Broadcom (AVGO)
Fabless semiconductor IP + co-design (custom XPUs sold to hyperscalers, merchant Ethernet switching/routing silicon) plus recurring infrastructure-software subscriptions (VMware/mainframe/security). Capital-light, outsources fabrication to TSMC; monetizes via per-unit chip sales, design wins, and software license/subscription revenue.
The thesis on this name
State of AI Compute
The toll-collector on the AI-silicon arms race across THREE shift-points — custom-XPU co-design (~95% with Marvell), first-to-volume Tomahawk 6 merchant-Ethernet, and CPO — at ~70% semi GM and the cheapest multiple of the networking/ASIC cluster (~26-37x); the most-bankable name, but verify warns the per-program map is 6-month-old priced consensus and the '…
State of AI Compute
Long AVGO over a 12-24mo horizon as the diversified arms-dealer toll-collector across the custom-XPU + merchant-AI-Ethernet build-out (FY27 AI guide >$100B on a $73B committed backlog), where the structural edge is a co-design + Ethernet-fabric moat the sell-side keeps treating as a single-customer ASIC bet — entered…
State of AI Compute
Diversified XPU+Ethernet arms-dealer toll; clip half ($100k) now on the ~19% pullback, second half only on a clean Q3 AI print or a re-test to ~$340-350. Hedge/pair is a HARD precondition (size an NVDA/semicap basket sh…
State of Nvidia
The custom-ASIC threat AND the comp: lower ASIC GM (~50-60s%) is the level toward which NVDA's erodible-inference pool reprices. Growing fa…
Earnings, margins, COGS & capex
Broadcom is in a step-change inflection: Q2 FY26 revenue hit a record $22.2B (+48% YoY) as AI semiconductor revenue reached $10.8B (+143% YoY), now ~49% of total (fact). Margins are exceptional — 77% non-GAAP gross, 67% non-GAAP operating, ~69% adj. EBITDA — funding ~$10B+ quarterly free cash flow on a near-fabless ~1% capex intensity (fact). Management guides Q3 FY26 to $29.4B (+84% YoY) with AI semis at $16.0B (+200%+ YoY) and reiterates >$100B FY27 AI revenue against a $73B committed backlog (fact/guide). The non-AI semiconductor cyclical lines and the VMware-anchored software segment (9% YoY growth, ~32% of mix) provide a high-margin, recurring counterweight to AI lumpiness.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~23¢ is cost of goods and ~10¢ operating expense, leaving ~67¢ of operating profit (~42¢ net).
Revenue trend
Margins
stable/up — mix shift to high-margin AI + VMware subscriptions
up — record operating leverage
up
stable — high cash conversion, capital-light
up — highest in company history (FY25 ~68%)
COGS structure
COGS is dominated by (1) TSMC wafer/foundry purchases for leading-edge custom XPUs and networking ASICs, plus advanced packaging (CoWoS) and HBM content passed through on XPU systems; (2) test/assembly and substrate costs; (3) software COGS is minimal (amortization of acquired VMware intangibles, support delivery). AI XPU 'systems' carry lower gross margin than Broadcom's merchant chips because third-party HBM/packaging is bundled — a headwind partially offset by very-high-margin software and the rising Ethernet networking mix. Estimate: blended GM held at 77% despite the AI ramp because of software + networking offset.
Capex
~$231M in Q2 FY26 (~1% of revenue) — fact. Capital-light fabless model: capex funds design tools/EDA, test infrastructure, lab/validation, and facilities, not fabs (fabrication is outsourced to TSMC). This is the structural FCF advantage vs. capital-intensive IDMs and vs. the hyperscaler customers spending hundreds of billions on data centers.
Latest earnings
Mixed beat: non-GAAP EPS $2.44 beat consensus ~$2.40; revenue $22.19B was a touch light vs ~$22.27B consensus. Stock fell ~15% post-print — not on the numbers but because management did NOT raise the $100B FY27 AI target despite the blowout AI quarter (expectations were running ahead of the guide).
Q3 FY26 revenue ~$29.4B (+84% YoY, above ~$28.5B consensus); AI semiconductor revenue ~$16.0B (+200%+ YoY); non-GAAP operating margin ~67%; adj. EBITDA ~68%. FY27 AI semiconductor revenue reiterated 'in excess of $100B' against a $73B backlog.
