AVGO — Research Report (2026-07-20)

TL;DR

• What it is: Broadcom designs custom AI chips + networking silicon, and (via VMware) sells the software that runs corporate data centers.
• Take: A rare "picks-and-shovels" AI winner with software-like margins — but priced for near-perfect execution.
• Revenue $63.9B FY25, +24% YoY (AI drove almost all the growth).
• AI chip revenue $56B guided for FY26, ~180% growth (the entire investment case).
• Free cash flow ~$32.8B (real cash, ~51% of revenue — extraordinary).
• Forward P/E ~20× on next year's earnings (sector average ~23× — not obviously expensive on forward numbers).
• Bull: Morgan Stanley $502 PT — "core AI winner," ~80% of Google's TPU custom-chip work.
• Bear: Trailing P/E ~62× with net debt $45B; a 12–13% drop after beating earnings shows how much is priced in.
• Watch next: Whether Q3 FY26 AI revenue actually hits the $16B guide (which was below the $17.2B Street hoped for).

Price $370.83, market cap ~$1.76T (7th most valuable company on earth) as of this report. Yahoo Finance · stockanalysis.com

1. Business overview

So what: Broadcom is two businesses fused — a high-margin custom-chip maker riding the AI wave, and a software cash machine (VMware) it squeezes for profit.

Broadcom sells into two segments:

Source: FourWeekMBA segment breakdown, from FY25 10-K.

Two pieces of jargon to unpack up front:

• XPU / custom ASIC — a chip designed for one customer's specific job (like Google's AI training). "ASIC" = Application-Specific Integrated Circuit. Because it's tailor-made, it beats a general-purpose GPU (Nvidia's product) on cost and power for that one job. Broadcom is the #1 designer-for-hire of these.
• Virtualization (VMware) — software that lets one physical server pretend to be many, so companies run more workloads on less hardware. VMware is the market standard; Broadcom bought it in 2023 for ~$69B.

The AI story is the whole story right now: in Q2 FY26 (reported June 3, 2026), AI semiconductor revenue hit a record $10.8B, up 143% YoY (Yahoo Finance).

Customer concentration is real but undisclosed by name — the big custom-chip customers are widely understood to be Google (TPU), Meta, and increasingly OpenAI/Anthropic-linked programs (CIO Dive).

2. Financials (5-year)

So what: Revenue and profit exploded — but the two step-changes (2021, 2024) are acquisitions, not pure organic growth. The constant is enormous, growing cash flow.

Source: EDGAR US-GAAP company facts (CIK 0001730168). *FY2020 uses a partial-year revenue tag; treat as directional. †FY24 operating income was depressed by VMware acquisition/integration charges (amortization of intangibles) — operating cash flow rose that year, confirming it's an accounting drag, not a cash problem.

• Revenue CAGR FY21→FY25 ≈ 23%/yr. Why it matters: this is faster than the semi market — Broadcom is gaining share, but VMware (2023) inflates the trend.
• Gross margin ~68%, operating margin ~40% in FY25. Why it matters: these are software-like margins on a company that's majority hardware — the sign of pricing power (custom chips + a software monopoly).
• FCF ~$27–33B depending on quarter timing (recent 8-Ks show trailing FCF ~$32.8B). Why it matters: FCF is cash left after keeping the business running — the "anti-story" number. AVGO converts >50% of revenue to free cash, which almost no hardware company does. (stocktitan)

Note the capex is tiny (~$0.6B on $64B revenue = ~1%). This is the crux of the business model: Broadcom is fabless — it designs chips and pays TSMC to manufacture them. It doesn't own the multi-billion-dollar factories. That's why FCF is so high (see §8).

Balance sheet

So what: Debt is large in dollars but comfortably covered by cash flow.

Source: EDGAR + stocktitan.

Why net debt/cash-flow matters: above ~3× is fragile territory. At ~1.6×, one year of operating cash could cover more than half the net debt — this is not a stretched balance sheet, despite the big absolute number.

One data caveat: the EDGAR CommonStockSharesOutstanding tag jumps from ~418M (FY22) to ~4,139M (FY23) — that's a 10-for-1 stock split in 2024, not real dilution. Share count is ~4.74B today; the growth from 4.14B→4.74B reflects the VMware deal (paid partly in stock) plus buybacks partly offsetting. Treat per-share history before mid-2024 as pre-split.

