How an AI Data Center Is Built

‹ Back

Read this page top to bottom and you have followed a real procurement chain: silicon → boards → racks → cables → cooling → power → software → model. The same trace works for any accelerator — swap the first stage and the rest of the chain barely changes.

1 The GPU itself is a supply chain

"An NVIDIA GPU" is really a package of parts from five countries. NVIDIA designs it and writes the software; nearly everything physical is made by others.

Logic dies (the compute) TSMC 4NP process fabricated in Taiwan on machines only ASML can build — ASML EUV Is TSMC the only choice? At the leading edge, effectively yes — every flagship AI chip is made there today. The two chasing alternatives: Samsung Foundry · Intel Foundry

Memory (HBM stacks) SK hynix Micron Samsung eight stacks per B200; sold out about a year ahead — HBM3E

Packaging (die + memory joined) TSMC CoWoS Amkor ASE the step that rationed world GPU supply in 2023–24

Board assembly Foxconn Quanta the SXM modules and superchip boards are built by Taiwanese manufacturers, not NVIDIA

Power delivery on the board Monolithic Power Infineon Delta stepping rack voltage down to ~1 V at over a thousand amps per die

Network interfaces NVIDIA BlueField / ConnectX one 400–800G NIC per GPU is standard equipment

Full spec sheets: B200 · GB200 · or trace any chip's own chain from its record page.

2 Servers and racks

GPUs ship inside engineered systems. At Blackwell scale the rack itself is the product: 18 compute trays, 9 switch trays, one copper spine.

The rack-scale system GB200 NVL72 HGX baseboards ~120 kW per rack — a hundred homes' draw in a wardrobe footprint

Who builds them Foxconn Quanta Wiwynn ODMs for the hyperscalers…

…and who brands them Supermicro Dell HPE Lenovo the enterprise channel, with service contracts attached

Connectors & retimers Amphenol TE Connectivity Astera Labs thousands of high-speed contacts per rack; retimers keep PCIe signals alive across the chassis

3 The cabling

The question everyone forgets to ask. Three kinds of wire, three supplier ecosystems — and at cluster scale the cabling bill runs to hundreds of millions.

Inside the rack: copper Amphenol TE Connectivity Credo DAC and active-electrical cables; the NVL72's NVLink spine alone is ~5,000 copper cables — over two miles of them per rack

Between racks: light 800G optics Coherent Lumentum Fabrinet pluggable transceivers by the hundred thousand per cluster; lasers from the photonics makers, assembly largely in Thailand

The glass itself Corning campus fiber counts now run to millions of kilometers

The switches it all plugs into NVIDIA InfiniBand Spectrum-X Arista Broadcom silicon the InfiniBand-vs-Ethernet choice is the biggest architecture decision after the GPU itself

Learn: how 100,000 GPUs become one computer →

4 Cooling — and why it decides the map

At 120 kW per rack, air is over. Liquid touches every chip, and how a campus rejects heat shapes where it can be built at all.

On the chip: cold plates Direct-to-chip CoolIT Boyd copper microchannel plates bolted to every GPU, CPU and switch ASIC

In the row: CDUs Coolant distribution units Vertiv Schneider (Motivair) Delta 2 MW-class pump-and-exchanger skids — a product category that barely existed in 2022

Rejecting the heat Rear-door HX Immersion Munters dry coolers, towers or chillers outside; closed loops fill once and barely sip water

The siting signal: cooling is why data centers historically followed rivers and lakes (evaporative designs), and why the new closed-loop generation can instead chase cheap power in dry country. Scouts read campuses off maps by exactly these tells — water rights, high-voltage corridors, fiber routes, cool climates. Learn: where data centers go, and why →

5 The power room

Between the grid and the racks stands a building full of equipment with multi-year lead times — usually the true critical path.

Transformers & grid tie Hitachi Energy Siemens Energy Interconnection 3–5 year lead times made these the scarcest boxes in the industry

Switchgear, UPS, busway Eaton Schneider ABB Vertiv the electrical spine; next-gen racks push toward 800-volt DC distribution

Backup & bridge power Cummins Caterpillar Gas turbines Bloom fuel cells gensets for outages; whole turbine plants where the grid queue is too slow

The megawatts themselves Power projects Nuclear restarts Solar + storage the deals feeding the campuses — mapped in the Power section

Run the arithmetic yourself: the Cluster Power Calculator →

6 The software you install

Racks arrive as metal. This is the stack, in the order it goes on — and yes, it starts with Linux.

Operating system Linux virtually every AI server on Earth runs it — Ubuntu, RHEL, or a hyperscaler's in-house build

Drivers & compute platform CUDA ROCm (AMD) XLA (TPU) the layer that makes the silicon programmable — and the deepest moat in the industry

Cluster plumbing NCCL Kubernetes Slurm Ray collectives to synchronize GPUs; schedulers to share the machine

Training frameworks PyTorch JAX Megatron DeepSpeed where the model code actually lives

Then: install your model Train one or pull open weights vLLM TensorRT-LLM serving engines turn the weights into tokens per second — the thing the whole building was for

Learn: how a frontier model gets trained → · what happens when you hit Enter →

Inside the giants: same problem, six answers

Everyone above sells to everyone below — but each big builder integrates the stack differently. These are the real configurations, assembled from the same records the rest of this database carries.

