The pod

A row of liquid-cooled racks. Each draws roughly as much power as a hundred homes and turns nearly all of it into heat.

Racks in view: 8
Per rack: 72 GPUs · 36 CPUs
Rack power: ≈120–130 kW (representative)

Field of view

≈20 m

zoom ×1

Specifications

The B200 at a glance

Representative Blackwell B200 specifications
GPU dies per package	2 (unified as one GPU)
Transistors	208 billion
Process	TSMC 4NP
Die-to-die link	10 TB/s NV-HBI
Memory	Up to 192 GB HBM3e (B200), 288 GB (B300)
Memory bandwidth	≈8 TB/s
NVLink 5	1.8 TB/s per GPU
Power	≈1.0–1.4 kW depending on SKU and system
Precision	FP64 · TF32 · BF16 · FP8 · FP6 · FP4
Packaging	CoWoS-L

Vendor-published or widely reported figures, representative of the class. Some popular animations of “Blackwell” zoom into the GB202 die (≈92 billion transistors, ≈750 mm², TSMC 4N), which is the consumer GeForce RTX 50-series chip. The data-center B200 uses two larger dies totaling 208 billion transistors on TSMC 4NP.

Supply chain

How one GPU gets made

01Design · United StatesArchitecture, logic and physical design by the fabless chip designer.
02EUV lithography tools · NetherlandsA single supplier builds the EUV scanners every leading-edge fab depends on.
03Wafer fabrication · Taiwan (US fabs ramping)4NP wafers processed over months; hundreds of steps per wafer.
04HBM3e · South Korea, US-headquartered makersDRAM dies stacked 8–12 high with through-silicon vias.
05Advanced packaging · TaiwanCoWoS-L interposer joins two dies and eight HBM stacks: historically the tightest bottleneck.
06Substrates · Japan, TaiwanABF build-up film substrates carry thousands of connections from the package to the board.
07Boards and trays · Taiwan, Mexico, Southeast AsiaODMs build superchip boards and compute trays with cold plates.
08Rack integration · Taiwan, Mexico, United StatesFull racks are integrated, filled, tested and shipped as one system.

Representative geography. Supplier relationships and plant locations change frequently.