§ 01Executive View
Broadcom is structurally fragile on the input side but commercially insulated by the same fragility — every credible AI ASIC and high-end Ethernet switch in the world routes through the same TSMC + CoWoS-L + HBM + ABF stack, so AVGO's supply risk is largely shared with the market it sells to. The single most concentrated risk is CoWoS-L allocation at TSMC, where AVGO is the #2 priority customer behind NVIDIA and is fighting Google/Meta/OpenAI demand against a ~150,000 wafer 2026 allocation that is already insufficient. Pass-through power on custom ASICs is exceptional (gross margin ~65% on chip revenue, ~77% blended), which means AVGO can absorb input cost shocks; what it cannot absorb is physical capacity denial, which is the binding constraint for the next 18–24 months.
§ 02Input Map
Tier 1
| Input | Supplier(s) | Concentration | Geography | Substitution | Notes |
|---|---|---|---|---|---|
| Leading-edge logic wafers (3nm, 2nm) for ASICs and Tomahawk 6 | TSMC | Sole | Hsinchu / Tainan / Kaohsiung (TW) | None at node — Samsung 2nm yield-impaired, Intel 18A unproven externally | TH6 confirmed N3; XPU programs migrating N3 → N2 across 2026–2027 (per Broadcom IR Tomahawk 6 release; Tom's Hardware on OpenAI 10GW co-design) |
| Advanced packaging — CoWoS-L | TSMC | Sole at the bridge-LSI variant; CoWoS-S is differently constrained | TW | None for CoWoS-L; CoWoS-S substitutable for some legacy designs | AVGO booked ~150k wafers for 2026 (~15% of TSMC CoWoS), split ~90k Google TPU / ~50k Meta MTIA / ~10k OpenAI; #2 behind NVIDIA's ~595k / ~60% (per DigiTimes, Morgan Stanley via Jukan, Tiger Brokers) |
| HBM3E / HBM4 | SK Hynix (primary), Samsung (secondary, growing on Google), Micron (entering) | Dual-leaning-single per program | Icheon / Cheongju (KR), Hiroshima (JP for Micron's leading-edge) | Within-tier substitutable but qualification 9–12 months | SK Hynix dominant on AVGO-Google in H1, Samsung expanding H2 (TrendForce); Samsung supplies >60% of Google TPU HBM3E (TrendForce Dec 2025) |
| ABF substrates (high-layer-count) | Ibiden (JP), Unimicron (TW), Shinko (JP), Kinsus (TW), Nan Ya PCB (TW) | Oligopoly; high-end tight | JP/TW | Limited — Ajinomoto controls 98% of ABF resin IP licensing | High-layer ABF "tight" through 2026; broader ABF "slightly undersupplied 2026" (DigiTimes April 2026) |
| Optical transceivers / DSPs (for switch ecosystem and CPO) | Innolight (CN), Eoptolink (CN), Coherent (US), Lumentum (US); AVGO supplies its own DSPs and CPO engines | Multi-source on modules; concentrated on lasers | China + US | Module-level substitutable; laser EML supply is the choke | TH6-Davisson uses TSMC COUPE photonic engines (per Broadcom investor news, March 2026) |
| Front-end CMOS for non-ASIC silicon (Tomahawk lower variants, Jericho, Trident, broadband, server PCIe) | TSMC majority; some legacy at GlobalFoundries / SMIC-class fabs not used | Concentrated TSMC | TW | Possible at trailing nodes only | Per AVGO 10-K FY25: "majority of front-end wafer manufacturing operations is outsourced to external foundries, including TSMC" |
| Assembly & test (non-CoWoS) | TSMC, ASE, Foxconn Tech, Amkor, SPIL | Multi-source | TW / KR / SG / VN / Philippines | Yes for standard packages | Per FY25 10-K Item 1 — explicit list |
| Internal fab inputs — FBAR filters (wireless), VCSEL/EEL lasers (GaAs/InP) | Broadcom internal (Fort Collins CO, Breinigsville PA) | Owned | US | N/A | These are the only inputs AVGO does not outsource |
| 224G / 448G SerDes IP | Broadcom internal | Owned | N/A | N/A | The actual moat under the ASIC business; not a supply-chain item |
Tier 2 chokepoints
This is where the real risk sits.
- EUV scanners at TSMC — every 3nm/2nm wafer AVGO buys traces back through ASML EUV (and from N2 onward, increasingly High-NA). ASML is single-source globally. A 6-month maintenance issue at Veldhoven or a single Cymer light-source disruption tightens supply for everyone, but AVGO's CoWoS-L bookings can't move to a different node faster than TSMC can produce the wafers.
