Nvidia's Vera Rubin goes beyond GPUs: the company now wants to own the CPUs, networking, and cooling inside every AI data center

Nvidia built its data center business on a simple proposition: we make the fastest GPU, you handle the rest. Vera Rubin, now ramping toward what CEO Jensen Huang calls "full production" for the second half of this year, scraps that model 1. The system is a bid to sell the CPU, the GPU, the networking, and the liquid cooling as one closed product, and OpenAI is already running a rack 1.

The spec sheet is where the strategy becomes concrete. A single Vera Rubin NVL72 rack pairs 36 ARM-based Vera CPUs with 72 Rubin GPUs, fully liquid-cooled 1. Nvidia claims it processes ten times as many tokens per watt as the previous Grace Blackwell superchip and delivers nearly three times the memory bandwidth 1. The company describes the racks as "cable-free compute" and "hot-swappable," reducing installation time per rack from a couple of hours to a few minutes 1.

If every rack a hyperscaler installs takes minutes instead of hours, and arrives with its CPU, GPU, network, and cooling pre-configured, Nvidia is not selling components. It is selling the data center, one rack at a time.

The CPU is the most revealing piece of this strategy. Nvidia's data center CPUs run on ARM, an architecture known for power efficiency, rather than the x86 architecture that still accounts for the vast majority of data center CPU revenue 1. AMD, which WIRED describes as "recognized as a pioneer of the modern chiplet architecture," has spent the past two years growing its share of the data center CPU market by stitching together smaller chip dies into a single package 1.

Nvidia is making the opposite bet. Vera Rubin uses a monolithic design, a single integrated circuit rather than assembled parts. Hannah Coutand, who runs Nvidia's CPU product marketing, argued that the chiplet approach imposes "a heavy tax on memory bandwidth and data movement" 1.

This is an architecture fight with enormous commercial stakes. If Nvidia is right that monolithic dies move data faster than assembled chiplets, AMD and Intel's modular approach becomes a performance constraint, not a cost advantage.

The full-stack ambition extends past the silicon. Nvidia is selling the Vera CPU as a standalone product and has reportedly told Chinese customers it could be ready as soon as August 1. The company is also sensitive to delay perceptions after its previous-generation Blackwell chips reportedly overheated in custom server racks, forcing design changes and pushed-back shipments 1. The marketing push is timed just ahead of AMD's annual conference, where AMD revealed more details about its Helios AI chip rack, designed to compete with Nvidia's system 1.

Ian Buck, Nvidia's longtime vice president of accelerated computing and the architect behind the company's CUDA software, framed the stakes without subtlety: "We're going to keep innovating, because it's do this or die in Silicon Valley" 1.

For builders and investors, the signal is in the customer list. Nvidia's early Vera Rubin customers include Microsoft, Oracle, and OpenAI 1. When the company that buys your GPU for training also buys your CPU for orchestration, your networking for interconnects, and your liquid-cooling system for thermal management, switching to AMD or Intel for any single layer means either accepting a fragmented stack or ripping out infrastructure that works.

Nvidia is no longer arguing it makes the best chip. It is arguing it makes the only stack. Whether AMD's chiplet architecture and Intel's x86 dominance can survive a competitor that controls the CPU, the GPU, the network, and the cooling as one integrated system will determine whether the AI infrastructure market stays competitive or consolidates around a single supplier.