- AI semiconductor revenue
- $10.8B, +143% YoY (~49% of total)
- AI networking share of AI rev
- ~40% (Ethernet fabric)
- Committed backlog
- $73B (~18mo visibility)
- Free cash flow
- $10.26B, 46% of revenue
Growth drivers
- Custom AI accelerators (XPUs) for hyperscalers — six core custom-silicon customers (Google, Meta, Anthropic, OpenAI, + two more incl. Apple per earlier disclosure); contracted gigawatt deployments (OpenAI 10GW by 2029, Meta 3GW through 2028, Anthropic 1GW+ in 2026 scaling to 5GW from 2027)
- Merchant AI networking — Ethernet switching/routing (Tomahawk, Jericho) scaling AI clusters; ~40% of Q2 AI revenue, the fabric that connects XPUs and GPUs alike
- VMware monetization — conversion of perpetual licenses to high-margin subscription bundles (VCF), driving software segment margin and recurring revenue
- $73B committed backlog providing ~18 months of AI revenue visibility toward the >$100B FY27 target
- Hyperscaler capex super-cycle + the structural push to diversify away from sole-source Nvidia toward custom silicon to lower $/token and own the roadmap
- Non-AI semiconductor cyclical recovery (broadband, server storage, wireless) as a secondary tailwind
Reported financials — SEC EDGAR
Audited GAAP figures pulled from SEC filings · latest filing 2025-12-18. The audited primary-source spine — not financial advice.
Revenue — annual (GAAP)
Margins & balance sheet — FY’25
Bull & bear
Long AVGO over 12-24 months as the diversified arms-dealer toll-collector on the AI build-out: it co-designs the custom XPUs AND sells the merchant Ethernet fabric, collecting on the data-center build regardless of which accelerator brand wins, with a $73B backlog underwriting a >$100B FY27 AI guide.
- Structural co-design + Ethernet-fabric moat: switching customer XPUs is multi-year and costly, and Broadcom's open-Ethernet stack (Tomahawk/Jericho) networks both its own XPUs and Nvidia GPUs — it wins the fabric either way
- $73B committed backlog (~18 months visibility) de-risks the >$100B FY27 AI target far more than typical semi forward estimates; contracted gigawatt deals (OpenAI/Meta/Anthropic) point beyond it
- Elite economics — 77% gross / 67% operating / ~69% EBITDA margins, ~46% FCF margin on ~1% capex — convert the AI ramp into prodigious cash with minimal reinvestment drag
- Diversification cushion: high-margin VMware/VCF software (recurring) + non-AI semis dampen the AI cyclicality that pressures pure-plays
- Customer base = the entities with the deepest pockets and the strongest incentive to custom-silicon away from Nvidia's margin, aligning Broadcom with the secular cost-down of AI inference
Avoid/short case: the stock prices a flawless multi-year hyperscaler-capex super-cycle into a name with extreme customer concentration; any AI-capex digestion, a lost design win, or simply failing to raise targets (as Q2's ~15% drop showed) re-rates a ~24x-forward, $1.9T cap hard.
- Expectations are the risk: Q2 FY26 beat on EPS yet the stock fell ~15% purely because the $100B FY27 target wasn't raised — the bar is set above the guide
- Customer concentration is acute — a few hyperscalers drive AI revenue; an in-sourcing decision, a logo loss to Marvell, or one customer's capex pause is materially damaging
- AI-capex cyclicality: gigawatt 'commitments' are multi-year and revisable; an AI-spend air-pocket would hit Broadcom's most-leveraged line first
- Margin mix risk — bundled HBM/packaging in XPU systems dilutes incremental gross margin as the AI mix grows past 50% of revenue
- Balance-sheet + valuation overhang: ~$45B net debt from VMware plus a premium multiple leave little room for a stumble; MRVL's ~218% YTD run shows where the marginal AI-beta dollar can flow
What it is worth
Forward P/E on non-GAAP EPS, cross-checked vs. FCF yield and the FY27 AI-revenue ramp. Trades ~24x forward EPS (~$9 FY-fwd consensus) at ~$379 / ~$1.87T cap (fact, late June 2026).
~$215-300
an AI-capex digestion, a lost design win to Marvell, or multiple de-rating toward the 52-wk low ($263) / the bearish ~$216 street minimum.
~$430-470
AI ramp on track to the guide, multiple modestly compresses from ~24x as growth normalizes; roughly in line with consensus average over 12-24mo.
~$525-670
sell-side consensus ~$524-526 average with a high of ~$671; assumes FY27 AI revenue meets/exceeds $100B and multiple holds on continued backlog growth.
Premium is underwritten by the $73B backlog and >$100B FY27 AI guide; the binary is whether AI capex sustains the ramp — Q2's ~15% drop on an unchanged target shows how much is priced in.
