Broadcom

Phase A — Understand the Business

Lens 1 · Company Overview

Two businesses fused by Hock Tan's M&A machine. Semiconductor Solutions — $15,009M Q2 FY26, +79% YoY, 68% of revenue, 62% segment op margin. Infrastructure Software — $7,178M, +9% YoY, 32% of revenue, 79% segment op margin. The semi half is two engines: custom AI accelerators (XPUs/ASICs) co-designed with hyperscalers, and the merchant AI-networking franchise (Tomahawk/Jericho switches + optical/PCIe). Software = the VMware roll-up (2023, ~$69B) + CA + Symantec, sold as VMware Cloud Foundation (VCF).

The model is not edge innovation — it's franchise consolidation: buy mission-critical, high-switching-cost technology, cut non-core R&D, re-price to the top ~1,000 accounts, harvest cash. Six core custom-chip customers named on the Q2 call: Google (TPU), Meta (MTIA), OpenAI, Anthropic + two unnamed. Investable identity for THIS call: the arms dealer to the companies trying to escape the Nvidia tax — the layer that owns the custom-silicon workflow (chip + interconnect + reference system), with a software annuity bolted on for cash stability.

Contract structure is the crux for the LONG: multi-year, firmly-committed, largely non-cancellable co-design programs. Note 2 confirms ~$164.6B of firmly-committed RPO and that ~64% of contract liabilities (a separate, smaller balance) carry termination-for-convenience — but the RPO figure specifically captures the committed, no-termination-right contracts. That distinction is the whole ballgame: the headline backlog is the sticky part.

Lens 2 · Supply Chain

Fabless, one layer below Nvidia. Upstream: TSMC (3nm/5nm) + CoWoS advanced packaging (the same chokepoint Nvidia/AMD queue for), HBM from SK Hynix/Samsung/Micron, EDA from Synopsys/Cadence. Downstream: unlike Nvidia (sells finished GPUs/racks), Broadcom sells the design + silicon to the hyperscaler, who owns the system and the integration margin — though the 10-K/10-Q now note Broadcom supplies "in some cases, racks," i.e. it is creeping up the stack toward systems (relevant to the Note 11 rack-lease backstop).

Capital intensity is trivially small — capex $481M for H1 FY26 vs $18,753M operating cash flow. TSMC carries the fab capital; Broadcom converts backlog to cash at ~46% FCF margin. This is the asset-light mirror of the fab-heavy names in the memory/foundry sweep — and a structural reason the ASIC layer can capture durable rent without the capex drag.

Chokepoint to watch (ties to the ASIC War dispatch): CoWoS + advanced-packaging allocation at TSMC is the shared bottleneck. If TSMC packaging capacity is the binding constraint, it caps how fast the $164.6B RPO converts — a supply-side falsifier distinct from a demand-side one. Meta–Samsung's $6.5B 2nm arrangement (the dispatch's second data point) is the hyperscalers' attempt to diversify foundry/packaging away from the single TSMC dependency.

Lens 3 · Competitive Advantages (moats)

1. Custom-silicon co-design lock-in. Once a hyperscaler co-designs a multi-year XPU roadmap with Broadcom, switching means re-architecting chip + interconnect + software stack across generations. The ~$164.6B committed RPO is the moat made legible — customers have contractually pre-committed years of spend. Broadcom holds ~60–70% of the custom-AI-ASIC co-design market (Marvell ~35%).
2. AI-networking dominance. ~40% of AI revenue is networking (Tomahawk/Jericho + optical). As clusters scale past 100k accelerators, the fabric becomes the bottleneck — and this is the one layer where Broadcom directly contests the "system" moat Nvidia calls its widening advantage (NVLink/Spectrum). Custom XPUs need an interconnect; Broadcom sells both sides.
3. Software switching costs. VMware/VCF at 79% segment op margin, mission-critical, re-priced to enterprise — a cash annuity that funds the AI build and smooths the semiconductor cyclicality.
4. vs Nvidia (the rent-map logic). Nvidia's edge is generality (CUDA) + rack integration; Broadcom's is specificity — a chip tuned to one inference workload, cheaper per useful FLOP. As inference standardizes (~2/3 of compute is now inference ), the CUDA generality premium erodes exactly where Broadcom's specificity wins. This is the "rent is moving from merchant GPU (high but peaking) to custom ASIC (rising, contested)" mechanism, stated at the company level.

