Nvidia's Vera Rubin Architecture: Moving the AI Bottleneck From Compute to Data

Nvidia's Vera CPU is built to keep data flowing to GPUs so memory doesn't become a bottleneck, and the company reports up to 3x improvement in operations as a result, allowing flash storage to operate at full potential TechCrunch. The chip is one of six that make up Nvidia's Rubin architecture, named for astronomer Vera Florence Cooper Rubin, which was launched in January 2026 TechCrunch.

The Rubin GPU itself is technically two GPUs in one. Paired with the Vera CPU, the combination can manage up to 50 petaflops during AI inference TechCrunch. At GTC 2025, Nvidia laid out a GPU roadmap spanning Blackwell Ultra, Vera Rubin, and Feynman, pointing to a multi-generation plan built around increasingly integrated silicon TechCrunch.

Nvidia VP of storage technology Jason Hardy framed the problem directly: Vera matters because there is only so much memory that can be put in a single server or compute platform TechCrunch. The bottleneck has shifted from raw compute horsepower to data delivery, and the Vera CPU's role is to manage that pipeline.

The architecture's interconnect layer is where the design philosophy becomes concrete. NVLink-C2C delivers up to 1.8 TB/s of coherent bandwidth between Vera CPUs and Nvidia GPUs Nvidia. The Vera Rubin POD integrates 72 Rubin GPUs and 36 Vera CPUs connected through an NVLink copper spine Nvidia Developer Blog. At a smaller scale, the Vera Rubin NVL4 connects four Rubin GPUs to two Vera CPUs over a bridge Nvidia. Nvidia also reports that its Vera Rubin NVL72 systems deliver up to 30x higher throughput per megawatt than the GB300 NVL72 Nvidia Blog. For configurations using PCIe rather than NVLink, the accelerated computing platform pairs Vera host CPUs with Rubin GPUs to enable balanced CPU-GPU performance across multiple server topologies Nvidia Developer Blog.

Beyond the single server, Nvidia has built out a three-tier networking model for what it calls AI factories. NVLink serves as the scale-up network, enabling GPUs inside a domain to behave as a single engine. Scale-out networks, built on Spectrum-X Ethernet, connect servers across the data center Nvidia Developer Blog. Spectrum-X delivers up to 1.6x faster AI network performance than traditional Ethernet, with congestion management and performance isolation across the storage path Nvidia. Spectrum-X Multiplane can scale AI networks to 128,000 GPUs in two tiers, 64x more than single-plane networks, by splitting each GPU's SuperNIC Nvidia. Spectrum-XGS Ethernet extends high-performance AI fabrics across racks, buildings, and entire campuses Nvidia Networking. At the host level, BlueField-4 powers what Nvidia calls scale-in network infrastructure for agentic AI factories Nvidia Developer Blog.

Demand is scaling with the architecture. Amazon tripled its Nvidia chip orders due to surging demand, with 2 million GPU chips being added to AWS starting in the third quarter, accompanied by an unspecified number of Vera CPUs TechCrunch. Nvidia is also building what it calls a Vera Rubin AI factory, a massive data center packed with next-generation chips TechCrunch. Google Cloud, meanwhile, launched two new AI chips of its own to compete with Nvidia, while also promising its cloud would have Nvidia's Vera Rubin available later in 2026 TechCrunch.

Nvidia is not the only company reasoning about data movement at the silicon level. OpenAI designed its Jalapeño chip to minimize data movement and communication delays, with a large domain that keeps an entire workload within one connected system TechCrunch. The design philosophy mirrors Nvidia's: the constraint is no longer floating-point throughput but the latency and bandwidth cost of moving data between compute and memory.

The financial context frames the stakes. Nvidia's market cap grew 10x between the start of 2023 and mid-2025, after which shares took a more modest trajectory for roughly a year, driven by concerns about GPU competition TechCrunch. Separately, CME began offering AI computing power futures, treating AI compute as a tradable asset class CNBC.

The broader context here is that the Vera Rubin architecture signals a transition from selling GPUs to selling full-stack AI compute systems. The six-chip Rubin design, the three-tier networking model, and the Vera CPU's data orchestration role collectively shift the battleground from single-chip performance to system-level throughput per watt. Nvidia reports 30x higher throughput per megawatt over the previous generation, and if that holds in deployment, it reframes the cost calculus for every hyperscaler evaluating build-versus-buy for AI infrastructure.

The competitive pressure is real. Google's custom TPU line continues to advance, and OpenAI's Jalapeño effort suggests that the largest AI workloads may eventually run on purpose-built silicon that is not Nvidia's. Nvidia's counter is not a better GPU in isolation. It is a vertically integrated system, from the CPU that feeds data to the GPU that consumes it, through the NVLink fabric that lets thousands of chips act as one, to the Ethernet that scales the network outward. Whether that stack-wide advantage holds against customers who increasingly want to design their own chips is the question that will define the next phase of this market.