EIGHT SYSTEMS AND THEIR BUILDERS
The AI server rack supply chain, part by part
Each part of the rack has a different job and a different set of public companies connected to it. Start with the hardware, then inspect the evidence. Rack builders are listed for their own AI systems, not as suppliers of the parts inside.
GPUs, HBM & packaging
The accelerator package is where the rack computes: GPU or custom-accelerator dies sit beside stacks of high-bandwidth memory on a silicon interposer, mounted on a large organic substrate. Each Blackwell GPU packs 208 billion transistors on two reticle-limited dies joined at 10 TB/s, and a GB300 NVL72 rack carries 20 TB of HBM3E across its 72 GPUs (NVIDIA).
21 companiesCPUs, memory & storage
Each compute tray pairs its GPUs with host CPUs that run the operating system and stage data, CPU memory, E1.S flash drives, a management controller and a high-layer-count, low-loss motherboard. A GB200 NVL72 rack carries 36 Grace CPUs with 2,592 Arm cores and up to 17 TB of LPDDR5X (NVIDIA).
26 companiesSwitches & network chips
An AI rack runs two networks. The scale-up fabric joins all 72 GPUs through NVLink Switch trays at 130 TB/s, and the scale-out network gives each GPU an 800 Gb/s SuperNIC link to top-of-rack and leaf switches that tie racks into a cluster (NVIDIA GB300 NVL72). DPUs, PCIe retimers and PCIe switches handle storage, security and the lanes inside each tray.
15 companiesCables, optics & connectors
Links move data between GPUs, trays and racks. Inside an NVL72-class rack the GPUs reach the NVLink switch trays over a copper cable cartridge, with no optics, while scale-out traffic leaves the rack through pluggable optical transceivers and fiber at 800 Gb/s per GPU in GB300 NVL72 (NVIDIA). Fujikura’s AFL says a leading-edge NVL72 rack can be fed with up to 1,152 fibers.
31 companiesPower shelves & delivery
Power shelves turn facility AC into about 50 V DC, a copper busbar carries it to every tray, and voltage regulators step it down to under 1 V at the chips, while capacitors and batteries ride through outages and load swings. In GB200/GB300-class racks a 33 kW shelf holds six hot-swap 5.5 kW supplies (Vertiv PowerDirect 3000). NVIDIA says a 1 MW rack at 54 V would need up to 200 kg of copper busbar, one reason it plans 800 V DC distribution from 2027.
31 companiesLiquid cooling
Direct-to-chip liquid cooling pumps coolant through copper cold plates on the GPUs and CPUs, then through rear manifolds and dripless quick disconnects to a coolant distribution unit that passes the heat to facility water. NVIDIA’s GB200 NVL72 design calls for 120 kW of cooling per rack and uses a blind-mate manifold with floating tray connections so the quick disconnects align as trays dock (NVIDIA, Oct 2024).
26 companiesFans & airflow
Fans still move air through power shelves, switch trays and any parts not on cold plates, and a rear-door heat exchanger can catch the warm exhaust in a water coil; temperature sensors and fan controllers set the airflow. In the Schneider Electric–AMD Helios reference design, liquid removes 84% of the rack’s heat and air handles the remaining 16% (Motivair, July 2026).
18 companiesRacks, chassis & rails
The rack is the steel structure that holds and services everything else: a 48U-class OCP/MGX frame, a sheet-metal chassis for every tray, telescopic rails so trays slide out, and casters and feet so the integrated rack can roll in and stand level. Supermicro’s GB300 NVL72 enclosure measures 600 × 1,068 × 2,236 mm (W × D × H), and Rittal rates its MGX rack for dynamic loads up to 1,400 kg.
12 companiesRack builders & platforms
The companies that design, manufacture, integrate or sell complete AI servers and racks: branded server makers (OEMs), original design manufacturers (ODMs) and contract manufacturers (EMS) building on platforms such as NVIDIA’s GB300 NVL72. This group is not a physical part; a builder is listed for its own servers and racks, and under a part only when separately documented as making it. Hon Hai expects its AI-rack shipments to more than double in 2026 (Aug 12, 2026).
20 companiesCompanies can participate in more than one sector. Sector counts overlap; the AI-rack universe contains 124 distinct listings.