Investment Thesis — Broadcom Inc. (AVGO)

§ 01One-Paragraph View

Broadcom is the cleanest "arms-merchant to the hyperscaler ASIC counter-leverage" name in the cohort: it sits at near-monopoly share (~70%) of top-tier custom XPU programs after Marvell's Trainium 3/4 stumble (per competitor.md), holds a one-generation Ethernet switch lead with Tomahawk 6 (per competitor.md, market.md), and has converted what should be classic single-supplier fragility into pass-through pricing power via cost-plus-flavored design-services contracts (per supply-chain.md). The thesis works because three structural facts compound: (1) every credible hyperscaler counter-Nvidia move flows through Broadcom IP, (2) the moat is rented over a 24–36 month physical commitment window but renewed continuously by next-generation tape-outs, and (3) a $73B AI backlog plus VMware annuity provides multi-year visibility that no merchant AI peer can match. What breaks it is not competition (low probability inside 24 months) but valuation × duration: at ~31x forward P/E and ~1.0% "real" FCF yield after honest SBC treatment, the equity is the highest-duration name in the cohort, and any combination of (a) Hock Tan's $100B 2027 AI number slipping by >25%, (b) a single hyperscaler defection, or (c) a coincident rate spike + AI capex digestion phase tightens the bull case fast.

§ 02Direction & Sizing

The screener gave a 5-conviction long. I am marking the PM conviction down to 4 and sizing medium, not large, for three reasons that the working hypothesis under-weighted: (1) the financial.md reverse DCF shows the current EV already prices ~22% blended CAGR at 42% FCF margin — "achievable but not generous" is precisely the language of a position that should not be concentrated, (2) honest SBC treatment cuts the FCF yield from 1.45% to ~1.0%, which puts AVGO's "real" cash yield below NVDA's despite higher leverage and slower top-line growth, and (3) the cohort is already heavy on Taiwan-exposed AI-cycle longs (NVDA large, TSM large), and AVGO is the narrowest customer-set inside that bucket — adding a third large-position concentrated AI-Taiwan name is poor portfolio construction even when the individual thesis is sound.

§ 03Bull Case

Five points form the integrated bull case, drawing across the seven memos:

1. The competitive moat strengthened in 2025–2026, not stabilized (per competitor.md). The Marvell Trainium 3/4 loss to Alchip on RDL interposer execution is the most important moat-trend datapoint of the AI cycle. The naive read had Marvell scaling into a co-leader; instead Marvell stumbled, NVIDIA bought 9% to hedge it into NVLink Fusion, and AVGO widened the gap on its top-tier accounts. Tomahawk 6 entered volume March 2026, a generation ahead of Spectrum-X1600 (H2 2026 at earliest). Two front lines, both strengthening.

2. The customer lock is physical, not contractual (per customer.md, supply-chain.md). ASIC tape-outs are 24–36 month physical commitments embedded in TSMC CoWoS-L allocation, hyperscaler datacenter buildout schedules, and 224G SerDes IP integration. Switching cost is measured in 18+ months of foregone roadmap and several $100M of cumulative engineering investment. The customer set went from 3 → 5 → 6 named hyperscalers in 18 months — concentration is increasing, but on materially better terms than the prior Apple-RF era.

3. Pass-through pricing power converts supply chain risk into customer risk (per supply-chain.md, financial.md). This is the most underappreciated structural feature of AVGO's business model. AI chip GM held ~65% in FY25 with massive HBM4/CoWoS-L cost inflation flowing through; Q1 FY26 blended GM held 77% with AI revenue +106% YoY. Five of six supply-chain risk vectors score 4–5, but the pass-through vector scores 1 — and that single feature recasts the whole picture. AVGO bears shipment risk (CoWoS allocation), not pricing risk.

4. The cycle is early-mid, not late (per market.md, macro.md). AI ASIC market HHI ~5,500 (effective duopoly), gross margins still at ~65% (no price war yet), backlog $73B (anchor for near-term steepness, not channel fill), customer count expanding (more hyperscalers choosing Broadcom, not fewer). Cycle-position triggers are 18+ months out even on conservative assumptions. Scale-up Ethernet is in early innings — barely begun shipping — and is genuinely additive TAM ($8B+ by 2030 per 650 Group).

