Broadcom Inc. (NASDAQ: AVGO) — The Backlog Showed Up; The Price Already Knew

Analytical framework: competitive advantage / capital allocation / capital cycle (Greenwald; Marathon) Company: Broadcom Inc., NASDAQ: AVGO | GICS: Information Technology — Semiconductors & Infrastructure Software Fiscal year: ends early November (FY2025 ended 2025-11-02) | Latest 10-K: filed 2025-12-18 | Latest 10-Q: Q2 FY2026, quarter ended 2026-05-03, filed 2026-06-09 Price reference: $420.565 (2026-08-06 close) | Market cap ~$2.00T | EV ~$2.05T | Diluted shares ~4.88B Date: 2026-08-07 | Type: update — supersedes my 2026-06-07 note on the same company Next catalyst: Q3 FY2026 results, 2026-09-02 (after close)

General information only, not investment advice. The analytical body of this article (sections 1–15) deliberately takes no position and carries no price target; it discusses valuation only as embedded expectations and scenarios. The single, deliberate exception is the opinion block immediately below, which is my own subjective view.

⚡ Claude’s Take

This block is my own subjective opinion. It is general information, not investment advice, and you should do your own work before acting on anything here. The analysis in sections 1–15 below deliberately takes no position and sets no price target.

Verdict: HOLD — still not chasing at ~$421, but the accumulation zone moves up, and my prior zone was too low. The defensible accumulation band is now roughly $300–345 (~17–20x my modelled FY2027 non-GAAP EPS, and the region between the rising 200-day EMA at ~$361 and the February-2026 low near ~$318); the base-case fair-value zone is ~$340–400; a genuine bull outcome — FY2027 AI revenue landing comfortably above $110B with FY2028 visibility intact — supports ~$480–540. In June I said I’d accumulate at $250–300. That was wrong, and I am saying so plainly: it under-weighted the possibility that the backlog was real and contractual rather than rhetorical. It was. The call itself does not change — I still will not pay the 90th percentile of this company’s own ten-year valuation history — but the price at which I stop being a spectator is materially higher.

Tag: “The backlog showed up; the price already knew.”

What changed. Two days after I published in June, Broadcom filed its Q2 FY2026 10-Q, and buried in the revenue note was the number that reframes the entire debate: firmly-committed, non-cancellable remaining performance obligations of ~$164.6 billion, up from $45.0B one quarter earlier and $33.3B at fiscal year-end — a +$119.6B, +266% sequential step-up, attributed by the filing to “a long-term contract for custom AI accelerators entered in the fiscal quarter.” My June memo was arguing against a verbal “~$73B backlog.” It is now an audited, contractually-committed $164.6B, roughly 30% of which converts inside twelve months. The single largest weakness in the bull case — that FY2027’s “>$100B” was a management hypothesis resting on multi-gigawatt letters of intent — has been substantially retired. Alongside it: Apple extended to 2031 including custom ASICs (retiring the biggest named wireless risk and adding a prospective seventh custom-silicon customer); OpenAI’s “Jalapeño” went from framework to reticle-sized shipping silicon in nine months; and a >$200B Samsung memorandum on HBM4/HBM4E, 2nm and advanced packaging attacks the exact TSMC-chokepoint vulnerability I flagged.

Why it’s still a HOLD. The market got there first. At ~$421 the stock trades at ~49x EV/EBITDA, ~27x EV/revenue and ~70x trailing earnings — higher absolute multiples than in June — and at the 90.2nd percentile of its own ten-year valuation range (P/B 97th, P/S 96th), essentially unmoved because the fundamentals grew into a 9% price rise. My reverse-DCF still requires a ~20% free-cash-flow CAGR sustained for a decade to justify today’s enterprise value at a 9% discount rate. The RPO underwrites roughly the first three years of that path with contracts; years four through ten remain an extrapolation of the most concentrated demand base in large-cap technology — and the 10-Q shows concentration got worse, not better: top-five end customers moved from ~40% to ~45% of revenue, one distributor to 42%. Meanwhile the European Commission’s VMware inquiry escalated from complaints into compulsory process that Broadcom just lost a court fight to resist; the infrastructure-software segment decelerated to +9% in Q2 before a guided +31% Q3 step-up it still has to deliver; and the company quietly stopped spending ~$3.8B a half in cash to net-settle vesting RSUs, moving that cost off its own cash-flow statement and onto shareholders as market share-supply.

The framing, evidence-based. This is not a momentum trade and not a falling knife. The factor model is unambiguous: AVGO carries a Growth loading of +2.42 and a Semiconductors-industry loading of −0.02 — effectively zero. The market does not trade this as a chip cyclical; it trades it as a high-beta (2.06) mega-cap growth vehicle. Momentum loading is only +0.18 despite +44.6% twelve-month relative strength, and the last quarter annualizes to +8.7% with a −25% intra-quarter drawdown. That is a high-beta consolidation inside an intact uptrend — price above a rising 200-EMA, 12.5% off the June all-time high of $480.81, having round-tripped $480 → $360 → $421 in nine weeks. You are not being paid a crowding premium to sell, and you are not catching a knife. You are simply paying full price for a franchise whose next three years are now contracted and whose next seven are not.

Conviction: medium (unchanged). The single piece of evidence that flips me bullish: an increase to the FY2027 AI number — >$110B guided on the 2026-09-02 call — with RPO holding above ~$150B after a quarter of ~$16B in AI shipments burns off, proving the backlog replenishes rather than merely front-loads. The single piece that flips me bearish: RPO falling materially in Q3 (backlog draining, not building), or a guided AI cut, or the European Commission opening a formal Statement of Objections that puts VMware pricing under remedy — the one event that would attack the 79%-margin annuity funding everything else.

📈 Stock Price Action — Five-Year Event Map

Broadcom has compounded roughly ten-fold in five years — from ~$44 (split-adjusted) in August 2021 to an all-time high of $480.81 on 2026-06-02, closing at $420.565 on 2026-08-06, −12.5% off that high. The 52-week range is $287.49–$480.81; the 200-day EMA sits at $360.96 and is rising, the 21- and 50-day EMAs are clustered at ~$390. Since the prior report’s $385.12 reference the stock is +9.2%, having first fallen to $360.45 (2026-07-02) and then rallied ~17% into August. (FACT — AZI adjusted price history, full series to 2026-08-06. Price moves are FACT; attributed drivers are INTERPRETATION.)

Cycle narrative. (1–2) The 2021–22 drawdown was macro, not idiosyncratic — a rate-driven multiple compression across semiconductors, compounded by widespread scepticism that a chip company should pay ~$61B (announced) for enterprise virtualization software. (3) VMware closed on 2023-11-22 and the market began to mark the repricing thesis to market; the July-2024 10-for-1 split coincided with, but did not cause, the re-rating. (4) The 2024-12-13 session is still the largest single-day move in the five-year record: management first framed custom XPUs as a multi-customer, multi-year, multi-gigawatt franchise rather than a single Google program. (5) 2025 delivered the two cleanest tests of how fast this stock de-rates on a narrative shock — DeepSeek’s efficiency claims (−17.4% in a day) and the April tariff round (−10.5%, then +18.7% six sessions later) — both fully retraced, which is precisely why the tape rewards patience rather than panic here. (6) The fourteen months to June 2026 were a near-unbroken re-rating on contracted AI wins: OpenAI (Oct-2025), Anthropic and Meta and the Google long-term agreement (Apr-2026, 8-K 2026-04-06). (7) The June-4 session is the single most informative price event in the file: a company that beat on revenue, margin and AI revenue lost ~$280B of value because it declined to raise. That is the definition of a price that has moved ahead of even exceptional delivery. (8) The two months since have been a violent range — a ~25% peak-to-trough drawdown inside a quarter that nonetheless annualizes to a merely positive +8.7% — with the August rebound driven by sector-wide hyperscaler-capex enthusiasm rather than anything Broadcom-specific (the Apple and Samsung announcements landed on 2026-07-06 and 2026-07-25, both inside the drawdown, not at its end).

0.5 Changes Since 2026-06-07

Nine weeks. No new quarter has printed — Q3 FY2026 reports 2026-09-02, after this memo — so the update is filings, contracts, regulation, insiders and tape. Eight items moved the thesis; three of them matter a great deal.

What was confirmed. The prior memo’s core reads all held: the two-engine structure, the exceptional segment economics (Semiconductor Solutions operating margin rose to 61.8% in Q2 from 57.2%; Infrastructure Software to 78.7% from 75.6%), the ~1%-of-revenue capex intensity, and the disciplined capital-allocation cadence.

What was falsified — including by me. My June accumulation zone of $250–300 was too low; it priced a probability of “the backlog is rhetoric” that the 10-Q substantially eliminated. Separately, the prior memo carried the Q2-call framing that a reiterated $56B FY2026 AI number implied a sequentially declining Q4. My own arithmetic says otherwise: H1 FY2026 AI was ~$19B (management’s figure on the call), Q3 is guided to $16B, so Q4 implies ~$21B — up ~31% sequentially, not down. The fear that helped produce the June-4 air-pocket does not survive the arithmetic. Finally, the prior memo named the incoming CFO “Amy Teiner,” a transcription artefact; the filings name Amie Thuener (O’Toole), formerly Alphabet’s VP, Corporate Controller and Chief Accounting Officer.

What is still open. The RPO is not disaggregated by customer or segment; the Samsung MOU has no disclosed binding terms; the Apollo/Blackstone ~$35B AI-XPU financing vehicle still has no definitive structure in a filing; and the guided +31% Q3 software acceleration off a +9% Q2 base is unproven.

1. Executive Summary

Broadcom is a ~$106B-forward-revenue (FY2026E, on management’s own guidance), ~$2.00T-market-cap designer of semiconductors and operator of infrastructure software, sitting at the centre of the AI data-centre build-out on the silicon side and running one of the world’s largest enterprise-software franchises on the other. It is fabless in chips and a serial roll-up in software (CA 2018, Symantec 2019, VMware 2023). Trailing twelve months through Q2 FY2026: revenue $75,465M, EBITDA $41,499M (55.0% margin), GAAP net income $29,317M, diluted EPS $6.006. Q2 alone delivered a 69.5% gross margin, a 48.9% GAAP operating margin, and $10,262M of free cash flow on $231M of capex.

The decisive new fact is contractual, not operational. The Q2 FY2026 10-Q discloses firmly-committed, non-cancellable remaining performance obligations of ~$164.6 billion — against $45.0B a quarter earlier, $33.3B at FY2025 year-end and $27.5B at Q3 FY2025. The filing attributes the step-up to “a long-term contract for custom AI accelerators entered in the fiscal quarter,” and expects ~30% (~$49B) to convert to revenue within twelve months. This is audited, customer-committed backlog, and it is roughly 2.2x trailing revenue. It substantially answers the question the prior memo left open — whether the FY2027 “>$100B” AI figure was underwritten or aspirational.

The moat portfolio is unchanged in structure but measurably wider at the edges. The VMware/CA customer-captivity annuity remains the strongest and most-proven moat (78.7% segment operating margin in Q2). The merchant-Ethernet scale franchise remains strong. The custom-XPU co-design platform — the narrowest and most concentrated moat, and the one carrying the valuation — was reinforced three ways in the interval: OpenAI’s Jalapeño went from concept to production silicon in nine months (a cadence no competitor has demonstrated), Apple signed through 2031 including custom ASICs, and a >$200B Samsung memorandum offers a second leading-edge foundry plus captive HBM4 supply against the TSMC chokepoint.

But the concentration got worse and the regulatory risk got sharper. Top-five end customers rose to ~45% of revenue (from 40%), with a single distributor at 42%. The European Commission’s VMware competition inquiry escalated: on 2026-08-03 the EU General Court refused to suspend a compulsory document demand, rejecting Broadcom’s US privilege objection. Infrastructure software grew only +9% in Q2 (+5% in H1) before a guided +31% Q3. And Broadcom stopped paying cash (~$3,802M in H1 FY2025) to net-settle vesting RSUs, instead selling vested shares into the market — a change that flatters cash returns while pushing the cost onto shareholders.

Capital allocation returned to offence. H1 FY2026 buybacks were $8,450M (versus $2,450M in all of FY2025), with $10.1B still authorized; dividends were $6,178M; and in June the company tendered for six senior-note series, upsizing the cap from $2.5B to $3.0B and retiring ~$2.9B of principal, with the two longest, highest-coupon series bought below par ($982.01 and $970.29 per $1,000).

Valuation remains the crux, and it has become more expensive in absolute terms while better-evidenced in substance. At ~$2.05T EV the stock trades at ~49.4x EV/EBITDA, ~27.1x EV/revenue and ~70x trailing GAAP earnings — versus ~44.7x/~24.9x in June — and at the 90.2nd percentile of its own ten-year valuation history. A reverse-DCF implies a ~20% FCF CAGR for a decade. Yet on management’s own FY2027 guidance the same price is roughly 24x forward non-GAAP earnings — the entire debate compressed into whether you underwrite the guide. Low business risk, high security risk. The fulcrum has moved from “is the backlog real?” to “does the backlog replenish?”

2. Business Overview

Structure. Two reportable segments, and the mix has shifted violently in twelve months. In Q2 FY2026, Semiconductor Solutions was $15,009M (68% of revenue) and Infrastructure Software $7,178M (32%) — versus 56%/44% a year earlier (FACT, 10-Q filed 2026-06-09, Note 9). The prior memo described a “58/42” company. Twelve months of AI scaling has made it a 68/32 company, and the guided Q3 (semiconductors ~$20.5B of ~$29.4B) makes it ~70/30. This matters analytically: the recurring, high-moat software ballast that made Broadcom structurally steadier than a pure-play AI-silicon name is becoming a smaller fraction of the whole even as it grows.

