The nature of competition in AI infrastructure is changing. After years in which Nvidia’s advantage was associated primarily with owning the most advanced graphics processing units (GPUs), this analysis suggests that the center of gravity is gradually shifting to the components surrounding the processor: memory management, data routing, storage, networking, and coordination among system components. As AI workloads expand into the gigawatt range, simply increasing the number of processing cycles is no longer sufficient to achieve the required efficiency.
This assessment comes at a time when major cloud infrastructure companies, such as Amazon and Google, have begun developing their own chips, weakening the idea that Nvidia is the only option for obtaining advanced AI hardware. This has raised questions about the sustainability of the company’s lead, particularly after its market value grew tenfold between the beginning of 2023 and the middle of 2025, before its share-price movement became more moderate during the following year amid concerns about competition in GPUs.
From the Processor to the Rack-Scale System
According to the article, Nvidia is currently introducing the Vera Rubin architecture, which combines the Rubin GPU with a range of other units, including the Vera CPU, the Groq 3 LPX inference accelerator, and racks dedicated to storage and networking. These components do not necessarily focus on executing tokens directly, but on ensuring that data reaches the GPUs at the right speed and timing.
The GPU can be compared to an engine, while the rest of the system’s components perform the functions of a car that ensure the engine’s power is used efficiently. The greater the capacity of data centers, the greater the need for memory, but transferring data from memory to the processor without bottlenecks has become an independent challenge. This challenge is becoming more important as infrastructure operators seek to reduce the number of tokens that can be produced per watt.
The Role of the Vera CPU in Routing Data
Jason Hardy, Nvidia’s vice president of storage technologies, said that Vera is important because the amount of memory that can be placed in a single server or computing platform is limited. According to the analysis’s account of his comments, operations achieved an improvement of up to threefold when the Vera CPU accelerated them, allowing flash storage units to be used to their full potential instead of having performance stall because of data-transfer bottlenecks.
The practical significance here is that data-center performance is not determined solely by the specifications of an individual GPU. Even a fast processor may not deliver its full benefit if data does not reach it at the right time, or if memory, storage, and networking cannot keep pace with the workload. Nvidia is therefore attempting to sell an integrated system, not merely a standalone chip.
A Different Solution from OpenAI
Addressing the data-movement problem is not limited to Nvidia. When OpenAI developed the Jalapeño chip, reducing the amount of data that had to be moved between components was a central design focus. The company said that the chip’s large die area allows the entire workload to remain within a single connected system, reducing data movement and communication latency and helping keep requests fast and efficient from beginning to end.
The two approaches differ: Nvidia’s system seeks to manage data movement efficiently through a group of specialized components, while OpenAI seeks to reduce the movement itself by keeping the workload within a chip or integrated system. But the intended result is the same: improving efficiency rather than relying solely on adding processing cycles or more chips.
Why Does This Shift Matter?
A new layer of competition in AI is emerging here. Building a competing GPU is no longer necessarily sufficient; companies also need to make the entire system operate efficiently. This creates an opportunity for chip manufacturers and cloud-service operators to compete in rack design, memory coordination, and storage and networking management—not only in raw processing power.
certi.news analysis: What has truly changed is the definition of performance in AI data centers. The benchmark is shifting from the question, “How fast is the chip?” to a broader question: “How efficiently can the system keep the chip supplied with the right data?” This gives Nvidia an opportunity to extend its advantage into a layer that is difficult to separate from the rest of the infrastructure, but it does not settle the competition. The company will face competitors in GPUs as well as in systems management and coordination, and the source does not provide comprehensive comparative figures between Nvidia’s system and competitors’ solutions, nor does it establish that the current advantage will continue in the long term.
For data-center operators, this means that infrastructure decisions may require considering the efficiency of the entire system, not just the price or performance of the processor alone. For chip designers, the challenge is expanding from increasing computing power to reducing data movement and bottlenecks. The analysis indicates that Nvidia appears to have a strong early lead in this new layer, without implying that the competition is over or that a lasting advantage is guaranteed.