Broadcom: The only merchant supplier of bespoke silicon to the AI frontier labs

Stevie AI on Broadcom Inc. (AVGO-USA | broadcominca)

10/26/2026

Summary

Broadcom has quietly become the indispensable second architect of frontier AI compute. While Nvidia sells merchant GPUs to everyone, Broadcom co-designs custom XPUs exclusively with the handful of hyperscalers and frontier labs — Google, Anthropic, OpenAI, Meta — that have decided bespoke silicon, tuned to their own model architectures, beats general-purpose GPUs on cost-per-token at scale. This is not a commoditized foundry relationship; it is a multi-year co-development partnership with enormous switching costs once a customer's model roadmap is baked into a chip generation. Combined with VMware's software base and a dominant position in AI networking silicon (Tomahawk, optical DSPs), Broadcom now sits at the center of two structurally growing, high-margin franchises rather than one cyclical semiconductor business. The financial inflection is already visible in the historicals: revenue jumped from $51.6B in FY2024 to $63.9B in FY2025, while net income nearly quadrupled from $5.9B to $23.1B and EPS went from $1.23 to $4.77, reflecting operating leverage from VMware integration and early AI silicon ramp. Management's guidance embeds a further step change — $118B revenue and $9.52 EPS in FY2026, scaling to $272B revenue and $25.39 EPS by FY2028 as AI semiconductor revenue compounds from a $58B run-rate to a guided $230B, with free cash flow climbing from $44.5B to $113.0B over the same window and the balance sheet flipping from $37B net debt to $63B net cash. We apply a 30x multiple to FY2028 EPS of $25.39, reflecting a business transitioning from semiconductor-cyclical to AI-infrastructure-compounder multiple territory, comparable to best-in-class platform semiconductor and software hybrids, yielding a price target of $761.70, discounted back to $692.40 for FY2027 and $629.50 for FY2026 to reflect the path of execution risk, and rolled forward to $913.60 for FY2029 on $30.40 EPS at a slightly moderated 30x as growth decelerates. This implies substantial upside from the current $325.81 price across the forecast horizon, with the stock still pricing in material skepticism about the durability of the AI silicon ramp beyond FY2027.

Thesis

1. **Uncontested position in custom AI silicon**: Broadcom is not fighting Nvidia for GPU market share; it has created an entirely different category — frontier-lab-specific XPUs co-designed with the end customer's model architecture in mind. With six customers anchoring the roadmap and Google, Anthropic, OpenAI, and Meta publicly committing to multi-gigawatt deployment scales through 2028, Broadcom has no direct merchant competitor in this niche. The Blue Ocean framing is apt: delivering GPU-comparable performance at a fraction of the power and cost profile for workloads that are increasingly inference-dominated is a genuinely differentiated value proposition, not an incremental one. 2. **Software stabilizes the model while silicon drives the growth**: The Infrastructure Software segment, anchored by VMware, provides a high-margin, subscription-like revenue base (~30% of revenue) that smooths the cyclicality historically associated with pure-play semiconductor names. As VMware acquisition-related amortization rolls off through FY2027-2028, this segment's contribution to operating margin expansion becomes a structural, low-risk tailwind independent of AI silicon execution. 3. **Financial trajectory is extraordinary by any semiconductor standard**: Revenue scaling from $63.9B (FY2025) to a guided $321.2B (FY2029) represents a 5x expansion in four years, with EPS moving from $4.77 to $30.40 over the same period. Free cash flow conversion is exceptional — $135.5B of FCF on $321.2B revenue by FY2029 — funding both the buyback program and a full deleveraging from $37B net debt to $148B net cash. This is a business self-funding an enormous capacity buildout while simultaneously returning capital and delevering, a combination rarely available in capital-intensive semiconductor franchises. 4. **Margin trough is transient, not structural**: The anticipated gross margin compression into the high-60s/low-70s through FY2026-2027 reflects a temporary mix shift toward lower-margin XPU hardware as AI revenue scales faster than software. This is a mix effect, not a competitive or pricing problem — margins recover as the software mix share stabilizes and as scale efficiencies in substrate and packaging offset custom silicon's inherently lower gross margin versus merchant chips. 5. **Market still underwrites a cyclical semiconductor multiple on a platform business**: At $325.81, the stock trades at roughly 34x FY2026 EPS of $9.52 and under 13x FY2028 EPS of $25.39 — a multiple compression path that assumes either AI revenue guidance proves overly optimistic or that margins never recover. Given management's disclosed visibility into customer deployment roadmaps (specific gigawatt commitments from each of the four largest AI customers), we think the FY2027-2028 guidance has a higher probability of being met or exceeded than consensus is pricing, creating the mispricing opportunity. 6. **Catalysts are near-dated and verifiable**: The December 2026 Q4 earnings report will be the first hard checkpoint against the $34.8B/$21.7B AI revenue guide, and subsequent customer deployment announcements (TPU v8i, Ironwood scaling) will provide continuous validation points through FY2027-2028. Each successful quarter de-risks the multi-year guidance and should compress the gap between current trading multiple and the platform-quality multiple we believe the business warrants.

Risks

1. **Customer concentration is extreme**: Six XPU customers drive the overwhelming majority of AI semiconductor growth, with Anthropic and OpenAI alone estimated to represent over $100B, or roughly 43%, of the $230B FY2028 AI revenue target. Any slowdown in capital spending, strategic pivot to in-house silicon, or renegotiation of these relationships would materially impair the growth trajectory that underpins our price targets. 2. **Execution risk on an unprecedented scaling cadence**: Management is guiding to AI revenue effectively doubling in both FY2027 and FY2028 off an already-large base. Historically, no semiconductor company has sustained this magnitude of dollar growth without supply chain, yield, or demand-timing disruptions. The FY2029 deceleration embedded in our own forecast acknowledges that this cadence cannot continue indefinitely, and any pull-forward of that deceleration into FY2027-2028 would materially reduce the price targets. 3. **Physical infrastructure constraints are explicitly flagged by management**: Land, power, and shell infrastructure constraints for data center buildouts were specifically called out as a risk to customers' ability to deploy the committed gigawatt capacity. If power availability or permitting delays slow hyperscaler buildouts, Broadcom's revenue recognition could lag even if underlying demand remains intact. 4. **Gross margin trajectory assumes a mix recovery that may not materialize on schedule**: The thesis depends on software mix and scale efficiencies offsetting XPU dilution by FY2028. If custom silicon mix continues to grow faster than software or if packaging/substrate costs for advanced optical interconnects remain elevated longer than modeled, margin recovery could be delayed, pressuring the EPS and multiple assumptions in later years. 5. **Capital intensity for substrate and optical capacity is front-loaded and could disappoint on ROI**: Significant capex for Singapore substrate production and more-than-tripled EML/CW/VCSEL/InP capacity through FY2027-2028 assumes sustained demand justifies this buildout. Any moderation in AI infrastructure spending industry-wide would leave Broadcom with excess capacity and depressed incremental returns on this capital. 6. **Valuation assumes multiple expansion that is not guaranteed**: Our 30x forward multiple on FY2028-2029 EPS assumes the market re-rates Broadcom from a semiconductor-cyclical multiple to a platform/compounder multiple. If AI capex cycles prove more cyclical than secular, or if competitors emerge in the custom silicon co-design market (from Marvell, or hyperscalers building fully in-house design teams), the multiple re-rating could fail to materialize even if the underlying earnings numbers are met.

📈 Price Targets