Broadcom
Provides data center Ethernet switching chips (Tomahawk/Jericho series) and custom AI accelerator (XPU) design services, serving as a core supplier of data center networking and cloud vendors' custom silicon.
The Tomahawk series Ethernet switching chips hold a dominant position in the data center market; Broadcom is the largest backend design partner for cloud vendors' custom AI chips such as Google TPU and Meta.
Company Overview
Broadcom Inc., formerly Avago Technologies, renamed after its $37 billion acquisition of the original Broadcom Corporation in 2016. Current CEO Hock Tan is known for his extremely shrewd M&A integration strategy, transforming Broadcom into the world's fifth-largest semiconductor company.
Broadcom's key positioning in the AI industry chain lies in two areas: "networking" and "custom chips." As AI clusters evolve from 1,000-card to 10,000-card or even 100,000-card scales, network communication efficiency between GPUs becomes the critical bottleneck determining training speed. Broadcom's Ethernet switching chips are the core components for building large-scale AI cluster networks, and it is also the largest design services partner for cloud vendors' custom AI chips such as Google TPU.
Core Business
Data Center Network Switching Chips | Core Business (~40% of revenue)
- Tomahawk 5: The industry's highest-bandwidth Ethernet switching chip, delivering up to 51.2 Tbps per chip, supporting 64 800G ports or 256 200G ports. In AI clusters, it is used to connect GPU nodes and storage systems.
- Jericho 3: A routing chip for data centers and carrier networks, supporting high-density 800G ports.
- Spectrum Series (Competition): NVIDIA's Spectrum-X Ethernet platform is a direct competitor to Broadcom's Tomahawk, and the two are vying for dominance in the AI networking market.
Custom AI Accelerator Chips (ASIC) | Fastest Growing (~20% of revenue)
Broadcom is the most important "behind-the-scenes contributor" to cloud vendors' custom AI chips:
- Google TPU: Broadcom is the core design partner for Google TPU (Tensor Processing Unit), responsible for chip back-end physical design and packaging integration.
- Meta: Broadcom provides design services for Meta's MTIA (Meta Training and Inference Accelerator) chip.
- Broadcom does not own the intellectual property of these chips, but each chip generates substantial NRE (Non-Recurring Engineering) fees and royalty revenue.
Fibre Channel and PCIe Retimers
High-speed signal transmission inside AI servers requires retimer chips to compensate for signal attenuation. Broadcom holds a dominant share in this niche market.
Technology Moat
First-mover advantage and accumulation in switching chips: Broadcom has invested in data center Ethernet switching chip R&D for more than 20 years. The Tomahawk series iterates every two years, consistently maintaining the industry's highest bandwidth record. The ecosystem compatibility testing and maturity of the software protocol stack (e.g., Broadcom SDK) create extremely high switching costs.
ASIC design services barrier: Designing a 5nm/3nm AI chip requires thousands of engineers and years of effort, plus extremely complex EDA toolchains and IP libraries. Broadcom has one of the world's largest chip back-end design teams and can simultaneously provide custom chip design services to several top cloud vendors including Google and Meta.
PCIe Retimer monopoly position: In AI servers, the massive number of high-speed PCIe lanes between GPU and CPU, and GPU and NIC, require retimer chips to ensure signal integrity. Broadcom holds more than 80% market share in this segment.
Market Landscape
| Dimension | Data |
|---|---|
| Data center switching chip share | ~60-70%, absolute leadership |
| Custom AI ASIC market | Largest design services provider |
| Main competitors | NVIDIA (Spectrum-X), Marvell, Intel |
| Core customers | Google, Meta, Amazon, Microsoft, Arista |
Financials and Growth
| Metric | Data |
|---|---|
| Revenue (FY2024) | Approx. $51 billion (including software business) |
| Semiconductor revenue | Approx. $30 billion |
| Net profit margin | Approx. 25% |
| Growth drivers | AI cluster network upgrades (400G→800G) driving a surge in high-end switch demand; cloud vendors' custom chips ramping up → growing ASIC service revenue |
Risks and Summary
Key Risks:
- NVIDIA is aggressively promoting its Spectrum-X Ethernet platform, which may erode Broadcom's share in AI data centers.
- Cloud vendors may shift their custom chip development in-house in the future, reducing their reliance on Broadcom.
Core Industry Value: Broadcom is the "hidden king" of AI network infrastructure. In the AI arms race, everyone's attention focuses on GPUs, but the network efficiency of 10,000-card clusters equally determines training speed and cost. Broadcom's switching chips are the "traffic hubs" of AI data centers; meanwhile, its custom ASIC business gives cloud vendors their own compute options beyond GPUs.