An assessment says Nvidia’s artificial intelligence competitiveness is expanding beyond the performance and supply capacity of GPUs themselves to system capabilities that coordinate entire data centres. As AI computing scales to gigawatt levels, the key factor shaping performance and cost is emerging as how efficiently memory, storage, networks and data flows are connected, rather than simply securing fast GPUs.

On Aug. 29 (local time), IT media outlet TechCrunch reported that after Nvidia’s recent earnings release, the market is analysing that Nvidia’s competitiveness no longer lies only in its advantage in GPU supply.

Nvidia’s core strength had been that it supplied cutting-edge GPUs virtually exclusively in the early days of the AI boom and reaped huge profits. But as hyperscalers such as Amazon and Google began developing their own AI chips, concerns grew over how long Nvidia’s GPU edge would last.

Those worries over intensifying competition also underpinned why Nvidia’s stock price gains have slowed relatively over the past year, after its market capitalisation rose nearly tenfold from early 2023 to mid-2025.

But as AI computing scales up, the competitive arena itself is changing. Operating large AI data centres efficiently requires more than boosting GPU computing performance alone. It also requires optimising the linking of memory, storage and networks, and delivering necessary data to GPUs on time.

In response to the shift, Nvidia is pursuing a strategy to build core hardware and systems needed for data centres together, beyond GPUs.

A representative example is the next-generation Vera Rubin architecture. Nvidia is showcasing an approach that combines Rubin GPUs with Vera CPUs and configures inference accelerators, storage devices and racks for networking as a single system.

The key is not simply adding computing units. The focus is on reducing situations where GPUs cannot deliver performance properly due to data shortages or network bottlenecks, and on improving the efficiency of the entire data centre.

In particular, the Vera CPU is tasked with reducing bottlenecks in data coordination. Jason Hardy (제이슨 하디), vice president for Nvidia storage technology, explained that there are physical limits to the amount of memory that can be installed in a single server or computing platform. Even if a data centre’s computing power and memory capacity increase together, overall system efficiency can fall if necessary data cannot be delivered to GPUs in time, he said.

Hardy said the Vera CPU showed up to a threefold improvement in some tasks responsible for acceleration. He explained that this made it possible to make full use of flash storage performance without bottlenecks.

As AI service companies seek to raise the number of tokens they can process per unit of power while lowering costs, this kind of data flow control capability could become an important competitive factor.

A similar awareness is also appearing at other AI companies. OpenAI said in a blog post earlier this month that it focused on reducing data movement and communication delays in its internally designed Jalapeno chip. It was designed to keep the entire workload within a single integrated interconnect system to minimise data movement, and to keep the full request processing process fast and efficient, it said.

The two companies’ approaches differ. While OpenAI processes workloads inside an integrated chip to reduce data movement itself, Nvidia’s approach is closer to raising efficiency by orchestrating the system overall, including GPUs, CPUs, storage and networks.

But their commonality is clear. As AI infrastructure expands at scale, moving data quickly and efficiently is becoming more important than simply boosting processors’ computing power.

Accordingly, the AI chip race is also shifting from comparing the performance of standalone GPUs to competing over system design capabilities for entire data centres.

Nvidia is expected to continue facing challenges from hyperscalers developing their own AI chips and from rival semiconductor companies. But in the current early stage of the AI infrastructure race, system capability to make entire data centres run efficiently is emerging as a more important variable than simply making GPUs that challenge Nvidia.

Whether Nvidia can evolve beyond the GPU market into a systems company that ties together data centres’ computing, memory, storage and networks is expected to be a key variable that will shape the competitive landscape of the AI infrastructure market.