You know the AI chip narrative by now. NVIDIA makes the picks and shovels, everyone else fights for scraps, and the whole market waits for Jensen to say something at GTC. That story's been good to a lot of people. But it's not the whole picture. There's a second AI chip revolution happening — one that's moving faster than most realize, and Marvell is sitting right in the middle of it.
Marvell is the number two custom AI ASIC player in the world. Only Broadcom is bigger. And while Broadcom has the scale advantage, Marvell has something arguably more valuable: relationships with the hyperscalers actively trying to break their dependence on NVIDIA. This isn't a narrative. It's happening right now, in volume.
Let's talk about who's building what. Amazon is Marvell's biggest customer — Trainium 2 and 3 for AI training, Inferentia for inference, and Nitro for networking controllers. Three separate chip programs with a single partner. Amazon doesn't do anything at small scale, and Trainium 2 is already deployed across AWS regions. Trainium 3 is on deck with a 3nm shrink designed to close the gap with NVIDIA's latest. When Amazon buys computing, it buys in fleets. Marvell designs the silicon those fleets run on.
Then there's Google. In April 2026, Google awarded Marvell the contract for its Axion ARM-based server processors — a deal CNBC reported Marvell won directly from Broadcom. That's a hand-to-hand combat win against the biggest player in custom silicon. Google didn't switch because Marvell was cheaper. They switched because Marvell could deliver something Broadcom couldn't — speed, flexibility, maybe a better engineering relationship.
Marvell is also competing for Microsoft's Maia and Meta's MTIA program. No confirmed wins yet, but being in the conversation for both tells you the hyperscalers treat Marvell as a legitimate alternative to Broadcom. When the four biggest buyers of compute hardware on the planet are all considering you as a primary partner, you're not a niche play. You're infrastructure.
The networking side is just as important. In June 2026, Marvell launched the Teralynx T100 — a 102.4 Tbps AI switch built on 3nm. That's a switch designed specifically for the scale of traffic AI clusters generate. As models get bigger, the fabric connecting GPUs and custom ASICs becomes a bottleneck. NVIDIA knows this — they bought Mellanox for this exact reason. Marvell's Teralynx is their answer, and it's already winning design wins in the same hyperscaler accounts that buy their compute silicon.
The numbers back it up. Marvell did $8.7 billion in revenue over the trailing twelve months, up 42% year-over-year. R&D spend runs about $2.4 billion annually. Free cash flow sits at $1.66 billion. The company is targeting $16.5 billion by fiscal 2028 — nearly double where they are today.
Here's where the structural thesis kicks in. This is not a cyclical product cycle. Custom AI silicon is a secular shift driven by a single incentive: hyperscalers don't want to be held hostage by NVIDIA's pricing power, allocation decisions, and roadmap timing. Every dollar Amazon saves running Trainium instead of H100s goes straight to their bottom line or gets reinvested into cheaper inference for customers. The same math works for Google, Microsoft, and Meta. They'll keep building custom silicon because it makes economic sense at their scale, and they'll keep needing design partners who can execute.
The market sizing backs this up. Bloomberg Intelligence puts the AI accelerator TAM at $600 billion-plus by 2033. Counterpoint says custom AI chip shipments triple by 2027. That's a massive, fast-growing segment where Marvell has established beachheads with the two most important hyperscalers in the world. Jensen Huang himself reportedly called Marvell "the next trillion-dollar company." When the CEO of the company you're supposedly competing against says that about you, it's worth paying attention.
Risks? Absolutely. Customer concentration is real — Amazon and Google represent a huge chunk of Marvell's custom silicon revenue. If one of those programs slows or shifts internal, the stock gets hit. Broadcom's scale advantage means they can throw more engineers at problems. And hyperscalers could always bring more design work in-house. But that last risk applies to Broadcom too, and in-house chip design is brutally hard at this scale.
Here's the bottom line. Marvell is not a GPU stock. It's not a networking stock. It's the custom silicon arm of the hyperscaler world — the company that designs the chips Amazon and Google use to reduce their dependency on a single supplier. That dynamic is not going away. If anything, it's accelerating. When your customers are the world's largest buyers of computing and they're explicitly trying to diversify their supply chain, you're in the right seat at the right time. Marvell isn't priced like a trillion-dollar company today. But the ingredients are on the table.
Disclosure: The Signal holds no position in MRVL. Positions may change. This is not financial advice.