Bargaining power is asymmetric and it's the key nuance: Broadcom has strong power over suppliers (it's TSMC's marquee customer) but weak power over its handful of hyperscaler customers — who own the chip IP/stack and could re-bid or in-source. The moat is real at the program level (switching mid-roadmap is painful) but thin at the renewal level (the customer owns the design). That asymmetry is the bear case's home.

Lens 4 · Segments

AI is ~49% of total revenue and the entire growth story. Trajectory: FY25 AI $20B (+65%) → FY26 guided $56B (+180%) → FY27 reiterated >$100B; split ~60% XPU / ~40% AI networking. The structure: a slow-growth, ultra-high-margin software annuity (32% of revenue, 79% margin) financing a hyper-growth, customer-concentrated AI-silicon business (49% of revenue) — the inverse of Nvidia's ~90%-Data-Center monoline. Geographic mix is heavily Asia-Pacific ($14,535M of $22,187M Q2), reflecting shipment/title transfer in Penang, Malaysia — not end-demand geography.

Phase B — Measure Performance

Lens 5 · Earnings Result (Q2 FY26, reported June 3 2026)

Revenue $22,187M, +48% YoY. AI semi $10.8B, +143%, above the ~$10.7B guide. GAAP: gross margin 69%, operating income $10,788M (48.6% GAAP op margin), net income $9,310M, diluted EPS $1.91 GAAP. Non-GAAP EPS ~$2.44 vs ~$2.32 est (beat ~5%). FCF ~$10.0B (Q2 OCF $10,493M − capex ~$0.48B) ≈ 45% of revenue. Guidance: Q3 ~$29.4B (vs ~$28.5B Street), Q3 AI $16.0B (>200% YoY), FY26 AI $56B (left unchanged), FY27 AI >$100B reiterated.

Market reaction: −12.6% next session (June 4), ~$280B erased — a beat-and-fall. The mechanism matters for the LONG's entry: the FY26 AI number was left unchanged (not raised) and software was soft, so a stock priced for an AI raise sold off on its absence. This is the same structural pattern as Nvidia's DeepSeek-day −17% — elevated AI expectations make even record prints a downside source. Since then the stock has recovered to $374 ($1.78T). The tape says: the market is not doubting the thesis, it is policing the pace of upgrades. That's a tradeable setup for a LONG that underwrites the multi-year RPO rather than the next single guide.

Balance-sheet flags in the print: inventory built to $4,328M (from $2,270M) and receivables to $10,830M (from $7,145M) — a $3.7B AR + $2.1B inventory build ahead of the H2 AI ramp. Consistent with ramping a backlog (building ahead of shipment), but worth watching that it's demand-led, not channel-stuffed; the 42%-distributor concentration makes the AR build a real (if presently benign) item.

Lens 6 · Earnings Calls (sentiment trend)

Hock Tan's tone has escalated across the last several calls from measured to near-evangelical: XPU/networking demand described as "insatiable," six mega-deal custom-chip customers, an OpenAI 10GW collaboration, and the reiterated >$100B FY27 AI target. The tell for a disciplined LONG: reiteration, not expansion, of the FY27 number — Tan is credible precisely because he did not raise into a hot tape (which triggered the June-4 drop). Recurring phrases: "custom AI accelerators," "AI networking," "insatiable," "leading-edge." What is conspicuously not volunteered: per-customer economics, XPU gross margin vs merchant, and the split of the $164.6B RPO by customer. Confidence on the call, concentration in the 10-K. (Note: transcripts/ dir on the shelf is still empty — ingest-transcript.ts could not clean-scrape the Q2 call from a paywall-free source this run; call figures are from CNBC/Yahoo/8-K. Flagged as a shelf gap.)

