Broadcom Inc. (AVGO) — Coverage File

Data as of: Q3 FY2026 (quarter ended August 2, 2026; 10-Q filed Sep 10, 2026); latest source Goldman Sachs Communacopia conference, Sep 8, 2026 File last updated: October 10, 2026 (initial build)

---

0. 30-second reopen

• Business: Custom AI accelerators (XPUs) and Ethernet networking for six frontier-model builders now drive 56% of revenue; VMware software (~84% segment margin) and non-AI chips are the stable base.
• Core debate: Is the ~$115bn FY2027 and ~$230bn FY2028 AI outlook contracted demand from durable customers, or growth Broadcom is increasingly financing, guaranteeing and pre-buying supply for?
• Base thesis: Labs want chips co-designed for their own models; Broadcom is the proven partner, so AI revenue roughly doubles each year while opex barely moves and operating margin holds ~66%.
• Main bear case: The two customers set to be largest in 2027–28 cannot fund their own compute; Broadcom carries a ~$29bn backstop, a $42bn notes line and $127bn of unconditional purchase commitments, so a lab stumble hits revenue, guarantees and inventory together.
• What to watch: AI revenue against the outlook, the 12-month share of RPO, new XPV tranches and backstops, purchase commitments vs. RPO, VMware ARR growth, XPU second-sourcing.
• Valuation: ~$1.73T market cap, ~45x trailing GAAP earnings and a ~2.3$30" FY2028 EPS target. The price leans on the FY2027–28 outlook being delivered.

---

1. Business snapshot

Broadcom reports two segments. Semiconductor solutions ($20,839M of Q3 FY2026 revenue, +127% y/y; $12,770M segment operating income, 61%) is now mostly AI: AI semiconductor revenue was $16.7B (+221% y/y), of which XPUs were 73% and AI networking (Tomahawk switches, PCIe, optical DSPs, lasers) the rest [Q3 FY2026 10-Q; Q3 FY2026 call]. Non-AI semiconductors (broadband, server storage, wireless, including Apple's custom ASICs under an LTA through 2031) were $4.2B, +5% y/y [Q3 FY2026 call; 8-K Jul 6, 2026].

Infrastructure software ($8,752M, +29%; $7,325M segment operating income, 84$30.8B of cash and 544M shares worth ~$53.4B [FY2025 10-K].

The AI book is six XPU customers. Management names Google (TPUs, every generation for over ten years, LTA through up to 2031), Anthropic (TPU-based, expected to be the largest XPU customer in 2027), OpenAI (first XPU "Jalapeño" shipped Q3, expected second-largest in 2028) and Meta (three MTIA generations through 2027) [Q3 FY2026 call; 8-K Apr 6, 2026]. Top five end customers were ~55% of Q3 revenue, and one distributor took 50% [Q3 FY2026 10-Q]. Broadcom is fabless: ~95% of its wafers come from TSMC [Q3 FY2026 10-Q].

2. Core debate

The first question is whether the outlook is contracted demand or financed demand. In its favour: RPO is $179.2B, up from $33.3B at FY2025 year-end after a long-term custom-accelerator contract signed in Q2, and Google signed a multi-year TPU and networking LTA [Q3 FY2026 10-Q; FY2025 10-K; 8-K Apr 6, 2026]. Against it: management says two of the six XPU customers lack the cash flow to fund their compute, so Broadcom built the AI XPV platform with Apollo and Blackstone to finance >20 GW for Anthropic and OpenAI through 2028. Broadcom backstops the first $35B tranche up to ~$29B, and the customer may issue Broadcom up to $42B of convertible notes to meet its lease payments [Q3 FY2026 call; Q3 FY2026 10-Q]. The April 8-K says Anthropic's consumption of its 3.5 GW "is dependent on Anthropic's continued commercial success."

The second question is whether Broadcom keeps its economics per gigawatt as customers grow more powerful. Management puts content at $20–30B per GW and argues co-design IP and first-pass execution are moats even against customers' own teams [Q3 FY2026 call; GS conference Sep 2026]. But Google already has MediaTek building a TPU v8 variant, the 10-Q says top customers demand leases and rack-level deals, and gross margin is falling as memory content rises [Q3 FY2026 call; Q3 FY2026 10-Q].

3. Base thesis

The few labs that buy most of the world's AI compute want accelerators designed around their own models, which management says cost less than half as much as a GPU for equal or better performance on those workloads [Q3 FY2026 call]. Broadcom is the partner with the deepest IP (SerDes, die-to-die links, HBM integration, advanced packaging) and a decade of shipping TPUs without respins. It also owns the Ethernet fabric that connects XPU and GPU clusters alike. Because design cost is largely fixed per programme while revenue scales with gigawatts deployed, AI revenue can roughly double each year while operating margin holds and free cash flow stays near 45% of revenue. VMware is a high-margin annuity underneath that funds the dividend and absorbs an AI air pocket.

