Thematic Exposure — 9/10
Broadcom is the purest large-cap proxy for two highest-conviction AI infrastructure themes: custom AI
accelerators (XPUs) and AI back-end networking. Dominant 60-70%+ share in custom AI ASIC inside a
Broadcom/Marvell duopoly. AI networking leadership (Tomahawk 6, Jericho 4). Plus a VMware near-monopoly
(~72-90%+ on-prem virtualization). Oligopoly gate PASS decisively in multiple segments. Held from 10
only because precise custom ASIC share estimates range 60-70%+ (not a clean >50% single number) and
customer concentration is a structural offset.
Weight: 35%
Custom AI Accelerators (XPU) — 60-70%+ Share
Broadcom/Marvell Duopoly — Co-Designed, IP-Locked, Multi-Year Engagements
Broadcom co-designs custom accelerators for Google (TPU), Meta, ByteDance, and Anthropic. Each
engagement is 3-5 year, multi-billion-dollar with deep IP entanglement (SerDes, packaging). $30B
of AI bookings in FQ2'26 with visibility through 2028. AI semiconductor revenue $10.8B in FQ2'26
(+143% YoY), 49% of semi revenue. FY27 AI guide >$100B.
Custom ASIC Share
60-70%+
Broadcom/Marvell duopoly ~95%
AI Bookings
$30B
In-quarter
FY27 AI Guide
>$100B
Management line of sight
AI Networking — First-to-Market Merchant Silicon
Tomahawk 6 — Jericho 4 — Ethernet Alternative to NVIDIA's Stack
Tomahawk 6 (102.4 Tbps) and Jericho 4 are first-to-market. Backlog of AI switches exceeds $10B.
Broadcom is the merchant-silicon standard for Ethernet AI fabrics — the credible open alternative
to NVIDIA's proprietary NVLink/Spectrum stack. 200T switch tape-out this quarter.
Infrastructure Software (VMware) — Near-Monopoly Annuity
Segment Mix (FQ2'26)
| Segment | Revenue | % of Total |
|---|---|---|
| AI Semiconductor | $10.8B | ~49% |
| Non-AI Semiconductor | $4.2B | ~19% |
| Infrastructure Software | $7.2B | ~32% |
FQ2'26 segment revenue. Data sourced from Daloopa (company_id 22).
Competitive Moat
Custom silicon co-design with 3-5yr IP lock-in.
Each XPU engagement entangles Broadcom's SerDes, advanced packaging,
and design IP so deeply that switching mid-program is technically infeasible. Customers cannot
replace Broadcom within 12 months.
Networking first-mover (Tomahawk/Jericho).
Tomahawk 6 at 102.4 Tbps and Jericho 4 have no peer in merchant
silicon. The credible open-Ethernet alternative to NVIDIA's proprietary NVLink/Spectrum stack.
$10B+ AI switch backlog.
VMware switching costs (multi-year migration).
Migrating off VMware is a 12-24 month enterprise undertaking.
Broadcom is a price-setter: VMware contract increases of 300%+ are sticking because the alternative
is a multi-year re-platforming project. Custom XPU is sole-sourced by design.
Oligopoly Gate
PASS. 60-70%+ in custom AI ASIC
(Broadcom/Marvell duopoly = ~95% combined). VMware near-monopoly (~72-90%+ on-prem virtualization).
Merchant switching silicon ~90% share. Multiple oligopoly/monopoly positions across segments.
Score Rationale
9/10 — Broadcom is the purest large-cap
proxy for AI infrastructure buildout. 60-70%+ custom ASIC share inside a duopoly that controls ~95% of
the market. First-to-market networking silicon (Tomahawk 6, Jericho 4) positions Broadcom as the
merchant-silicon standard for Ethernet AI fabrics. VMware adds a near-monopoly annuity with 78% operating
margins and ~93% gross margins. Held from 10 because precise custom ASIC share estimates range rather
than converge on a single clean number, and customer concentration (Google remains the largest single
customer) is a structural offset that cannot be dismissed. But the breadth of dominance across custom
compute, networking, and infrastructure software is unmatched.
Data sourced from Daloopa (company_id 22), Counterpoint/JPMorgan ASIC share estimates, IDC/Gartner virtualization market data.