Layer 1 · AVGO — Broadcom Inc.
One-line thesis
Broadcom is the only credible non-NVIDIA design partner for hyperscaler custom AI silicon at scale — Google TPU, Meta MTIA, ByteDance, the rumoured OpenAI XPU — and a confirmed $73B AI backlog into a $100B 2027 AI revenue target makes the June 3 print a structurally bullish setup.
What Broadcom physically does
Broadcom sits in two physically distinct semiconductor businesses that together form one of the cleanest AI bottlenecks in the stack. The first is AI custom silicon — the co-design of hyperscaler-specific accelerators (Google's TPU v5e, v5p, v6 Trillium; Meta's MTIA v1 and v2; ByteDance's internal accelerator; and the strongly rumoured OpenAI Titan / XPU due 2026-27) where Broadcom owns the SerDes, the packaging integration, and the back-end physical implementation while the customer owns the architecture and the compiler. Roughly 50% of the world's hyperscaler custom-accelerator volume rolls through Broadcom's Avago / LSI design teams in San Jose, Singapore, and Bangalore. The second business is AI networking — the Tomahawk 5 and Tomahawk 6 Ethernet switch silicon (51.2T and 102.4T per ASIC respectively), the Jericho 3 routing silicon, the Bailly 51.2T co-packaged optics module, and the 800G / 1.6T SerDes / PAM4 DSPs that move data inside and between AI clusters. This is the "second silicon" that NVIDIA Spectrum-X competes with — and where Cisco / Arista / Marvell / Broadcom compete with each other for the merchant share that hyperscalers refuse to single-source to NVIDIA InfiniBand.
In plain English: every hyperscaler AI rack that is not pure NVIDIA-NVLink contains Broadcom silicon, usually multiple Broadcom dies, in the switching layer. Every credible non-NVIDIA training cluster (Google's TPU pods, Meta's MTIA-based RSC, the AWS Trainium clusters at Anthropic) contains Broadcom-co-designed accelerators or Broadcom networking or both. The company is the only Tier-1 fab-less semiconductor vendor with a meaningful design-services franchise for AI silicon — Marvell is the only direct comp, and Broadcom's share of the custom-XPU market is materially larger.
The non-AI half of Broadcom — VMware, mainframe software, broadband, wireless — generates roughly half of corporate cash flow but is structurally low-growth. The AI half is what re-rates the stock and is the entirety of the equity thesis.
The financial print
Q1 FY26 (reported March 4 2026, quarter ended Feb) printed $14.92B of revenue against $14.62B consensus, with AI semiconductor revenue at $8.4B (+106% YoY) and total semiconductor revenue at $11.7B (+62% YoY). Non-GAAP EPS of $2.10 beat consensus by 9 cents. The headline that moved the stock was the disclosed $73B AI backlog and CEO Hock Tan's guidance that the AI revenue line should hit a $100B annualized run-rate by end-2027. Software revenue at $5.4B grew 47% YoY (VMware re-pricing still flowing through).
Q2 FY26 prints June 3 after-hours. Consensus sits at approximately $15.05B of revenue and $2.40 of non-GAAP EPS, with whisper expectations as high as $2.55 on the back of continued AI semiconductor strength. The Goldman PT raise to $500 from $475 on May 12 and the broader Wall Street move toward $500-550 PTs ahead of the print reflect a tape that is set up for a beat — the risk is that the bar is now high enough that an in-line print causes consolidation.
The stock at $423 sits roughly 8% below the all-time high of $462 set in mid-April. 1m +12%, 3m +18%. Market cap above $2T puts AVGO firmly in the top-5 US semiconductor capitalisation tier alongside NVDA, TSM and the AMD/QCOM cohort below.
