Nvidia's GTC keynote

Nvidia's GTC keynote yesterday had a lot of chip announcements, but the real story is vertical integration.

The Vera Rubin platform combines GPU, CPU, networking, and inference acceleration into one integrated stack. When a GPU vendor launches its own CPU and secures inference capability in the same breath, the intent is obvious: they want to be the only name on your AI invoice.

I watched AWS expand from EC2 to everything over the years. Same playbook. Start with compute. Add networking. Add storage. Add orchestration. Keep adding until switching costs become prohibitive and your negotiation power evaporates.

The Vera CPU launch tells you everything. Nvidia is removing Intel and AMD from your stack, owning one more layer, controlling one more pricing lever. What's particularly interesting: these CPUs appear optimised for single-threaded coordination workloads with massive memory access, not traditional multi-threaded compute. They're designed to orchestrate data movement between GPUs and manage control plane operations for AI pipelines—essentially becoming the conductor while GPUs do the heavy lifting. This architectural choice signals that Nvidia sees the CPU not as competing with traditional server processors, but as the essential glue that locks the entire AI stack together. Vertical integration playbook, running at scale.

For anyone thinking about AI infrastructure governance, some uncomfortable questions emerge. What percentage of your AI spend goes to one vendor? Can you articulate cost per inference, cost per training run, cost per business outcome? What happens to your economics if Nvidia adjusts pricing next year? What are your walk-away terms?

Without unit economics, you're guessing while vendors chase trillion-dollar revenue targets [1].

The cloud parallel feels familiar. Remember when "AWS will help us optimise" felt like a strategy? Same dynamic, different domain, higher margins for the vendor. Architecture designed with exit options helps you maintain negotiation posture. If your AI stack can only run on Nvidia, you've already lost the pricing conversation.

Building financially sustainable AI infrastructure means understanding vendor strategies alongside chip specifications. Context is everything, and Nvidia's context is vertical integration. What's yours?

Sources and first draft AI, fine-tuning Frank.

[1] https://www.reuters.com/technology/nvidia-posts-record-revenue-strong-ai-demand-2024-02-21/