- Hybrid bonding tools for HBM4 base die — BESI and Shibaura are the chokepoints. HBM4 base die is now a logic-process item consuming TSMC capacity, which loops the memory chokepoint back through the foundry chokepoint AVGO is already exposed to.
- Ajinomoto ABF resin — single Japanese supplier of the build-up film resin chemistry. The five substrate makers all license from Ajinomoto. A single-site disruption in Kawasaki has no global substitute.
- Ultra-pure neon, krypton, xenon — TSMC's gas inputs. Pre-2022 ~70% of neon for litho came from Ukraine/Russia; that risk has been partially diversified post-2022 but not eliminated.
- EML lasers for 800G/1.6T optics — Coherent and Lumentum are the only Western suppliers at volume. CW-EML for CPO is genuinely tight; this is a Tier-2 issue for the Tomahawk 6 ramp specifically.
- Yangtze base for Innolight — China-located transceiver assembly is exposed to export-control retaliation. AVGO sells silicon into Innolight's modules; a Chinese export ban on optical modules would not stop AVGO's switch ASIC shipments but would stop the system-level deployment its customers need.
§ 03Risk Scoring
| Risk vector | Score (1–5) | Why |
|---|---|---|
| Single-source exposure | 5 | TSMC is sole-source for both leading-edge wafers AND CoWoS-L. ASML upstream is also sole-source. Two stacked single points of failure. |
| Geographic concentration | 5 | Taiwan-anchored on logic, packaging, and substrates; Korea/Japan-anchored on memory. The geopolitical surface is the entire East Asian first island chain. |
| Geopolitical exposure | 4 | Taiwan strait is the dominant tail; AVGO has no direct China revenue concentration, but customer-side optical modules go through China supply chains. |
| Capacity tightness | 5 | CoWoS-L is sold out; AVGO's 150k allocation is the binding constraint on Google TPUv7, Meta MTIA, OpenAI accelerator. The customer wants more than the supplier can produce — for years. |
| Inventory cushion | 3 | AVGO discloses no usable supply-of-days metric for AI accelerators; given lead times and CoWoS allocation cycles, effective forward inventory is months of demand, not a true cushion. |
| Pass-through power | 1 | Among the strongest in semis. ASIC NRE is paid by hyperscalers; per-unit pricing reflects it (TPU at ~$13k/chip, other ASIC programs ~$5k); blended GM ~77%, AI chip GM ~65%. AVGO gets to push input shocks through. |
Synthesis: Five out of six structural vectors score 4–5, but the pass-through vector — usually a footnote — is the one that recasts the whole picture. AVGO does not bear the financial cost of supply-chain shocks the way a fabless competitor like AMD does, because its hyperscaler customers contractually absorb wafer-cost movement and packaging-cost movement on cost-plus-margin custom programs. What AVGO bears is shipment risk: missed CoWoS-L allocation translates one-for-one into revenue not booked. The vector AVGO cannot hedge is physical capacity. The vector it hedges naturally is price.
§ 04Pass-Through Power
Custom ASIC contracts at AVGO are structured closer to a design-services + cost-plus-silicon model than a merchant chip sale. NRE is paid up front, per-unit pricing escalates with packaging and HBM cost, and the customer carries inventory risk. The direct evidence: Q1 FY26 AI revenue was $8.4B (+106% YoY), AI semiconductor GM ~65%, blended GM ~77% (down only ~210 bps YoY despite massive AI mix shift) — input cost did not impair margin. Compare to a reference-design NVIDIA SKU, where pricing power comes from CUDA + integrated rack but COGS exposure is fully on NVIDIA. AVGO's structure is the more defensive of the two on supply shocks; NVIDIA's is the more offensive on demand surprises.
§ 05Stress Scenarios
Scenario 1: Taiwan invasion / blockade (existential, shared with cohort)
- Probability (subjective): Low (3–5% over 3 years), but tail-shaped — small probability, total impact.
- Mechanism: TSMC fabs go offline. CoWoS-L goes offline. Substrate suppliers (Unimicron, Kinsus) go offline. Optical module assembly partially in-China. ASE/SPIL go offline.
- Financial impact: 70–90% revenue loss for 18–36 months until Samsung Foundry / Intel 18A / a hypothetical recovery operation can absorb leading-edge ASIC + advanced packaging. AVGO has no fab. The internal Fort Collins / Breinigsville fabs make FBAR filters and lasers, not accelerators.
