Broadcom
The Custom-Silicon & VMware Compounder

Snapshot: Executive Summary

Broadcom is the world's only at-scale AI custom-silicon partner outside Nvidia, and the operator of the most consequential infrastructure-software franchise in enterprise (VMware). Q2 FY26 (quarter ended May 3, 2026) set a fresh record at $22.19B revenue (+48% YoY)¹, non-GAAP gross margin of 77.1%, adjusted EBITDA of $15.24B (69% of revenue), non-GAAP EPS of $2.44 (+54% YoY), and free cash flow of $10.26B (46% of revenue). AI semiconductor revenue reached $10.8B (+143% YoY)², now roughly half of total revenue, up from ~31% in FY25. Q3 FY26 is guided to ~$29.4B (+84% YoY) with AI semi revenue of ~$16.0B.

Three franchises, three valuation regimes:

• AI custom silicon (~$10.8B/qtr run-rate): XPUs (Google TPU, Meta MTIA, plus confirmed Anthropic and OpenAI, with ByteDance/Fujitsu reported) + AI-Ethernet networking, ramping toward the company-guided >$100B FY27 AI revenue.
• Non-AI semiconductors: Wireless (Apple), broadband, server storage, industrial. Mature, FCF-rich, cyclical.
• Infrastructure software (~$7.2B/qtr): Post-deal VMware re-stacked onto VCF/VVF subscription bundles; high-90s gross margin, mid-60s EBITDA margin, fully ratable.

At 31.4x forward earnings (FY26E non-GAAP EPS $11.62) versus a 5-year median of ~22x, AVGO trades at a premium, but with AI now ~half the company, the ~46% FCF-to-revenue conversion, and the recurring-software EBITDA, the re-rate is fundamental. Net debt of ~$45B is down from a ~$73B post-VMware peak. The stock sits at $365, ~26% below its $495 ATH after a "priced-for-perfection" post-Q2 selloff, which de-risked the entry. Reiterate Buy, 12-month PT $483 via SOTP (bull $560 / bear $360).

Tactical: AVGO is trading at $365, ~32% below our $483 12-month target, the post-Q2 ~26% drawdown removed the "priced-for-perfection" premium while the AI ramp accelerated (Q2 AI +143% YoY, FY27 guide >$100B). Rating Buy; note the bear case ~$360 sits near spot, an attractive entry asymmetry.

Investment Thesis

Rating: Buy. The probability-weighted SOTP (30% bull / 50% base / 20% bear) blends to ~$482, essentially the $483 base, so the base case is the probability-weighted central value, not an optimistic tail. The base requires only that Broadcom deliver against its own reiterated FY27 >$100B AI guide at a multiple where it already trades. Crucially, the bear case ($360) sits right at the current $365 price, today's entry roughly equals the downside-scenario fair value, an attractive asymmetry after the post-Q2 ~26% drawdown removed the "priced-for-perfection" premium.

Business Overview

Broadcom reports two segments, Semiconductor Solutions ($15.01B in Q2 FY26, 68% of total, +79% YoY) and Infrastructure Software ($7.18B, 32%, +9% YoY). For valuation, the more useful frame is three franchises, because the AI semis subset trades at radically different multiples than the rest of semis, and AI is now ~half of total revenue, up from ~31% in FY25:

Franchise commentary

• AI semis: Two pillars: custom XPUs (Hock Tan's "ASIC compute", Google TPU and Meta MTIA are the foundational customers; Anthropic and OpenAI are newly confirmed, with ByteDance and Fujitsu widely reported as the fifth and sixth) and AI networking (Tomahawk 6 and Jericho 4-5 Ethernet switches displacing Infiniband for hyperscale AI fabrics). Both ramp on multi-year, multi-generation contracts with high switching costs. Networking was ~40% of Q2 AI revenue, expected to trend toward ~30% over time.
• Non-AI semiconductors: Wireless (Apple FBAR/PA content), broadband (cable modem, PON), server storage (HDD controllers + custom storage SoCs), and industrial. This pool is mature, FCF-rich, cyclical, and recovering from a 2024 trough.
• Infrastructure software: Post-VMware, this bucket is ~95% software (VMware re-stacked onto VCF/VVF, plus Symantec EntSec, CA mainframe, Brocade fibre-channel SAN). Recurring, ratable, high-90s gross margin, mid-60s EBITDA margin on a normalized basis.