(FACT — 10-Q, Note 9 and MD&A segment tables.)

Semiconductor Solutions — five end markets, one engine. Networking Connectivity houses both AI franchises: custom silicon / XPUs (accelerators co-designed with hyperscalers and frontier labs) and AI networking (Tomahawk switching, Jericho routing, NICs, PHYs/SerDes, optics, co-packaged optics). Management disclosed that networking was ~40% of Q2 AI revenue and guided that share down toward ~30% as XPU volume scales — relevant because networking carries the richer margin. The other four markets — Wireless Device Connectivity (overwhelmingly Apple), Servers & Storage, Broadband, and Industrial — collectively produced $4.2B of non-AI semiconductor revenue in Q2 (+6%) against bookings of >$6B, guided to ~$4.5B (+12%) in Q3, which management characterises as “the path towards a full cyclical recovery” (FACT — Q2 FY2026 earnings call, 2026-06-03).

The Apple relationship was materially extended and re-scoped. On 2026-07-06 Broadcom filed an 8-K disclosing that it and Apple “have agreed to expand their long-standing technology collaboration through 2031 by entering into new multi-year long-term agreements for Broadcom to develop and supply a range of custom ASIC silicon products for use in multiple generations of Apple products” (FACT — 8-K, Item 8.01, 2026-07-06). Press reporting ties the ASIC element to Apple’s “Baltra” AI server programme and puts Apple at ~20% of Broadcom revenue (Bloomberg, 2026-07-06). INTERPRETATION: this does two distinct things. It removes, for five years, the single most-cited bear point on the legacy wireless franchise — Apple insourcing RF/connectivity — and it promotes Apple from a mature, cyclical cash-cow customer to a prospective seventh custom-silicon engagement. The share price rose ~5.3% on the disclosure.

AI revenue — the scale and the shape. FY2025 AI revenue was ~$20B. FY2026 is guided to ~$56B (+~180%); Q2 FY2026 AI semiconductor revenue was a record $10,800M (+143% YoY, ~49% of total); Q3 is guided to ~$16.0B (+>200%). Management reiterated FY2027 at “>$100B” and said FY2028 will be “a substantial growth from what we are forecasting in 2027” (FACT — Q2 FY2026 call). INTERPRETATION on the internal arithmetic: H1 FY2026 AI was ~$19B; $56B less $19B less the $16B Q3 guide leaves ~$21B implied for Q4 — up ~31% sequentially. The widely-repeated concern that the reiterated full-year number implied a down Q4 does not survive the arithmetic; it implies a materially up Q4, which is why the “reiterate, don’t raise” posture reads as conservatism rather than deceleration.

Custom-silicon customer set. Six core customers, now with more specificity than in June: Google (TPU, multi-generation long-term agreement announced April 2026), Anthropic (>1GW of Broadcom TPU-based compute in 2026, a further 5GW of next-generation compute from 2027), OpenAI (silicon delivered, production late 2026; contractual commitment to deploy 1.3GW in 2027 inside a 10GW-by-2029 agreement), Meta (multiple generations of MTIA X XPUs; 3GW through end-2028, with an initial 1GW order received, deliveries from H2 2027), plus two additional customers shipping from late 2026 with $6B of purchase orders already received (FACT — Q2 FY2026 call). Management plans to ship ~10 gigawatts in 2027, back-half weighted (FACT — call Q&A, Rasgon).

Infrastructure Software — five portfolios. Private Cloud (VMware/VCF), Mainframe (CA), Cybersecurity (Symantec/Carbon Black), Enterprise Software, and FC SAN (Brocade). VMware Cloud Foundation 9.1 shipped in the quarter, adding heterogeneous GPU/CPU support across AMD, Intel and NVIDIA platforms to run AI inferencing on-premises alongside traditional virtualized workloads. Management asserts ARR growth of ~17% and guides Q3 software to ~$8.9B, +31% YoY (FACT — call). INTERPRETATION: the reported Q2 growth was +9% and H1 +5%. Getting from +9% to +31% in one quarter is a very large step, and while the mechanics are explicable (renewal timing, up-front recognition on multi-year VCF subscription conversions), the burden of proof sits squarely on the 2026-09-02 print.

Geographic mix. Q2 FY2026 revenue by region: Asia-Pacific $14,535M (66%), Americas $5,202M (23%), EMEA $2,450M (11%) — versus APAC $8,100M (54%) a year earlier (FACT — 10-Q). INTERPRETATION: the APAC surge is a shipment/title-transfer artefact of the XPU ramp (Broadcom recognises most product revenue on title transfer in Penang, Malaysia) rather than a statement about end-demand geography, but it does concentrate logistical and trade-policy exposure.

Customer concentration — worse. Top-five end customers were ~45% of net revenue for both the quarter and half ended 2026-05-03, versus ~40% in the prior-year periods; one direct customer, a distributor, accounted for 42% of net revenue versus 29% (FACT — 10-Q, MD&A). The 10-K’s own language applies: “The loss of, or significant decrease in demand from, any of our top five end customers could…” INTERPRETATION: this is the structural cost of the AI ramp. Broadcom is becoming more profitable and more fragile at the same time, and the fragility is now quantified and rising.

Verdict (Business Overview). A higher-quality, higher-margin, and more concentrated franchise than three months ago. The software annuity that justified the “not really a semiconductor company” framing is now only 32% of revenue and falling as a share; the AI silicon engine is now ~49% of revenue and rising. The company has simultaneously bought itself real supply-chain optionality (Samsung), a five-year customer extension (Apple), and ~$164.6B of contractual visibility — while allowing its top-five revenue concentration to rise five points. Both things are true, and an honest reading must hold them together.

3. Industry Dynamics

Broadcom straddles two industries with opposite structural characters. The prior memo’s structural analysis stands; three developments in the interval refine it.

A. Semiconductors — the AI accelerator capital cycle

Position in the value chain, and a new second source. Broadcom occupies the fabless-designer rung: capital-light (capex ~1.0% of revenue), IP-intensive, and historically dependent on TSMC for leading-edge fabrication and CoWoS advanced packaging — competing for the same constrained capacity as Nvidia, AMD and every hyperscaler ASIC programme. The prior memo named this the single most under-priced structural dependency.

The Samsung memorandum attacks it directly. On 2026-07-25 at the San Francisco AI Summit, Samsung Electronics and Broadcom signed a memorandum of understanding valued at more than $200 billion through 2030, spanning (a) Samsung supply of next-generation HBM4 and HBM4E to Broadcom, (b) manufacture of Broadcom products on Samsung’s 2nm-and-below processes at Pyeongtaek, and © advanced packaging integrating memory and logic (FACT — Samsung Global Newsroom, 2026-07-25; CNBC, 2026-07-25). The arrangement forms part of a broader ~$950B Korea–US semiconductor cooperation package. INTERPRETATION: if it converts, this is the most consequential structural development of the interval. It gives Broadcom (i) a second credible leading-edge logic source, (ii) captive HBM allocation in the tightest input market in the industry — cross-read prior Micron coverage, where HBM4 pricing and allocation are the defining variable of the memory cycle — and (iii) a bundled memory+logic+packaging offer that a pure-play foundry structurally cannot match. Discount heavily: it is an MOU, not a binding contract; no volume, price or take-or-pay terms are disclosed; Samsung’s leading-edge yield record is unproven at 2nm; and this is a supplier-side arrangement that must never be conflated with the customer-side RPO. It reduces a chokepoint risk; it does not add a dollar of revenue.

The capital cycle (Marathon lens). Unchanged and still the master risk. Broadcom carries almost no capacity risk directly — it is a fabless arms dealer with ~$231M of quarterly capex. Its exposure is that its customers over-build. The Marathon caution is that high returns attract capital, capacity overshoots, and order rates mean-revert violently in project-lumpy businesses. What the interval added is a partial defence: ~$164.6B of firmly-committed, non-cancellable obligations convert some of that order flow from “revocable demand” into “contractual demand.” That is a genuine, if partial, insulation — the contracts do not eliminate the cycle, they change who bears the first loss when it turns.

The merchant-GPU versus custom-ASIC split. The structural logic is intact: for known, stable, high-volume workloads (scaled inference, ranking, recommendation) a purpose-built XPU wins on performance-per-watt and per-dollar; for frontier training and fast-moving architectures, merchant GPUs win on flexibility and ecosystem. The interval’s evidence favours the custom side. OpenAI and Broadcom unveiled “Jalapeño” on 2026-06-24 — OpenAI’s first custom LLM-inference processor, a reticle-sized ASIC taken from design to production in roughly nine months with Celestica as manufacturing partner, with engineering samples already running production workloads (including GPT-5.3-Codex-Spark) at intended frequency and power, and initial deployment planned by end-2026 through data-centre partners including Microsoft (FACT — Broadcom IR release and OpenAI, 2026-06-24; CNBC and Tom’s Hardware same date). INTERPRETATION: the nine-month cycle is the story. That cadence collapses the classic objection to custom silicon — that fixed-function ASICs cannot keep pace with model architecture — and it is itself the clearest available evidence for the strength of Broadcom’s co-design platform.

Regulation. Two live vectors. (1) China / export controls: the 10-K risk factors remain, and the interval added reporting that a US push to exclude Chinese components from the domestic AI supply chain could raise cost and complexity for optical transceivers and silicon photonics — a Broadcom-relevant input (Benzinga, 2026-08-05). This is press-reported policy risk, not enacted rule; treat as OPEN QUESTION. (2) Trade/tariffs: unchanged as a demand risk.

Verdict (Semiconductors): Structurally attractive and now better insulated on the supply side than in June, but still the textbook cyclical industry in the boom phase of a capital cycle. Good industry, late-cycle posture, with a contractual buffer that is new and real but does not repeal the cycle.

B. Infrastructure / Private-Cloud Software

Structure. Unchanged: an oligopoly-to-monopoly at the core. VMware has no at-scale on-premises equal; the alternatives (public cloud, Nutanix, Red Hat/OpenShift, KVM) each carry multi-year migration friction. Mainframe (CA) and FC SAN (Brocade) are mature near-monopoly niches with captive bases. These are the “toll-road” franchises Broadcom targets, and the value is in margin and pricing power, not unit growth.

The regulatory picture materially worsened. The prior memo characterised VMware regulatory exposure as complaints (EU CISPE alleging 800–1,500% increases) plus private litigation (AT&T, settled December 2024; UnitedHealth, filed April 2025). It is now a live European Commission competition investigation with compulsory process. On 2026-08-03 the EU General Court refused to suspend a Commission request for information requiring Broadcom to produce documents created outside the EU, rejecting Broadcom’s argument that US attorney-client privilege shielded them — the court holding that the Commission has authority to determine what information its investigation requires (FACT — Reuters, 2026-08-03; Concurrences case note, August 2026). The divergence is structural: US law broadly protects confidential lawyer-client communications, while EU privilege extends only to external counsel, not in-house. INTERPRETATION: two consequences. First, procedurally, Broadcom has lost its shield and must hand over internal legal communications about VMware pricing and licensing strategy to a regulator actively investigating it. Second, substantively, this is the highest-probability path by which the 78.7%-margin software annuity — which funds the dividend, the buyback and the deleveraging — comes under externally-imposed pricing constraint. Important limits: the ruling addressed interim relief only, did not decide the merits of Broadcom’s challenge, and the Commission has neither issued a Statement of Objections nor concluded. It is an escalation of process, not yet of remedy.

Verdict (Software): Structurally excellent markets that Broadcom monetises with rare pricing power — but the pricing power is now being examined by a regulator with compulsory powers and a demonstrated willingness to reach across the Atlantic for evidence. The long-run question is no longer only “does the price-shocked base churn?” but “will Broadcom be permitted to keep pricing this way in its second-largest region?”

Overall Industry Verdict: A favourable blend that improved on the semiconductor side (supply diversification, contractual demand, custom-ASIC validation) and deteriorated on the software side (regulatory escalation), with the overall weight shifting toward the cyclical engine as software falls to 32% of revenue. The dominant industry risk is unchanged: the AI capital cycle turning.

4. Competitive Position

The question is whether a durable advantage shows up in financial outcomes that would deteriorate without it. It does — and it remains a portfolio of three moats of unequal strength, with the market still paying for the narrowest. Two of the three were reinforced in the interval; one was attacked.

Moat 1 — Custom-silicon co-design platform (proprietary intangibles + customer captivity). Strengthened.

Mechanism. Broadcom does not sell a finished accelerator; it sells a co-design platform — the industry’s highest-speed SerDes IP (200G and 400G shipping, enabling direct-attach copper for scale-up within racks), HBM integration, IP cores, physical design and advanced-packaging expertise — letting a customer build its own XPU to spec. In Greenwald’s taxonomy this is proprietary intangibles fused with customer captivity: after three-plus generations of co-design, switching means redoing physical design, requalifying IP, retaping-out at tens of millions per mask set, and absorbing yield-ramp risk on a multi-billion-dollar programme.

New evidence for the mechanism. The nine-month OpenAI Jalapeño design-to-production cycle is the strongest single datapoint the file has ever contained for this moat. Building a reticle-sized inference ASIC in nine months is not a procurement outcome; it is the output of a platform where the hard, slow parts — SerDes, packaging, physical design, HBM integration — are already solved and reusable. Marvell has demonstrated nothing comparable. The Apple extension to 2031 including custom ASICs is the second datapoint: the most engineering-capable, most vertically-integrated customer in consumer electronics, one that has systematically insourced silicon for a decade, chose to extend rather than displace.

