Nvidia’s AI advantage is moving beyond the GPU. The company’s market cap increased 10x between the start of 2023 and mid-2025, but Nvidia shares have followed a more modest trajectory for the past year. That shift puts more attention on the parts of an AI system that sit around the main processor: memory, data movement, communication, and the hardware that keeps everything working together.
“Nvidia’s advantage goes far beyond GPUs.” That idea now reaches into the company’s approach to data centers, where performance depends on more than raw computing power. A powerful GPU still matters, but the full system also needs to move data to the right place and coordinate operations across different units.
Vera Rubin Treats Orchestration as a Hardware Problem
Nvidia has built hardware around this broader view through the Vera Rubin architecture. The design pairs the Rubin GPU with other units, including the Vera CPU and the Groq 3 LPX inference accelerator. Each part addresses a different task within the system, creating a hardware platform focused on how AI work flows through a data center.
The Vera CPU focuses on the problem of orchestrating data. Jason Hardy, Nvidia’s VP of storage technology, described the reason for that focus in direct terms: “Vera is important because there’s only so much memory that you can put in a single server or any sort of compute platform.” When memory cannot all sit in one place, the system must coordinate where data goes and when it moves.
Nvidia saw upwards of 3x improvement in operations where the Vera CPU allowed for acceleration. That figure points to a part of AI performance that can remain hidden when attention stays fixed on GPU specifications. The result depends on the interaction between processing units, memory, and the movement of information through the full system.
Other Nvidia technologies, including Nvidia CCCL and Nvidia Nemotron, sit within this wider AI hardware and software picture. Nvidia is also supporting Chinese open AI models, showing that its reach extends across different model efforts and computing environments.
Why Data Movement Shapes AI Performance
AI systems do not spend every moment performing calculations on a GPU. They also transfer data, coordinate tasks, and handle communication between parts of the platform. Those steps can add time to a workload, especially when data must travel between a host system and a device or between separate units in a server.
The performance figures connected with Nvidia’s optimization work show how much those steps can matter. Median computation time per image fell from 2.1 seconds to 773 microseconds, with a listed speedup of 2,717x in median computation. Total pipeline time dropped by 2.6x, while host-to-device transfers accelerated 10x.
The full three-image run time dropped to about 25 milliseconds. The final runtime reached 23 milliseconds, which was 300x faster than the original 6.8 seconds. These numbers describe a complete pipeline rather than a single processor measure, so they connect computation, transfers, and the work needed to finish the run.
That distinction matters because a fast processor cannot remove every delay by itself. Nvidia’s approach places system orchestration and data management alongside GPU performance, treating them as connected parts of the same AI problem. The Vera CPU addresses that coordination task, while the Rubin GPU and Groq 3 LPX inference accelerator contribute their own roles within the architecture.
OpenAI Takes a Similar Aim at Communication Delays
OpenAI developed its Jalapeño chip with a related goal. The company said, “We designed Jalapeño to minimize data movement and communication delays.” That focus places the movement of information at the center of chip design, not at the edge of the discussion.
Nvidia and OpenAI are working with different hardware names and designs, but the stated challenge is closely connected: AI performance depends on how quickly a system can coordinate data as well as how much computation a processor can perform. The figures tied to Nvidia’s pipeline show the impact of reducing those delays, while OpenAI’s description of Jalapeño states the design goal in simple terms.
This focus also helps explain why Nvidia’s story now includes more than GPU sales. Its market cap rose 10x between the start of 2023 and mid-2025, and its shares have taken a more modest path for the past year. The next stage of its AI advantage rests on the complete platform: processors, accelerators, storage technology, memory limits, and the software that coordinates them.
Russell Brandom has been covering the tech industry since 2012, and Kai Nicol-Schwarz is listed as a discussant or author connected with this discussion. The startup community will gather in San Francisco from October 13 to 15, creating a setting for more attention on the companies and technologies shaping AI systems.
The direction is clear from the hardware and performance figures. Nvidia is building around the idea that AI speed comes from the whole system, not one chip in isolation. When data movement and orchestration receive dedicated hardware, the gains can show up across the full pipeline.
Based on
- The Modern CUDA Toolbox in Practice: A Step-by-Step Optimization Walkthrough — developer.nvidia.com
- Here are 2 underappreciated positives from Nvidia’s stellar earnings — cnbc.com
- Nvidia’s AI advantage is moving beyond the GPU | TechCrunch — techcrunch.com
- Nvidia is bolstering support for Chinese open AI models — cnbc.com