SWOT
Strengths
- Dominant custom AI ASIC franchise (~70% of the custom-accelerator market by estimate) co-designing XPUs with the largest AI spenders — Google, Meta, OpenAI, Anthropic
- Industry-standard merchant Ethernet switching/routing silicon (Tomahawk/Jericho) — the open-standard fabric that scales AI clusters, capturing networking spend regardless of whose accelerator wins
- Best-in-class financial model — 77% gross, 67% operating, ~69% adj. EBITDA margins with ~46% FCF margin on ~1% capex intensity (fact, Q2 FY26)
- $73B committed backlog gives rare multi-quarter revenue visibility for a semiconductor company
- High-margin recurring software (VMware/VCF) diversifies away from chip cyclicality and funds the dividend/buyback
Weaknesses
- Severe customer concentration — a handful of hyperscalers drive the AI franchise; loss or in-sourcing by any one is material
- AI XPU 'systems' carry bundled third-party HBM/packaging that dilutes gross margin vs. pure merchant silicon
- ~$64.9B total / ~$45B net debt from the VMware acquisition (fact); rate/refinancing and integration overhang
- VMware price/bundling changes have alienated some enterprise customers, capping software-segment upside
- Revenue is increasingly tied to discretionary hyperscaler capex — a single budget pause cascades fast
Opportunities
- Contracted gigawatt roadmaps (OpenAI 10GW by 2029, Meta 3GW, Anthropic 5GW from 2027) imply AI revenue well beyond the $100B FY27 marker if they convert
- Capturing a larger share of the AI rack — co-packaged optics, scale-up/scale-out Ethernet, and NICs as clusters grow
- New custom-silicon logos (sovereign AI, additional hyperscalers/AI labs) expanding the six-customer base
- Mix shift toward higher-margin networking and software within the AI bill of materials
- Non-AI semis cyclical recovery (broadband, wireless, storage) as a secondary leg
Threats
- Nvidia's full-stack (GPU + NVLink + Spectrum-X Ethernet/InfiniBand) competing for the same networking and accelerator dollars
- Marvell as the credible #2 custom-ASIC house, intensifying competition for hyperscaler design wins (MRVL up ~218% YTD vs AVGO ~10%)
- Hyperscaler multi-sourcing / in-housing — customers deliberately diversify away from any single XPU vendor
- An AI-capex air-pocket or 'digestion' phase would hit the most cyclical part of the story hardest
- Valuation/expectations risk — the ~15% post-Q2 drop on an unchanged target shows the bar is set very high
Moats, dependencies & bottlenecks
Moats
high (multi-year per design cycle) Hyperscalers co-architect silicon with Broadcom over 2-3 year cycles; switching vendors mid-roadmap forfeits sunk IP and time-to-market — sticky once designed in.
Tomahawk/Jericho are the de-facto open-standard switching silicon for scale-out AI; ~40% of AI revenue and brand-agnostic — networks Nvidia GPUs as well as Broadcom XPUs.
Volume and engineering depth secure priority allocation of leading-edge wafers and advanced packaging; hard for a sub-scale entrant to match on cost and supply.
VMware/VCF is embedded in enterprise private-cloud estates; migration is costly, underpinning recurring high-margin revenue — though aggressive repricing risks goodwill.
Deep SerDes, networking and connectivity IP raises the bar for new custom-ASIC entrants, but Marvell shows the moat is contestable at the high end.
Dependencies
Single leading-edge foundry dependency; wafer and CoWoS capacity gate AI XPU output and any disruption is unhedgeable near-term.
A handful of customers drive AI revenue; a budget pause, in-sourcing, or logo loss is material. Concentration is the dominant risk.
XPU systems bundle third-party HBM; tight HBM supply/pricing affects deliverability and margin on AI systems.
Revenue increasingly tied to discretionary AI infrastructure spend; an air-pocket hits the most-leveraged line first.
US export rules and any antitrust scrutiny (VMware) can constrain end markets and customer geographies.