AI Custom Silicon: The >$100B FY27 Call

The single most important slide in this report. Hock Tan's framing has escalated from a $60-90B FY27 SAM to an explicit company guide of >$100B FY27 AI revenue³ (he said "significantly above $100B" on the Q2 FY26 call, reiterated, not raised). The customer cohort has expanded from three "lead customers" to six core custom-silicon customers: Google (TPU), Meta (MTIA), plus newly confirmed Anthropic and OpenAI (which signed a ~10-GW custom-accelerator deal), with ByteDance and Fujitsu widely reported as the remaining two. AI revenue hit $10.8B in Q2 FY26 (+143% YoY), ~half of total revenue; Q3 FY26 is guided to ~$16.0B (+200%+ YoY). Management explicitly ruled out entering the full-rack/system business.

What the custom-silicon model actually is

Each customer brings a target workload (Google: TPU for training + inference of internal models; Meta: MTIA for ranking/recommendation; OpenAI/Anthropic: frontier-model training and inference). Broadcom does the SerDes IP, the package design, the foundry interface, the manufacturing yield ramp, and the supply-chain orchestration. The customer owns the architecture and the compiler. XPU gross margin is lower than a merchant GPU, but the volumes are committed multi-year and the design-win incumbency is sticky, the architecture is co-evolved with the customer's stack and a re-bid is a 24-36 month project. The higher-margin networking silicon (Tomahawk 6 / Jericho) lifts the blended franchise margin.

AI Ethernet: the quieter story

The AI networking line (Tomahawk 6 switches at the leaf and Jericho 4-5 at the spine) is the displacement story for Infiniband. Hyperscalers want a single network fabric, Ethernet, for both their general cloud and AI clusters. Networking was ~40% of Q2 FY26 AI revenue and is expected to trend toward ~30% as XPU volume scales. This is a higher-margin business than XPU.

The capex backdrop

VMware Integration: The Cash-Compounder

The Hock Tan playbook on VMware was identical to CA (2018) and Symantec EntSec (2019): radically simplify the SKU stack, push the customer base onto subscription bundles, re-paper the top ~10,000 strategic accounts at an ARR uplift, and let the long tail churn. Infrastructure Software posted $7.18B in Q2 FY26 (+9% YoY), ~32% of total revenue, a high-90s gross-margin, mid-60s EBITDA-margin recurring-revenue business, fully ratable. Growth has decelerated to high-single-digit YoY as the subscription re-base laps; the segment is now the steady cash-compounder while AI drives growth.

The SKU re-stack

• VMware Cloud Foundation (VCF): flagship full-stack bundle (compute + storage + network + management). Priced by core; standardizes the top-tier of customers onto a single contract.
• VMware vSphere Foundation (VVF): smaller-tier bundle for mid-market.
• Standalone perpetual licensing was discontinued. ~80% of the legacy SKU portfolio was retired or wrapped into VCF/VVF.

What's actually driving the revenue line

The top ~10K strategic VMware customers re-papered at meaningful ARR uplift on like-for-like renewals, per management commentary. The long-tail SMB and channel customers, by management's own framing, not strategic, are the locus of any churn. The economic impact is modest because that tail was already a small share of revenue.

The risk lens

The bear case is that net dollar retention slips below 90% as Nutanix, Red Hat OpenShift Virtualization, and Proxmox accumulate share at the renewal cliff. Our position: the strategic-customer renewal pipeline is largely visible; the marginal churn happens in tiers that were never margin-meaningful. The harder question is the FY27 first-cohort renewal wave, when the largest first-cohort 3-year contracts come up for re-papering.

Financial Health & Trends

FY15-FY26E: the M&A arc

Broadcom's revenue line cannot be read without the M&A overlay. The current entity is the product of: Avago/Broadcom Limited merger (Feb 2016), Brocade (Nov 2017), CA Technologies (Nov 2018), Symantec Enterprise Security (Nov 2019), and VMware (Nov 2023). Every step-change in revenue corresponds to a closed deal; the steady-state organic growth between deals has been mid-to-high single digits, with gross margin steadily expanding as software mix has grown.

Q2 FY26 print highlights

Note: capex was just $231M (fabless), so FCF tracks EBITDA closely. The stock fell ~13-15% in the days after this beat because management reiterated, but did not raise, the FY27 >$100B AI target: a textbook "priced for perfection" reaction.