Customer mix
The AI custom-silicon revenue is heavily concentrated in three hyperscaler customers — Google (TPU), Meta (MTIA), and ByteDance — which together represent approximately 80% of the AI custom-silicon line. The rumoured fourth and fifth customers (OpenAI / Titan, and either Apple or Microsoft for inference silicon) move from rumour to confirmed during 2026-27 on the current sell-side consensus path. On the networking side, the customer mix is broader — every hyperscaler buys Tomahawk and Jericho silicon for some portion of their fleet, with Cisco / Arista / Juniper as the system-integration vendors in between. Concentration risk in AI silicon is genuine; concentration risk in networking is much lower because the customer set spans hyperscalers, enterprise, and the AI-native cloud build-outs (CoreWeave, Lambda, Oracle Cloud).
Google is likely Broadcom's largest single customer at roughly 25-30% of AI revenue, given the TPU v6 (Trillium) volume that ramped through 2025 and the v7 production now ramping through 2026. Meta MTIA v2 is the second-largest contributor at roughly 15-20% of AI revenue. ByteDance — the dark-horse customer that Western sell-side typically under-attributes — likely represents 10-15% of the AI line, tied to TikTok and Doubao internal AI build-out. The remaining ~30-40% of the AI custom-silicon line is distributed across smaller wins and the design-services engagements that have not yet shipped at production volume.
The networking franchise has a different structure: every hyperscaler buys Tomahawk silicon at multiple data-center generations, but the volume is mediated through systems vendors (Cisco using the Silicon One competing platform, Arista using Tomahawk as the primary platform, Juniper using mixed solutions). The software business (VMware, mainframe, broadband, wireless) goes to enterprise and operator customer sets that are entirely separate from the AI silicon book — this provides the cash-flow ballast that allows Broadcom to fund the AI custom-silicon R&D investment without the equity equivalent of the pure-AI-cycle volatility of Marvell.
Competitive context
Marvell is the primary competitor in custom AI silicon — AWS Trainium 3 is Marvell-designed, the Microsoft Maia speculation is Marvell-flavoured, and the broader custom-ASIC design-services market is essentially a Broadcom-Marvell duopoly. Broadcom's structural advantage is scale (Avago's back-end IP library is materially larger than Marvell's), packaging integration (Broadcom owns more of the 2.5D / CoWoS-S interface IP), and the bundled networking offering. In switching silicon, Cisco's Silicon One and Marvell's Teralynx are the direct competitors at the merchant 51.2T / 102.4T tier; Broadcom Tomahawk maintains roughly 60% merchant share and is the default reference platform every hyperscaler validates against.
The VMware acquisition is now 24 months into integration and the once-controversial price-up strategy has delivered the cash-flow accretion Hock Tan promised, with very little customer churn beyond the long-tail enterprise base. That cash flow is what funds the AI design-services investment.
Terminal risk
The terminal risk is the hyperscaler-bypass scenario: if a Google or a Meta moves to fully in-house design teams and cuts Broadcom from the TPU / MTIA roadmap on a future generation, the AI custom-silicon line gets a 25-30% air-pocket on a single product transition. There is precedent — Apple did exactly this with its modem effort against Qualcomm — and the only reason it has not happened with Broadcom is the design complexity at 3nm and 2nm nodes, where hyperscaler internal teams genuinely cannot execute alone. If the AI-stack node moves to a model where the customer fully owns physical design (Google has hinted at this for v7), Broadcom's value-add compresses materially.
Bull case
The disclosed path is $73B of contracted AI backlog converting through 2026-27 plus net new design wins lifting the FY27 AI revenue run-rate to $100B, on company guidance. If that lands, total FY27 revenue prints near $80-85B (FY25 was $51.6B), non-GAAP EPS at $10.50-12.00, and at a 35x forward — below the 5-year average and consistent with the AI-quality re-rate that has happened across the cohort — the stock is a $370-420 print, +5-25% off spot but with the option value of an OpenAI Titan announcement.
The upside case is OpenAI Titan confirmed as a Broadcom design partnership in 2026 (rumoured but unconfirmed), Apple inference silicon design-win added in 2027, and the AI revenue line passing $120B run-rate by FY28. That scenario prints the stock toward $550-620 in 2027, +30-46% from spot. Goldman's $500 PT and the broader $500-550 sell-side cohort are tracking this case at a discount.