- Response options: Long-dated Samsung Foundry qualification (currently SF3/SF2 yield-impaired), Intel Foundry 18A external qualification (Microsoft and DoD are early customers — AVGO would be in line behind). Both are 18–30 month migration timelines from cold start. There is no fast lever.
Scenario 2: CoWoS-L allocation squeeze (the live case, base case for 2026–27)
- Probability: High — already happening. Morgan Stanley/Jukan and DigiTimes both confirm AVGO is at ~150k wafers vs. >200k of latent demand from its three big ASIC programs.
- Mechanism: TSMC prioritizes NVIDIA (60% of capacity) on Rubin/Vera. AVGO's allocation is bounded by what NVIDIA does not take. Each incremental TPU/MTIA/OpenAI unit is a CoWoS-L wafer not produced.
- Financial impact: Quantifiable as the gap between 2027 AI revenue ambition ($100B target by Tan) and what allocation actually permits. If TSMC ramps CoWoS to 127k–150k wafers/month by end-2026 and AVGO sustains 15% share, that's ~225k–270k wafers/year, which mathematically maps to ~$25–35B of AI silicon revenue at current ASPs. The $100B 2027 target requires both more allocation share AND HBM4-driven ASP expansion.
- Response options: Multi-year capacity reservations with TSMC (already in place); some non-CoWoS workloads onto CoWoS-S; internal lobbying to expand TSMC capacity faster than the published 150k roadmap. AVGO has no escape valve outside TSMC — Samsung's I-Cube and Intel's EMIB are not qualified for the 224G SerDes + HBM4 packages AVGO ships.
Scenario 3: HBM tightness on TPUv7 / OpenAI accelerator ramp
- Probability: Mid (40–50% over the 2026 ramp window).
- Mechanism: SK Hynix HBM4 yield ramp is the gating item. Samsung HBM3E qualified into Google's TPU stack (now ~60% of TPU HBM3E per TrendForce) is the swing variable; if Samsung HBM4 slips, AVGO loses a second source on the program with the largest unit volume.
- Financial impact: ~$2–4B revenue deferral risk if HBM4 slips a quarter; gross-margin impact muted by pass-through but the unit shipments lag.
- Response options: Continue dual-sourcing (SK Hynix + Samsung + Micron entering); HBM3E carry-over if HBM4 slips (lower bandwidth, customer-acceptable on some workloads); internal pressure on Samsung to pull-in.
Scenario 4: ABF substrate disruption (Ajinomoto resin or Ibiden site issue)
- Probability: Low–mid (10–15%).
- Mechanism: High-layer ABF (>20 layers) is the package layer for TPU/MTIA/OpenAI accelerators. Ajinomoto is a true single point in the resin step. Ibiden's Ogaki site is the single largest high-layer ABF facility globally.
- Financial impact: ~$1–3B per quarter of disruption.
- Response options: Switch to Unimicron / Shinko / Kinsus for high-layer; degraded yield in transition; some package re-spins (3–6 months).
Scenario 5: Optical-module disruption (China-side, Innolight/Eoptolink)
- Probability: Mid (25–35%) given U.S.–China export-control trajectory.
- Mechanism: A retaliatory Chinese ban on optical-module exports, or a tariff/license regime that adds 6–12 months to import cycles.
- Financial impact: Small at AVGO's silicon line (which does not depend on the modules), but material at the system-deployment line — Tomahawk 6 deployments stall, slowing follow-on switch silicon orders. Estimate ~5–10% AI segment revenue impact at peak.
- Response options: Coherent / Lumentum can scale Western EML and module supply but at 9–12 month qualification lag; CPO at TH6-Davisson reduces module-vendor dependency over time. AVGO is the only switch-silicon supplier that has CPO at production volume in 2026.
§ 06Bull Points
- CoWoS-L #2 priority is structural, not cyclical. TSMC explicitly allocates AVGO behind only NVIDIA. That ranking does not move with a quarter — it reflects multi-year hyperscaler design-win commitments. AVGO is the supply-side gatekeeper of the entire hyperscaler custom-ASIC market.
- Pass-through power is the rare semi profile that converts supply chain risk into customer risk. Hyperscalers absorb wafer/HBM/substrate cost shocks via cost-plus contract structure. AVGO realizes ~65% AI chip GM into a sold-out market; that's the inverse of fragile.
- CPO at TH6-Davisson reduces optical-module Tier-2 dependence sooner than competitors. Broadcom + TSMC COUPE = supply chain step-change. Marvell, Cisco, Arista lag here by 12–18 months.
- Internal manufacturing for FBAR + lasers is a hidden defensive asset. Wireless and optics components that would be supply-chain risks for competitors are owned by AVGO. Fort Collins / Breinigsville give negotiating leverage on optics partners.