3. Unit economics — the one number

So what: The business runs on the gap between what a custom AI chip costs to design-and-source and what a hyperscaler will pay to avoid buying Nvidia — and that gap is currently enormous.

There's no single "take rate" like a marketplace, but the number that runs Broadcom is AI semiconductor gross margin combined with backlog visibility:

• Q2 FY26: shipping $10.8B of AI silicon per quarter against ~$30B of AI bookings — meaning contracted revenue is visible out toward 2028 (Investing.com).
• On the software side, the operating number is even starker: Infrastructure Software gross margin 93%, operating margin ~77–78% after Broadcom forced VMware's 10,000 largest customers off perpetual licenses onto subscriptions (Futurum).

Why this matters: 78% operating margin means for every $1 of VMware revenue, ~78¢ is profit. That's the cash engine funding the dividend and buyback while the AI chip business scales.

4. Metrics table (with the "why")

Sources: stockanalysis.com, EDGAR, FY25 press release.

5. Ownership & management

So what: No controlling owner, standard index-fund top holders — but heavy, consistent insider selling, led by the CEO.

Source: whoistheownerof.com, TIKR.

• Insider signal is negative: ~64% of insiders are net sellers over the past year; CEO Hock Tan sold ~$50M at ~$337/share (Yahoo Finance).
• How to read this: At companies with big equity-comp packages, routine insider selling is normal (diversification, taxes) and isn't automatically bearish. But when 64% are selling and none are buying on the open market, it's at least not a vote of confidence at these prices. Hock Tan is a legendary operator (he built Broadcom through serial acquisition), so his continued leadership is itself the bull's asset; his selling is the caveat.

6. Market & value-chain position

So what: Broadcom is a bottleneck, not a commodity — it's the near-monopoly designer of custom AI chips — but it sits between two more powerful players: TSMC (who makes the chips) and the hyperscalers (who could in theory design their own).

The value chain for a custom AI chip:

Hyperscaler (Google/Meta/OpenAI) — wants a cheaper alternative to Nvidia
        ↓ pays to design
   BROADCOM — designs the custom XPU + networking  ← bottleneck, pricing power
        ↓ pays to manufacture
   TSMC — fabricates the silicon                    ← the real physical bottleneck
        ↓
   Deployed in AI data centers

• Where AVGO has power: Designing a leading-edge custom AI chip requires IP (intellectual property blocks), networking know-how, and a decade of relationships. Morgan Stanley estimates Broadcom keeps ~80% of Google's TPU business even as MediaTek tries to enter (TipRanks). That's duopoly-to-monopoly structure — exactly the "bottleneck with monopoly structure" this vault's discovery method hunts for.
• Who has power over AVGO: (1) TSMC — the actual factory bottleneck; if TSMC capacity is tight, Broadcom queues like everyone else. (2) The customers themselves — a hyperscaler is Broadcom's partner and its potential competitor (they could hire their own chip team). Concentration is the risk: a handful of customers drive AI revenue.

Competitors: Marvell (the other custom-ASIC house), Nvidia (the general-purpose GPU alternative customers are trying to escape), MediaTek (entering TPU). On software: no real VMware substitute at scale, which is why Broadcom can raise prices.

7. Valuation

So what: On forward earnings and cash flow AVGO looks reasonable; on trailing earnings it looks extreme. Which lens you trust depends entirely on whether you believe the AI revenue guide.

Sector benchmark source: bear-case analysis cites forward P/E "~35× vs ~23× sector average" (Bitget/valuation summary); note different data providers compute AVGO's forward P/E anywhere from ~20× to ~35× depending on whether they use GAAP or non-GAAP and which forward year — a real reason to be careful.

What a reasonable buyer might pay: This is not a company anyone acquires (it's $1.76T). For a shareholder, the honest read: at 20× a credible next-year earnings number with 20%+ growth and 50%+ FCF conversion, the price is defensible if the AI guide holds. The Street consensus 12-month target is **$523** (~41% above current) with a Strong Buy rating (23 buys, 3 holds) (MarketBeat). That consensus is itself a bet on the FY27 "$100B+ AI revenue" guide — which is ~2.7× the entire FY25 semiconductor segment. That's the number to be skeptical of.

Bull / base / bear

Catalysts (bull): FY27 $100B+ AI guide confirmed by new customer wins; VMware price hikes continue to lift software margin; a fourth/fifth major custom-chip customer disclosed.