Advantages
- Wins on both sides of the AI rack — custom accelerators AND the merchant Ethernet fabric — so it collects regardless of whose GPU/XPU brand prevails
- Capital-light fabless model — ~1% capex intensity and ~46% FCF margin while customers shoulder hundreds of billions in data-center capex
- $73B committed backlog provides revenue visibility almost no other semi company can claim
- Diversified high-margin software (VMware) recurring revenue cushions chip cyclicality and funds capital return
- Best-in-class margins (77% gross / 67% op / ~69% EBITDA) give pricing and reinvestment headroom rivals lack
- Deep, multi-year co-design relationships with the deepest-pocketed AI buyers, aligned with their incentive to custom-silicon away from Nvidia
Weaknesses
- Acute customer concentration in the AI franchise — a few hyperscalers drive the growth
- Increasing exposure to a single cyclical demand driver (AI capex) with limited hedge if it digests
- ~$64.9B total / ~$45B net debt from VMware creates leverage and refinancing sensitivity
- Bundled third-party HBM/packaging dilutes incremental gross margin as AI mix exceeds 50% of revenue
- VMware repricing/bundling has strained enterprise goodwill, capping software upside
- Premium valuation + sky-high expectations leave little margin for error (Q2 ~15% drop on an unchanged target)
Bottlenecks
- TSMC leading-edge wafer + CoWoS/advanced-packaging capacity caps how fast XPU and networking volume can ramp
- HBM supply (SK Hynix/Samsung/Micron) gating AI XPU system shipments and pressuring system gross margin
- Customer concentration — growth is bottlenecked by the spend and roadmap decisions of a few hyperscalers
- Design/engineering cycle time — each custom XPU is a multi-quarter co-design effort, limiting how fast new logos convert to revenue
- Data-center power/gigawatt availability at the customer end (deals are quoted in GW) can pace real-world deployment vs. the order book
Top signals & trends
Top signals
An upward revision or a 7th custom customer would validate the bull case; the Q2 ~15% drop shows the market is waiting on a raise.
Backlog growth = forward visibility extending; a flattening backlog would be an early warning of digestion.
Rising Ethernet mix is higher-margin and brand-agnostic — a healthy diversifier within AI.
Broadcom's AI revenue tracks customer capex; any cut to 2026/27 capex plans is a direct read-through.
MRVL up ~218% YTD; new Marvell wins at Broadcom accounts would signal share erosion.
Watch whether bundled HBM/packaging compresses blended GM as AI passes 50% of mix.
Trends
The core secular tailwind for Broadcom's custom-ASIC franchise; aligned with customers' incentive to reduce Nvidia dependence.
Open-standard Ethernet (Tomahawk/Jericho) gaining vs InfiniBand as clusters reach 100k+ accelerators; ~40% of AI revenue.
OpenAI/Anthropic gigawatt deals show non-hyperscaler AI labs entering custom silicon — new TAM beyond the original cloud four.
Customers diversify XPU suppliers (Marvell, internal teams) to avoid sole-source dependence — caps Broadcom's share even as TAM grows.
Constrains AI system output and margin; a bottleneck Broadcom shares with the whole AI supply chain.
Any sign of an AI-spend digestion phase would hit Broadcom's most-leveraged revenue line and its premium multiple.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Leading-edge foundry + CoWoS advanced packaging for all custom XPUs and networking ASICs; the critical single dependency.
HBM memory bundled into AI XPU systems; supply/pricing affects deliverability and margin.
HBM and memory supplier into AI systems; alternate foundry/packaging source.
US HBM/DRAM supplier into AI compute systems.
OSAT assembly/test partners for packaging and final test of Broadcom silicon.
EDA tools and IP underpinning custom-silicon design flows.
Long-standing TPU custom-silicon partner and core XPU customer; also runs in-house design.
MTIA XPU partner; ~3GW deployment planned through 2028, first 1GW order delivering H2 2027.
Custom silicon in production late 2026; contracted 1.3GW in 2027 within a 10GW-by-2029 agreement (private company).
TPU-based compute access deal — 1GW+ in 2026 scaling to 5GW from 2027 (private company).
Custom-silicon/connectivity customer (disclosed among the six core custom accounts) and large wireless-component buyer.
Thousands of enterprises consuming VMware Cloud Foundation subscriptions — the recurring software customer base.
Full-stack rival — merchant GPUs plus NVLink and Spectrum-X/InfiniBand networking compete for both accelerator and fabric dollars; the incumbent Broadcom's custom-XPU thesis is built to undercut.
The credible #2 custom-ASIC house (Trainium, Maia, Alphabet work); pure-play AI profile, up ~218% YTD vs AVGO ~10%. Direct competition for hyperscaler design wins.
Merchant MI-series GPUs and (via acquisitions) networking; an alternative to both Nvidia GPUs and custom XPUs for some workloads.
Competes in AI data-center Ethernet switching at the systems level (using merchant silicon); a fabric competitor at the box layer.
Google (TPU), Amazon (Trainium/Inferentia via Annapurna), Microsoft (Maia) build internal teams that can reduce reliance on Broadcom over time — customer and competitor.
Connectivity/interconnect silicon (retimers, PCIe/CXL fabric) competing for parts of the AI-rack interconnect BOM.