Lens 7 · Comps

Lens 8 · Stock-Price Catalysts (moves >5%, last ~18 months)

• Dec 2024 ~+24% — the SAM disclosure ($60–90B AI SAM by 2027, 1M-XPU clusters) launched the re-rate; AVGO crossed $1T.
• Oct 13 2025 ~+9% — OpenAI 10GW custom-accelerator collaboration announced.
• Jan 2025 (DeepSeek) / Apr 2025 (tariffs) — correlated complex-wide drawdowns.
• June 4 2026 −12.6% / ~−$280B — beat-but-no-AI-raise + soft software.
• Pattern: up-moves from AI disclosure (new deal, new SAM), down-moves from AI expectations (a good print without a raise). For the LONG, the actionable read: the durable re-rates come from new committed backlog, and the RPO jump to $164.6B is exactly that kind of catalyst — but one the market has only partly priced (the stock is below its 52wk high).

Phase C — Judge People & Books

Lens 9 · Management (Hock Tan)

The most accomplished serial acquirer in semis — the "Broadcom Playbook" executed across LSI/Brocade/CA/Symantec/VMware. Capital return is enormous and shareholder-aligned: H1 FY26 dividends $6,178M + buybacks $8,450M, 15th+ consecutive dividend raise, large insider equity, performance-weighted comp. Capital-allocation archetype: franchise consolidator + cash-return machine, not a reinvest-everything founder.

The distinct tension vs Nvidia: the model runs on debt. VMware pushed total debt to $64,907M, and ~$20B/yr FCF splits between dividends, buybacks, and debt paydown rather than Nvidia's net-cash pure-buyback. Red flags: (1) the growth algorithm historically required an ever-larger next acquisition (accretion treadmill) — though the AI-organic story now carries growth, reducing that dependency; (2) VMware per-core re-pricing alienated customers and software grew only +9% — if churn accelerates the "annuity" weakens (it contributed to June-4); (3) the Note 11 $29B lease backstop is a new capital-allocation wrinkle — using the balance sheet to finance customers' access to compute, which juices deployment but socializes customer credit risk onto Broadcom. Best-in-class allocator, but levered and now taking on contingent customer-financing exposure to keep the AI ramp on schedule.

Lens 10 · Forensic Red Flags (quality-of-earnings — now primary-verified)

1. Goodwill/intangibles + amortization add-back — HIGHEST FLAG. Three decades of deals: Goodwill $97,801M + Intangibles-net $28,333M = $126.1B = 70% of the $179,158M balance sheet. Non-GAAP EPS adds back $1,967M/quarter of acquisition-related amortization ($1,461M COGS + $506M opex). That add-back represents real capital spent on acquisitions. Like-for-like, Nvidia's headline earnings are cleaner (its gap is SBC; near-zero acquisition amortization). This does not break the LONG — it caps the multiple and is why the seed underwrites GAAP-plausible economics, not the flattered non-GAAP.
2. Debt $64,907M — genuine leverage/refi exposure Nvidia (net cash) doesn't carry; effective rates 1.4%–5.8%, laddered, $47.8B due "thereafter" (post-FY30). Manageable given ~$20B/yr FCF but a real anchor if AI growth disappoints.
3. Customer concentration — WORSE than prior read. One distributor = 42% of net revenue (from 29%); top-5 end customers ~45% (from 40%). Broadcom's biggest XPU customers overlap Nvidia's biggest GPU customers — so a single hyperscaler capex pause hits both. Explicit mgmt material-adverse-effect warning.
4. AI-revenue disclosure — amber. "AI semiconductor revenue" ($10.8B / $56B / $100B) is management-defined, not a clean GAAP segment; the audited statements show only two reportable segments. The RPO ($164.6B) is GAAP-disclosed and auditable, which is the stronger anchor — lean on RPO over the AI-metric where possible.
5. SBC $2,092M/qtr; $20,106M unrecognized — material, but buyback+dividend offset dilution (shares roughly flat at ~4,758M).
6. Note 11 backstop ($29B max exposure) — off-balance-sheet-flavored contingent exposure to a customer's lease obligations; disclosed as a subsequent event, not yet a booked liability. Watch whether it recurs/scales across deals.

Regulatory findings. No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR + AAER, 0 hits), and 10-Q Item 1/Note 10 (Contingencies) as of 2026-07-07. Note 10 states no material amounts accrued for loss contingencies; ordinary-course IP/commercial/tax litigation only, none deemed individually or collectively material. (Historical note: Broadcom's merchant-chip contracting practices have drawn antitrust scrutiny in the past — worth a standing watch, but nothing material in the current filings or SEC enforcement record.)