Pressure-test (where it remains weak). Three vulnerabilities persist and one worsened. (1) Concentration — worse. Top-five end customers moved to ~45% of revenue; the per-customer AI split remains undisclosed; Tan himself conceded on the Q2 call that Google will see “some diversity of sources.” (2) Merchant-GPU substitution. If frontier architectures keep moving, hyperscalers may keep buying flexible GPUs — though Jalapeño’s nine-month cadence is direct evidence against this. (3) The customer is also the competitor’s customer. Hyperscalers play Broadcom against Marvell and can credibly threaten to insource design; Amazon’s Annapurna already does more in-house. INTERPRETATION: real, currently lucrative, and now better-evidenced — but still the most concentrated of the three, and still the moat that the valuation is underwriting.

Moat 2 — Networking-silicon scale (economies of scale + intangibles). Unchanged, strong.

Mechanism. Tomahawk/Jericho switching and routing, NICs, PHYs, DSPs and optics rest on decades of merchant-switch volume amortizing an R&D base no smaller rival can match, plus proprietary high-speed SerDes/DSP IP, with Broadcom as the open-Ethernet standard-bearer against proprietary fabrics. Management’s Q2 position statement: shipping the industry’s only 100-terabit Ethernet switch (Tomahawk 6) for over a year, taping out a 200-terabit successor this quarter, Jericho 3/4 for cross-data-centre fabric, and “de facto standard” status in 1.6T DSPs, CW and EML lasers for co-packaged optics (FACT — Q2 FY2026 call). Networking was ~40% of Q2 AI revenue, guided toward ~30% as XPU volume scales.

Pressure-test. The threat is unchanged: Nvidia bundling NVLink and Spectrum-X into full rack systems and capturing networking content, plus Marvell and others in optics/DSP. Broadcom’s explicit refusal to move up-stack — Tan on the Q2 call: “No racks. It is all chip business only. We only do chips” — is a deliberate choice to stay in the highest-return, lowest-capital rung and cede systems economics. INTERPRETATION: wider and more durable than Moat 1 — multi-customer, scaling with cluster size, with a genuine one-generation technology lead — and it is the part of the AI franchise carrying the richest margin.

Moat 3 — VMware / infrastructure-software installed base (customer captivity). Attacked.

Mechanism. The textbook Greenwald customer-captivity moat with the cleanest financial proof: VMware runs the mission-critical virtualization layer of most of the Fortune 500; migrating off it is a multi-year, high-risk re-platforming; so when Broadcom moved to subscription-only, bundled into VCF/VVF and raised prices dramatically, the base largely paid. The proof is in the margin: 78.7% segment operating margin in Q2 FY2026, up 310bp YoY.

What changed. Two things, in opposite directions. Positive: VCF 9.1’s heterogeneous GPU/CPU support is a genuine product extension into on-premises AI inferencing, and management claims strong server demand is accelerating VMware growth — asked directly whether generative AI was hurting software renewals, Tan said “we are not seeing it… it is driving some accelerated growth of our VMware business” (FACT — Q2 call). Negative and more important: the European Commission investigation escalated (section 3B), and reported growth decelerated to +9% in Q2. INTERPRETATION: the moat’s mechanism is intact — nothing has demonstrated that customers can leave. What is now in question is whether Broadcom will be permitted to keep exercising it at this intensity in Europe. That is a different, and for a regulated-outcome investor a worse, kind of risk than churn: churn is gradual and observable; a remedy is discrete and imposed.

Direct competitor comparison (updated)

Sources: 10-Q 2026-06-09; 8-K 2026-07-06; Samsung Global Newsroom 2026-07-25; Q2 FY2026 call; prior notes on NVDA (2026-06-09) and MRVL (2026-06-10) for peer multiples and framing. Competitor figures are external estimates or prior-report values, not re-derived this session.

INTERPRETATION: Against Nvidia, Broadcom remains the structural beneficiary of hyperscalers diversifying away from merchant GPUs — and Jalapeño is the sharpest expression of that yet, since OpenAI is Nvidia’s largest customer building its own inference silicon with Broadcom. Against Marvell, the gap widened: Broadcom now discloses $164.6B of committed backlog, holds Apple through 2031, and trades at the 90th percentile of its own history versus Marvell’s 97th — Broadcom is both the better business and, relative to its own record, the less extended security.

Verdict (Competitive Position)

Durable advantage — yes, and the evidentiary base improved. Remove the VMware captivity and a 78.7%-margin, ~$28B-annualized software annuity disappears. Remove the SerDes/packaging/physical-design platform and neither a nine-month reticle-sized ASIC nor a five-year Apple ASIC commitment is possible. Remove networking scale and the open-Ethernet position erodes to proprietary fabric. All three pass the test that a moat must be tied to a financial outcome that would deteriorate without it. The honest framing is unchanged in shape but sharper in detail: a genuinely wide-moat software-plus-networking business with a narrower, more concentrated, now-contractually-underwritten AI-XPU franchise on top — and the valuation is still underwriting the franchise, not the fortress. What changed is that the franchise now has $164.6B of signatures behind it, and the fortress has a regulator inside the gate.

5. Growth History and Forward Opportunities

The trajectory, and the quality question

Reported revenue: $35,819M (FY23) → $51,574M (FY24, +44%) → $63,887M (FY25, +24%), with H1 FY2026 at $41,498M (+39%) and Q3 guided to ~$29.4B (+84%). On management’s guidance the full FY2026 lands near ~$106B, roughly +66% (INTERPRETATION — my build: H1 $41.5B + Q3 $29.4B + a Q4 of ~$35B comprising ~$21B AI, ~$4.7B non-AI semiconductor and ~$9.3B software).

The analytical asterisk from the prior memo stands: most of the FY23→FY24 step-up was bought, not earned (VMware closed 2023-11-22, one month into FY2024), and a meaningful slice of FY2025 software growth was price extraction from a captive base. What has changed is that this asterisk is now shrinking into irrelevance, because the incremental dollars are overwhelmingly organic AI silicon. Software is 32% of revenue and grew +9%; semiconductors are 68% and grew +79%.

The two engines, on two clocks

1. AI Semiconductors — the organic engine, now contract-backed. Q2 delivered $10.8B (+143% YoY), ~49% of total revenue, against >$30B of AI bookings in the quarter alone. Q3 is guided to ~$16.0B; FY2026 to ~$56B; FY2027 to “>$100B”; FY2028 to “substantial growth” from there. Management plans ~10 gigawatts of 2027 shipments, back-half weighted, and states that visibility now runs “all the way to 2028” versus 2027 three months earlier.

The evidentiary upgrade is the RPO series:

(FACT — 10-K filed 2025-12-18; 10-Qs filed 2025-09-10, 2026-03-11 and 2026-06-09, revenue notes. The Q2 filing states the balance “include[s] obligations under a long-term contract for custom AI accelerators entered in the fiscal quarter ended May 3, 2026.”)

INTERPRETATION: this is the most important table in the memo. These are contracts “where customers do not have termination rights” containing “firmly committed amounts” — not pipeline, not bookings, not letters of intent. At ~$164.6B the committed book is ~2.2x trailing twelve-month revenue and roughly 1.6x my FY2026 revenue build. The ~30%-in-twelve-months disclosure implies ~$49B converting by May 2027, which is broadly consistent with the guided AI trajectory plus software. It does not prove FY2027 exceeds $100B — the RPO spans both segments and multiple years, and is not disaggregated — but it moves the FY2027 number decisively out of “management hypothesis” and into “contractually supported.” The one thing it cannot tell us is whether the book replenishes. A single April contract created ~$120B of it; if Q3 shows RPO falling as shipments burn it off without new signings, the interpretation reverses from “durable annuity” to “one large front-loaded order.”

2. Infrastructure Software — repriced, and now decelerating before a promised re-acceleration. Q2 revenue $7,178M (+9%); H1 $13,974M (+5%); operating income +13% in Q2 on a 78.7% margin. Management asserts ARR growth of ~17% and guides Q3 to ~$8.9B (+31%). INTERPRETATION: the gap between +9% reported and +31% guided is the largest single unproven assertion in the guidance stack. The plausible mechanics are renewal-cohort timing and up-front licence recognition on multi-year VCF subscription conversions — both real, both lumpy. But a committee should note that the reported trend in the repricing engine is decelerating, that the repricing cycle mathematically cannot compound at these rates once it laps, and that the European Commission is now examining the pricing behaviour that produced it.

Forward opportunities

• Custom XPU — content per gigawatt rises even as gigawatts dominate. Tan’s framing on the Q2 call is worth quoting because it is the crux of the FY2027–28 model: dollars per gigawatt are “relatively stable” because higher-power chips mean fewer chips at higher ASPs, but “the number of gigawatts will keep accelerating,” and content per XPU rises generation-to-generation as chips add SRAM, embedded CPU cores, multi-die construction and more HBM. INTERPRETATION: this reframes the growth algorithm from unit growth to gigawatt growth times rising content — and it is why the FY2027 arithmetic (10GW at “$15–20B per gigawatt” historical framing) sits above “>$100B,” which is itself the reason the non-raise reads as conservatism.
• Apple as a seventh custom customer. The 2031 agreement explicitly covers custom ASICs across multiple product generations. Broadcom has not sized it. OPEN QUESTION — no revenue figure is disclosed in the 8-K.
• AI networking. Tomahawk 6 (100T) shipping, 200T taping out this quarter, Jericho 3/4 for cross-data-centre scale-out, 1.6T DSPs and CPO lasers. Guided from ~40% to ~30% of AI revenue — a mix headwind to gross margin, not a demand statement.
• VCF “Private AI.” Genuine optionality; unproven as a durable second software leg.
• Non-AI semiconductor cyclical recovery. $4.2B (+6%) in Q2 with >$6B of bookings, guided +12% in Q3. A modest tailwind, not a thesis driver.

Quality of growth — Verdict

Upgraded from “mixed-to-high” to “high, with concentrated tail risk.” In June I docked the grade principally because the FY2027 number was a hypothesis. It is now substantially contract-backed, and the software-acquisition asterisk is being diluted to irrelevance by organic AI dollars. The remaining deductions are two, and they are real: first, the growth is more concentrated than ever (top-five ~45%, one distributor 42%, per-customer AI split undisclosed), so the same contracts that create visibility also create a small number of counterparties whose capex decisions determine everything; second, the software engine — the ballast that makes the whole thing defensible — is decelerating and now faces a European regulatory examination of the very pricing that drove it. This is high-quality growth resting on a narrow base, which is a materially different thing from high-quality growth resting on a broad one.

6. Financial Quality

Verdict up front: Economics continue to improve dramatically with scale, and Q2 FY2026 is the cleanest evidence yet — a 48.9% GAAP operating margin, a 55.0% EBITDA margin and $10.3B of free cash flow on $231M of capex. Three qualifiers, one of them new: heavy reliance on non-GAAP that adds back ~$8B of intangible amortization and now ~$8.5B annualized of stock-based compensation; a still-large ~$65B debt load against ~$130B of acquired goodwill and intangibles; and — new this quarter — a change in how RSU withholding taxes are settled that improves the optics of cash returns while transferring the cost to shareholders.

Trailing-twelve-month and quarterly trajectory

(FACT — ROIC.ai quarterly series reconciled to the 10-Q filed 2026-06-09 and the 10-K filed 2025-12-18. Q4 FY2025 net income and EPS are flattered by a −$1,649M tax benefit; see below.)

The operating-leverage story is unambiguous. GAAP operating margin has expanded ten points in four quarters — 39.4% → 48.9% — while gross margin rose only 1.5 points. That is almost entirely operating leverage: consolidated operating expenses were $2.2B in Q2 (of which $1.6B R&D on a non-GAAP basis) against revenue that grew 48%. Segment-level, semiconductor operating margin rose 460bp to 61.8% and software 310bp to 78.7%. Management’s own instruction is worth heeding: “We highly recommend that investors model semiconductor and infrastructure software margins separately to properly reflect the impact of changes in total revenue mix” (Q2 call) — because consolidated gross margin is guided down to ~74% in Q3 purely on mix, with operating margin held flat at ~67% non-GAAP.

Quality of earnings — three items, one new and adverse

(1) Stock-based compensation is large, growing, and now settled differently. SBC was $2,092M in Q2 and $4,268M in H1 (+40% YoY), running ~$8.5–8.8B annualized (~11–12% of revenue). Unrecognized compensation cost is $20,106M over a 3.0-year weighted-average period — $4,078M in the FY2026 remainder, $7,235M in FY2027, $5,137M in FY2028 (FACT — 10-Q equity note). This is down from $23,833M at FY2025 year-end, but only because ~41M RSUs vested.

The new item, and it deserves attention. In H1 FY2025 Broadcom “paid $3,802 million in employee withholding taxes due upon the vesting of net settled equity awards and withheld 17 million shares.” In H1 FY2026 that line is zero: “In the current year fiscal period, we settled withholding taxes upon the vesting of employee equity awards using proceeds from the sale of a portion of the vested shares” (FACT — 10-Q, MD&A cash-flow discussion). INTERPRETATION: mechanically, Broadcom has stopped absorbing ~$3.8B per half-year of cash to buy back the shares it would otherwise have withheld, and instead lets employees sell vested shares into the open market to fund their own tax bills. The consequences are three: (i) financing cash outflow falls by ~$3.8B per half, flattering net cash returns and the debt-paydown/buyback capacity; (ii) share count is no longer reduced by ~17M shares per half through withholding, so effective dilution is higher than the prior run-rate implies; and (iii) roughly $3.8B per half of additional share supply hits the market. This is not accounting manipulation and it is fully disclosed — but any comparison of FY2026 to FY2025 “cash returned to shareholders” that ignores it is wrong, and it modestly degrades the quality of the FY2026 capital-return story. Flagged as a new quality-of-earnings item not present in the prior memo.