- Multi-program ASIC portfolio dilutes single-customer concentration in the supply chain. Top-five customers = ~40% of revenue per FY25 10-K — high but lower than a pure-play merchant accelerator vendor would carry. AVGO captures margin on TPU AND MTIA AND OpenAI AND Maia; NVIDIA is one architecture.
§ 07Bear Points
- Two stacked single points of failure. TSMC + CoWoS-L creates correlated supply risk that no portfolio diversification can hedge. Anything that takes TSMC offline takes AVGO offline.
- The CoWoS-L allocation is a ceiling, not a floor. AVGO's $100B 2027 AI revenue target requires TSMC capacity expansion that is not yet contracted. If TSMC's 150k/month roadmap doesn't ramp to ~200k+ by end-2027, the math doesn't work.
- Customer concentration on the buy side has a supply-side mirror. Google + Meta + OpenAI are >70% of the ASIC book by allocation. Each is a strategic customer who could vertically integrate further (Google could pull more design in-house; Meta has a known habit).
- Samsung HBM3E is a swing variable. Loss of Samsung as a second HBM source on Google TPU narrows AVGO's only redundancy in memory. If SK Hynix has a yield event with no Samsung backstop, AVGO cannot cover.
- Ajinomoto + Ibiden are unsexy but real chokepoints. Substrate risk is the kind that hits during a Q3 earnings call with no warning.
§ 08Conviction (1–5)
4 / 5. On the supply-chain dimension specifically: AVGO is exposed to every chokepoint in the AI compute stack — but it is also prioritized at every chokepoint, and its commercial structure converts supply risk into customer risk via pass-through. That asymmetry is rare and durable. Conviction is not 5 because the Taiwan tail and the CoWoS-L ceiling are both real, and because AVGO has zero internal manufacturing optionality on the AI accelerator side (unlike Intel or Samsung).
§ 09Key Risks to This Read
- The single most load-bearing assumption is that TSMC continues to allocate AVGO at #2 priority. If a hyperscaler walks (e.g., Google moves more design in-house and pulls allocation directly), AVGO's CoWoS share drops without compensating program-level wins.
- Pass-through power on AI ASICs is asserted from gross margin observation, not contract disclosure. AVGO does not publish ASIC contract structure. If contracts are fixed-price rather than cost-plus, supply shocks would impair margin in a way the current data does not show.
- The $100B 2027 AI revenue target is the company's own framing. It assumes capacity expansion, ASP expansion (HBM4), and continued program wins. Independent capacity math (~150k–270k wafers × ASP) does not yet support it without further allocation.
- Disclosure that would change my view: (1) explicit CoWoS-L wafer allocation by program in TSMC's earnings, (2) AVGO breaking out per-program AI revenue / unit volumes, (3) any disclosed ASIC contract pricing structure (cost-plus vs fixed). Currently working from triangulated reporting — the conviction grade reflects that triangulation.
Works cited
- Broadcom FY24 10-K (avgo-20241103)
- Year-over-year comparison of supplier disclosures
- Broadcom FY25 10-K (avgo-20251102, filed Dec 18 2025)
- Sources of supply (TSMC + ASE + Foxconn + Amkor + SPIL)
- internal vs outsourced manufacturing (FBAR, GaAs/InP lasers internal
- CMOS outsourced)
- + 1 more
- Broadcom FY25 10-K (investor static-file copy)
- Backup link to FY25 10-K
- Broadcom Inc. 10-K (FY2024, FY2025) — Legal Proceedings, Risk Factors, Geographic Revenue
- Litigation roster baseline
- export-control / regulatory risk-factor language
- geographic revenue split for tariff/sanctions exposure quantification
- Broadcom Inc. 2025 Annual Report
- FY2025 GAAP financials
- segment results
- deleveraging update
- Broadcom Inc. Form 10-K (filed 12/20/2024, FY2024)
- FY2024 income statement, balance sheet, cash flow
- VMware purchase accounting impact