5. Regulatory net-tailwind plus VMware ballast (per regulatory.md, macro.md, financial.md). AVGO sits inside the moat BIS is building around US hyperscaler compute, benefits indirectly from CHIPS via TSMC Arizona, and faces no structural antitrust risk on the AI franchise. VMware's ~$27B ARR (+26% YoY) at 60–70% segment operating margin dampens semi-cycle troughs by 200–400bps and is the reason AVGO's equity has lower amplitude than NVDA or AMD even on the same AI capex curve. The Hock Tan playbook on VMware is working as advertised.

§ 04Bear Case

Five points, in priority order:

1. Valuation is the bear case in one line. At ~31x forward P/E, ~16x forward EV/Sales, and ~1.0% "real" FCF yield post-SBC, the stock has zero margin of safety. Every dollar of Hock Tan's "$100B+ AI by 2027" line is already capitalized in the EV. The reverse DCF (per financial.md) requires 22% blended CAGR through FY30 at 42% FCF margins — achievable but not generous. That phrasing is the entire bear thesis: a long without asymmetry is a beta trade, not an alpha trade.

2. Highest-duration equity in the cohort meets a leveraged balance sheet (per macro.md). Two stacked features that didn't exist pre-VMware: (a) the AI ASIC + VMware long-tail recurring stream pulls cash flows out into 2027–2030+, giving AVGO a 1.3–1.5x SPX rate beta versus net-cash NVDA; (b) post-VMware ~$65B gross debt with refi waves in 2027–2029 cross-overs the AI revenue ramp. A coincident rate spike + AI capex digestion phase delivers numerator and denominator pain simultaneously — unlikely but unhedgeable in this name.

3. SBC at 11.8% of revenue is a structural dilution drag the headline FCF understates (per financial.md). $7.6B FY25 SBC vs $26.9B headline FCF means "real" FCF after dilution is closer to $19B, putting "real" FCF yield ~1.0%. Buybacks ($6.3–12B annually) are largely sterilizing dilution rather than compounding share count down. This is the silent re-rating that happened post-VMware and that the bull narrative does not price.

4. Customer concentration is real and one-defection severe (per customer.md, market.md, competitor.md). Three named hyperscalers (Google, Meta, ByteDance) drive the bulk of XPU revenue. A single hyperscaler in-housing more of the design stack on a TPUv9 / MTIA v4 horizon is the structural existential risk the entire ASIC partner model has to live with. No evidence of it today, but the asymmetric severity is high — losing Google would be a 20–30% multiple re-rate.

5. NVLink Fusion + Spectrum-X bundling is the silent share-shift (per market.md, competitor.md). Every dollar of Rubin GPU sold with bundled Spectrum-X attach is a dollar that does not enter Broadcom's TAM. NVIDIA's $2B Marvell investment + NVLink Fusion ecosystem (MediaTek, Alchip, Astera Labs, Synopsys) is a strategic pincer that pulls the open-fabric/UALink/Ultra Ethernet thesis sideways over 2027–2028. Lower probability than the rate / valuation risks, but a 3-year disruption that the Q1 FY26 earnings cadence will not surface in time to react cleanly.

§ 05Where the Analysts Disagreed

These are the cross-memo tensions and how I resolved each. This is the most valuable section of the thesis.

Tension 1 — Competitor (5: "moat strengthening, near-monopoly restored") vs. Customer + Market (both 4: hyperscaler in-housing as 5–10y existential threat). Is the moat permanent or rented? Resolution: The moat is rented in 24–36 month chunks but renewed continuously. Each tape-out cycle is a physical commitment that cannot be unwound mid-cycle, and the relationship-level engineering integration (224G SerDes IP, RDL interposer methodology, CoWoS-L co-allocation with TSMC) compounds across generations. The Marvell-Alchip evidence cuts the right way: it shows that even with a willing alternative partner and a motivated customer (AWS), execution risk on the back-end physical IP is real and AVGO's accumulated process know-how is not commoditized. The competitor analyst is correct on the 24-month horizon; the customer + market analysts are correct on the 60-month horizon. Both are right. The PM-level read: position sized for a 12–24 month thesis with a kill criterion that surfaces hyperscaler in-housing if it accelerates.

Tension 2 — Financial (4: reverse DCF "achievable but not generous") vs. Macro (4: highest-duration name). How does that pair with valuation discipline? Resolution: This is the core sizing question, and it pushes hard against a large position. Reverse DCF says the current EV is fairly priced for the management roadmap. Macro says the equity has the highest interest-rate beta in the cohort plus a leveraged balance sheet that can't deploy capital aggressively in a downturn. A fairly-priced fairly-leveraged long-duration equity is a 'medium' position, not a 'large' one. It is also the explicit reason to overrule the screener's conviction-5: the screener's conviction came from competitive position quality, which is correctly captured by competitor.md's 5. The PM has to integrate that across financial discipline + duration risk, which mark the position down. This is exactly the kind of trade-off the screener cannot make.