The vertical integrator

Google

The only player integrated from self-designed silicon to frontier model — no NVIDIA required anywhere in Gemini's path.

Compute TPU v7 Ironwood Trillium Its own TPUs — Ironwood for the inference era, Trillium in volume — plus NVIDIA GPUs rented out to Cloud customers.

Fabric TPU pods Optics 3D-torus inter-chip links reconfigured by in-house optical circuit switches; the Jupiter network carries the data center.

Racks & systems TPU pods Self-designed: the 9,216-chip pod, not the server, is the unit of hardware.

Cooling Direct-to-chip Direct liquid cooling on TPU pods since 2018 — the earliest big liquid fleet.

Power Kairos fleet Nuclear A giant renewables book, the first corporate SMR fleet order, and restart interest.

Software JAX XLA Linux + Borg orchestration underneath; JAX and XLA compile straight to TPU.

Model & site Gemini 3 Council Bluffs Gemini 3 trained and served on TPUs end to end.

The alliance at NVIDIA scale

Microsoft + OpenAI

The biggest NVIDIA buyer builds 'AI factories' for its partner's models — with its own silicon waiting in the wings.

Compute GB200 H100 Maia 100 Frontier NVIDIA fleets — GB200 today atop huge H100 estates — with in-house Maia 100 in supporting roles.

Fabric InfiniBand The world's largest InfiniBand deployments, flat networks spanning whole sites.

Racks & systems GB200 NVL72 Foxconn Quanta ODM-built NVL72 racks and HGX servers from the Taiwanese manufacturers.

Cooling Fairwater Direct-to-chip Closed-loop direct liquid at Fairwater — fill once, near-zero water after.

Power Crane restart Constellation The Three Mile Island restart PPA (835 MW, 20 years) plus a 10.5 GW renewables framework.

Software DeepSpeed ONNX Runtime Linux + Azure's stack; DeepSpeed for training scale; ONNX for deployment.

Model & site GPT-5 Fairwater Stargate GPT-5 trains and serves on this estate; Stargate adds OpenAI's own campuses.

The open-everything builder

xAI

The existence proof that a frontier cluster can be improvised in months — factory shell, gas turbines, batteries and all.

Compute H100 H200 ~100,000 H100s in 122 days, doubled since; GB200-class capacity at Colossus 2.

Fabric Spectrum-X The flagship Spectrum-X Ethernet deployment — NVIDIA networking without InfiniBand.

Racks & systems Supermicro Dell Supermicro and Dell liquid-cooled systems, delivered at sprint pace.

Cooling Direct-to-chip Direct liquid throughout — the density made air a non-starter on day one.

Power Gas turbines Colossus On-site gas turbines bridging the grid queue, Tesla Megapacks smoothing training spikes.

Software CUDA A lean in-house training stack on Linux and CUDA — small team, big machine.

Model & site Grok 4 Colossus Grok 3 and 4 trained at Colossus; Grok ships inside X.

The custom-silicon bet

AWS + Anthropic

The clearest test of whether a hyperscaler's own chips can carry a frontier lab — half a million Trainiums and counting.

Compute Trainium2 UltraServer Trainium2 UltraServers at Project Rainier scale, beside NVIDIA fleets for customers.

Fabric UltraServer NeuronLink inside the UltraServer; AWS's own EFA networking and Nitro offload beyond it.

Racks & systems Annapurna/AWS Annapurna-designed, self-integrated — the cloud primitive as the building block.

Cooling Direct-to-chip Liquid-and-air mix tuned per hall across the Rainier sites.

Power Susquehanna X-energy Nuclear at Susquehanna, X-energy SMR backing, and the largest corporate renewables book.

Software PyTorch vLLM The Neuron SDK is the whole ballgame — porting the open stack to its own silicon.

Model & site Claude 4.5 Project Rainier Claude trains at New Carlisle — and, uniquely, also runs on TPUs and GPUs elsewhere.

The embodied-AI factory

Tesla

Trains on video from millions of cars, ships the result back to them — with its own batteries steadying its own cluster.

Compute H100 Cortex The ~50,000-GPU Cortex cluster at Giga Texas; Dojo wound down for AI5/AI6 chips.

Fabric 800G optics Ethernet-fabric clusters sized for video-scale data ingest.

Racks & systems Supermicro Supermicro and Dell systems, plus purpose-built data-hall extensions of the factory.

Cooling Direct-to-chip A dedicated liquid-cooling plant built alongside the cluster.

Power Gas turbines Grid supply buffered by its own Megapacks — the manufacturer as its own customer.

Software PyTorch Vision-only training pipeline on PyTorch; models compile down to the car's FSD computer.

Model & site Tesla FSD Samsung FSD v13+ ships over the air; AI5/AI6 silicon comes from Samsung and TSMC.

And who's coming next

The full Watchlist →

Challenger silicon, the 2026–27 platform generation, photonic interconnects, the nuclear wave and the architecture bets beyond the transformer — tracked with status labels on the Watchlist.

Etched Sohu Lightmatter UALink HBM4 Rubin MI400 SMRs Watchlist →