- debt schedule
- Broadcom Inc. Form 10-K (FY2024)
- Customer concentration risk factor
- segment disclosure (Semiconductor Solutions vs Infrastructure Software)
- >10% customer pattern
- Broadcom Inc. Form 10-K (FY2024) — debt schedule, geographic revenue mix, FX disclosure
- Debt maturity ladder framing for refi-rate sensitivity
- revenue by geography (Americas/China/Asia/EMEA splits)
- functional-currency disclosure
- + 1 more
- Broadcom Q1 FY2026 Financial Results press release (Mar 4, 2026)
- Q1 FY26 revenue $19.3B (+29%), AI semi $8.4B (+106%), Q2 guide $22B incl. AI $10.7B
- $100B+ AI 2027 line of sight
- Broadcom Q4 and FY2025 Financial Results press release (Dec 11, 2025)
- FY2025 revenue $63.9B (+24%), AI semi $20B (+65%), software $27B (+26%), segment splits, EBITDA
- Broadcom Ships Tomahawk 6: World's First 102.4 Tbps Switch — Broadcom Investor Relations
- Volume shipment confirmation
- co-packaged optics support
- unified scale-up + scale-out fabric
- Broadcom Q4 2025 Earnings Call Transcript — The Motley Fool
- AI backlog $10B → $73B q/q
- XPU customer #5 signed at $1B
- 3-5 year recurring revenue per XPU customer
- Broadcom Q4 FY2025 earnings call transcript
- Three named hyperscaler ASIC customers language
- Hock Tan AI SAM commentary
- VMware ELA conversion progress
- 650 Group - Ethernet to Surge in Scale-Out and Ramp in Scale-Up
- Scale-up TAM split: NVLink $25B+, Ethernet $8B+, UALink $3B+ by 2030
- AI Server Compute ASIC Shipments to Triple by 2027 — Counterpoint Research
- Custom ASIC market growing ~45% in 2026
- ~$118B by 2033
- share commentary
- AInvest - Mapping the 2026 Semiconductor Cycle
- Inventory cycle commentary
- GB200 build-up partial resolution
- 800G/1.6T cycle ahead-not-behind framing
- AWS Trainium3 Deep Dive — A Potential Challenger Approaching — SemiAnalysis
- Marvell loss of Trainium 3 socket to Alchip on RDL interposer execution
- Trainium 3 design split (Annapurna front-end, Alchip back-end physical and package)
- Bloomberg Intelligence - AI Accelerator Market $600B by 2033
- Custom ASIC 27% CAGR to $118B by 2033
- ASIC share trajectory 8% (2024) -> 19% (2033)
- structural CAGR cross-check
- Broadcom Custom Chip Strategy Targets $60B Hyperscaler AI Infrastructure Market — Alphastreet
- Broadcom ~70% of custom AI ASIC market
- anchor programs at Google TPU, Meta MTIA, Microsoft Maia
- Broadcom Q4 FY 2025 Earnings: AI And Software Drive Beat — Futurum Group
- Q4 FY25 AI semi revenue $6.5B (+74% YoY)
- Q1 FY26 guide $8.2B (~100% YoY)
- ~100bps gross-margin compression on AI mix
- + 1 more
- Broadcom vs Marvell: Custom AI Silicon Battle 2026 — HeyGoTrade
- AVGO ~70% / MRVL ~25-35% ASIC co-design share split
- Marvell design wins (Trainium, Maia, DPU, Axion)
- Deloitte Insights - 2026 Semiconductor Industry Outlook
- $975B 2026 industry size
- cycle position framing
- structural shift away from traditional cyclicality
- Futurum Group - Broadcom Q1 FY 2026 Earnings Driven by XPU Momentum
- Q1 FY26 AI revenue $8.4B (+106% YoY)
- networking division +60% YoY
- customer expansion narrative
- Futurum — Broadcom Q4 FY25 earnings recap
- AI revenue trajectory and gross margin commentary
- InfiniBand vs Ethernet: Broadcom and NVIDIA Scale-Out Tech War — TrendForce
- Tomahawk 6 102.4 Tbps shipping ahead of NVIDIA Spectrum-X1600 (H2 2026)
- generational lead positioning
- Inside the Custom AI Chip Race: Google, AWS, Microsoft, Meta, OpenAI — Hashrate Index
- Hyperscaler-to-design-partner mapping
- ByteDance + OpenAI 2027 attribution to Broadcom
- Intel Market Research - Data Center Ethernet Switch Chips Market Outlook 2026-2034
- Ethernet switch silicon TAM $3.8B (2025) -> $4.2B (2026) -> $7.1B (2034)
- 6.8% CAGR
- top-5 concentration
- io-Fund - Broadcom: Silent Winner in the AI Monetization Supercycle
- Gross margin ~65% on AI chip sales