(2) Tax is noisy in both directions. Q4 FY2025 carried a −$1,649M tax benefit (a negative provision) that inflated that quarter’s net income and EPS; FY2025 as a whole benefited from a ~$397M net discrete benefit partly offset by a $1,321M CAMT valuation allowance following the July 2025 tax legislation. Management guides the non-GAAP tax rate to ~16% for Q3 and FY2026, up from FY2025 on global-minimum-tax and geographic-mix effects (FACT — Q2 call). INTERPRETATION: trailing GAAP EPS of $6.006 embeds a genuinely favourable tax year; normalise to ~16% before extrapolating.

(3) Working capital is building, appropriately but worth watching. Q2 saw a −$2,572M working-capital drag: receivables/unbilled −$2,370M and inventories −$1,366M. Days of inventory rose to 86 from 68 in Q1, which management attributes to securing supply for the H2 AI ramp. Current contract assets rose to $5,918M (from $5,005M at FY25-end) and contract liabilities to $10,038M current plus $4,204M long-term (from $9,469M/$3,547M) (FACT — 10-Q). INTERPRETATION: the inventory build is a deliberate, disclosed pre-position ahead of a guided doubling of AI shipments — the right decision if the ramp lands, a classic cyclical trap if it slips. The contract-asset growth is the expected signature of the VCF subscription transition (revenue recognised ahead of billing on non-cancellable software contracts), consistent with, not contradictory to, the RPO disclosure. Neither is a red flag today; both belong on the Q3 checklist.

Balance sheet and leverage

(FACT — 10-Q debt note and balance sheet; tender adjustment is INTERPRETATION applying the ~$2.9B principal accepted per the 8-K of 2026-06-18, and assumes no offsetting issuance before quarter-end.)

Deleveraging continued through the quarter and then accelerated in June. Against TTM EBITDA of $41.5B, gross leverage is ~1.6x and net ~1.1x (~1.07x pro-forma) — comfortable for a business generating ~$33B of annual free cash flow. Q2 interest expense was $776M, down from $807M a year earlier despite a larger balance sheet, and the June tender retired the two longest, highest-coupon series (4.926% due 2037 and 4.900% due 2038) at $982.01 and $970.29 per $1,000 — below par, generating a small extinguishment gain and removing duration at a discount.

The balance-sheet caveat is structural and unchanged: purchased intangibles dominate. Amortizable intangibles alone carry a remaining balance of $27,583M with weighted-average lives of 5 years (purchased technology), 6 (customer contracts), 11 (trade names). Combined with ~$97.8B of goodwill, roughly three-quarters of the asset base is purchase accounting against a trivial tangible plant.

Returns on capital

• ROE remains strong (FY2025 ~31%; the TTM figure is higher on rising net income against equity inflated by the $71.3B of VMware-issuance paid-in capital).
• ROIC including goodwill ~18–20% on TTM NOPAT (~$33.3B EBIT at a ~16% normalized rate ≈ $28B) over ~$140B of invested capital — improved from the ~15–17% in the prior memo, purely because earnings grew against a static acquired-capital base. INTERPRETATION: this is the M&A thesis proving out in the denominator. The ~$130B Broadcom paid to assemble these franchises is now earning a genuinely attractive return, and the gap between reported ROIC and the extraordinary return on tangible capital (only ~$2.5B of net PP&E) is narrowing as the acquired base amortizes.

Verdict — do economics improve with scale?

Emphatically yes, and more so than three months ago. Operating margin expanded ten points in four quarters on 1.5 points of gross-margin improvement — textbook operating leverage. FCF margin reached 46.3% in Q2 on 1.0% capex intensity. ROIC including goodwill has moved into the high teens. The honest qualifiers: non-GAAP optics overstate true economics by adding back ~$8.5B of annualized SBC with $20.1B more committed; trailing GAAP EPS embeds a favourable tax year against a guided ~16% forward rate; ~$47B of net debt, while shrinking, still consumes ~$3B a year of interest; and the switch in RSU-withholding settlement makes FY2026 cash returns look better than the underlying economics warrant. None breaks the thesis that this is an exceptional cash machine. All of them price it.

7. Capital Allocation

Verdict up front: capital allocation returned to offence in FY2026 after the deliberate FY2025 deleveraging year, and the execution has been good — but the mix of that offence, the scale of CEO compensation, and an off-balance-sheet AI-financing pivot that still has no definitive filed structure keep this from being an unqualified endorsement.

The Hock Tan playbook, and where it stands

Broadcom under Hock Tan is a private-equity-style roll-up wearing a semiconductor ticker: acquire a mature franchise with a sticky installed base, strip operating expense, reprice, harvest cash, deleverage, repeat. LSI (2014), Broadcom Corp (2016), Brocade (2017), CA (2018), Symantec enterprise (2019), VMware (2023, ~$84.2B actual consideration — $30,788M cash plus 544M shares at $53,398M fair value). No new acquisition was announced in the interval, which is itself notable: with the stock at 90th-percentile own-history valuation and ~$33B of annual FCF, management is choosing buybacks and debt reduction over M&A. Given where the equity trades as acquisition currency, that is the right call.

FY2026 capital returns — the restart, quantified

(FACT — 10-Q cash-flow statement and MD&A, filed 2026-06-09; FY2025 figures from the 10-K.)

Three reads. First, the buyback restarted decisively — $8,450M in one half against $2,450M in the whole of FY2025, with $10.1B of authorization still available at 2026-05-03. Since FY2025 year-end remaining authorization was $7,550M and $8,450M was spent, the board must have added roughly $11B of new authorization during FY2026. INTERPRETATION: buying back ~$8.5B of stock at an average price in the $300–450 range, at the 90th percentile of the company’s own valuation history, is the one capital-allocation decision this quarter I would challenge. Marathon’s discipline is that repurchases create value when the capital cycle is against you and destroy it when the tape is with you; Broadcom bought heavily into strength. It is not value-destructive at these growth rates — but it is not the counter-cyclical, high-conviction buyback that distinguishes the best allocators.

Second, the tender was excellent liability management. Announced 2026-06-11, priced and closed 2026-06-17 with settlement 2026-06-18/23: cash tenders across six senior-note series, the consideration cap upsized from $2.5B to $3.0B, ~$5.5B tendered and ~$2.9B accepted, with acceptance priority given to the 4.926% notes due 2037 and 4.900% notes due 2038 — the longest duration and highest coupons — purchased at $982.01 and $970.29 per $1,000 (FACT — 8-K 2026-06-18, EX-99.1 and EX-99.2). Retiring your most expensive, longest paper below par with excess cash is textbook.

Third, the dividend is steady and secondary. $0.65 per quarter ($2.60 annualized), ~0.62% yield at $420.57 — this is a compounding vehicle, not an income name, and the dividend is the stable leg while buybacks flex.

The Apollo / Blackstone AI-XPU platform — still the biggest open item

Tan described it on the Q2 call in his own words: “we are creating the AI XPU platform with Apollo and Blackstone and other leading investors to deploy more than 20 gigawatts of compute capacity… the first tranche of this platform valued at $35 billion is in fact currently being launched by Apollo.” Pressed on the model, he was explicit that Broadcom remains a chip vendor: “we are providing chips, technology in the form of chips… What we are doing to enable some of these LLM players to be able to get access to the volume of compute capacity… is we are creating in partnership with guys with the best balance sheets around a vehicle to basically have these chips funded for these LLM players who otherwise might have difficulty getting access to our technology.”

INTERPRETATION and status: no definitive structure has appeared in a Broadcom filing in the nine weeks since. The 10-Q filed 2026-06-09 discloses no related commitment, guarantee or variable-interest-entity consolidation. That absence is mildly reassuring on recourse — if Broadcom had assumed material obligations, disclosure would be required — but it leaves the central question open. The economic substance is vendor financing with extra steps: capital that Broadcom helped arrange, deployed by third-party investors, to buy Broadcom chips for customers who might not otherwise afford them. It is not fraudulent or unusual (Cisco and Lucent both did versions of this in the late 1990s, with instructive outcomes), but it does mean a slice of demand in the $164.6B RPO may ultimately be underwritten by private credit rather than by customer operating cash flow. This remains the single most important forward capital-allocation watch-item, exactly as in June, and it has not advanced.

Incentive alignment — read from the 2026 proxy (DEF 14A, filed 2026-03-02)

Unchanged from the prior memo’s read, and it remains the most analytically useful disclosure Broadcom makes. Hock Tan’s FY2025 total compensation was $205,278,006 against median employee compensation of $378,281 — a 543:1 ratio. Salary was $1.2M; stock awards $202.35M; cash bonus 0% (a 0% annual-plan target). Say-on-pay passed with 92% support in 2025.

• Annual plan (other NEOs) pays on revenue (threshold $54,429M / target $57,294M / max $60,159M) and adjusted non-GAAP operating income as a percentage of revenue (61.7% / 63.7% / 67.7%) — the correct metrics for this model, and note that Q2’s 67.3% non-GAAP operating margin sits at the maximum payout level.
• 2023 Tan PSU mega-grant (FY2023–27) vests on stock-price hurdles plus employment to 2027-10-31; tracking above target.
• 2025 Tan PSU (FY2028–30, extending to FY2030): 610,521 target shares, 0–300% payout on AI Revenue — ≤$60B = 0%, $90B = 100%, $105B = 200%, ≥$120B = 300%.

INTERPRETATION: the pay level is indefensible in absolute terms; the pay structure is genuinely aligned and, in the light of the $164.6B RPO, newly informative. Tan earns nothing on the 2025 grant unless AI revenue holds above $60B and earns the maximum only above $120B — thresholds that the FY2027 “>$100B” guide brackets almost exactly. The CEO’s personal economics and the central variant-perception variable for the stock are the same number. That is either excellent alignment or a powerful incentive to guide and sign aggressively, and a sceptical reader should hold both readings.

Insider behaviour — read precisely, not generically

The Form 4 corpus for 2026-06-11 → 2026-07-14, verified against the raw XML including the Rule 10b5-1 affirmation flag:

(FACT — EDGAR Form 4 filings, CIK 0001730168, parsed from source XML.)

INTERPRETATION: two signals worth naming precisely rather than dismissing as “routine.” The positive: Harry You’s 1,000-share open-market purchase at $373.57 is his third such buy in roughly ten months (September 2025, December 2025, June 2026) and is not plan-driven. The dollar amount is trivial (~$374k); the repetition is not. The negative, and it is the sharper one: the Chief Legal & Corporate Affairs Officer sold 75,000 shares for ~$29.2M across three discretionary, non-10b5-1 blocks in sixteen trading days, cutting his reported holding from 277,949 to 194,989 shares — a ~30% reduction in under a month, by the officer with the clearest visibility into the European Commission proceeding. Samueli’s ~$250M is plan-driven diversification from a ~66.5M-share position (~1% of his stake) and carries essentially no information. Neither insider datapoint is thesis-determinative. The Brazeal pattern belongs on the watch-list rather than in the conclusion.

Verdict (Capital Allocation)

Good, disciplined, and one notch below the prior memo’s “best-in-class.” The tender was excellent, the deleveraging is real, restraint on M&A at this equity valuation is correct, and the incentive structure ties the CEO’s pay to precisely the variable investors care about. But $8.45B of buyback executed into the 90th percentile of the company’s own valuation history is capital deployed with the cycle rather than against it; the RSU-withholding settlement change quietly improves reported cash returns at shareholders’ expense; and the Apollo/Blackstone vehicle — which injects vendor-financing circularity into the demand base — has gone nine weeks without a definitive filed structure. The playbook still works. It is being run slightly less conservatively than it was a year ago.

8. Changes and Headwinds — Last Two Years, and the Last Nine Weeks

The two-year arc (unchanged framing)

Over roughly twenty-four months Broadcom transformed from a diversified semiconductor-plus-software cash harvester into the most highly-valued pure-play on the AI capital cycle. VMware closed (2023-11-22) and was repriced to a ~77–79% segment operating margin; the EUC business was divested to KKR for $3.5B (July 2024); a 10-for-1 split executed (July 2024); AI revenue went from a sideline to ~49% of revenue; and multi-gigawatt customer agreements were signed with Google, Anthropic, Meta and OpenAI.

The last nine weeks — eight developments

1. The RPO disclosure (2026-06-09) — the most important. ~$164.6B of firmly-committed, non-cancellable obligations, up from $45.0B. Covered in section 5 and section 10. Strengthens the thesis materially.

2. OpenAI “Jalapeño” (2026-06-24). OpenAI’s first custom LLM-inference processor, co-designed with Broadcom, built by Celestica, taken from design to production in ~9 months; reticle-sized; engineering samples running production workloads at intended frequency and power; initial deployment planned by end-2026 through data-centre partners including Microsoft; intended to scale across generations to gigawatt scale (FACT — Broadcom IR release and OpenAI, 2026-06-24). Strengthens Moat 1.

3. Apple through 2031, including custom ASICs (8-K, 2026-07-06). New multi-year agreements for custom RF, Wi-Fi/Bluetooth and custom ASIC silicon across multiple Apple product generations; press reporting links the ASIC element to Apple’s “Baltra” AI server programme. Apple is ~20% of Broadcom revenue. Shares +~5.3%. Retires the largest named wireless risk; adds a prospective seventh custom customer.

4. Samsung MOU >$200B through 2030 (2026-07-25). HBM4/HBM4E supply, 2nm-and-below manufacturing at Pyeongtaek, advanced packaging. Attacks the TSMC/CoWoS chokepoint. MOU, not contract.

5. Customer concentration deteriorated (2026-06-09). Top-five end customers ~45% of revenue (from 40%); one distributor 42% (from 29%). Weakens the thesis.

