What Broadcom makes for data centers
public · HQ Palo Alto, United States · broadcom.com
Broadcom co-designs custom AI accelerator ASICs with hyperscalers, including Google's TPU and Meta's MTIA, and makes scale-out switch silicon (Tomahawk, Jericho) and scale-up Ethernet switch chips (Tomahawk Ultra) used in GPU cluster networking. It also supplies NICs, PCIe/CXL switches, optical transceivers and VCSEL lasers, optical DSPs, and co-packaged optics for data center networking.
Units and subsidiaries: OpenAI
Parts, by layer
Layer 1 — Compute silicon & packaging
| Part | Role | Confidence | Competitors | Note |
|---|---|---|---|---|
| Hyperscaler custom ASICs | design partner | inferred | Alphabet (Google), Microsoft, Alchip, Amazon (AWS), Marvell, MediaTek | Google TPU; Meta MTIA — partner on Google TPU; partner on Meta MTIA |
| ASIC design services, IP & EDA | service | inferred | Alchip, Arm, Cadence, Credo, Faraday, Global Unichip | |
| NICs / SuperNICs | manufacturer | inferred | AMD, Intel, Marvell, Nvidia | |
| Hyperscaler custom ASICs (via OpenAI) | manufacturer | inferred | Alphabet (Google), Microsoft, Alchip, Amazon (AWS), Marvell, MediaTek | custom accelerator |
Layer 2 — Memory & storage
| Part | Role | Confidence | Competitors | Note |
|---|---|---|---|---|
| PCIe/CXL switches & retimers | manufacturer | inferred | Astera Labs, Marvell, Microchip, Xconn | PLX |
Layer 3 — Networking & interconnect
| Part | Role | Confidence | Competitors | Note |
|---|---|---|---|---|
| Scale-up fabric | manufacturer | inferred | Alphabet (Google), AMD, Astera Labs, Nvidia, UALink consortium | scale-up Ethernet (Tomahawk Ultra) |
| Scale-out switch ASICs | manufacturer | disclosed | Cisco, Marvell, Nvidia, Huawei | Tomahawk/Jericho |
| Optical transceivers | manufacturer | inferred | Accelink, Applied Optoelectronics, Cisco, Coherent, Eoptolink, HG Genuine | |
| Lasers | manufacturer | inferred | Coherent, Lumentum, Mitsubishi Electric, POET, Sivers, Sumitomo Electric | VCSELs |
| Optical DSPs, drivers, TIAs | manufacturer | inferred | Credo, Marvell, MaxLinear | |
| Silicon photonics & co-packaged optics | manufacturer | inferred | TSMC, GlobalFoundries, Intel, Marvell, Nvidia, Tower Semiconductor | CPO switches |
| Timing & telemetry silicon | manufacturer | inferred | Marvell, Microchip, Renesas, SiTime | |
| Retimers inside AECs | manufacturer | inferred | Astera Labs, Marvell, Spectra7 |
Layer 9 — Building shell, site & construction
Where Broadcom sits in the stack
Broadcom's placements in this catalog cover 13 parts, from layer 1 (compute silicon & packaging) to layer 9 (building shell, site & construction). None of them currently carries a sourced supply constraint tag. Across these parts it overlaps with other companies most often on the lines below.
What these parts are
- Hyperscaler custom ASICs — AI accelerator chips (often called XPUs) designed in-house by a hyperscaler for its own fleet, typically co-developed with an outside ASIC design partner and fabbed by a merchant foundry.
- ASIC design services, IP & EDA — Outside firms and licensors — design-services houses, IP blocks, and EDA software — that hyperscalers and accelerator startups use to turn a chip specification into a manufacturable design.
- NICs / SuperNICs — Network interface cards, including Nvidia's higher-bandwidth "SuperNIC" variants, that connect a server's CPU and GPUs to the cluster network.
- PCIe/CXL switches & retimers — Fabric silicon — switches that fan a single PCIe or CXL root out to many downstream devices, and retimers that clean up the signal over longer runs — connecting accelerators, SSDs, and pooled memory back to the host.
- Scale-up fabric — The fabric linking GPUs to each other inside a single rack or pod, separate from the switching fabric that connects racks to each other.
- and 8 more in the table above.
Closest peers in this catalog
Seen in real machines
Broadcom is named in 4 of the rack teardowns and supercomputer breakdowns on this site. Each line below is a piece of that machine.