- rack-scale margin compression risk
- Anthropic $21B Ironwood Rack order
- Marvell's Custom XPU Pipeline Is A Declaration Of AI Independence — The Next Platform
- Marvell competitive positioning, customer roster context, XPU pipeline vs Broadcom
- Maximize Market Research - Hypervisor Market 2026-2032
- vSphere 84% hypervisor share
- Hyper-V 4%, Nutanix AHV 2%
- market structure HHI inputs
- Mordor Intelligence - Data Center Switch Market
- Broader DC switch system TAM $19.4B (2026) -> $28.5B (2031) at 8.0% CAGR
- Mordor Intelligence - Virtualization Software Market 2025-2031
- Virtualization software TAM $110B (2026)
- VMware ~35% share
- competitive shift commentary
- Morgan Stanley CoWoS analysis (via Jukan / X)
- CoWoS allocation framework
- 40-50% cloud AI chip surge by 2026
- Omdia - AI data center chip market forecast (Aug 2025)
- $286B AI DC chip TAM by 2030
- ASIC share-of-TAM cross-check
- SemiAnalysis (via cohort synthesis) — Broadcom as structurally defensible AI bet
- Margin-capture-on-every-hyperscaler-ASIC thesis
- relative attractiveness vs. pure-play AI silicon names
- SemiAnalysis - Ultra Ethernet UEC vs UALink vs Broadcom Scale-Up Ethernet (SUE)
- Scale-up Ethernet (SUE) framing
- Broadcom withdrawal from UALink for own path
- StockTitan — Broadcom FY25 10-K AI / VMware / chip risk summary
- Customer concentration (top-5 ~40%, distributors 48% of revenue)
- Tikr — Broadcom CEO Hock Tan $100B AI 2027 framing
- AVGO management ambition for 2027 AI revenue
- used to stress-test CoWoS-L capacity math
- TrendForce - InfiniBand vs Ethernet: Broadcom and NVIDIA Scale-Out Tech War
- Spectrum-X $2B+/quarter at GCP/Meta/Azure/OCI
- back-end network share dynamics
- Worldwide Ethernet Switch Market Q1 2025 — IDC Quarterly Tracker
- Ethernet switch market +32.3% Q1 2025
- NVIDIA Ethernet revenue +760% YoY to $1.46B
- 12.4% total / 21.2% DC share
- Amazon AI Chip Gambit Backed by Taiwan Supply Chain — CommonWealth Magazine
- Morgan Stanley 1.5M+ Trainium chip estimate for 2026 with Alchip as principal design partner
- AWS CoWoS wafer growth 5k → 70k
- Broadcom takes a Tomahawk to Nvidia's AI networking empire — The Register
- Tomahawk 6 architecture details
- 224G SerDes leadership for scale-up + scale-out unification
- Caproasia — Broadcom completes $69B VMware acquisition (Nov 22, 2023)
- VMware deal close date, $61B cash + $8B debt structure for goodwill/intangibles step-up
- Marvell Stock Downgraded as Concerns Grow Over Amazon Trainium Transitions — Yahoo Finance
- Benchmark 'high-conviction' downgrade on Trainium 3/4 loss
- JPMorgan dissenting Overweight as the contested view
- OpenAI taps Broadcom to build its first AI processor — NBC News
- OpenAI-Broadcom $10B custom chip program
- 10 GW deployment target H2 2026 onward
- The Great Ethernet Pivot: Broadcom Begins Volume Shipments of 102.4 Tbps AI Switch — FinancialContent
- March 2026 volume-ship date
- Broadcom ~80% high-end Ethernet share
- Why Nvidia just poured $2 billion into AI ASIC competitor Marvell — Tom's Hardware
- $2B NVIDIA-Marvell equity investment March 2026
- soft ecosystem lock-in framing vs UALink
- Alphastreet - Broadcom Custom Chip Strategy Targets $60B Hyperscaler AI Infrastructure Market
- Original Hock Tan $60-90B by FY27 SAM framing
- 3-customer baseline
- Anthropic — Google TPU $10B / 400k-unit commitment (public disclosure 2025)
- Pull-through demand quality validation for TPUv7 ramp
- Astute Group — NVIDIA secures 60% of CoWoS capacity
- NVIDIA CoWoS dominance leaving AVGO at #2 priority
- AT&T Services Inc. v. Broadcom Inc. and VMware LLC (NY Sup. Ct., Commercial Div., filed Nov 21, 2024)
- Active material litigation on VMware perpetual-license forced migration
- discovery flow into EC/CMA enforcement files
- H1–H2 2026 calendar
- BIS Advanced Computing and Semiconductor Manufacturing Items export rule (Oct 2022 initial)