Forensic bottom line: Broadcom's earnings quality is lower than Nvidia's — but the gap is inside the earnings construction (amortization, leverage, self-defined AI metric), not in the reality of the backlog, which is now primary-confirmed at $164.6B firm RPO. The LONG is underwritten on the RPO + GAAP cash generation, discounting the flattered non-GAAP; the EQ flags become the stop discipline, not a reason to pass.

Phase D — Project & Stress-Test

Lens 11 · Forward Projection (non-GAAP EPS, next 3 fiscal years)

Bottom-up from FY26 guide + RPO conversion; Broadcom FY ends ~early Nov (FY26 = Nov'25–Nov'26). Share count ~4,876M diluted.

FY26 revenue base (~$66B) reflects the two segments' run-rate + AI ramp, above the prior stub's ~$98B typo-level figure which conflated multi-year RPO with a single year. The honest FY26 total is ~$62–67B (Q1 $19.3B + Q2 $22.2B actual + Q3 ~$29.4B guide implies a very strong H2; ~$66B full-year is the reconciled base). FY27 >$100B is the AI-segment guide, not total revenue — total FY27 could reach ~$100–115B including software + non-AI semi.

Log the base call (Brier-tracked): npx tsx scripts/research/forecast.ts create --topic hardware --question "AVGO FY27 AI semi revenue >= $100B" --p 0.62 --resolves 2027-12-15 --tags broadcom,deep-dive — p=0.62 reflects: RPO makes the number deliverable (high confidence it's contracted) but conversion timing + the "does it slip a quarter" risk keeps it below 0.7.

Lens 12 · Bull vs Bear

Bull. Broadcom is the structural winner of the exact rent-shift the thesis names — commodity inference migrating from merchant GPU to custom silicon. The evidence is no longer a slide, it's a $164.6B firmly-committed RPO including the OpenAI 10GW contract, growing AI +143%, ~70% co-design share, a networking franchise that Nvidia must defend, and a 79%-margin VMware annuity funding it — all at ~46% FCF margin with near-zero capex. If FY27 AI lands at >$100B (and the RPO says the demand is contracted), the premium is earned and the SOTP (below) supports meaningful upside.

Bear (impairment vectors). (1) Premium multiple (~41.5x EV/EBITDA) on lower earnings quality than Nvidia leaves no room for a merely-good quarter — the −12.6% June-4 drop proved it. (2) Earnings quality: $1.97B/qtr amortization add-back, $64.9B debt, a self-defined AI metric, and a new $29B lease backstop. (3) Concentration: 42% single-distributor / 45% top-5, overlapping Nvidia's customers — doubly-levered to one capex cycle, plus single-program-reset risk because the customer owns the design.

Pre-mortem (18 months out, thesis broke): A hyperscaler re-bid or in-sourced one XPU program (there's no CUDA switching cost — the customer owns the stack), CoWoS packaging capped the RPO conversion so FY27 AI printed ~$85B not $100B, VMware churn accelerated as enterprises finished escaping per-core repricing, and the $29B backstop drew as a customer's compute-lease economics wobbled. On a ~41x EV/EBITDA priced for the raise, the stock de-rated 25–35% and the $64.9B debt turned a growth story into a deleveraging one.

Contrarian view (what the market refuses to see): The consensus frames Broadcom as the safer AI bet than Nvidia ("it has software, it owns the custom-silicon future"). The data say it's the higher-multiple, more-levered, lower-EQ, more-concentrated name — the riskier of the two, not the hedge. The LONG works anyway, but only if you size it as a thesis position (own the rent-shift) with a hard valuation/backlog stop — NOT as a safe-compounder substitute for Nvidia.