Assumptions this rests on:

1. Customers deploy roughly as guided: ~$58B of AI revenue in FY2026, ~$115B in FY2027, ~$230B in FY2028 [Q3 FY2026 call].
2. Broadcom stays the XPU partner of choice across generations, keeping content near $20–30B per GW and networking growing as fast as XPUs [Q3 FY2026 call].
3. Non-GAAP operating margin holds ~66% while gross margin falls on memory mix (Q4 guide: ~66% and ~73%) [Q3 FY2026 call].
4. Anthropic and OpenAI grow into their commitments, so XPV backstops and the notes line are never called.
5. VCF renewals hold after the post-acquisition price reset, keeping software above ~$8B a quarter at 80%+ margins.

4. Main bear case

Broadcom used to be an asset-light chip designer that bought on purchase orders and sold to the biggest balance sheets in technology. In 2026 it became something else: a supplier that has signed $126.8B of unconditional purchase commitments ($0.13B at FY2025 year-end), guarantees a lessor's exposure to one customer, and may lend to that customer to keep lease payments current [FY2025 10-K; Q3 FY2026 10-Q]. The customers set to be largest in 2027–28 are the two that cannot fund themselves. Management calls this "not circular financing" [GS conference Sep 2026]; the risk factors describe the same structures as exposing Broadcom to counterparty credit risk and lower free cash flow and gross margin [Q3 FY2026 10-Q].

How it breaks: a lab's revenue disappoints or its funding slips; deployments slide because land, power and shell aren't ready (management's own stated constraint); Broadcom still pays for wafers, HBM and substrates; racks repossessed under the backstop are worth less because a new generation ships every year. At the same time, concentrated customers keep pushing price and terms (leases, racks, a second source), and VMware's growth fades as the price reset laps.

5. Business quality

The AI franchise has real technical moats: networking leadership (first 100 Tbps Tomahawk 6, now in nearly all AI hyperscalers including GPU-based ones; Tomahawk 7 at 200 Tbps taped out) and XPU co-design with every TPU generation since the first [Q3 FY2026 call; GS conference Sep 2026]. A respin costs about six months and $30M, which raises the cost of switching to a less proven partner [GS conference Sep 2026]. The weakness is bargaining power: six customers, top five at ~55% of revenue and rising (40% a year ago), each large enough to second-source or insource [Q3 FY2026 10-Q].

Software is a high-quality cash business once acquired. FY2025 segment operating income was $20,765M on revenue of $27,029M (77%), up from 65% in FY2024 [FY2025 10-K, computed], achieved by cutting cost and moving customers to VCF subscriptions. Retention data (churn, customer count) is not disclosed. Non-AI semiconductors are a mature, mid-single-digit-growth portfolio.

6. Key value drivers

• Gigawatts deployed × content per GW: management outlines ~30 GW across its customers for FY2027–28 at $20–30B of Broadcom content per GW but guides to "somewhat less" being deployed; analysts' arithmetic on the guide implies ~$11–12B per GW [Q3 FY2026 call].
• XPU vs. networking mix: XPUs 73% of Q3 AI revenue; management expects networking to grow as fast as XPUs [Q3 FY2026 call].
• Gross margin mix: non-GAAP gross margin 78% a year ago, 75% in Q3, ~73% guided for Q4, as XPUs with rising HBM content dilute; semiconductor segment gross margin ~67% in Q3 vs ~70% in Q2 [Q2 and Q3 FY2026 calls].
• Opex leverage: semiconductor opex +22% y/y against revenue +127%; GAAP R&D −5% y/y [Q3 FY2026 call; Q3 FY2026 10-Q].
• VMware ARR and licence timing: ARR growth 19% → 17% → 15% through FY2026; Q3 upfront licence revenue $3,465M (from $1,916M) on non-cancellable contracts [Q1–Q3 FY2026 calls; Q3 FY2026 10-Q].
• Supply: wafers (TSMC ~95%), CoWoS and substrates (own Singapore substrate fab from FY2027), HBM, and lasers (InP capacity more than tripling) [Q3 FY2026 10-Q; Q3 FY2026 call].

7. Financial profile

FY2025: revenue $63,887M (+24%), GAAP operating income $25,484M (40%), net income $23,126M, operating cash flow $27,537M, capex $623M, free cash flow $26,914M (computed), SBC $7,568M; adjusted EBITDA $43B (67%) [FY2025 10-K; Q4 FY2025 call]. FY2024 operating income was $13,463M, depressed by VMware integration and intangible amortization.