- Foreign Direct Product Rule baseline for advanced computing exports to PRC
- framing AVGO networking-silicon and direct-China-channel exposure
- BIS export-control update (Oct 17, 2023)
- FDPR expansion and performance thresholds — calibration for which AVGO SKUs fall inside vs outside controls
- BIS HBM and AI-chip package rule (Dec 2, 2024)
- HBM rule and packaging-level controls
- secondary effect on Chinese ASIC competitor capability vs AVGO
- Broadcom Inc. - Q4 FY2025 / Q1 FY2026 earnings releases & calls
- Q1 FY26 AI revenue $8.4B (+106% YoY)
- customer count 3 -> 5 -> 6
- $73B AI backlog with 18-month delivery
- + 1 more
- Broadcom Inc. — Avago/Broadcom Singapore-to-Delaware redomicile (2018) public record
- HQ/IP domicile classification (US-jurisdictional, no Singapore/Cayman exposure)
- Broadcom IR — Tomahawk 6 / TH6-Davisson shipping in production volume (March 12 2026)
- TH6 production start
- TSMC 3nm process
- 102.4 Tbps switching capacity
- Broadcom IR — Tomahawk 6 Davisson CPO Ethernet switch announcement
- TSMC COUPE photonic engine integration
- CPO disintermediating optical-module supply chain
- China SAMR conditional approval of Broadcom / VMware (Nov 21, 2023)
- Five-year behavioral commitments (non-discrimination, interoperability) baseline
- 2027 sunset / renegotiation catalyst
- CHIPS Program Office — TSMC Arizona Preliminary Memorandum of Terms (April 8, 2024)
- Indirect AVGO benefit via tier-one TSMC partner
- CHIPS conditions (China guardrails, claw-back, dividend/buyback restrictions) bound on TSMC, not AVGO
- Chipstrat — Coherent's vertical integration strategy
- Optical component-layer competition vs Broadcom/Lumentum
- CW-EML supply tightness
- Cignal AI — 800GbE optics shipments to grow 60% in 2025
- Optical-module market sizing and supplier ranking (Innolight, Coherent, Eoptolink, Lumentum)
- DigiTimes — ABF substrate sells out for Unimicron, Kinsus, Nan Ya PCB (April 2026)
- Substrate tightness on AVGO accelerator packaging
- warping/thermal issues in CoWoS
- DigiTimes — Advanced packaging drives ABF substrate expansion (Dec 2025)
- ABF capacity dynamics across Ibiden, Unimicron, Shinko
- DigiTimes — Kinsus ABF view, high-end ABF tight by 2026
- Differentiation between high-layer ABF (tight) and base ABF (adequate)
- DigiTimes — TSMC CoWoS capacity expansion with NVIDIA booking >50% for 2026-27 (Dec 10 2025)
- CoWoS allocation context, NVIDIA majority share, AVGO #2 priority
- EU Corporate Sustainability Reporting Directive (Directive 2022/2464/EU) and ESRS technical standards
- CSRD in-scope status via VMware EMEA
- FY2025 reporting cycle compliance lift
- European Commission, Case M.10806 — Broadcom / VMware
- EU clearance with interoperability commitments
- ongoing post-clearance monitoring file
- baseline for any Article 8(5) modification analysis
- HeyGoTrade - Broadcom vs Marvell: Custom AI Silicon Battle 2026
- Broadcom 60-80% / Marvell 20-25% AI ASIC share split
- market HHI input
- HPCwire - Upscale AI Eyes Late 2026 for Scale-Up UALink Switch
- UALink 1.0 silicon timing: 2H 2026 lab samples, 2027 in product
- I/O Fund — Broadcom Silent Winner of AI Monetization Supercycle
- Pass-through power evidence: AI chip GM ~65%, blended GM ~77%, ASIC ASP differentiation ($13k TPU vs $5k other)
- IRC Section 48D Advanced Manufacturing Investment Credit — final Treasury/IRS regulations (2024)
- 25% ITC pass-through to wafer pricing
- indirect tailwind to AVGO COGS
- Macrotrends — AVGO P/E ratio history 2012–2026
- 5-year P/E range and average for historical valuation comparison
- NVIDIA Unveils NVLink Fusion for Industry to Build Semi-Custom AI Infrastructure — NVIDIA Newsroom
- NVLink Fusion launch May 2025
- partner roster (MediaTek, Marvell, Alchip, Astera Labs, Synopsys, Cadence
- later Samsung Foundry, Arm)
- Orchestrator cohort context (customer dimension brief)