Capital Allocation & Returns

Hock Tan's stated capital priority hierarchy is unchanged: (1) grow the dividend; (2) reduce debt; (3) opportunistic buyback; (4) M&A. With FCF now running ~$40B+ annualized, all three of the first priorities run concurrently. Q1 FY26 alone returned $10.9B to holders ($3.1B dividends + $7.8B buybacks).

Why the dividend matters

Broadcom has raised the dividend every year since 2011, through the 2019 semiconductor downturn and the VMware closing year. The quarterly dividend stands at $0.65/share (declared with Q2 FY26). We expect the next dividend raise (typically announced in December) to land in the +10-15% range; management has been explicit that the dividend grows in line with normalized FCF, which is now compounding fast on the AI ramp.

The M&A question

The most-asked question on every Hock Tan investor call. Net debt has fallen to ~$45B (from ~$52B at Q1 and a ~$73B post-VMware peak) on the AI-driven FCF; deleveraging is fast. Our read: he is unlikely to do a large software deal while the AI ramp is the priority, but a bolt-on chip M&A (small, accretive, custom-silicon-adjacent) is possible at any time. A large re-leveraging deal would test the dividend-first priority.

Valuation Overview

AVGO trades at 31.4x fwd P/E (FY26E non-GAAP EPS $11.62) vs a 5-year median of ~22x. Read at face value that looks rich. But the headline multiple obscures the three-franchise structure, applying a single P/E to a business that is now ~half AI custom-silicon (deserves a premium multiple), part non-AI semis (deserves a market multiple), and part recurring software (deserves a software multiple) is the wrong frame. Peer fwd P/E: NVDA ~20x, QCOM ~23x, TXN ~35x, AMD ~60x, MRVL ~85x. The right frame is SOTP, and we develop the full mechanics in the next section, which reconciles to the $483 base.

Why the SOTP

A blended P/E argument always loses to a fundamentals-aware SOTP when the underlying business mix is heterogeneous. We separate AVGO into three buckets, apply scenario-appropriate multiples to each, sum the enterprise values, subtract net debt, and divide by shares. That work is in the next section, with an interactive scenario tool.

SOTP: Sum-of-the-Parts

The SOTP is the heart of the report. Each franchise is valued on its own revenue, EBITDA margin, and EV/EBITDA multiple. Toggle between Bull, Base, Bear, and Reverse-DCF scenario anchors; drag the AI and Software multiples directly to see the implied PT update.

Sensitivity grid: AI multiple × Software multiple (PT in $)

PT calculator (alternative simple-multiple view)

Risk / Reward calculator

Ask the Thesis AI-assisted checking…

Describe a scenario in natural language; the assistant returns a structured impact analysis against this dashboard's thesis, scenarios, and SOTP math. Powered by Claude via a Cloudflare Worker proxy (Anthropic key held server-side; same pattern as the live-quote feed).

Note: The assistant reasons from the dashboard's data snapshot and thesis sections, it does not browse the web, hit Bloomberg, or access real-time fundamentals beyond what's in data.js. Treat its responses as scenario-modeling support, not as primary research. Author judgments on rating, PT, and probabilities remain with the analyst.

Upcoming Catalysts

Risk Factors

• "Priced for perfection." The pre-eminent risk. The stock fell ~13-15% after the Q2 FY26 beat because Hock Tan reiterated, but did not raise, the FY27 >$100B AI target. At ~31x forward earnings the bar is extreme; any AI-guide stumble is asymmetric and re-rates the multiple fast.
• Hyperscaler ASIC in-housing. AWS Trainium ramping at scale shows hyperscalers can do this themselves; Google is internalizing more of the SerDes stack. If a lead customer (e.g. Meta) pulls a future MTIA generation fully in-house, the AI revenue trajectory disappoints.
• AI capex digestion year. A hardware-spend digestion year would compress both AI revenue growth and the EBITDA multiple simultaneously, the bear-case mechanism.
• VMware net dollar retention slipping. The FY27 first-cohort renewal wave is the test. If Nutanix / Red Hat OpenShift / Proxmox accumulate enough share, NDR slips and the software multiple compresses.
• Customer concentration + counterparty risk. Google plus the top AI customers are a large share of AI revenue. Critically, OpenAI and Anthropic are pre-profit counterparties funding multi-gigawatt commitments that depend on continued capital raising.
• China export-control expansion. A broader regime including custom XPU shipments to China-headquartered hyperscalers would directly affect ByteDance revenue.
• Non-AI semi cyclicality / M&A re-leveraging. Wireless (Apple content) and broadband (cable capex) remain cyclical. Separately, a large software deal at premium pricing would force re-leverage and test the dividend-first capital priority.

