AI Infrastructure Database

The chips, data centers, power, software and models that AI runs on, and the companies behind each piece.

What's behind AI Chips, servers, memory, cooling and power, each as a record you can open. How it all goes together The build chain from silicon to a running model, in the order it is put together. How it's evolving Six shifts under way now, from air to liquid cooling and from copper to optics.

New here? This site is a database, not a news feed: every chip, company, data center and model gets its own record page, and records link to each other the way the real supply chain does. Start with the build sheet below, or just search.

Start here

How an AI data center actually gets built

Building around NVIDIA GPUs? Follow the real procurement chain: who fabricates the die, who supplies the memory, who makes the cables and racks, how it is cooled, and the software you install before the first token. Every step links to the companies that make it.

01DieTSMC

02MemorySK hynix · Micron

03PackagingCoWoS · TSMC

04Boards & racksFoxconn · Quanta

05CablingAmphenol · Credo

06OpticsCoherent · Fabrinet

07CoolingVertiv · CoolIT

08PowerEaton · Hitachi

09SoftwareLinux · CUDA · PyTorch

10Install your modelLlama · DeepSeek · GPT

Read the full build sheet → Inside the giants' stacks Who's next: the Watchlist

Key inflection points

The stack is mid-transition: six shifts to watch

Each hardware generation is rewriting a layer of the industry. These are the transitions happening right now — follow them here and you can see the next era arriving before it lands.

Air cooling→Liquid everything

An H100 rack drew about 40 kW and fans coped. GB200 racks draw ~120 kW and ship liquid-only — and ~600 kW racks are announced for 2027. Every new hall is being replumbed around cold plates.

Direct liquid cooling Why AI got wet Vertiv

The server→The rack as one computer

The unit of design jumped from an 8-GPU box to a 72-GPU NVLink rack that software sees as a single accelerator. The open UALink camp is racing to answer before the pattern locks in.

GB200 NVL72 NVLink UALink

The training era→The inference era

Reasoning models moved the compute to answer time. Serving now outweighs training across the industry and is pulling chips designed for nothing else.

Test-time compute Inference Rubin CPX

One giant die→Chiplets & packaging

Dies hit the lithography limit, so the frontier moved into the package — multi-die GPUs, 3D stacking, and CoWoS capacity acting as the world's GPU ration.

CoWoS Chiplets NVIDIA B200

Grid power→Owned power

Campuses stopped waiting in interconnection queues: shut reactors restarted under 20-year deals, turbines behind the meter, fuel cells at the fence line.

The nuclear revival Three Mile Island restart Gas turbines

Copper + HBM3E→Optics + HBM4

The 2026 component turn: HBM4 doubles the memory interface while co-packaged optics start replacing pluggable transceivers — the parts lists behind Rubin, MI400 and the next TPUs.

HBM4 Co-packaged optics The Watchlist

What's coming next: the Watchlist → How we got here: the Timeline The numbers: Trends