9M FY2026: revenue $71,089M (+55%), operating income $35,306M (49.7%), net income $29,747M, operating cash flow $32,950M, capex $1,013M, SBC $6,287M [Q3 FY2026 10-Q]. Q3 alone: revenue $29,591M (+86%), GAAP operating margin 53.9%, non-GAAP 67.9%, GAAP diluted EPS $2.68 (non-GAAP $3.32), free cash flow $13.7B (46% of revenue). Q4 guide: revenue $34.8B, AI $21.7B, non-GAAP operating margin ~66%, capex $1.4B [Q3 FY2026 call].

Balance sheet at August 2, 2026: cash $23,975M; debt principal $61,079M ($2,252M due within 12 months); net debt ~$37.1B (computed); fixed-rate debt at a 4% average coupon and 7.4-year average maturity, with another $1.5B repaid after quarter-end [Q3 FY2026 10-Q; Q3 FY2026 call]. Working capital is building: inventory $4,523M (from $2,270M) and receivables $13,707M (from $7,145M) since FY2025 year-end. Off the income statement: RPO $179.2B (~25% due within 12 months), unconditional purchase commitments $126.8B ($52.7B in FY2027, $73.0B in FY2028), a backstop of up to ~$29B, and a notes line of up to $42B (none issued, nothing paid) [Q3 FY2026 10-Q].

8. Valuation

On trailing numbers the stock is a premium compounder: ~45x TTM GAAP earnings ($38.3B), ~20x EV/TTM revenue ($89.1B) and a ~2.3% TTM free cash flow yield ($39.4B). On management's targets it looks very different: under 12x the ">$30" of FY2028 EPS Hock Tan says Broadcom is "very much on target to exceed" [Q3 FY2026 call]. The gap between the two is the debate. The price assumes most of the FY2027–28 AI outlook arrives and that margins hold. It does not price in any loss on backstops or purchase commitments.

Market data callout — as of market close October 9, 2026 (source: stockanalysis.com web market data, outside sources/; not company filings). Share price $361.54; 52-week range $289.96–$495.00; shares outstanding ~4.77B; market capitalisation ~$1.73T; quoted P/E 46.2x trailing, 20.9x forward; dividend $2.60 a share (0.72$37.1B at August 2, 2026, so EV ≈ $1.76T → ~19.8x TTM revenue; TTM figures are FY2025 + 9M FY2026 − 9M FY2025. Market data is not a company source and is used nowhere else in this file.

9. Management & capital allocation

Hock Tan has been CEO since March 2006; Henry Samueli is Chairman [2026 proxy]. Charlie Kawwas runs semiconductors and Ram Velaga software. Kirsten Spears retired as CFO on June 12, 2026, and Amie Thuener succeeded her [8-K Apr 2, 2026]. Tan beneficially owned 908,474 shares (<1%) at February 2026 [2026 proxy]. His 2025 PSU award, designed to keep him through fiscal 2030, vests only on AI revenue: the best four quarters in FY2028–30 must exceed $60B to pay anything, with target at $90B and maximum (300$120B [8-K Sep 9, 2025]. That pays on revenue, not on returns or on how the revenue is financed, and the FY2027 outlook (~$115B) is already near the maximum. Say-on-pay passed in April 2026 with ~2.43B votes for and ~1.23B against, about a third opposed [8-K Apr 21, 2026, computed].

The capital-allocation record is the core of Broadcom's history: buy durable franchises, cut cost hard, pay out cash. VMware cost ~$84.2B in cash and stock and earned $20.8B of segment operating income in FY2025 (~25% pre-tax on the purchase price, computed) [FY2025 10-K]. FY2025 returned $11.1B in dividends and $2.45B in buybacks; 9M FY2026 returned $9.3B and $8.45B, with $6.1B of net debt repayment [FY2025 10-K; Q3 FY2026 10-Q]. Buybacks have not shrunk the count: shares outstanding rose to 4,774M from 4,741M in 9M FY2026 as SBC outran repurchases [Q3 FY2026 10-Q]. Tan says FY2027 cash will go to higher dividends and buybacks rather than repaying low-cost debt; the board decides in December [GS conference Sep 2026]. The new uses of the balance sheet (XPV backstops, the notes line, purchase commitments, substrate and laser fabs) are a break from this asset-light pattern.

Claims to verify over time: AI revenue ~$115B in FY2027 and ~$230B in FY2028, with supply "secured"; ">$30" FY2028 EPS; $20–30B of content per GW, "sustaining"; XPV backstops are "low risk"; XPUs at "half the cost" of GPUs; VCF "cannot be disintermediated" by AI [Q1 and Q3 FY2026 calls; GS conference Sep 2026]. Credibility marker: Q3 non-AI semiconductor revenue was $4.2B against a $4.5B guide [Q2 and Q3 FY2026 calls], while AI beat ($16.7B vs $16B).