The structural reason the upside case is credible is that hyperscaler custom-silicon spend is converging on a multi-vendor procurement structure where every major operator validates designs at both Broadcom and Marvell. As the number of custom-XPU customers expands from three (Google, Meta, ByteDance) to a credible five-to-seven (adding OpenAI, Apple, Microsoft for inference, possibly Tesla for Dojo successor), the design-services TAM materially expands. Broadcom is positioned to capture 50-60% of this expanded TAM given the SerDes IP advantage and the depth of packaging-integration capability. The conservative-base-case versus upside-case differential is principally the question of whether one or two of these rumoured wins convert during 2026.
Gap / bear case
Two things the market may be missing the wrong way. First, the $73B AI backlog figure is a multi-year revenue book, not a one-year guide — converting that to a $100B FY27 run-rate requires net-new wins beyond what is already booked, and the OpenAI Titan / Apple inference contributions are in the consensus model but not yet contractually papered. Second, the VMware software re-pricing tailwind is now in the base, and the FY27 software comp gets harder as VMware ARR has now substantially re-rated to subscription pricing.
The third concern is concentration: if any one of Google, Meta, or ByteDance pulls a design generation in-house, the AI custom-silicon line gets choppy in a way the bull case currently dismisses.
Optionality
OpenAI Titan / custom XPU disclosure during 2026 — this is the largest single discrete option in the name, and the rumours have moved from "speculative" to "concrete enough that Hock Tan is responding 'no comment'" on the earnings calls. The reported timing on Titan first silicon is late 2026, with mass-production volume in 2027 — a $4-6B annual revenue line at peak if the program scales to multi-100k-chip-per-year deployment. Apple inference silicon partnership (rumoured for Apple's data-center inference build-out around Apple Intelligence) is the second-largest option — Apple historically does not use external design partners, but the publicly disclosed pivot to data-center inference compute for the Apple Intelligence rollout requires custom silicon that Apple's internal team has not been resourced to deliver in time for the 2026-27 deployment.
The slower-burn option is hyperscaler co-packaged optics adoption ramping in 2027-28 (Bailly 51.2T CPO) — a Broadcom-led category that could add a $3-5B incremental annual revenue line within five years. CPO eliminates the pluggable optical transceiver and integrates the optical engine directly into the switch ASIC, dramatically reducing power per bit at the expense of higher manufacturing complexity. Broadcom is the share leader at the CPO development tier; first commercial deployments are expected at Microsoft / Google in 2027-28.
The fourth option is the broader networking displacement of NVIDIA InfiniBand. As hyperscalers move toward open-Ethernet AI clusters (the UALink / Ultra Ethernet Consortium effort, jointly backed by Broadcom, AMD, Cisco, Microsoft, Meta and others), the Tomahawk 6 silicon at 102.4T is positioned to capture share that NVIDIA Spectrum-X and Quantum InfiniBand currently hold. Even modest share displacement in the AI networking spend ($30B+ annually by 2027 on consensus) materially accretes to Broadcom.
The trade
AVGO — BUY 8.5/10. Entry: starter at $420-425 spot; size up on $390-400 pullback into the 50-day moving average if the broader semiconductor cohort consolidates ahead of the June 3 print. Position size: 3-4% of NLV as the second-largest AI Layer-1 position behind NVDA. Stop: daily close below $370 (50-day MA breach plus reversal of the April highs). Catalyst date: June 3 Q2 FY26 print after-hours; any OpenAI Titan / Apple inference disclosure during 2026-27; Q3 FY26 print early September. Trim/exit triggers: AI semiconductor YoY growth decelerating below 70% on a single quarter; verifiable loss of Google TPU v7 design partnership; software margin contraction signalling VMware integration fatigue. Conviction: the cleanest non-NVIDIA Layer-1 AI exposure in the public market, with a software cash-flow ballast that makes the equity less GPU-cycle-sensitive than peers. The June 3 print is the binary.