Tension 3 — Supply chain (4: ASIC contracts are cost-plus design-services) vs. Customer (4: ASIC contracts are 24–36 month physical commitments). Are these consistent or in tension? Resolution: They are consistent and reinforcing — the most undervalued feature of the business. AVGO has both pass-through pricing economics (supply shocks flow to the customer P&L, not AVGO's GM) AND deep customer lock-in (customers cannot exit mid-program because the physical commitment in TSMC capacity, HBM allocation, and tape-out cycle is non-fungible). This is the rare semiconductor profile where supplier and customer carry asymmetric risk — AVGO captures the upside of design-services margin, hyperscalers carry the downside of input cost inflation, and neither party can walk away inside the program window. The combination is what justifies the high multiple. The bear has to attack one or the other; attacking both simultaneously is internally inconsistent.

Tension 4 — The SBC re-rating (financial.md): 11.8% of revenue is real cost. Does this change sizing from medium to small? Resolution: It pushes toward small but doesn't cross the line. The right way to think about it: pre-VMware AVGO was a 4.6%-of-revenue SBC business; post-VMware it's an 11.8%-of-revenue SBC business — a structural step-change of ~700bps. That changes the "real" FCF yield from 1.45% to ~1.0%, which is below NVDA's and not far above the 10Y Treasury. But the same VMware integration is what creates the recurring-software ballast that lowers cyclical amplitude (per macro.md), and the SBC dilution is somewhat sterilized by ongoing buybacks ($12B TTM run-rate). Net: SBC is the single biggest reason this is a 4 not a 5 on PM conviction, and a medium not a large on sizing. It is also a key kill criterion below — if SBC keeps drifting up while buybacks pull back, the thesis is compromised.

Tension 5 — Regulatory mild tailwind (regulatory.md, 4) vs. Macro Taiwan tail + China decoupling (macro.md, 4). Is the regulatory frame net positive or net negative? Resolution: Net mildly positive on a probability-weighted basis. The regulatory analyst's read on BIS / CHIPS / CFIUS is correct: AVGO sits inside the US compute fence and benefits from controls that constrict Chinese ASIC competitors. The macro analyst's Taiwan tail is real but unhedgeable and shared with NVDA, TSM, and the entire fabless cohort — it doesn't differentiate AVGO. The genuinely AVGO-specific regulatory risks are EC VMware behavioral observance and the AT&T litigation; both are bounded (100–300bps segment margin compression in worst case, settlement most likely). Section 232 is the live binary inside 12 months. The regulatory + macro composite is supportive of the long, not a barrier.

§ 06Catalyst Calendar

The single highest-information catalyst is the Q4 FY26 print (Dec 2026) — it will either confirm the implied AI revenue cadence ($40B+) toward the $100B 2027 number, or surface the gap. Position size should reflect a deliberate review checkpoint at that print.

§ 07Asymmetry

Upside (if right): ~30–45% over 12–24 months, driven by AI revenue compounding through the $73B backlog at the implied cadence, multiple holding at current levels (no expansion required — the bull case is Hock Tan's $100B 2027 number coming through, not the multiple expanding from already-stretched levels), VMware free cash flow continuing to delever the balance sheet, and a fifth/sixth XPU customer (OpenAI 2027, Apple AFM optionality) adding a structurally additive program. The right way to think about upside is dollar-of-FCF growth at flat multiple: $26.9B FY25 FCF growing to $40B+ in FY27 at flat 16x EV/Sales = ~30–40% equity appreciation.

Downside (if wrong): ~25–35% over 12–24 months, driven by some combination of (a) reverse-DCF math falling short — implied 22% CAGR coming in at 14–16% triggers multiple compression to ~22x P/E (–25% just on the multiple), (b) a single hyperscaler defection (–20–30% on the news), (c) coincident rate spike + AI capex digestion compressing both numerator and denominator (–30–40% in the tail). The honest read is that 25–35% is a realistic 12-month drawdown in the bear scenarios; a sharper drawdown is tail.

Ratio: ~1.3:1 to ~1.5:1.