Scenario Stress Tests

All stress-test PTs are derived from the same SOTP framework used in the interactive tool, each shock changes only the inputs noted. Each 1x on the AI EBITDA multiple is ~±$13 PT; each $10B of FY27 AI revenue is ~±$37 PT. The base case is the central anchor we maintain; note the bear composite lands ~flat to the current spot.

Bull vs Bear Debate

Technical Analysis

Trader's view

• Two tapes in twelve months: a run from ~$292 (Jul 2025) to a $495 all-time high (Apr-May 2026), then a ~26% correction to $365 after the Jun-3 Q2 print. Price is now below the 50-DMA, consolidating beneath the highs.
• Key support: the post-earnings $360-365 zone (also the bear-case fair value). A sustained break below invalidates the near-term setup.
• Key resistance: the $446 May close, then the $495 ATH overhead. Reclaiming $446 reopens the path toward the $483 PT.
• Momentum: MACD constructive after the reset; RSI neutral, not extreme.

Glossary & Methodology Notes

XPU (custom AI accelerator)

A workload-specific AI processor co-designed with a single hyperscale customer. Google's TPU and Meta's MTIA are the foundational examples; Broadcom's six-customer cohort now also includes Anthropic and OpenAI (with ByteDance/Fujitsu reported). Distinct from a merchant GPU (Nvidia) and a standardized ASIC.

SerDes

Serializer/Deserializer, the high-speed I/O block that connects a chip to off-chip signaling (PCIe, Ethernet, custom interconnects). Broadcom's SerDes IP is the bedrock of its custom-silicon program.

SOTP (Sum-of-the-Parts)

A valuation method that values each business segment of a multi-line company separately, then sums the enterprise values, subtracts net debt, and divides by shares to get a per-share fair value. Used when applying a single blended multiple obscures the underlying mix.

EV/EBITDA

Enterprise value divided by EBITDA (earnings before interest, taxes, depreciation, amortization). A capital-structure-neutral measure of valuation; appropriate for cross-comparing business with different debt loads.

VCF / VVF (VMware bundles)

VMware Cloud Foundation (the flagship full-stack bundle) and VMware vSphere Foundation (mid-market). The post-deal SKU simplification reduced ~80% of legacy SKUs into these two.

Net dollar retention (NDR)

The expansion or contraction of revenue from a fixed cohort of customers year-over-year, including upsells, downgrades, and churn. NDR >100% indicates the cohort expands; >90% is healthy for an enterprise software business.

Hyperscaler

Refers to the largest cloud and AI infrastructure operators: Amazon (AWS), Microsoft (Azure), Google (GCP), Meta (FAIR + ads infrastructure), and increasingly Oracle (OCI).

Tomahawk / Jericho

Broadcom's flagship Ethernet switch silicon families (current generation: Tomahawk 6 / Jericho 4-5). Tomahawk is the leaf-switch product (high port-count, lower-latency); Jericho is the spine/core switch (deep buffers, longer-haul). Both are the backbone of Ethernet-based AI cluster fabrics displacing Infiniband.

10-GW deal

A capacity commitment sized in gigawatts of compute power rather than chip units, e.g. OpenAI's ~10-GW custom-accelerator agreement with Broadcom. Gigawatts have become the unit of account for frontier-AI buildouts because power, not silicon count, is the binding constraint.

Methodology

• Snapshot anchor: June 26, 2026 close ($365.02), post the Q2 FY26 print (reported Jun 3, 2026; quarter ended May 3, 2026). Live price patches via the Cloudflare-Worker quote proxy on page load.
• Only the FY27 AI revenue line (>$100B) is company-guided; the FY27 non-AI and software splits are author estimates calibrated to the SOTP. All FY26-FY27 figures are estimates.
• SOTP scenarios use FY27E revenues; the EBITDA-margin and multiple anchors are the author's view, not consensus. The base reconciles to $483.
• Conclusions are the author's view. Illustrative, not investment advice.