Broadcom and Marvell are positioned to benefit from the shift to AI inference, where custom ASICs are gaining share versus GPUs; Goldman Sachs sees ASIC demand catching up to GPU demand by 2027, with a 50/50 split in AI servers by next year. The article argues both companies could nearly double earnings over the next two years, implying potential stock gains of about 74% for Marvell to $229 and 81% for Broadcom to $687. Overall, the piece is bullish on custom AI processor demand and the long-term earnings outlook for both chipmakers.
The market is increasingly pricing a two-layer AI capex stack: training remains GPU-led, but inference is becoming a custom-silicon procurement cycle where hyperscalers optimize for unit economics, latency, and power. That shifts bargaining power away from the incumbent GPU supplier at the margin and toward the design houses that can embed workload-specific performance into the customer’s own architecture. The second-order winner set extends beyond the obvious two names: cloud titans that monetize lower inference cost through margin expansion, and advanced packaging / foundry ecosystems that benefit from a broader custom-chip mix rather than a single-vendor GPU pull-through. The key competitive nuance is that share gains in ASICs do not have to come from outright displacement of GPUs; they can come from silicon being reserved for the highest-throughput, most repeatable inference jobs while GPUs remain the flexible layer for frontier model development and burst workloads. That makes the addressable market more durable than a simple substitution story, but it also means the upside is more linear and less explosive than the market’s “next Nvidia” framing. For Broadcom, the risk is valuation and customer concentration; for Marvell, execution and design-win conversion matter more than headline socket counts, because delayed ramps can compress multiple expansion even if the long-term TAM is intact. The consensus appears to underappreciate timing risk: design wins are not revenue, and the next 2-3 quarters can still look lumpy if hyperscalers stagger deployment schedules or re-spec chips. The bigger overhang on both names is not technology but procurement discipline—cloud customers can slow spend if AI inference monetization lags internal expectations, especially if they decide to sweat existing hardware longer. That creates a classic “good story, noisy quarterly prints” setup where the stock can outrun cash flow in the near term, then de-rate on any evidence of slower production ramps. From a portfolio perspective, the cleanest expression is to own the custom-silicon beneficiaries while hedging against a relative reset in the GPU complex if the market broadens beyond training. Over a 6-12 month horizon, the best risk/reward likely comes from a pair trade that captures inference share gains without relying on the entire AI hardware basket moving higher in tandem. If earnings revisions continue to inflect upward, these names can compound; if not, the downside is mostly multiple compression rather than permanent thesis breakage.
AI-powered research, real-time alerts, and portfolio analytics for institutional investors.
Request a DemoOverall Sentiment
moderately positive
Sentiment Score
0.66
Ticker Sentiment