Lens 13 · Devil's Advocate (short-seller)

A custom-silicon "win" is a single-customer dependency dressed as a moat: if Google re-bids its TPU or Meta re-insources MTIA, an entire revenue line resets overnight, and there is no CUDA-style switching cost protecting Broadcom because the customer owns the IP and the stack. The $164.6B RPO looks bulletproof until you remember it's concentrated in a handful of counterparties whose own AI-monetization is unproven — if OpenAI's or Anthropic's economics disappoint, the "firmly committed" contracts get renegotiated regardless of the ink (and the $29B Note 11 backstop tells you Broadcom is already financing a customer's ability to pay). The non-GAAP EPS is flattered by $1.97B/qtr of real-money amortization; strip it and the leverage and pay a premium to Nvidia for the privilege. The AI number is management-defined. And the 42%-single-distributor line is the kind of concentration that, historically, precedes a nasty air-pocket when one whale pauses. The most dangerous competitor the bulls underestimate isn't AMD — it's the customers themselves, who are Broadcom's partners today and its replacements tomorrow.

Lens 14 · Management Questions (ordered by information value)

1. Of the $164.6B RPO, what is the split by customer, and how much sits with counterparties whose own AI revenue is still pre-scale?
2. What is XPU gross/operating margin vs the merchant-networking and vs the corporate 62% semi-segment margin — is AI accretive or dilutive to semi margins?
3. If a hyperscaler re-bids or in-sources an XPU program at renewal, what is the revenue-line reset risk and are there take-or-pay/termination protections beyond the current program generation?
4. What are the terms, triggers, and recurrence expectation of the Note 11 $29B lease backstop — is customer-financing now a standing part of winning deals?
5. What is the quarterly acquisition-amortization add-back trajectory and GAAP AI-segment profitability?
6. How CoWoS/advanced-packaging-constrained is the RPO conversion — is FY27 >$100B supply-capped or demand-capped?
7. VMware net-retention / churn since per-core repricing — is the annuity stable?
8. Debt paydown path and refi schedule on the $47.8B due post-FY30?
9. Networking attach rate as clusters standardize — does the fabric grow with or faster than XPU units?
10. How correlated is your demand to the same hyperscaler capex cycle as Nvidia, and what's your exposure if 2027 CSP capex growth decelerates from ~40%?
11. What is the next acquisition (if any) the model needs, and does the organic AI story now fund growth without one?
12. How do you defend the networking franchise against Nvidia Spectrum-X + in-house fabrics?
13. What share of the RPO is firm PO vs framework/LOI, and how is "firmly committed" tested for the RPO disclosure?
14. Geographic/tariff exposure given the 66% Asia-Pacific shipment mix?
15. What single scenario would most impair the AI franchise, and how is it hedged?

SOTP — the AI-ASIC segment (the number the brief asked for)

, shown with arithmetic; anchors: Marvell ~63x adj-EBITDA on ~41% growth (pure ASIC comp), de-rated for Broadcom's scale/maturity.

• AI franchise (FY27 basis): ~$100B AI-semi revenue guide. Apply ~55% incremental/segment EBITDA margin (semi op margin 62%, AI the richest slice, less allocated overhead) → ~$55B AI EBITDA.
• At a de-rated 18–22x EBITDA (vs Marvell's ~63x — Broadcom is larger, higher-share, but slower-growing at scale): ~$1.0–1.2T for the AI franchise alone.
• Core semi (non-AI) + Infrastructure Software: ~$45–50B revenue, ~$20B EBITDA (software 79% margin + core semi) → at high-teens EBITDA multiple → ~$350–450B.
• SOTP enterprise value ≈ $1.35–1.65T vs current EV ~$1.83T (cap $1.78T + net debt ~$45.3B).

Read (honest): On a conservatively de-rated AI multiple, the SOTP sits slightly below the current EV — the current price already capitalizes much of the FY27 AI franchise. The LONG is therefore NOT a "cheap on a depressed base" call — it's a bet that (a) the RPO converts and lifts the AI-EBITDA base toward FY28 (~$130B revenue → SOTP re-rates above current EV), and/or (b) the market keeps paying Marvell-adjacent multiples (>22x) for a franchise with better share and cash conversion. Upside case (FY27 lands + AI multiple holds ~22–25x + FY28 visibility): EV ~$2.1–2.4T, ~+15–30%. Base: roughly fair-to-modestly-higher as backlog de-risks, ~+5–15%. This is why the seed is LONG/MEDIUM, not high-conviction — the thesis is right, but the entry is not a fat pitch on valuation; it's a fair price for a structurally-winning franchise with a hard stop.