Verdict on asymmetry: The asymmetry is positive but not the >2:1 we want for a high-conviction add. This is the second mathematical reason for sizing medium not large — alongside the cohort-fit argument. A 1.3:1 long is a position you carry because the underlying business is durable, not one you size up because the entry is asymmetric. The asymmetry is better than NVDA's at current levels (NVDA is closer to symmetric) but worse than TSM's (TSM is the cleanest 2:1 in the cohort because the bottleneck math is unambiguous and the multiple has not stretched as far).

§ 08Kill Criteria

The thesis is invalidated if any of the following observable, time-bounded conditions trigger:

• A named hyperscaler explicitly cancels or deprioritizes a multi-year ASIC program with Broadcom in any of the next 4 quarters. Specifically: Google publicly announces TPU vN+1 with a non-Broadcom physical-design partner, Meta moves MTIA v3+ in-house in a way that excludes Broadcom IP/services, or ByteDance pulls allocation. Any one of these is a thesis crack.
• AI semiconductor gross margin compresses by more than 500bps in any single quarter without a clearly disclosed mix-shift explanation. A hard floor of 60% AI chip GM through FY26; below that, the pass-through-pricing-power thesis is empirically broken and the supply-chain analyst's conviction-4 has to be marked down.
• AI revenue cadence comes in below $36B for FY26 (implied run-rate is ~$40B+ from Q1's $8.4B and Q2 guide). A miss of >10% versus that implied cadence with no cogent explanation breaks the reverse-DCF math.
• SBC drifts above 13% of revenue or buyback pace falls below dilution pace for two consecutive quarters. This would mean the structural SBC re-rating is widening rather than stabilizing, which materially impairs the "real" FCF yield further.
• EC initiates formal Article 8(5) modification proceedings on VMware behavioral commitments. This is the regulatory channel that flips from "bounded behavioral remedy" to "structural pricing constraint" and would cap VMware segment margin trajectory at ~55%.
• Section 232 outcome includes a Taiwan-origin tariff at >15% with no CHIPS-Arizona carve-out. Below 15% or with structured carve-outs is digestible; above 15% with no carve-out is 300–600bps GM compression that does not fully renegotiate within the FY27 review window.
• Hock Tan signals retirement / transition before announced succession plan with capital allocation continuity. The entire VMware-style playbook is one CEO; an unmanaged transition would meaningfully re-rate the multiple.

§ 09Conviction Distribution Across Analysts

The cluster at 4 with one 5 is informative: every analyst sees a real long, only one (competitor) sees an unambiguous best-in-class long. That distribution is consistent with a medium long, not a large one. If the financial analyst had come back at 5 (i.e., undervalued) the case for sizing up would be clean; the financial 4 (fairly priced) is the binding constraint.

§ 10Open Questions for Next Round

1. What is the actual cost-plus contract structure on AI ASIC programs? The supply-chain analyst inferred pass-through from observed gross margin behavior, not contract disclosure. A leaked or disclosed contract structure (cost-plus markup percentage, NRE recoupment schedule, ASP escalator clauses) would either confirm the conviction-4 supply-chain read at 5 or knock it down to 3. Highest-leverage research question.
2. Per-program AI revenue / unit volume disclosure or estimate. AVGO does not break out per-customer AI revenue. If the bear case is "loss of Google = 20–30% multiple re-rate," the position-sizing correctly should adjust to actual program-level concentration. A clean Hashrate Index-style triangulation is the missing input.
3. Hock Tan succession signal. The single largest under-discussed risk to the model. Any public commentary, board action, or compensation-plan modification that shifts the transition timeline materially affects sizing.
4. NVLink Fusion ramp pace at MediaTek / Alchip / Marvell. Quarterly tracking of which programs (ByteDance, Tier-2 hyperscalers, sovereign AI) commit to NVLink Fusion vs. UALink/Ultra Ethernet. The 2027–2028 disruption risk hinges on this.
5. VMware ELA renewal cadence post-2027. The "harvesting → maintenance" inflection is the bear's anchor for software margin. A sample of Top-100 ELA renewal terms (price, length, displacement vs. Nutanix/Microsoft) would either de-risk or accelerate the headwind.
6. Position sizing relative to cohort. Given the cohort is heavy on Taiwan-exposed AI longs (NVDA large, TSM large), should AVGO be sized as a replacement for some of the NVDA exposure (it captures the counter-leverage trade) or as a complement? The two are partial substitutes when hyperscaler ASIC is the share-shift mechanism. Worth a separate cohort-level sizing memo.

§ 11Cross-References

• Cohort synthesis: ../synthesis.md
• Analyst memos: competitor.md, supply-chain.md, customer.md, financial.md, market.md, regulatory.md, macro.md