The traditional data center was relatively straightforward.
Fill a building with racks of servers, connect them to networking and storage, keep everything cool and supply enough electricity to keep the machines running.
Artificial intelligence is rewriting that model.
Today’s most advanced AI infrastructure is increasingly being engineered as one enormous computing system, with processors, memory, networking, software, cooling and electrical systems designed together. The change is so significant that hardware companies have begun describing entire racks—and increasingly entire data centers—as the computer.
From Servers to Rack-Scale Computing
Nvidia’s GB300 NVL72 provides a good example.
The system integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs in a liquid-cooled rack connected through Nvidia’s NVLink technology.
Rather than treating each GPU-equipped server independently, NVLink enables the processors to operate inside a large, tightly connected computing domain.
Nvidia then connects racks through technologies including Spectrum-X Ethernet or Quantum InfiniBand while software coordinates workloads across the infrastructure.
The architecture is designed around a basic reality of modern AI: very large models need more computing resources than a single processor—or often a single server—can provide.
AMD Builds the Same Way
AMD’s new Helios platform follows the same philosophy.
Helios integrates 72 Instinct MI455X GPUs, EPYC Venice CPUs and Pensando networking into a rack-scale system with 31 terabytes of HBM4 memory.
The accelerators are joined through a scale-up fabric providing 260 terabytes per second of aggregate bandwidth, while separate scale-out networking allows workloads to expand across additional racks.
AMD’s description of Helios captures the industry’s direction neatly: the rack is no longer simply where the AI system is installed.
The rack is the AI system.
Networking Becomes Part of the Computer
That architectural shift makes networking far more important.
Traditional applications can often tolerate delays moving information between servers. Distributed AI training and inference can be much less forgiving.
Thousands of GPUs may work simultaneously on parts of the same problem, continuously exchanging data to keep calculations synchronized.
Research published this year continues to explore ways to move some AI communication operations directly into networking hardware, reducing unnecessary data transfers between GPUs and switches. One recent architecture proposed accelerating collective communications inside the network itself, illustrating how increasingly blurred the line between “computing” and “networking” has become.
The network is no longer just connecting computers.
It is becoming part of the computer.
Cooling and Power Join the Architecture
There is another consequence.
Packing dozens of high-end accelerators into tightly integrated racks produces enormous heat loads and electrical demand.
That is pushing AI infrastructure toward liquid cooling and forcing data-center operators to think about electrical systems at the rack and campus level.
Nvidia’s GB300 NVL72, for example, is built as a fully liquid-cooled rack-scale architecture.
Those requirements influence everything from building design to power distribution.
An AI data center therefore cannot simply install a new generation of processors without considering whether its networking, cooling and electrical infrastructure can support them.
The Data Center Becomes an AI Factory
Nvidia has increasingly adopted the term AI factory for these installations.
The phrase is marketing, but it describes a genuine architectural shift. Modern AI infrastructure takes electricity and data as inputs and produces training, inference and ultimately tokens as outputs. The performance of that factory depends not just on GPUs but on how efficiently every component works together.
