XPU (Custom AI Accelerator) Co-Design and 102.4 Tbps Ethernet — Ruler of the Custom-Silicon Track
Broadcom sells the road that is not NVIDIA's in AI infrastructure. It holds two rails at once — XPU (custom AI accelerator) co-design and open Ethernet networking — and defends that road with seven-generation customer relationships and fabless margins. Execution is near-flawless, and the only question left is how far a market cap of roughly ~$1.8T has already bought that flawlessness.
FQ2 FY2026 AI semiconductor revenue set a record at $10.8B (+143% YoY), and FQ3 guidance was set above $16B (+200%+). Management projected fiscal 2027 AI chip revenue in excess of $100B, while new customers OpenAI and Anthropic plus the Apollo-Blackstone $35B financing (June 9, 2026) tied demand to the capital markets, lifting the custom-silicon track into an investment theme.
Design power — Broadcom (#1 in the ~95% custom-ASIC co-design duopoly, leader in Ethernet switches) and Marvell (#2, chasing share via the Google tie-up). Market companion — NVIDIA (about 80% of GPUs; both tracks grow together as inference explodes). Offsets — top-customer concentration (top five about 45%), hyperscaler in-housing risk, VMware China regulation, and an already-embedded valuation (forward P/E about 26–33x range).
The chip market splits into two tracks. The first is general-purpose GPUs, where NVIDIA holds roughly 80%. Anyone can buy them, and the software ecosystem (CUDA) runs deep. The second track is custom ASICs, application-specific integrated circuits. Here a hyperscaler designs a chip tuned only to its own workload.
Why do giant cloud companies move to the second track? The answer is cost and power. When inference runs at massive scale, a general-purpose GPU carries features the customer never uses. A custom chip strips that excess away. It does the same job at a lower total cost ownership (TCO) and with less power. As inference volume explodes past training, that gap converts directly into margin.
The custom- co-design market is effectively a duopoly. Broadcom and Marvell together hold about 95%. Broadcom is the runaway leader, Marvell is second. The barrier to entry is not chip design alone. It is core IP such as (serializer/deserializer) for high-speed data movement, packaging know-how, and the ability to lock up leading-edge TSMC capacity.
Custom silicon is not a substitute for the general-purpose GPU but a parallel market for the inference era, and the design power in that market is concentrated in just two firms: Broadcom and Marvell.
Sign in with Google to unlock the full body of every free report instantly.
Sign in with GoogleThis site runs on ads — the tier system rewards community contributions.
This report is provided for informational purposes only and does not constitute a recommendation to buy or sell any financial instrument. Investment decisions should be made based on your own judgment and responsibility. The analysis and opinions contained herein are based on information available at the time of writing and are subject to change.