10. Risks ranked by damage

1. Customer credit and vendor financing Damage rank: High · Permanence: High · Fixability: Low Mechanism: Broadcom's largest future customers are labs that depend on outside capital, and Broadcom now guarantees part of that capital and may lend the rest. Evidence today: ~$29B maximum backstop (85% of unpaid lease balance less rack resale value, five-year leases); up to $42B of customer notes; Anthropic's 3.5 GW "dependent on Anthropic's continued commercial success"; risk factors on counterparty credit risk [Q3 FY2026 10-Q; 8-K Apr 6, 2026]. What would confirm it: further XPV tranches with Broadcom backstops, any note issued, any backstop payment, or a lab's financing delayed. Impact: one backstop at its maximum equals about a year of FY2025 free cash flow; revenue, guarantee and inventory losses arrive together.

2. Supply pre-commitments meet deployment slippage Damage rank: High · Permanence: Medium · Fixability: Medium Mechanism: $126.8B of non-cancellable purchases sit against deployments timed by customers' land, power and shell readiness. Evidence today: commitments up from $0.13B; the share of RPO due within 12 months fell from ~35$49B to ~$45B between Q2 and Q3 (computed); inventory and receivables roughly doubled in nine months [Q3 FY2026 10-Q]. What would confirm it: inventory outgrowing revenue, or the outlook cut while commitments stand. Impact: cash tied up in inventory and write-downs, reversing operating leverage.

3. Customer bargaining power and second-sourcing Damage rank: High · Permanence: Medium · Fixability: Medium Mechanism: six customers, each able to split a generation or take design in-house, push price and terms. Evidence today: top five ~55% of Q3 revenue (45% in Q2, 40% a year ago); MediaTek on Google's TPU v8t; customers demanding leases and racks; "customer-owned tooling" named in risk factors [Q2 and Q3 FY2026 10-Qs; Q3 FY2026 call]. What would confirm it: a customer's next XPU generation going elsewhere, or semiconductor gross margin falling faster than memory mix explains. Impact: content per GW and segment margin fall; lost share at a six-customer business doesn't return within a cycle.

4. AI capex cycle Damage rank: High · Permanence: Low · Fixability: Medium Mechanism: demand is a few companies' data-centre budgets; if the economics of frontier models disappoint, orders are cut together. Evidence today: Tan's own figures put global token-generation cost (~$200B a year) above model revenue (~$150B) [GS conference Sep 2026]; the 10-Q warns of a "perceived" AI spending bubble [Q3 FY2026 10-Q]. What would confirm it: customers cutting capex guidance or delaying site builds. Impact: a cyclical downturn in 56% of revenue with fixed R&D for multi-die chips.

5. VMware decay after the price reset Damage rank: Medium · Permanence: Medium · Fixability: Medium Mechanism: growth came from repricing and converting the base to VCF; once lapped, growth reverts and customers migrate at renewal. Evidence today: ARR growth 19% → 17% → 15%; ~79% of Q3's $1,966M y/y software growth was upfront licence revenue (computed); Q4 guided down sequentially to ~$8.7B [Q1–Q3 FY2026 calls; Q3 FY2026 10-Q]. What would confirm it: ARR growth below ~10% or software revenue excluding upfront licence turning flat. Impact: erodes the ~$29B-a-year operating-income cushion (annualised Q3) that funds the dividend.

6. TSMC and Taiwan concentration Damage rank: High · Permanence: Low · Fixability: Low Mechanism: ~95% of wafers and the leading-edge packaging come from TSMC, which also serves competitors and has raised prices [Q3 FY2026 10-Q]. Evidence today: no change in concentration; Broadcom now says it "may enter into long-term capacity commitments" with its CMs [Q3 FY2026 10-Q]. What would confirm it: allocation shortfalls or a regional disruption. Impact: missed shipments across the whole AI book; low probability, severe.

11. Metrics to monitor

12. Thesis tripwires

13. Open questions

• Which customer is backstopped and which may issue the $42B of notes? The 10-Q names neither; the call ties the first tranche to Anthropic.
• How much of the $179.2B RPO is AI, how much is VMware, and how much belongs to customers that need XPV financing?
• What are XPU prices and gross margins by customer, and how much HBM passes through at little or no margin?
• What are VMware churn, customer count and the renewal pricing on the second cycle? Not disclosed.
• What will the Singapore substrate fab and laser expansion cost in capex, and does Broadcom carry utilization risk on them?
• What will the December 2026 board decide on dividends and buybacks? (See §9.)

14. Update log

15. Source map