- Nvidia GB300 NVL72 · rack teardown
- 800G / 1.6T OSFP transceivers (scale-out)
- AMD Helios (MI455X) · rack teardown
- Scale-up switch tray (2x Broadcom Tomahawk 6 ASICs per tray, UALink-over-Ethernet, 21.6 TB/s bidirectional per switch)*
- Google TPU7x (Ironwood) · rack teardown
- TPU7x (Ironwood) ASIC die (co-designed with Broadcom, fabricated on TSMC N3P)* (co-design, SerDes and ASIC implementation)
- Fugaku · supercomputer
- On-die SerDes, HBM I/O and SRAM IP in the A64FX* (SerDes, HBM I/O and SRAM blocks per a single independent report)
Disclosed customers
Latest feed
- Commentary: Samsung's US 2nm ramp faces no room for error — DigiTimes, 2026-10-20Samsung's ability to successfully ramp production of its 2nm process at its US fab is crucial for its foundry recovery and servicing major orders from clients like Tesla and Broadcom.
- Co-Packaged Optics Moves from Promise to Production Planning — EE Times, 2026-10-14Major industry players including Meta and Broadcom discussed the roadmap for co-packaged optics (CPO) production and the standardization of the OCI MSA to meet anticipated demand in 2028.
- AI chip competition shifts to data-center efficiency as memory, interconnect limits bite — DigiTimes, 2026-10-14The competition in AI chip development is pivoting toward data-center efficiency to overcome scaling limitations in high-bandwidth memory and interconnect speeds.
- Broadcom executive says chipmakers must tackle packaging, power limits together — DigiTimes, 2026-10-14A Broadcom executive emphasized that semiconductor manufacturers must simultaneously address advanced packaging challenges and power consumption limits to support next-generation hardware.
- Micas Networks Unveils 3 102.4T Open Ethernet Platforms for AI Factories — HPCwire, 2026-10-12Micas Networks has launched three 102.4T Open Ethernet platforms designed for high-scale AI factory interconnects.
- Beijing escalates domestic chip drive with Broadcom audits — DigiTimes, 2026-09-27China is increasing its domestic push to replace foreign technology in critical computing infrastructure, focusing on high-speed networking hardware and AI accelerators.
- Google TPU pays back investment in 1 year, creating upside for MediaTek, Broadcom ASIC revenues — DigiTimes, 2026-09-14Google Cloud reported that its AI server investments, specifically using its in-house TPUs, have a rapid payback period of less than one year, benefiting ASIC design partners like MediaTek and Broadcom.
- Broadcom says CPO fits scale-out better than scale-up — DigiTimes, 2026-09-03Broadcom stated that co-packaged optics is better suited for scale-out networking rather than scale-up, citing potential power savings of up to 70% versus pluggable modules.
- Broadcom puts 30GW, US$350 billion of custom AI silicon on the calendar through fiscal 2028 — DigiTimes, 2026-09-03Broadcom outlined that its six XPU customers require 30GW of accelerated compute capacity between fiscal 2027 and 2028, totaling US$350 billion in semiconductor shipments.
- Broadcom's custom AI silicon takes over the income statement as revenue climbs 86% — DigiTimes, 2026-09-03Broadcom reported strong third-quarter revenue of US$29.6 billion, with custom AI silicon revenue reaching US$16.7 billion and now exceeding all other product segment sales.
Questions about Broadcom
- What does Broadcom make for AI data centers?
- This catalog places Broadcom on 14 parts: Hyperscaler custom ASICs (design partner), ASIC design services, IP & EDA (service), NICs / SuperNICs (manufacturer), PCIe/CXL switches & retimers (manufacturer), Scale-up fabric (manufacturer), Scale-out switch ASICs (manufacturer), Optical transceivers (manufacturer) and Lasers (manufacturer), and 6 more.
- Is Broadcom publicly traded?
- Yes — Broadcom is listed on US exchanges (NYSE/Nasdaq) under AVGO. Prices shown on this page are a nightly snapshot, not real-time, and nothing here is investment advice.
- Where is Broadcom based?
- Palo Alto, United States, per the company record in this catalog.
- Who else makes hyperscaler custom ASICs for AI data centers?
- Alphabet (Google), Microsoft, Alchip, Amazon (AWS), Marvell and MediaTek are also placed on Hyperscaler custom ASICs in this catalog.
- How well sourced are these placements?
- Every part→company placement carries a confidence tag. For Broadcom the mix is 1 disclosed and 13 inferred. "disclosed" means the company or a filing says so; "reported" comes from trade press or a reference-design list; "inferred" is a lead from the taxonomy with no citation yet. Treat reported and inferred as leads, not facts.
Answers are generated from this catalog's own rows — see sources & method.
Verification: enriched, last verified 2026-08-27. Sources: www.broadcom.com · www.sec.gov.· Markdown · JSON