130 parts · 1,202 companies
Prices as of2026-10-25
Search

Google TPU7x (Ironwood): every supplier in the pod

A pod, not a rack: Google's own numbers are 9,216 TPU7x chips per superpod, 42.5 exaFLOPS FP8, 192 GB of HBM3E and 7.37 TB/s of memory bandwidth per chip. Chips wire into 64-chip "cubes" over a direct 3D-torus interconnect (no separate switch chip inside a cube), and cubes wire into pods and superpods over a reconfigurable optical circuit switch fabric Google calls "Apollo" — architecturally the most different of the three racks on this site from NVL72's NVSwitch trays and Helios's UALink-over-Ethernet switches.

Unmarked companies come from Google's own materials (cloud.google.com, blog.google, docs.cloud.google.com); reported companies come from named trade press or industry-analyst reporting whose article body was read in full, not just the headline. This is a map of who plays where, not a confirmed bill of materials — Google is far more closed about its hardware supply chain than AMD or Nvidia, so most of what's below comes from industry-analyst reporting rather than Google's own partner disclosures.

  1. Pod architecture and reference design (custom ASIC, not a merchant GPU rack)
    Rack-scale systems 4.01
    Alphabet (Google)
  2. Board assembly and rack integration
    Rack-scale systems 4.01 · EMS, system integration & rack-level test 4.02
    Celestica* (primary board assembly and rack integration partner)
  3. Compute host (4x TPU7x chips per host; 64 chips per rack form a "cube")
    GPU baseboards & compute trays 4.03
    Alphabet (Google)
  4. TPU7x (Ironwood) ASIC die (co-designed with Broadcom, fabricated on TSMC N3P)
    Hyperscaler custom ASICs 1.02 · Leading-edge foundry 1.09
    Alphabet (Google), Broadcom* (co-design, SerDes and ASIC implementation), MediaTek* (I/O module design and peripheral components), TSMC* · host CPU vendor (x86 host confirmed to exist per Google's own docs — 224 vCPUs / host — but no vendor named in any source read)
  5. HBM3E (192 GB per chip, 8 stacks, 7.37 TB/s bandwidth)
    HBM 2.01
    Samsung Electronics* (primary supplier, >60% of shipments in 2025), SK Hynix* (dominated Broadcom-Google HBM orders in H1 2025; secondary since)
  6. Optical circuit switch fabric ("Apollo" — MEMS mirrors, connects cubes into pods and superpods)
    Scale-up fabric 3.01
    Lumentum* (sole-source per the source below; no qualified alternative in production), Celestica* (also manufactures the OCS units themselves)

Liquid cooling is confirmed as the rack's cooling method by Google's own materials, but no specific component supplier for it could be confirmed to reporting-grade — one source explicitly labeled its own cooling-vendor guesses "LIKELY", not confirmed, so none are listed here. Facility-side power and cooling outside the rack sit in Layers 6–7 same as for the other teardowns — see Layer 6 and Layer 7. Compare with the GB300 NVL72 and AMD Helios teardowns.