- Hyperscaler ASIC customer roster (Google, Meta, Microsoft, ByteDance, OpenAI 2027, Apple AFM)
- 'three named hyperscaler ASIC' disclosure pattern
- Orchestrator cohort context (macro dimension brief)
- Post-VMware ~$70B debt and de-leveraging glide path
- revenue/cost FX mix (USD-invoiced revenue, Asia TWD/JPY/KRW cost base)
- VMware counter-cyclical software cushion
- + 2 more
- SEC Climate-Related Disclosures rule (March 6, 2024) and subsequent rescission (2025)
- US climate-disclosure compliance burden — currently low net cost following rescission
- StockAnalysis — AVGO balance sheet (FY22–FY25)
- Cash, debt, goodwill, intangibles, equity, working capital line items
- StockAnalysis — AVGO cash flow statement (FY22–FY25, TTM Feb '26)
- OCF/capex/FCF/SBC/dividends/buybacks/debt activity for cash quality table
- StockAnalysis — AVGO income statement (FY22–FY25)
- Revenue/margin/EPS time series FY22–FY25 for trajectory table
- StockAnalysis — AVGO statistics & valuation
- Current EV, EV/Revenue, EV/EBITDA, P/E (trailing & forward), ROIC, debt/equity, FCF yield
- StockAnalysis — peer multiples (NVDA, AMD, MRVL)
- Peer EV/EBITDA, forward P/E, EV/Sales for valuation comparison table
- techovedas — How Broadcom and TSMC are dominating custom AI chips 2026
- CoWoS-L (Local Silicon Interconnect) bridge-LSI technology framing
- The Register - Broadcom Tomahawk 6 102.4 Tbps coverage (June 2025)
- Tomahawk 6 102.4 Tbps technical specs
- Spectrum-X1600 timing roughly one year behind
- Tiger Brokers / itiger — TSMC CoWoS to 127,000 wafers/month, NVIDIA / Broadcom / AMD ranking
- AVGO ~150k wafers / 15% share, ~90k Google TPU / ~50k Meta MTIA / ~10k OpenAI
- Tom's Hardware — OpenAI and Broadcom 10GW custom AI chip co-development, deployments 2026
- Scale of OpenAI ASIC program (10GW) and 2026 deployment timing
- TrendForce — Broadcom $10B OpenAI custom AI chip order, 2026 AI sales lift
- OpenAI custom chip program scale and 2026 AI revenue trajectory
- TrendForce — Samsung supplies >60% of Google TPU HBM3E in 2026
- HBM dual-sourcing on AVGO ASIC programs (Samsung primary on Google TPU
- SK Hynix dominant on Broadcom in H1)
- TweakTown — SK Hynix HBM order to Broadcom (Google, Meta, ByteDance ASICs)
- SK Hynix as primary HBM supplier into Broadcom ASIC programs
- UK CMA, Broadcom / VMware Phase 1 decision (ME/7022/23, Aug 21, 2023)
- CMA clearance baseline
- subsequent CMA information request 2025
- Section 18 abuse-of-dominance scoping risk
- Ultra Ethernet Consortium - UEC 2025 in Review (2026 roadmap)
- UEC 1.0 spec out
- 2026 priorities (PCM, CSIG)
- standardization horizon and silicon-vs-protocol layer dynamics
- US Department of Commerce — Section 232 semiconductor investigation (initiated April 2025)
- Tariff catalyst on Taiwan-fabbed wafers
- 270-day statutory clock
- CHIPS-Arizona carve-out scenario analysis
- ValueInvesting.io — AVGO EV/EBITDA history
- 10-year EV/EBITDA min/median/max for historical band
- Cohort note: 'The Semiconductor Industry: A Beginner's Companion'
- TPU/MTIA/Maia/OpenAI/Apple-AFM hyperscaler ASIC partnership roster
- Cohort synthesis.md (sections 3.4 three bottlenecks, 3.7 custom silicon, 3.8 China parallel stack, 5 tailwinds/headwinds table)
- AI capex super-cycle scale (~$600B 2026 hyperscaler capex)
- TSMC/CoWoS chokepoint framing
- Taiwan tail-risk framing
- + 2 more
- Cohort synthesis.md — sections 3.6 (CUDA moat) and 3.7 (custom silicon counter-leverage)
- Hyperscaler counter-leverage framing
- pricing-power constraint logic
- Cohort synthesis.md — sections 3.7 (custom silicon) and 5 (tailwinds/headwinds)
- $60–90B AI SAM 2027 framing
- SemiAnalysis 'AVGO captures margin on every TPU/MTIA/OpenAI chip' thesis
- companies.json — AVGO entry (id 5)
- Sentiment +2
- supporting quotes
- catalysts (Tomahawk 6, 224G/448G SerDes, OpenAI 2027)
- + 5 more