Supplier Rings

Who makes each part of nine AI platforms · Where it is made

An AI rack is assembled from parts made by dozens of companies in a handful of countries. Across nine platforms and 300 supplier rows, almost every chip, memory stack, substrate and power shelf is made in Asia; US production so far is wafers in Arizona and final assembly in Texas and California.

Matrix of nine AI platforms against 18 components, each cell shaded by US production status of its best-known supplier: most cells are made abroad or not disclosed, with US production concentrated in rack building and Cerebras systems US PRODUCTION BY PLATFORM AND COMPONENT Accelerator CPU HBM Packaging PCB, substrate Tray ODM Rack build Scale-up switch Scale-up cable NIC Optics Power shelf Busbar Battery backup CDU Cold plates Quick disconnects Enclosure GB200 NVL72 GB300 NVL72 Vera Rubin NVL72 TPU v7 Ironwood Trainium3 Maia 200 AMD Helios CloudMatrix 384 Cerebras CS-3 Made in US Some US volume US planned Made abroad Not disclosed
Platforms
9
Supplier rows
300
Suppliers
79
Made in US now
8 rows
Hub
Supply Chain

What it is

Who Makes an AI Rack

A supplier ring lists every company that supplies a component of one AI platform, from the accelerator chip to the quick disconnects on the cooling loop, with where each part is made and whether any of it is made in the US.

  • Scope: nine rack-scale and pod-scale platforms from Nvidia, Google, AWS, Microsoft, AMD, Huawei and Cerebras, each broken into 18 components.
  • US lens: each row is marked made in the US, some US volume, US production planned with a date, made abroad, or not disclosed.
  • Evidence: each row carries its source and whether it rests on a company statement, a teardown or an analyst report.
  • Not a directory: suppliers appear only where they are tied to a named platform.

Platforms

Nine Platforms, 18 Components Each

AI platforms and US production status by component
PlatformAcceleratorAcceleratorsRowsComponents with US productionUS plannedNot disclosed
Nvidia GB200 NVL72NvidiaB2007240103
Nvidia GB300 NVL72NvidiaB300 (Blackwell Ultra)7252302
Nvidia Vera Rubin NVL72NvidiaRubin7252105
Google TPU v7 IronwoodGoogle with BroadcomTPU v7 Ironwood9216320110
AWS Trainium3 UltraServerAmazon (Annapurna Labs)Trainium314436108
Microsoft Maia 200MicrosoftMaia 200Not disclosed190013
AMD HeliosAMDInstinct MI450 / MI455X72300012
Huawei CloudMatrix 384HuaweiAscend 910C38421000
Cerebras CS-3CerebrasWSE-3 wafer-scale engine1182015

Findings

Assembled in America, Made in Asia

  • Rows: 300 supplier rows; two made in the US, six with some US volume, two planned, 161 made abroad and 129 not disclosed.
  • Headquarters: 130 of the named rows are Taiwanese companies, 59 American, 20 Chinese, 14 Korean and five Japanese.
  • US production now: TSMC Arizona GB300 wafers, Wistron's Fort Worth GB300 compute trays, Supermicro rack builds in San Jose and Flex's Cerebras lines in Milpitas. Coherent makes lasers in Sherman, Texas, but where its transceivers are assembled is not disclosed.
  • Still offshore: advanced packaging, HBM, PCBs and substrates, power shelves and cold plates; Amkor Arizona packaging starts in 2028 and SK hynix Indiana HBM in late 2029.
  • China: Huawei's CloudMatrix 384 runs on TSMC dies and Samsung HBM from pre-ban stockpiles, with SMIC and CXMT as domestic replacements; Chinese firms also supply optics and PCBs to US platforms.

Record

Dataset Record

Dataset
Supplier rings for nine AI platforms: 300 rows, 79 suppliers
Sources
Company filings and press releases, SemiAnalysis, TrendForce, DigiTimes, KGI and Morgan Stanley via Taiwan press, TechInsights; source link on each row
Last verified
October 2026