What would prove the thesis wrong: (1) A big hyperscaler in-sources its chip design or defects to Marvell/MediaTek — the concentration risk fires. (2) AI capex among hyperscalers slows (the whole "insatiable demand" premise breaks). (3) The Q3 FY26 AI guide of $16B — already below the $17.2B Street hoped for — missing, which would crack the growth narrative the entire valuation rests on. The June post-earnings drop of 12–13% despite a beat is the market rehearsing exactly this fear.

📚 What this company teaches

1. Fabless > fab-owning for cash generation. Broadcom makes ~$33B of free cash flow on ~$0.6B of capex because it designs chips and lets TSMC own the factories. The lesson: who owns the capital-heavy step of a value chain earns lower returns on capital than who owns the high-IP, asset-light step. Compare Broadcom's ~1% capex/revenue to a chip manufacturer's 30%+.

2. "Custom" can be a moat, not a cost problem. Conventional wisdom says customization kills margins (you can't reuse the work). Broadcom flips this: because designing a leading-edge AI chip is so hard, being the only credible design partner gives it ~80% share and pricing power. Specialization created a bottleneck.

3. A trailing P/E of 62 and a forward P/E of 20 can describe the same stock. When earnings are ramping fast, the trailing multiple looks insane and the forward multiple looks normal. The gap between them is the growth expectation — and the risk. Always ask which multiple a bull is quoting.

Self-test question: Broadcom's forward P/E (~20×) is below its sector average (~23×), yet its trailing P/E (~62×) is nearly 3× the sector. How can both be true, and which one should worry a skeptical investor more — and why?

(Answer sketch: earnings are growing so fast that next year's E is much larger than last year's, collapsing the forward multiple. The skeptic should worry about the trailing/forward gap itself — it means the entire valuation depends on the forecasted earnings actually arriving. If growth disappoints, the stock re-rates toward the ugly trailing number.)

📖 Glossary

• P/E (price-to-earnings) — Stock price divided by earnings per share; how many dollars you pay for $1 of annual profit. "Forward" uses next year's expected earnings; "trailing" uses the last reported year.
• EV/EBITDA — Enterprise value (market cap + debt − cash) divided by operating cash proxy; the standard for comparing companies with different debt loads. (Not the primary lens here — AVGO is profitable, so P/E fits better.)
• FCF (free cash flow) — Operating cash flow minus capex; the cash actually left over after keeping the business running. The "anti-story" number because it's hard to fake.
• OCF (operating cash flow) — Cash generated by the core business before investment spending.
• Capex (capital expenditure) — Money spent on physical assets (factories, equipment). Low capex = asset-light.
• Fabless — A chip company that designs chips but outsources manufacturing to a foundry (like TSMC). Broadcom is fabless.
• Foundry — A contract chip manufacturer (e.g., TSMC). Owns the multi-billion-dollar factories ("fabs").
• ASIC / XPU — Application-Specific Integrated Circuit; a chip built for one customer's specific task. Broadcom's "XPU" is its custom AI accelerator. Contrast with a general-purpose GPU (Nvidia).
• TPU — Tensor Processing Unit; Google's custom AI chip, which Broadcom co-designs.
• Hyperscaler — A giant cloud/AI operator (Google, Meta, Amazon, Microsoft, OpenAI-scale) that buys chips in enormous volume.
• Virtualization — Software (VMware's core product) that splits one physical server into many virtual ones.
• Net debt — Total debt minus cash; the debt that would actually remain if you used all cash to pay it down.
• Net debt / OCF — How many years of operating cash flow it would take to clear net debt; a leverage stress test.
• Backlog / bookings — Contracted future revenue not yet delivered; the more of the market cap it covers, the more "already sold" the future is.
• Revenue CAGR — Compound annual growth rate of revenue; average yearly growth over a period.
• Perpetual vs subscription license — Old software model: pay once, own forever. New model: pay every year. Broadcom forced VMware customers onto subscriptions, which raised recurring revenue and margin.

Not financial advice — personal research per the vault's [[investment-research]] process. Primary source: EDGAR company facts (CIK 0001730168); all forward figures and prices are as of 2026-07-20 and cited inline. No position taken here — a buy/sell/watch call requires a journal entry per §6 of the equity-research workflow.

Terminal state: DONE.