6. EU antitrust escalation (2026-08-03). The EU General Court refused to suspend the European Commission’s compulsory document demand in the VMware-related competition inquiry, rejecting Broadcom’s US-privilege objection. Interim relief only; merits undecided; no Statement of Objections issued. Weakens the thesis — and attacks the highest-margin segment.

7. Debt tender, upsized (2026-06-11 → 2026-06-23). ~$2.9B of principal retired, longest and highest-coupon series bought below par. Modestly strengthens.

8. CFO transition executed (2026-06-12). Kirsten Spears retired after twelve years, remaining an adviser for nine months; Amie Thuener (O’Toole) became CFO, having served since 2018 as Alphabet’s VP, Corporate Controller and Chief Accounting Officer. INTERPRETATION: an orderly, pre-announced transition (announced April 2026, joined May 4), and the hire is interesting on two levels — deep technical accounting and controls experience at exactly the moment Broadcom’s revenue recognition is becoming more complex (multi-year non-cancellable XPU contracts, VCF subscription conversion, a $164.6B RPO), and a background inside Broadcom’s single largest AI customer. Tan’s stated rationale cites “financial reporting, corporate governance, AI-related transactions.” Key-person dependence on Tan remains the more important governance exposure; there are no written employment contracts and no key-person insurance per the 10-K.

Headwinds inventory

• AI-capex digestion — the master risk, partially buffered by contracted RPO but not eliminated.
• Customer concentration — quantified and rising; ~45% top-five, one distributor at 42%.
• VMware regulatory — European Commission investigation with compulsory process; private litigation (UnitedHealth) outstanding.
• Software deceleration — +9% reported in Q2 against a guided +31% in Q3.
• Mix-driven gross-margin compression — guided to ~74% in Q3 from 77.1% non-GAAP, structural as XPUs outgrow software and networking.
• China / export controls and optical-component policy — press-reported risk that a US ban on Chinese AI-supply-chain components raises transceiver and photonics cost.
• Taiwan / TSMC concentration — mitigated in intent by the Samsung MOU, not yet in fact.
• SBC and share supply — ~$8.5B annualized, $20.1B unrecognized, plus the withholding-settlement change adding ~$3.8B/half of market supply.
• Vendor-financing circularity — Apollo/Blackstone, still unstructured in filings.

Verdict: The interval strengthened the thesis on balance — contract-backed backlog, a five-year Apple extension, supply-chain diversification and shipping OpenAI silicon materially outweigh a regulatory escalation whose remedies are speculative and a concentration increase that was the arithmetic consequence of success. But the improvements are concentrated in the numerator of the story (demand, visibility, supply) while the deteriorations sit in the risk structure (concentration, regulation, share supply). The company is better; the risk profile is not simpler.

9. Risk Analysis

Risk verdict. The distribution shifted, it did not shrink. Business risk fell (contracted backlog, supply diversification, a five-year Apple extension, lower leverage). Security risk and structural risk rose (a higher multiple, five more points of customer concentration, a regulator with compulsory powers inside the software franchise, and a new source of share supply). The single dominant risk remains that this is a wonderful company at a price that requires a decade of near-flawless compounding — the risk of permanent loss to a buyer at $420 is far higher than the risk of impairment to the business.

10. Valuation Discussion — Embedded Expectations

No price target and no recommendation appears in this section. What follows is what the current price requires, and what evidence exists for and against it.

Where the stock trades (rebuilt at the 2026-08-06 close of $420.565)

(FACT — ROIC.ai enterprise-value and income-statement series reconciled to the 10-Q filed 2026-06-09; AZI valuation_index accessed 2026-08-06. Note the P/E percentile understates richness relative to P/B and P/S because trailing GAAP earnings are flattered by the Q4 FY2025 tax benefit; per the standard treatment for tax-distorted earnings, read P/B and P/S as the truer own-history signal.)

The most striking fact in this table is what did not change. The stock is 9% higher, the absolute multiples are 8–10% higher — and the own-history percentile is essentially flat at ~90th. Trailing fundamentals grew into the price almost exactly. This is a stock that is expensive against its own record and has been for a year, and which is currently neither cheapening nor further extending on that measure.

Forward multiples on management’s own guidance

The absolute trailing multiples are close to meaningless for a business guided to grow revenue ~66% this year and roughly 50% again next. The relevant question is what the price implies against the guided path.

(ASSUMPTION throughout: FY2027 non-AI semiconductor ~$18.5B and software ~$39B; interest expense ~$2.7–3.0B; a 16% tax rate per management’s guidance; ~4.94–4.95B non-GAAP diluted shares per management’s Q3 guidance. These are my constructions from disclosed guidance, not consensus estimates, and no consensus figure was obtainable from a primary source this session — OPEN QUESTION.)

INTERPRETATION — and this is the whole debate in one table. If management’s FY2027 guide is met, today’s price is roughly 24x forward earnings for a business compounding at 50%+ with 66% operating margins and 43% free-cash-flow margins. That is not expensive; it is arguably cheap. If FY2027 lands at, say, $70B of AI revenue — still enormous growth — the same price is roughly 34x on ~$12.30 of EPS, and the multiple has to carry the disappointment as well. The entire valuation reduces to how much weight you place on the guide. What changed since June is that the guide now has ~$164.6B of firmly-committed, non-cancellable customer obligations sitting behind it, where nine weeks ago it had a verbal “~$73B backlog.” That is a genuine, material upgrade in the quality of the evidence — and it is precisely why the stock is not cheaper despite being 9% higher.

Reverse-DCF — what the price requires

Discounting TTM free cash flow of $32.8B at a 9% cost of capital with a 3% terminal growth rate, the enterprise value implied by a range of ten-year FCF CAGRs:

(INTERPRETATION — my model; sensitive to the discount rate and terminal assumption. A 10% WACC raises the required CAGR by roughly two points; an 8% WACC lowers it by roughly two.)

The price requires a ~20% free-cash-flow CAGR sustained for ten years — unchanged in magnitude from the prior memo’s ~20–24% range, because the 9% price rise was offset by TTM FCF growth. For context, that means growing free cash flow from ~$33B to roughly ~$200B by FY2036. Only a handful of companies in history have done anything comparable, and Broadcom is plausibly one of the few that could. But the load-bearing question is not the first three years — those are now largely contracted — it is years four through ten, which rest entirely on the AI capital cycle not mean-reverting.

Scenario framework

(No price target is derived from these scenarios in this section; the directional zones appear only in Claude’s Take.)

What the market is underwriting correctly, and incorrectly

Correctly: that this is a genuinely exceptional business — 68% gross margins, 49% GAAP operating margins, 43% FCF margins on 1% capex intensity, high-teens ROIC on a heavily acquired capital base, and three distinct moats of which two are demonstrably wide. That AI silicon is a real, contract-backed franchise, not a narrative. That management is an exceptional operator with genuinely aligned long-dated incentives.

Possibly incorrectly, in three places: (1) Duration. The market appears to be capitalising a capital-cycle boom as a secular annuity. The RPO underwrites ~3 years; the price underwrites 10. (2) Concentration. Top-five customers at ~45% and rising, with one distributor at 42%, is not priced as the fragility it is — a single flagship insourcing decision would break more of the model than the diversification narrative implies. (3) The software leg. The market treats the 78.7%-margin annuity as the stable, de-risked part of the story. It grew +9% last quarter and is under active investigation by the European Commission — it may be the least de-risked part of the story, not the most.

11. Variant Perception

Consensus

The sell-side and financial media consensus is overwhelmingly constructive and has become more so through the drawdown. Sample coverage in the interval: a consensus price target implying ~33.7% upside (Zacks, 2026-07-31); “Broadcom Is Down 20% From Its All-Time High. Here’s Why It’s a Screaming Bargain” (Motley Fool, 2026-08-06); “The $30 Billion Order Book Nobody’s Pricing In” — Strong Buy (Seeking Alpha, 2026-07-30); a Buy upgrade citing the Jalapeño processor and the Samsung lock-in with a “$4.3T market cap by FY2030” case (Seeking Alpha, 2026-08-05); against a minority Hold view citing “limited near-term upside at current valuation” and VMware renewal risk (Seeking Alpha, 2026-07-30). INTERPRETATION: consensus is that Broadcom is the highest-quality way to own the AI capital cycle, that the July drawdown was a sector-wide de-risking rather than anything company-specific, and that the FY2027 guide is conservative. Nobody is arguing the business is bad. Almost nobody is arguing the price is wrong.

The strongest bull case

Broadcom is the only company that combines (a) the dominant position in the fastest-growing part of AI compute — custom accelerators, growing far faster than merchant GPUs — with (b) the merchant-Ethernet networking franchise that every cluster requires regardless of whose accelerator wins, and © a ~$28B-annualized, 78.7%-margin recurring software annuity that no semiconductor peer possesses. It now discloses ~$164.6B of firmly-committed, non-cancellable customer obligations — an audited backlog worth 2.2x trailing revenue that did not exist as a disclosed figure nine weeks ago. It has extended Apple to 2031 including custom ASICs, put OpenAI’s first inference processor into production in nine months, and signed a >$200B memorandum with Samsung that would break its single-source dependence on TSMC. On management’s own FY2027 guidance the stock trades at ~24x forward earnings — for a business compounding at 50%+ with 66% operating margins. The CEO’s own pay plan pays zero unless AI revenue holds above $60B and maximum above $120B; he has every incentive to have guided conservatively, and his guides have repeatedly proven low.

The strongest bear case

You are paying the 90th percentile of a company’s own ten-year valuation history for a business whose growth depends on the discretionary capital spending of roughly six counterparties — now ~45% of revenue, up from 40%, with a single distributor at 42%. The reverse-DCF requires a ~20% free-cash-flow CAGR for a decade; the contracts cover perhaps three years of it. The demand is partly self-financed: an Apollo/Blackstone vehicle Broadcom helped create is funding chips for cash-burning frontier labs, and nine weeks after being announced on an earnings call it still has no definitive structure in any filing. The highest-margin segment — the one funding the dividend, the buyback and the deleveraging — grew +9% last quarter and is the subject of a European Commission competition investigation in which Broadcom just lost a court fight to withhold its internal legal communications. The company spent $8.45B buying its own stock at the top of its valuation range in one half-year. It stopped paying ~$3.8B a half in cash to net-settle vesting equity, pushing that supply onto the market. And the tape already told you the answer once: on 2026-06-04 the stock fell 12.6% and lost ~$280B of value on a quarter that beat, because it declined to raise. When a price requires perfection, merely being excellent is a disappointment.

The 3–5 assumptions that matter most

1. The AI capital cycle does not digest before FY2029. Everything rests here. The RPO buffers roughly three years; the valuation underwrites ten.
2. The ~$164.6B RPO replenishes rather than merely front-loads. One April contract created ~$120B of it. If Q3 shows the balance falling as shipments burn it off, the interpretation inverts from “durable annuity” to “one big order pulled forward.”
3. No flagship customer materially dual-sources or insources. Tan has already conceded Google will see “some diversity of sources.” The per-customer split is undisclosed, so the market cannot price this precisely — which is itself the risk.
4. The software annuity holds at these prices and margins. Requires both that price-shocked renewals stick and that the European Commission does not impose a remedy.
5. The multiple does not mean-revert. At ~49x EV/EBITDA and the 90th own-history percentile, even a perfectly-executed deceleration produces a de-rating.

The factor-positioning read — where consensus may be offside

The factor model gives an empirical answer that fundamental analysis cannot (FACT — FactorsToday, 2026-08-06/07; ElasticNet betas, all factors vol-scaled to 10%, R² 0.769 on the Base + Sector + Industry model; loadings read within one model only):

Risk-adjusted track record (annualized, FACT): y10 +41.8% / Sharpe 1.00 / max drawdown −48.3%; y5 +56.9% / Sharpe 1.25; y3 +69.7% / Sharpe 1.36; y1 +40.4% / Sharpe 0.81 / max drawdown −28.7%; m6 +84.8% / Sharpe 1.71; m3 +8.7% / Sharpe 0.12 / max drawdown −25.0%. Idiosyncratic volatility 21.6% annualized. Relative strength: rs_12m +44.6, rs_6m +31.8, rs_peak −12.5.

INTERPRETATION — three genuine variant reads. First, the “semiconductor cyclical” label is wrong in the market’s own pricing. A −0.018 semiconductor-industry loading against a +2.42 Growth loading says AVGO is priced as a high-beta mega-cap growth vehicle, full stop. Investors who own it as diversified semiconductor exposure, or who hedge it with semiconductor indices, are mis-specified — and it explains why the July drawdown tracked the AI-growth complex rather than the chip cycle. Second, the market does not give Broadcom credit for its software. The Software-industry loading is negative (−0.336) despite 32% of revenue and roughly 40% of segment operating income coming from a 78.7%-margin recurring franchise. The ballast that the fundamental case leans on is not in the price as ballast. If the AI narrative cracks, the software floor the bulls invoke will not be supported by the stock’s own factor behaviour — it will fall as growth. Third, and most useful: this is not a crowded momentum trade. A +0.18 momentum loading on a stock up 44.6% in twelve months, combined with a quarter that annualizes to +8.7% with a −25% intra-quarter drawdown, describes a high-beta consolidation inside an intact uptrend — price above a rising 200-day EMA, 12.5% below its high, having traversed $480 → $360 → $421 in nine weeks. The crowding risk that a “momentum unwind” bear case relies on is not visible in the loadings. Neither is the capitulation that a “falling knife” bull case would want.

