For years, the battle for artificial intelligence computing could be summarized in three letters: GPU.

That description is becoming outdated. Graphics processing units remain at the center of the AI boom, but the industry’s competition is expanding beyond individual processors. The new battleground includes CPUs, high-bandwidth memory, networking, interconnects, software and entire racks of tightly integrated hardware.

In other words, Nvidia, AMD and their competitors are no longer simply fighting over who builds the fastest AI chip. They’re competing over who builds the best AI computing system.

The GPU Is Only One Piece

Training and running large AI models requires enormous amounts of computation, but adding more GPUs doesn’t automatically translate into equivalent performance.

Those processors must constantly exchange data.

They need access to memory fast enough to keep their computing cores occupied. They must communicate with CPUs, storage systems and thousands of other accelerators. And they increasingly need specialized networking capable of moving enormous quantities of information across an AI cluster.

That has shifted the industry’s focus from individual chip performance to what is known as scale-up and scale-out computing.

Scale-up technology makes accelerators inside a system behave more like one enormous processor. Scale-out networking connects those systems into much larger computing clusters.

Both are becoming critical.

Nvidia Is Selling the Rack

Nvidia’s GB300 NVL72 illustrates the transition.

Rather than presenting Blackwell Ultra as simply another GPU, Nvidia packages 72 GPUs and 36 Grace CPUs into a liquid-cooled rack-scale architecture connected through its NVLink interconnect.

The system also incorporates high-speed networking and Nvidia’s Mission Control management software.

Nvidia describes the architecture as purpose-built for increasingly compute-intensive AI reasoning workloads, where models may perform more computation while generating an answer.

The important part isn’t merely that Nvidia has a faster GPU.

Nvidia is increasingly controlling the architecture around the GPU.

AMD Answers With Helios

AMD is attacking the same problem.

Its new Helios rack-scale AI platform combines 72 Instinct MI455X GPUs, sixth-generation EPYC “Venice” CPUs and Pensando networking into one system.

AMD says Helios provides 31 terabytes of HBM4 memory and 1.7 petabytes per second of aggregate high-bandwidth-memory bandwidth, along with 260 terabytes per second of scale-up bandwidth connecting the GPU domain.

The company is explicitly positioning Helios as a complete alternative to Nvidia’s rack-scale platforms rather than merely pitching MI455X against an Nvidia GPU.

As AMD puts it, the rack itself is becoming the AI system.

Memory and Networking Join the Chip War

This shift creates opportunities well beyond GPU manufacturers.

High-bandwidth memory has become one of the most important components in AI hardware because processors must move enormous datasets in and out of memory extremely quickly.

Micron’s latest HBM4 technology, for example, can deliver more than 2.8 terabytes per second of bandwidth per stack—more than double the company’s previous-generation HBM3E implementation.

Networking is similarly important because thousands of accelerators are useless if they spend too much time waiting for data from one another.

That means the AI hardware race now includes companies specializing in memory, switches, optical connections and interconnect technologies.

The Winner May Own the Architecture

Nvidia’s advantage was built partly on making its GPUs extraordinarily useful for AI.

Its next challenge is larger: making every major component surrounding those GPUs work together better than competing systems.

AMD is trying to do the same thing with a more open architecture.

That changes what it means to win the AI chip war. The fastest processor still matters. But increasingly, the winner may be the company that best connects compute, memory, networking, software, cooling and power into a single machine capable of operating at enormous scale.