The AI factory looks different from a traditional data center — denser, hotter, and cabled for east-west traffic instead of north-south

1. From Training to Inference: AI's New Center of Gravity

The first chapter of AI infrastructure was a compute-capacity race: build the biggest clusters, train the largest models. That chapter, defined by the xPU, is giving way to a second one defined by inference — the continuous, real-time processing that actually puts AI to work.

Two forces are driving the shift. First, reasoning models that "think" before answering consume dramatically more inference compute than a single forward pass. Second, agentic AI — autonomous systems that plan and act — and physical AI (robots, software-defined vehicles) turn inference from a batch job into a constant stream. Industry estimates suggest humanoid robots alone could approach a billion units by mid-century.

The consequence for infrastructure is simple and profound: the workload that dominates the data center is no longer the one it was originally built to run.

2. The AI Factory: Data Centers That Manufacture Intelligence

NVIDIA's framing captures the shift in one phrase: the data center is becoming an AI factory. A traditional data center stores and processes data; an AI factory manufactures intelligence, and its unit of output is the token — a real-time prediction or decision that drives an application.

Dimension	Traditional data center	AI factory
Core function	Store and process data	Manufacture intelligence
Primary output	Processed data	Tokens (real-time predictions)
Workload	Diverse, general-purpose	AI lifecycle: ingest, train, fine-tune, infer

This is not marketing language — it is an engineering reality. AI factories are being treated as national infrastructure on par with telecom and energy, and the economics are unforgiving: at a gigawatt scale, a single day of downtime can cost over $100 million.

3. Three Scaling Laws That Break Traditional Data Centers

What makes the AI factory different is that three scaling laws are now pulling compute demand upward at once:

Scaling law	What it means	Compute impact
Pretraining	Bigger models on bigger datasets	~50 million x growth over 5 years
Post-training	Fine-tuning for specific tasks	~30 x pretraining compute
Test-time reasoning	Models "think" through many responses	Up to 100 x traditional inference

Traditional data centers were not built for this. The AI factory is — and the redesign shows up first in the two things every facility fights hardest: power and cooling.

4. The Power Wall: 1 MW Racks & 800V DC

The most concrete number in the whole conversation: next-generation AI racks can draw more than one megawatt each — enough for over a thousand homes. A few rows of those and a facility's original power design is obsolete.

The industry's answer is a power-architecture overhaul:

• Higher-voltage distribution (800V DC) — raising voltage lowers current, cuts conduction losses, and restores density as rack power climbs.
• Vertical power delivery — modular, high-density power stages placed close to the processor to shorten the current path.
• Gallium-nitride (GaN) switching — efficient high-voltage conversion for the new architecture.

As one analysis put it: if the xPU defined the first chapter of AI infrastructure, power architecture is shaping the next. For the power-side pressures on AI data centers, see our AI data center power guide.

5. Liquid Cooling: Thermal Goes Mainstream

Pack a megawatt into a rack and air simply cannot carry the heat away. Direct-to-chip liquid cooling has moved from a niche option to a first-class design requirement, because thermal management now limits how much compute you can fit as much as electrical performance does.

This matters for more than the chillers: liquid cooling changes the physical layout of the data hall, the routing of every cable, and the maintenance procedures. A cabling plant designed for air-cooled racks does not simply survive the move to liquid — it has to be planned around it from day one. See our in-rack cooling and cable architecture guide.

Liquid cooling and high-density fiber are no longer separate concerns — they share the same rack and must be planned together

6. The Network: Back-End Fabric, 800G/1.6T & Million-GPU Scale

AI training and inference both depend on a dedicated back-end fabric — the east-west, GPU-to-GPU network that is separate from the front-end network serving users. That fabric is what lets tens of thousands of accelerators act as one machine.

• NVLink binds GPUs tightly within and across nodes; InfiniBand or Spectrum-X Ethernet handles the scale-out network.
• Interconnects are moving to 800G and 1.6T, with the cabling plant shifting toward high-density MPO single-mode fiber.
• The target is million-GPU systems — a cabling and optics challenge on a scale the industry has never faced.

For the 800G/1.6T cabling roadmap, see our 800G/1.6T cabling trends guide.

7. What It Means for Cabling: A Checklist

Every layer of this shift lands on the physical plant. The cabling implications are concrete:

For the structured-cabling re-think this demands, see our structured cabling for AI data centers.

Key Questions (FAQ)