather than just sounding good in a strategy deck.