What is an AI data center?
A data center houses computers alongside networking, power and cooling. An AI-focused facility devotes substantial capacity to accelerated workloads. Its GPUs still depend on CPUs, storage and conventional infrastructure.
AI Data Center INTERACTIVE 3D
Explore the GPUs, networks, electricity, and cooling that power AI.
FROM PROMPT TO RESPONSE
Follow my prompt
Send “Why does water evaporate?” and follow its journey from your device to a server and back.
Simplified inference example. One representative serving system; allocation and networking vary.One request. Many connected systems.
How is it different from a general-purpose data center?
AI clusters often concentrate more accelerators and high-speed networking in each rack. That can demand denser power and cooling. The boundary is not absolute: general-purpose data centers can also host AI workloads.
Why does AI use GPUs?
Many AI calculations apply similar arithmetic to large arrays of numbers. GPUs execute many operations at once, while nearby high-bandwidth memory feeds them data. CPUs remain important for orchestration and tasks that do not benefit from this parallelism.
How do GPUs communicate?
Within a server, dedicated GPU interconnects or PCIe move data between devices. Between servers, network adapters and switches form a cluster fabric, commonly using Ethernet or InfiniBand. Communication patterns depend on the model and workload.
What happens when someone sends an AI prompt?
Serving software receives the input, prepares tokens and schedules computation. GPUs use trained model weights to calculate output. Many language models generate tokens through repeated steps. Responses can stream back before the entire answer is complete.
Why do AI servers produce so much heat?
Processors and memory consume electrical energy while switching and moving data. Almost all of that energy becomes heat inside the facility. Cooling must carry it away to keep equipment within its operating range.
How does liquid cooling work?
In this example, sealed cold plates absorb GPU heat. A CDU pumps server coolant through those plates and transfers heat through a heat exchanger into a separate facility loop. Outdoor equipment rejects the heat. Fans and contained aisles handle the remaining air-cooled equipment. Designs differ between facilities.
What happens if utility power fails?
UPS batteries provide a bridge while backup generation starts and transfer equipment accepts the new supply. Cooling and network systems also need continuity planning. Runtime, redundancy and transition times depend on the installation; no timing guarantee is implied here.
How does AI training differ from inference?
Training updates model parameters using many examples and often extensive communication between accelerators. Inference uses those learned parameters to produce results for new inputs. A normal prompt does not retrain the underlying model.
About this model
A representative AI facility, not a named company’s site. Rack count, equipment size and pipe routes are simplified for readability. The separate hardware scene enlarges a generic four-GPU server; it is not a manufacturing specification. Server coolant and facility water exchange heat in a CDU without mixing. Air cooling handles other components. Redundant paths, security systems and detailed plumbing are omitted. Workload indicators are qualitative and do not estimate energy per prompt.
Controls: drag to orbit, pinch or scroll to zoom, right-drag or use the hand tool to pan. Reset returns to the facility cutaway and exits the guide. Use component buttons for keyboard access to every part.
All components and their roles
Overhead power distribution
These electrical feeds carry power from the UPS distribution system to each rack. The yellow lines represent electrical cables or busways, not coolant pipes.
Foundation and server hall floor
The building foundation and floor support the racks and heavy service equipment. This model simplifies the structural design.
Floor tiles and aisle markings
Floor joints and marked routes organize access around equipment. These are access markings, not cables or coolant pipes.
Facility walls
The enclosure separates the equipment hall from the outside environment. Selected walls are removed in this cutaway.
Facility roof
The roof protects the server hall. It can be removed to inspect the infrastructure inside.
Building frame and overhead lighting
Structural columns support the enclosure, and steady overhead lights illuminate service aisles. These are separate from blinking equipment status indicators.
Overhead network cable trays
Trays support organized cable runs above the racks. Drops connect rack equipment to the cluster network.
Server rack
A cabinet organizes servers and their connections. Rails hold equipment, while cables and coolant manifolds run alongside it.
GPU server
A complete computer pairs GPUs with CPUs, memory, storage, power supplies and cooling. It receives work from serving software.
GPU module
A GPU runs many mathematical operations in parallel. AI models repeatedly multiply and combine arrays of numbers.
CPU
The CPU runs the operating system and coordinates input, networking and work sent to GPUs.
System memory
System RAM holds programs and working data for the CPU. It is distinct from memory beside the GPU.
GPU memory
High-bandwidth memory holds model weights and intermediate results close to the GPU. A model can be split across devices when needed.
Storage
Solid-state drives keep software and model files when power is off. Data is loaded into working memory before use.
Rack switch
A top-of-rack switch connects servers to the wider cluster network. It forwards packets toward their destination.
Cluster fabric switch
The cluster fabric connects racks. Fast links help cooperating compute systems exchange data without spending all their time waiting.
High-speed links
Short board-level links connect GPUs inside this server; network adapters connect servers through the cluster fabric. These are different interconnect layers.
External network connection
Requests enter through the facility network edge. Serving software routes accepted requests to available compute; this cable does not make scheduling decisions.
Utility connection
The site receives electrical energy from the grid. Protection and conversion equipment prepare it for use inside.
Transformer
A transformer changes the supply voltage. This model omits exact voltage ratings and redundant supply paths.
Switchgear
Switchgear protects circuits and controls which supply feeds downstream equipment. Fault isolation is as important as switching sources.
UPS and batteries
An uninterruptible power supply conditions power and uses stored battery energy to bridge a loss of supply. Battery duration is finite.
Backup generator
A generator can supply power during a longer outage after it starts and its output is accepted. Transfer behavior varies by design; batteries bridge the gap.
Rack power distribution
Rack distribution units deliver protected electrical power to server power supplies. Electricity follows cables, not coolant pipes.
Server fans
Fans move air past memory, power electronics and other air-cooled parts. Liquid cooling does not remove every need for airflow.
Contained air aisles
Containment separates warm exhaust from cool intake air. Air handlers return cooled air to the hall; GPU heat also leaves through a liquid loop.
GPU cold plates
Metal plates transfer heat from GPU packages into circulating coolant. The liquid stays inside sealed passages.
Coolant distribution unit
The CDU circulates the server coolant and exchanges its heat with a separate facility loop. The two fluids do not mix in the heat exchanger.
Facility cooling loop
A separate water loop carries heat from CDUs and air handling to the outdoor cooling plant. Pipes shown are schematic supply and return routes.
Outdoor heat rejection
Outdoor equipment releases collected heat to ambient air. This example uses a dry cooler with supplemental chilling when conditions require it.