Marvell (MRVL) vs Broadcom: Inside the Custom AI Chip War

Marvell (MRVL) stock jumped roughly 25% after Nvidia CEO Jensen Huang suggested it could become the next trillion-dollar company — a stunning move for a chipmaker that does not sell a single flagship GPU. The reason cuts to the heart of how AI compute is being built in 2026.

What Just Happened to Marvell?

Marvell stock jumped roughly 25% after Nvidia CEO Jensen Huang publicly suggested the company could be the next trillion-dollar chipmaker. For a stock that lives in Nvidia's shadow, that was a powerful endorsement.

But the price move is a symptom, not the story. The story is that Marvell (MRVL) sits at the center of the fastest-growing corner of AI hardware: custom silicon designed for, and only for, the world's largest cloud companies.

The market spent two years assuming Nvidia GPUs were the only way to buy AI compute — the Marvell move is the market pricing in a second path. That second path is the application-specific integrated circuit, or ASIC.

Why Are Hyperscalers Designing Their Own Chips?

Because owning the silicon is cheaper and stickier than renting it. When a cloud giant runs the same AI workload millions of times a day, a chip tuned to that exact task can beat a general-purpose GPU on cost and power.

So Amazon (AMZN) built Trainium, Microsoft (MSFT) built Maia, Alphabet (GOOGL) has its TPU line, and Meta (META) built MTIA. None of these companies fabricates chips alone — they co-design them with a silicon partner.

That is where Marvell and Broadcom come in. They supply the design expertise, the high-speed interconnect, and the optical networking that stitch thousands of these accelerators into a single training cluster.

Broadcom vs Marvell: Who Owns the ASIC Market?

Broadcom is the clear leader. It is cited as holding roughly 70% of the custom AI accelerator market, anchored by long-running programs with Alphabet on TPUs and Meta on MTIA chips.

Broadcom (AVGO) carries an AI backlog of around $73 billion and has guided toward roughly $100 billion in annual AI revenue by 2027. Marvell is the focused challenger, projecting up to about $11 billion in AI ASIC revenue for 2026 on the back of Amazon and Microsoft wins.

Together, Broadcom and Marvell control close to 95% of the ASIC co-design market — a near-duopoly in a TAM that could reach $30-50 billion by 2028. Concentration like that is rare, and it is exactly why investors are paying up.

Is This a Threat to Nvidia?

Less than the headlines suggest. Custom ASICs and merchant GPUs are increasingly parallel markets, not a zero-sum fight — and the clearest proof is that Nvidia (NVDA) itself invested around $2 billion in Marvell via its NVLink Fusion architecture.

Nvidia still dominates training for cutting-edge, rapidly changing models, where flexibility beats specialization. ASICs win the steady-state, high-volume inference and training workloads where the model is fixed and cost-per-query is everything.

The smarter framing is that the AI compute pie is growing fast enough for both. AMD (AMD) is fighting for the merchant-GPU alternative slot, while the custom-silicon duo feeds the hyperscalers' in-house ambitions.

The Picks-and-Shovels Read

The cleanest way to think about this theme is infrastructure, not end-product. Whoever wins the model race, the clusters still need chips, interconnect, and networking — and that demand flows to a small set of suppliers.

Arista Networks (ANET) sits alongside the ASIC names as the switching layer that connects accelerators into a fabric. For readers new to evaluating these businesses, our guide to fundamental analysis covers how to judge backlog quality and revenue durability, and the investment strategies primer explains how to size a position in a high-volatility theme.

The watch item is design-win conversion. Backlog and partnership announcements are leading indicators; the lagging confirmation is whether that pipeline actually shows up as booked revenue over the next several quarters.

The Bear Case: Where the ASIC Story Breaks

Critics argue the enthusiasm has outrun the fundamentals. A CEO calling Marvell a future trillion-dollar company is sentiment, not a financial result, and the stock re-rated well ahead of any booked numbers.

The structural risk is customer concentration. ASIC revenue is lumpy and tied to a handful of hyperscalers, so a single customer reallocating its capex — or pulling a program in-house — can erase a year of expected revenue. These chips also tend to carry lower margins than merchant GPUs.

The risk, then, is paying a GPU-like multiple for a more concentrated, more cyclical, lower-margin business. The theme is real; the price you pay for it still matters.

Ready to analyze these stocks yourself? Search any ticker on MainRatios to see valuations from 6 legendary investors - free.