Comp cross-check caveat: FactorsToday’s factor-similar peer list for AVGO is dominated by leveraged single-stock ETFs on AVGO itself (AVGX, AVL, AVGU, AVGG at 0.93–0.96 similarity), with the nearest genuine names being Korea-exposure vehicles (EWY, KORU) and ASML/TSMC proxies. INTERPRETATION: the derivative-ETF dominance is a mechanical artefact and carries no information; the Korea and ASML/TSM proximity is mildly interesting given the Samsung MOU, but it is too thin to build a comp set on. The fundamental comp set (NVDA, MRVL, TSM, ANET) remains the right one.

12. Fact vs. Interpretation

13. Open Questions

1. Does the RPO replenish? The ~$164.6B balance was created substantially by one contract signed in Q2. The Q3 print on 2026-09-02 is the first observation of whether new signings offset ~$16B of AI shipments. This is now the single most informative disclosure in the filing.
2. How is the RPO distributed by customer and by segment? Broadcom does not disaggregate. Without it, the market cannot distinguish “six diversified multi-year franchises” from “one enormous Google contract plus five smaller ones.”
3. Does the Samsung MOU convert, and on what terms? Volume, price, take-or-pay, yield commitments and timing are all undisclosed. A memorandum that does not convert changes nothing about the TSMC dependency.
4. What is the scope and likely remedy of the European Commission’s VMware inquiry? Pricing? Bundling? Tying VCF? The Commission has not issued a Statement of Objections. The range of outcomes runs from closure with no action to imposed pricing or unbundling commitments in Broadcom’s second-largest region.
5. Does software actually re-accelerate from +9% to +31%? And if it does, how much is genuine demand versus renewal-timing and up-front recognition on subscription conversions?
6. What is the definitive structure of the Apollo/Blackstone AI-XPU platform? Recourse, guarantees, residual-value exposure, consolidation treatment, and how much of the RPO it ultimately funds. Nine weeks on, nothing has been filed.
7. Is the RSU-withholding settlement change permanent policy? If so, the run-rate of shares reaching the market rises by roughly $7–8B a year in value, and prior-year cash-return comparisons need permanent restating.
8. How much of the Apple 2031 agreement is custom ASIC versus legacy RF/connectivity? No dollar value or mix is disclosed in the 8-K.
9. What did the Chief Legal Officer’s three discretionary sales reflect? A ~30% position reduction in sixteen trading days by the officer closest to the European Commission proceeding is worth watching, though it is not evidence of anything on its own.
10. What is the per-gigawatt revenue realisation in 2027? Management says content per gigawatt rises generation-to-generation while dollars per gigawatt are “relatively stable.” Reconciling ~10GW of 2027 shipments with “>$100B” implies a per-gigawatt figure at the low end of the historically cited $15–20B range — which is either conservatism or a mix statement, and management has not clarified which.

14. What Must Be True

Bull case — what must be true

1. AI demand compounds past FY2028, not merely to it. Hyperscaler and frontier-lab capex keeps growing through FY2029–FY2032 rather than digesting. The RPO covers roughly three years of a ten-year requirement.
2. The committed backlog replenishes. RPO must hold near or above current levels as shipments burn it off — proving a rolling multi-year annuity rather than one large front-loaded order.
3. Custom-XPU share, pricing and margin hold at ~62% semiconductor segment operating margin as customers dual-source, insource and negotiate against a supplier whose backlog they can see.
4. The software annuity survives both renewal and the regulator — VMware/CA renewals hold at the higher prices and the European Commission closes without a pricing remedy.
5. The multiple does not mean-revert. A ~49x EV/EBITDA / 90th-own-history-percentile valuation must persist through an inevitable deceleration.

Falsification test (bull): RPO falling materially in the Q3 FY2026 10-Q; a guided AI-revenue cut; a flagship customer publicly dual-sourcing or insourcing its XPU; software failing to reach the guided +31% in Q3; or a European Commission Statement of Objections targeting VMware pricing. Note the asymmetry: the price already embeds the bull path, so merely failing to exceed guidance — as on 2026-06-04 — is sufficient to disappoint.

Bear case — what must be true

1. AI capex digests or decelerates within one to three years (Marathon capital-cycle mean reversion), and the ~45% top-five concentration amplifies the revenue hit rather than cushioning it.
2. The committed backlog proves less binding than it reads — renegotiated, deferred, or partially funded by private credit that withdraws (the Apollo/Blackstone circularity).
3. The multiple de-rates toward the broader large-cap semiconductor cohort as growth decelerates — even if revenue still grows.
4. The software leg cracks — an imposed EU remedy, accelerating churn, or simply the arithmetic exhaustion of the repricing cycle now visible in the +9% Q2.

Falsification test (bear): A raised FY2027 AI trajectory (>$110B) on the 2026-09-02 call, with RPO holding above ~$150B after a ~$16B AI shipment quarter; the two newer XPU customers and the Apple ASIC programme ramping ahead of plan; the Samsung MOU converting to binding supply agreements; and software printing the guided +31%. Any three of those together would break the bear by converting the current bull case into the base case.

Where the prior memo’s tests stand. In June the stated bull-flip trigger was “a raised, contract-backed FY27/FY28 AI trajectory.” The interval delivered contract-backed (RPO $45.0B → $164.6B) but not raised (still “>$100B”). The trigger is half-met — which is precisely why the accumulation zone moves up in Claude’s Take while the HOLD stands.

15. Source Appendix

See Appendix B below for the full source list with access dates. Primary sources: the Broadcom 10-Q for the fiscal quarter ended 2026-05-03 (filed 2026-06-09), 10-Qs for Q1 FY2026 (2026-03-11) and Q3 FY2025 (2025-09-10), the FY2025 10-K (filed 2025-12-18), 8-Ks of 2026-06-11, 2026-06-18 (with EX-99.1/EX-99.2) and 2026-07-06, the DEF 14A filed 2026-03-02, and the Form 4 corpus parsed from source XML (EDGAR CIK 0001730168). Quantitative figures reconciled to those filings via the ROIC.ai MCP. Management commentary from the Q2 FY2026 earnings call (2026-06-03) is treated as hypothesis per section 0 rule 8. Third-party: Samsung Global Newsroom and CNBC (2026-07-25); Broadcom IR / OpenAI / CNBC / Tom’s Hardware (2026-06-24); Bloomberg (2026-07-06); Reuters and Concurrences (2026-08-03). Market and factor data: AZI adjusted price history and valuation_index (2026-08-06) and the FactorsToday factor model (2026-08-06/07). Prior coverage: my own AVGO note of 2026-06-07 (the update baseline) and prior notes on NVDA, MRVL, TSM, ANET, MU, CLS and DELL for peer cross-read.

No investment recommendation or price target appears in sections 1–15; the only position taken is in the clearly-labeled opinion block at the top. This article is general information and is not investment advice.

APPENDIX A — Standard Diligence Questionnaire

Broadcom Inc. (NASDAQ: AVGO) — Standard Diligence Questionnaire Appendix

Companion to the main analysis above | Price reference: $420.565 (2026-08-06 close) | Date: 2026-08-07 Type: update — supersedes the 2026-06-07 version

A standard due-diligence checklist worked through for this company. Labels — FACT / INTERPRETATION / ASSUMPTION / OPEN QUESTION — applied where the distinction matters. Greenwald (moats) and Marathon (capital cycle) frameworks applied where they add insight. No price target and no buy/sell verdict outside the opinion block above.

General — What thoughtful questions have other investors asked?

The debate has shifted materially in nine weeks. In June the sophisticated questions were about whether the AI backlog was real. The Q2 FY2026 10-Q, filed 2026-06-09, largely answered that — and the questions moved downstream.

• “Does the $164.6B RPO replenish, or did one contract front-load three years of revenue?” (the new central question) Firmly-committed, non-cancellable remaining performance obligations went $27.5B (Q3 FY25) → $33.3B (FY25) → $45.0B (Q1 FY26) → $164.6B (Q2 FY26), with the filing attributing the step-up to “a long-term contract for custom AI accelerators entered in the fiscal quarter” (FACT). INTERPRETATION: the Q3 FY2026 balance, reported 2026-09-02, is the first read on whether new signings offset ~$16B of AI shipments. A flat-to-rising balance validates a rolling annuity; a falling one recasts it as a single pulled-forward order.
• “Why reiterate $56B rather than raise, after a $30B booking quarter?” Still unresolved as intent, but the arithmetic deserves correction: H1 FY2026 AI was ~$19B and Q3 is guided to $16B, so the residual implies Q4 AI ~$21B, up ~31% sequentially — not the sequential decline that helped drive the 2026-06-04 selloff (INTERPRETATION on management’s own figures).
• “Is the concentration priced?” Top-five end customers rose to ~45% of revenue from 40%, and one distributor to 42% from 29% (FACT — 10-Q). No per-customer AI split is disclosed (OPEN QUESTION).
• “What does the European Commission want, and what can it impose?” After the 2026-08-03 General Court ruling, Broadcom must produce internal legal communications about VMware to a regulator with compulsory powers. Scope and remedies undisclosed (OPEN QUESTION).
• “Is the software leg actually growing?” Reported +9% in Q2 and +5% in H1 against a guided +31% for Q3 (FACT). The gap is the largest unproven assertion in the guidance stack.
• “How much of the demand is self-financed?” The Apollo/Blackstone ~$35B first tranche of a >20GW “AI XPU platform” was described on the Q2 call but has no definitive structure in any filing nine weeks later (FACT/OPEN QUESTION).
• “Is ‘FCF’ clean?” SBC runs ~$8.5–8.8B annualized (~11–12% of revenue) with $20,106M unrecognized over 3.0 years — and Broadcom has stopped spending ~$3.8B a half in cash to net-settle vesting RSUs, selling vested shares into the market instead (FACT — 10-Q MD&A).

Cyclicality & Earnings Nature

• Cyclical high or low? INTERPRETATION: A structural earnings high on the silicon side, with the non-AI portfolio still recovering off a trough. GAAP operating margin has expanded ten points in four quarters (39.4% → 48.9%); EBITDA margin reached 55.0%; AI revenue is ~49% of the total. The Marathon read is unchanged: the AI-accelerator sub-segment is in the boom phase of a capital cycle whose demand side is itself a boom. The genuinely new mitigant is that ~$164.6B of that demand is now contractually committed and non-cancellable — which changes who absorbs the first loss when the cycle turns, without repealing the cycle. Non-AI semiconductors (wireless, broadband, industrial, server/storage) remain near a cyclical trough recovering: $4.2B in Q2 (+6%) against >$6B of bookings, guided +12% in Q3.
• External environment or internal actions? Both, and the balance has tilted external. The AI capex super-cycle is the external driver of ~49% of revenue. The internal drivers — VMware repricing, cost-out, subscription migration — produced the 78.7% software margin but only +9% software revenue growth last quarter. A year ago the earnings recovery was roughly half “the world got better” and half “we extracted it.” Today it is predominantly the former.
• Revenue stability. INTERPRETATION: bifurcated, and the durable half is shrinking as a share. ~32% is recurring/subscription software with high renewal economics; ~49% is AI silicon that is sticky-but-project-lumpy (now with multi-year committed contracts behind much of it); the remaining ~19% is classically cyclical wireless/broadband/industrial/storage. Note the direction of travel: software was 44% of revenue a year ago and is 32% today.
• Outlook for products/services. FACT (management guidance, a hypothesis per section 0 rule 8): FY2026 AI ~$56B (+~180%); Q3 AI ~$16B; FY2027 AI “>$100B”; FY2028 “substantial growth” from there; ~10GW of 2027 shipments, back-half weighted; Q3 software ~$8.9B (+31%); Q3 consolidated revenue ~$29.4B (+84%) at a ~67% non-GAAP operating margin.
• Market size — growing/shrinking, domestic/international? Growing on both engines. Custom accelerators continue to take share from merchant GPUs, and the interval added direct evidence (OpenAI’s Jalapeño reaching production in nine months). Networking content scales with cluster size. Software markets are mature-but-very-high-moat. International: Q2 revenue was 66% Asia-Pacific / 23% Americas / 11% EMEA (FACT — 10-Q), though the APAC weighting reflects title transfer in Penang rather than end-demand geography. Leading-edge fabrication is concentrated at TSMC in Taiwan, with a Samsung MOU (2026-07-25) offering a prospective second source at Pyeongtaek.

