Broadcom

Broadcom Inc. is an American semiconductor designer that has become the second-largest vendor of AI-specific chips by revenue after NVIDIA, through two business lines: custom accelerators (which it brands XPUs) built to a hyperscaler's own specification, and the AI networking gear — Ethernet switches, optical components and fabric software — that ties thousands of those accelerators together.^[1]

History

The company's roots go back to 1961 as a component maker, but the present entity was formed in 2016 when Avago Technologies acquired Broadcom Corporation and adopted its name; the 2023 acquisition of VMware for approximately US$69 billion added a large enterprise software division to the chip business. Broadcom entered the AI era through networking: its Tomahawk and Jericho switch families already carried a large share of data-centre traffic, positioning it as the natural supplier when AI clusters pushed Ethernet to the limits of available bandwidth.^[1]

Custom Silicon and the TPU Partnership

Broadcom's defining AI relationship is with Google: the company has been the design partner for Google's Tensor Processing Unit family since the early generations, and current TPU deployments designed with Broadcom are manufactured on TSMC's 3-nanometre process and power the training and serving of Gemini.^[2] Because a fully designed AI chip — ASIC — is expensive to develop, only a handful of companies can afford one, so Broadcom's role is to industrialise that process for customers who want an alternative to NVIDIA's margins.

The customer list widened through 2025 and 2026. Broadcom publicly confirmed multiple XPU customers during its fiscal 2025 quarters, reported an OpenAI custom-silicon programme (codenamed Titan, targeting roughly 10 gigawatts of deployments), and expanded a Google–Broadcom compute partnership that also supplies multiple gigawatts of TPUs to Anthropic.^[3] Analyst coverage in early 2026 placed Broadcom's committed AI backlog at roughly US$73 billion.^[4]

Financial Scale

Broadcom reported fiscal 2025 AI revenue of US$20 billion, up 65 percent year over year, with a first-quarter fiscal 2026 AI quarter of US$8.4 billion.^[5] At its March 2026 investor day the company guided fiscal 2026 AI revenue to around US$58 billion and signalled more than US$100 billion of AI chip sales in 2027 — a trajectory that makes custom silicon credible as structural competition to NVIDIA rather than a niche.^[4]

Why It Matters for AI

Broadcom's rise is the clearest evidence that hyperscalers want negotiating leverage over their most important input. Every gigawatt of XPUs deployed is a gigawatt not rented from NVIDIA, and Ethernet-based Broadcom fabrics interoperate with merchant switches, lowering the switching cost of leaving an NVIDIA-only stack. The result is a gradual shift from a single-supplier accelerator market toward a multi-architecture one.

>See Also

• Tensor Processing Unit
• NVIDIA
• TSMC
• Penang Semiconductor and AI Ecosystem
• AI Chip Export Controls
• GPU Cluster