Business Quality & Competitive Moat

• Industry more or less competitive? INTERPRETATION: two answers, both refined this quarter. Custom AI silicon continues to concentrate around Broadcom and Marvell, and Broadcom’s position strengthened — a nine-month design-to-production cycle for OpenAI’s first inference ASIC, and Apple extending through 2031 including custom ASICs. The intensifying threat vectors (hyperscaler insourcing, merchant-GPU substitution) are unchanged in kind. Infrastructure software remains less competitive at the core — VMware has no at-scale on-premises equal — but the constraint on Broadcom’s pricing is now regulatory rather than competitive.
• How profitable (ROIC, ROE)? FACT/INTERPRETATION: TTM ROE remains strong (FY2025 ~31%, higher on a TTM basis) though equity is inflated by ~$71.3B of VMware-issuance paid-in capital. ROIC including goodwill has improved to ~18–20% (TTM EBIT $33,253M at a ~16% normalized tax rate ≈ ~$28B NOPAT over ~$140B of invested capital), up from the ~15–17% in the prior appendix — purely because earnings grew against a static acquired-capital base. The Greenwald/Marathon tension is resolving in Broadcom’s favour: the ~$130B it paid to assemble these franchises is now earning a genuinely attractive return, and the gap to the extraordinary return on tangible capital (only ~$2.5B net PP&E) is narrowing as intangibles amortize ($27,583M of amortizable intangibles remain).
• Industry profitability / competitor count / barriers. Few credible competitors in each franchise, and the segment margins prove the barriers: Semiconductor Solutions 61.8% and Infrastructure Software 78.7% operating margins in Q2 FY2026. Custom XPU: a duopoly with Marvell, barrier = SerDes/packaging/physical-design IP, multi-generation co-design lock-in, tens of millions per mask set. Networking: merchant-Ethernet scale leadership (100T Tomahawk 6 shipping, 200T taping out). Software: near-monopoly niches where migration is a multi-year re-platforming project.
• Easily understood? INTERPRETATION: moderately, and slightly harder than a quarter ago. The two-engine structure is comprehensible, but three things require work: the per-customer AI split is undisclosed; the GAAP-to-non-GAAP wedge is ~$4.1B a quarter of unallocated expense; and the new $164.6B RPO is not disaggregated by segment, customer or year beyond the “~30% within 12 months” disclosure.
• Undermined by foreign low-cost labour? INTERPRETATION: No. This is IP-and-design value, not labour arbitrage. The relevant foreign exposure runs the other way — dependence on TSMC in Taiwan for leading-edge wafers and CoWoS packaging, which the Samsung MOU is explicitly intended to reduce.
• Do brands matter? INTERPRETATION: Partially, and not in the consumer sense. In software the installed base and switching costs matter far more than brand — customers stay because migration is painful, and several have sued over price. In silicon the relevant “brand” is engineering reputation and IP leadership; the clearest expression of it this quarter is that OpenAI and Apple both chose Broadcom for first-of-kind custom programmes. This is a switching-cost / scale / intangibles moat, not a brand moat.
• Nature of competition. Custom XPU: head-to-head design-win competition with Marvell and with customers’ own in-house teams for multi-year, effectively single-source-per-programme engagements; versus Nvidia it is custom-ASIC-vs-merchant-GPU substitution. Networking: open Ethernet against proprietary NVLink/InfiniBand/Spectrum, with Broadcom explicitly refusing to move up-stack (“No racks… We only do chips” — Tan, Q2 FY2026 call). Software: incumbency against slow secular migration to public cloud.
• Customers’ switching costs. FACT (revealed in pricing power): very high in software — several-fold price increases with the base largely retained and a 78.7% segment operating margin. High but customer-specific in custom XPU — and the interval produced a new, cleaner proof: ~$164.6B of firmly-committed obligations “where customers do not have termination rights.” Customers do not sign non-cancellable multi-year commitments of that scale to a supplier they believe they can readily replace.

Financial Condition & Balance Sheet

• Assets not fully recognized on the balance sheet? INTERPRETATION: Yes — the moats, and now the backlog. The VMware/CA installed-base captivity and the SerDes/packaging IP that generate 62–79% segment margins are economic assets carried only partly as (amortizing) acquired intangibles and partly not at all. The ~$164.6B of firmly-committed remaining performance obligations is a disclosed contractual asset that appears in the revenue note, not on the balance sheet. Conversely the balance sheet is over-weighted with purchase accounting: ~$97.8B goodwill plus ~$27.6B of remaining amortizable intangibles against only ~$2.5B of tangible net PP&E.
• Off-balance-sheet liabilities — the Apollo/Blackstone vehicle. FACT/OPEN QUESTION: management described a >20GW “AI XPU platform” with Apollo, Blackstone and other investors whose first tranche of ~$35B “is currently being launched by Apollo” (Q2 FY2026 call). No definitive structure, commitment, guarantee or variable-interest-entity consolidation appears in the 10-Q filed 2026-06-09, nine weeks after the announcement. INTERPRETATION: the absence of disclosure is mildly reassuring on recourse, but the economic substance is vendor financing with extra steps — capital Broadcom helped arrange, deployed by third parties, to buy Broadcom chips for customers who might not otherwise afford them. Other off-balance-sheet items are ordinary: operating-lease right-of-use assets, $20,106M of unrecognized SBC over 3.0 years, and $14,242M of contract liabilities ($10,038M current, $4,204M long-term) against $5,918M of current contract assets from the VCF subscription transition.
• How conservative is the accounting? INTERPRETATION: GAAP is clean; two items warrant flags this quarter. Cash conversion is strong and unmanipulated — Q2 operating cash flow $10,493M against $9,310M of net income (1.13x), with the difference explained by non-cash add-backs net of a deliberate working-capital build. Flag 1 (new): the RSU-withholding settlement change. In H1 FY2025 Broadcom paid $3,802M in cash withholding taxes and withheld 17M shares; in H1 FY2026 that line is zero because withholding is now settled “using proceeds from the sale of a portion of the vested shares.” Fully disclosed, not manipulation — but it flatters reported cash returns and shifts ~$3.8B a half of share supply to the market. Flag 2: tax. Q4 FY2025 carried a −$1,649M benefit (negative provision) that inflates trailing GAAP EPS; management guides a ~16% non-GAAP rate for Q3 and FY2026. Normalize before extrapolating $6.006 of trailing EPS. Watch item: days of inventory rose 68 → 86 in one quarter (a deliberate pre-position for the guided H2 AI ramp), with a −$2,572M Q2 working-capital drag.
• How CapEx-hungry? FACT: Structurally capital-light and getting more so relative to revenue. Capex was $231M in Q2 on $22,187M of revenue — 1.0% — against a 46.3% free-cash-flow margin. The fabless model pushes fab capital to TSMC (and prospectively Samsung). Capital intensity has instead migrated to M&A historically, and now potentially to the off-balance-sheet AI-financing vehicle.

Capital Allocation & Management

• How much FCF, how used, what philosophy? FACT: TTM free cash flow ~$32.8B (~43% margin); Q2 alone $10,262M. FY2026 marks a return to offence after the FY2025 deleveraging year: H1 buybacks $8,450M (versus $2,450M in all of FY2025) with $10.1B of authorization remaining; dividends $6,178M; a June cash tender retiring ~$2.9B of senior-note principal (cap upsized $2.5B → $3.0B), with the longest, highest-coupon series bought below par. Philosophy remains the Hock Tan private-equity roll-up: acquire sticky franchises, strip cost, reprice, harvest cash, deleverage, repeat. INTERPRETATION: the tender was excellent; the buyback is the one decision to challenge — $8.45B deployed at the 90th percentile of the company’s own ten-year valuation range is capital committed with the cycle rather than against it, which is the opposite of the Marathon discipline.
• Significant acquisitions recently? FACT: None announced in the interval. The chain remains LSI (2014), Broadcom Corp (2016), Brocade (2017), CA (2018), Symantec enterprise (2019), VMware (2023, ~$84.2B actual consideration: $30,788M cash plus 544M shares at $53,398M fair value). INTERPRETATION: restraint is the right call with the equity at 90th-percentile own-history valuation as acquisition currency, and the VMware thesis is proving out — a 78.7% segment operating margin two-and-a-half years post-close, against ~$84B of consideration, with the EUC business already carved out to KKR for $3.5B.
• Buying back shares? FACT: Yes, aggressively and newly so — $8,450M in H1 FY2026 versus $2,450M in the whole of FY2025, with $10.1B still authorized (implying roughly $11B of new authorization added during FY2026). Important change: the separate “shares repurchased for tax withholdings on vesting of equity awards” line, which consumed $3,802M in H1 FY2025 and $3,860M in FY2025, is zero in H1 FY2026 — so headline buyback growth overstates the improvement in true net capital return.
• Issuing large amounts of new shares to insiders? FACT: Yes, materially — via SBC, and now with less offset. SBC was $2,092M in Q2 and $4,268M in H1 (+40% YoY), ~$8.5–8.8B annualized (~11–12% of revenue), with $20,106M unrecognized over a 3.0-year weighted-average period ($4,078M remaining FY2026, $7,235M FY2027, $5,137M FY2028). RSUs outstanding fell 229M → 183M as ~41M vested. Diluted share count is ~4.876B and drifting down on buybacks, but the withholding-settlement change means vested shares now reach the market rather than being withheld.
• Compensation policy of directors/management. FACT: Hock Tan’s FY2025 total compensation was $205,278,006; CEO pay ratio 543:1 (median employee $378,281); salary $1.2M, stock awards $202.35M, 0% cash bonus (0% annual-plan target). Say-on-pay passed with 92% support in 2025. NEO annual-plan metrics are revenue ($54,429M / $57,294M / $60,159M) and adjusted non-GAAP operating margin (61.7% / 63.7% / 67.7%) — note Q2’s 67.3% non-GAAP operating margin sits at the maximum payout level. The 2023 Tan PSU (FY2023–27) vests on stock-price hurdles plus employment to 2027-10-31, tracking above target. The 2025 Tan PSU (FY2028–30) covers 610,521 target shares paying 0–300% on AI Revenue: ≤$60B = 0%, $90B = 100%, $105B = 200%, ≥$120B = 300%. INTERPRETATION: pay level indefensible in absolute terms; pay structure genuinely aligned — and newly informative in light of the RPO disclosure, because the CEO’s personal economics and the single most important variable for the stock are now literally the same number. A sceptical reader should hold both readings: excellent alignment, and a powerful incentive to sign and guide aggressively.
• Motivations of management. INTERPRETATION: structurally aligned to multi-year stock price and a tripling of AI revenue; five-year TSR through FY2025 was +1,082%. Watch-items: (i) the Apollo/Blackstone pivot as a capital-discipline test, unresolved; (ii) key-person dependence on Tan, with no written employment contract and no key-person insurance per the 10-K; (iii) the CFO transition executed 2026-06-12 — Kirsten Spears retired after twelve years (adviser for nine months) and Amie Thuener (O’Toole), formerly Alphabet’s VP, Corporate Controller and Chief Accounting Officer since 2018, took the role. Note: the Q2 call transcript renders her name “Amy Teiner”; the filings and the company release do not. The hire brings deep technical-accounting and controls experience precisely as revenue recognition becomes more complex, and comes from inside Broadcom’s largest AI customer. Insider behaviour (FACT, Form 4s parsed from source XML, 2026-06-11 → 2026-07-14): director Harry You made a 1,000-share open-market purchase at $373.57 on 2026-06-11 (code P, not 10b5-1) — his third such buy in ~10 months; co-founder Henry Samueli sold ~654,000 shares (~$250M) on 2026-06-24 under a 10b5-1 plan adopted 2025-12-16, roughly 1% of a ~66.5M-share position; and Chief Legal & Corporate Affairs Officer Mark Brazeal sold 75,000 shares (~$29.2M) across three discretionary, non-10b5-1 blocks on 2026-06-25, 2026-07-08 and 2026-07-10, reducing his reported holding from 277,949 to 194,989 shares — a ~30% cut in sixteen trading days by the officer with the clearest visibility into the European Commission proceeding. None of this is thesis-determinative; the Brazeal pattern belongs on the watch-list.

Valuation & Market Data

• ADR, MLP, or K-1 issuer? FACT: None of these. Broadcom Inc. is a Delaware C-corporation, single class of common stock, listed on NASDAQ (AVGO), issuing a standard Form 1099-DIV. No ADR ratio complication, no partnership structure.
• Dividend policy. FACT: a regular, growing cash dividend of $0.65 per quarter ($2.60 annualized) — ~0.62% yield at $420.565; $6,178M paid in H1 FY2026 against $11,142M in FY2025. This is a compounding vehicle, not an income name; the dividend is the stable leg while buybacks flex with the deleveraging cycle.
• How profitable is the business? FACT: exceptionally. Q2 FY2026 GAAP gross margin 69.5%, operating margin 48.9%, net margin 42.0%, EBITDA margin 55.0%, FCF margin 46.3%; segment operating margins Semiconductor Solutions 61.8% and Infrastructure Software 78.7%. TTM: revenue $75,465M, EBITDA $41,499M, net income $29,317M, FCF ~$32,762M. (Trailing GAAP figures are flattered by the Q4 FY2025 −$1,649M tax benefit; management guides a ~16% forward non-GAAP rate.)
• Is net income diverging from cash from operations? FACT/INTERPRETATION: No — converged and high quality. Q2 operating cash flow $10,493M against $9,310M of net income (1.13x); H1 OCF $18,753M against H1 net income $16,659M (1.13x). The divergence is explained by non-cash add-backs (D&A $2,165M, SBC $2,092M in Q2) net of a deliberate −$2,572M working-capital build (receivables −$2,370M, inventories −$1,366M). Watch items: days of inventory 68 → 86, and a $5,918M current contract-asset balance from the VCF subscription transition. Neither is a red flag; both belong on the Q3 checklist.
• Where does it trade? FACT (2026-08-06 close, $420.565): market cap ~$2.00T (basic) / ~$2.05T (diluted); net debt $47.1B (~$44.2B pro-forma for the June tender); EV ~$2.05T. That is ~49.4x EV/EBITDA, ~27.1x EV/revenue, ~61.6x EV/EBIT, ~62.5x EV/FCF and ~70.0x trailing GAAP earnings — up from ~44.7x / ~24.9x / ~64x at the prior memo’s $385.73. Against its own ~10-year history the stock sits at the 90.2nd percentile composite (P/B 97.1st, P/S 96.0th, P/E 77.4th), essentially unchanged from June despite a 9% price rise, because trailing fundamentals grew into it. (Per standard treatment: the P/E percentile understates richness because trailing GAAP earnings are tax-flattered; read P/B and P/S as the truer own-history signal.) On forward guidance the picture inverts — my build puts the stock at ~37x FY2026 and ~22–24x FY2027 non-GAAP earnings (ASSUMPTION, from management’s guided revenue and stated margin/tax/share structure; no primary-source consensus was obtainable this session). A reverse-DCF implies a ~20% free-cash-flow CAGR for a decade at a 9% discount rate.

Risks & Downside

• What factors would cause the stock to decline? INTERPRETATION (top four, interlinked): (1) Valuation/expectations — at ~49x EV/EBITDA and the 90th own-history percentile, anything short of a raise de-rates it; the empirical proof is 2026-06-04, when the stock fell 12.6% and lost ~$280B on a quarter that beat, because FY2026 AI was reiterated rather than raised. (2) RPO failing to replenish — the new and most informative single disclosure; a sequential decline in the Q3 10-Q would recast ~$164.6B from annuity to pulled-forward order. (3) Concentration — top-five at ~45% and one distributor at 42%; a flagship insourcing or dual-sourcing decision would break more of the model than the six-customer narrative implies. (4) An EU remedy on VMware pricing — the one event that attacks the 78.7%-margin annuity funding the dividend, the buyback and the deleveraging. Secondary: merchant-GPU substitution, Apollo/Blackstone circularity, China/export controls and a possible US ban on Chinese optical components, TSMC/Taiwan supply, SBC dilution plus the new open-market share supply, and key-person risk.
• Risk of a catastrophic loss? INTERPRETATION: Low at the company level, and lower than in June. Diversified profitable franchises, ~$33B of annual free cash flow, ~1.1x net leverage (~1.07x pro-forma) falling, interest expense declining, and ~$164.6B of committed customer obligations. There is no realistic path to permanent capital impairment from operations. The one plausible catastrophic vector remains a Taiwan/TSMC shock — geopolitical or natural — which is industry-wide rather than Broadcom-specific, and which the Samsung MOU would partially mitigate if it converts. The dominant risk is security risk, not business risk: a violent de-rating in a digestion scenario is a permanent loss for a buyer at $420.57 even if the company thrives.
• Chance of a total loss? INTERPRETATION: Negligible (very low likelihood / very high impact). A total loss would require a Taiwan-scale supply catastrophe compounded by a simultaneous collapse in AI demand and an inability to refinance ~$65B of debt against $33B of annual free cash flow — not a base-case combination. The honest asymmetry is “outstanding company, possibly the wrong price,” not “risk of zero.”

Recent News & Events

• Has the business environment changed recently? FACT: Yes, and materially, in nine weeks — mostly for the better. The Q2 FY2026 10-Q (filed 2026-06-09) disclosed firmly-committed remaining performance obligations of ~$164.6B, up from $45.0B one quarter earlier, converting the FY2027 “>$100B” AI guide from a management hypothesis into a substantially contract-backed trajectory. OpenAI’s “Jalapeño” inference processor was unveiled on 2026-06-24 — designed to production in ~9 months with Celestica building, initial deployment by end-2026 through partners including Microsoft. Apple extended through 2031 including custom ASIC silicon (8-K, 2026-07-06), retiring the largest named wireless risk and adding a prospective seventh custom-silicon customer. And a >$200B Samsung memorandum (2026-07-25) covering HBM4/HBM4E supply, 2nm manufacturing at Pyeongtaek and advanced packaging attacks the single-source TSMC dependency. Against those: customer concentration rose to ~45% (from 40%), and the European Commission’s VMware inquiry escalated — on 2026-08-03 the EU General Court refused to suspend a compulsory document demand, rejecting Broadcom’s US attorney-client-privilege objection.
• Significant acquisitions? FACT: None. Capital went to buybacks ($8,450M in H1), dividends ($6,178M) and debt reduction (a ~$2.9B tender, upsized, with the longest and highest-coupon notes purchased below par).
• Change in accounting policies? INTERPRETATION: No policy change, but one settlement-mechanics change with real economic effect. Broadcom stopped net-settling vesting RSUs with cash ($3,802M in H1 FY2025) and now settles withholding taxes “using proceeds from the sale of a portion of the vested shares.” Fully disclosed; it flatters reported cash returns and adds ~$3.8B a half of share supply. The ongoing dynamic remains the VMware perpetual-to-subscription conversion, which continues to build contract assets ($5,918M current) and contract liabilities ($14,242M total) — a transition effect, not manipulation.
• Recent changes — new markets, facilities, management? FACT: the CFO transition completed on 2026-06-12 — Kirsten Spears retired after twelve years (remaining an adviser for nine months); Amie Thuener (O’Toole), previously Alphabet’s VP, Corporate Controller and Chief Accounting Officer since 2018, became CFO, with a 50,000-share new-hire award granted 2026-06-15. New “markets”: VMware Cloud Foundation 9.1 added heterogeneous GPU/CPU support (AMD, Intel, NVIDIA) for on-premises AI inferencing; Tomahawk 6 (100T) is shipping with a 200T successor taping out this quarter. Geographic mix shifted sharply toward Asia-Pacific (66% of Q2 revenue, from 54%) as XPU shipments scaled. The most informative “event” of the interval, however, was not an announcement but a filing — the RPO line in the Q2 10-Q, published two days after the prior report and largely unremarked in the press coverage that followed.

APPENDIX B — Source Appendix

Source Appendix — Broadcom Inc. (NASDAQ: AVGO)

Update report, 2026-08-07. Primary sources first. Quantitative figures anchored to SEC EDGAR filings; third-party data providers used for computation and orientation and reconciled to filings. Access dates: 2026-08-07 unless noted.

A. Company SEC filings (primary) — EDGAR CIK 0001730168

Form 4 corpus (insider transactions), parsed from source XML including the Rule 10b5-1 affirmation flag (aff10b5One):

* Accession/filer mapping verified per-document from the parsed XML reporting-owner name.

• Filing corpus enumerated from the SEC EDGAR filings index for CIK 0001730168 from 2026-05-15 forward; the trailing 60-month corpus was reviewed in the prior engagement.

B. Earnings-call transcript

• Broadcom Q2 FY2026 earnings call — quarter ended 2026-05-03, call held 2026-06-03. Retrieved in full this session from the ROIC.ai MCP (get_latest_earnings_call, NASDAQ:AVGO → FY2026 Q2). Speakers: Hock E. Tan (President & CEO), Kirsten Spears (outgoing CFO), Ji Yoo (IR), with Charlie Kawwas and Ram Velaga. Source of: FY2026 AI ~$56B, Q3 AI ~$16B, FY2027 “>$100B”, ~10GW of 2027 shipments, >$30B of Q2 AI bookings, the six named XPU customers and their gigawatt commitments, the ~40%→~30% networking-mix guide, the ~$35B Apollo/Blackstone first tranche, the “No racks… We only do chips” positioning, the Q3 gross-margin (~74%) and tax-rate (~16%) guides, and the ~4.94B Q3 share-count guide. Treated as management hypothesis per section 0 rule 8 and validated against the 10-Q wherever possible.
• Q3 FY2026 results are scheduled for 2026-09-02 after market close (PRNewswire, 2026-08-03) — i.e. after this report date. No newer call exists.

C. Industry, corporate and legal — third-party (named publisher, dated)

• Samsung Global Newsroom, “Samsung Electronics and Broadcom Expand Strategic Collaboration Across Memory and Foundry Technologies,” 2026-07-25 — MOU >$200B through 2030; HBM4/HBM4E supply; 2nm-and-below manufacturing at Pyeongtaek; advanced packaging. Corroborated by CNBC (“Samsung Electronics wins $200 billion Broadcom AI chip partnership,” 2026-07-25) and Bloomberg/Yahoo Finance (2026-07-25). An MOU, not a binding contract.
• Broadcom Investor Relations, “OpenAI and Broadcom Unveil LLM-Optimized Intelligence Processor,” 2026-06-24; OpenAI blog, same date; CNBC, “OpenAI and Broadcom reveal Jalapeno, first AI chip in partnership,” 2026-06-24; Tom’s Hardware, 2026-06-24 (reticle-sized ASIC, ~9-month design cycle, Celestica as manufacturing partner, end-2026 deployment via partners including Microsoft).
• Bloomberg, “Broadcom, Apple Extend Tie-Up to 2031 With New Custom Chips,” 2026-07-06 — Apple ~20% of Broadcom revenue; ASIC element linked to Apple’s “Baltra” AI server programme. Underlying primary source: the 8-K of 2026-07-06.
• Reuters, “Broadcom loses court bid to suspend EU antitrust request for US legal papers,” 2026-08-03; Concurrences case note, August 2026 — EU General Court refuses interim relief against a European Commission Article 18 request for information in the VMware-related competition inquiry; US attorney-client privilege does not shield documents from EU compulsory process. Interim relief only; merits undecided; no Statement of Objections issued.
• Benzinga, “Broadcom, Marvell And Nvidia Could Feel the Impact of US Ban on Chinese AI Components,” 2026-08-05 — press-reported policy risk to optical transceivers and silicon photonics. Policy proposal, not enacted rule.
• PRNewswire, “Broadcom Inc. to Announce Third Quarter Fiscal Year 2026 Financial Results on Wednesday, September 2, 2026,” 2026-08-03.
• CFO Dive / Bloomberg / company release (2026-04-02) — appointment of Amie Thuener as CFO effective 2026-06-12; previously Alphabet VP, Corporate Controller and Chief Accounting Officer since 2018; Kirsten Spears to advise for nine months.
• Sell-side and financial-media sentiment sample used only to characterise consensus in section 11: Zacks (consensus target implying ~33.7% upside, 2026-07-31); Motley Fool (2026-08-06); Seeking Alpha (“The $30 Billion Order Book Nobody’s Pricing In,” 2026-07-30; Buy upgrade citing Jalapeño and the Samsung lock-in, 2026-08-05; “Safer Choice For AI Hardware” Hold, 2026-07-30); Invezz (2026-08-05); Investopedia (chip pullback on AI-bubble concerns, 2026-07-28). Cited as evidence of consensus positioning only, never as evidence of fact.
• VMware repricing and litigation background carried forward from the 2026-06-07 report: Network World and licensing-timeline reporting (2024–2026) on the AT&T dispute (alleged increases up to 1,050%; settled December 2024), the UnitedHealth suit (April 2025), EU CISPE complaints (800–1,500%), and the reversed 72-core minimum.

D. Quantitative data feeds (computation and orientation; reconciled to filings)

• ROIC.ai MCP (NASDAQ:AVGO), accessed 2026-08-07 — get_income_statement, get_cash_flow and get_enterprise_value (quarterly, 6–9 periods) for the TTM rebuild: revenue $75,465M, EBITDA $41,499M, EBIT $33,253M, net income $29,317M, diluted EPS $6.006, TTM FCF $32,762M, and the market-cap/EV/net-debt components. Also get_company_news (2026-06-01 → 2026-08-07, limit 50) and get_latest_earnings_call. Third-party aggregated data, not primary; every material figure reconciled to the 10-Q filed 2026-06-09. Note: the identifier must be exchange-qualified (NASDAQ:AVGO); a bare ticker is rejected.
• AZI adjusted price history — https://azitrading.com/controls/download-data.php?t=AVGO, full series to 2026-08-06 (4,276 rows). Source of the five-year event map: 5-yr low $40.35 (2022-10-14), all-time high $480.81 (2026-06-02), close $420.565 (2026-08-06), 52-week range $287.49–$480.81, 21/50/200-day EMAs ($391.18 / $390.36 / $360.96), beta 2.064, alpha 0.288, and the largest single-day moves.
• Own-history valuation percentiles, 2026-08-06 — composite 90.169th percentile of own ~10-year history; P/B 97.076th; P/S 96.042th; P/E 77.387th; n_components 3; TTM EPS $6.007, book value/share $17.98, TTM sales/share $15.47. Per the standing caveat, AZI statement-line fundamentals are unreliable for AVGO (stale legacy-Avago data) and were not used; only valuation_index.
• FactorsToday factor model (https://www.factorstoday.com/api), 2026-08-06/07 — /stock-loadings/AVGO (four nested ElasticNet models; loadings read within a single model only, per the published methodology; R² 0.769 on Base + Sector + Industry), /leaderboard/AVGO (annualized returns, Sharpe, Sortino and max drawdown by horizon), /stock-info/AVGO (beta, alpha, relative strength, market cap $1.99T), /stock-specific-vol/AVGO (21.6% annualized idiosyncratic vol), /related-stocks/AVGO. Third-party statistical estimates; loadings and realized returns are reportable facts, forward inference is labeled interpretation.
• SEC EDGAR company-filings index and document archive (CIK 0001730168), queried with a compliant contact user-agent — filing enumeration and document retrieval.

E. Prior coverage of the same company and its peers

• My own 2026-06-07 Broadcom note — the baseline thesis for this update. Read in full; its falsification tests are explicitly re-tested in section 14 above.
• Prior notes on same-sector peers, read for comparison and cross-check: NVDA, MRVL, TSM, ANET, MU (HBM pricing and allocation context for the Samsung MOU), CLS (the Jalapeño manufacturing partner, and the AI-hardware capital-cycle analog), DELL, AMD, QCOM, TXN, ADI, NXPI, AMAT, LRCX, KLAC, COHR, CRDO, ALAB, SMCI and VRT. Used for framework and peer context; independent primary research was performed for this note. Peer multiples quoted in section 4 are prior-note values, not re-derived this session.
• No third-party sell-side semiconductor primer was available to this analysis. Industry framing is therefore sourced independently from filings, company disclosure and named public reporting.
• Nothing in this article should be read as disclosing a position in AVGO. It is written position-agnostic.

F. Analytical frameworks

• Greenwald & Kahn, Competition Demystified (barriers to entry; the moat taxonomy of proprietary intangibles, customer captivity/switching costs and economies of scale; market-share-stability and ROIC tests) and Marathon / Chancellor, Capital Returns (supply-side capital-cycle analysis; the asset-growth anomaly; high returns attract capital and mean-revert). Applied in section 3 (capital cycle and the contractual buffer), section 4 (moat taxonomy and pressure-tests), section 7 (buyback timing against the cycle) and the diligence appendix.