Marvell Technology holds its investor day in New York on Tuesday, and the company has to show that a custom AI chip story worth roughly $120 billion in potential Google orders converts into contracted revenue rather than staying a ceiling nobody is required to reach.
The stock, up more than 200% this year, trades near $272 after analysts at BofA and B. Riley raised their targets to $200 and $205 in the run-up. That is the easy part of the story. The hard part is arithmetic: whether the custom silicon business, currently about $1.5 billion a year, can scale fast enough to justify a multiple built on AI optimism rather than booked sales, and whether Marvell can keep winning hyperscaler sockets against Broadcom and an increasingly ambitious Nvidia.
The Google warrant, explained
The centerpiece is an arrangement struck in August. Marvell handed Google warrants to buy as many as 58.97 million shares at $206.58 apiece, worth about $12.2 billion if exercised in full. Google pays no cash upfront. Every $500 million of chip orders unlocks another tranche, tying Google’s equity position directly to how much silicon it actually buys. If purchasing targets through fiscal 2033 are hit, Marvell could book roughly $120 billion in custom chip sales over that stretch, and Google would end up holding close to 7% of the company, making it the fifth-largest shareholder.
The structure is a hedge for both sides. Google commits to nothing it can walk away from, while Marvell gets a customer whose incentives are now aligned with its stock price. Broadcom, whose custom-silicon relationship with Google this deal is widely read as challenging, fell roughly 5% the day the warrant news landed, while Marvell jumped as much as 14% before closing up about 8%.
| Marvell custom silicon, at a glance | Figure |
|---|---|
| Custom revenue run rate | about $1.5 billion a year, roughly 18% of fiscal 2026 sales |
| Q2 fiscal 2027 revenue | $2.7 billion, up 37% |
| Previous custom target | more than $10 billion by fiscal 2029 |
| Potential Google custom sales through fiscal 2033 | roughly $120 billion |
| Google warrant position if fully exercised | close to 7% of shares |
| XPU design wins disclosed | three, plus nine XPU-attach programs |
What Marvell actually sells
Marvell is not trying to out-design Nvidia’s merchant GPUs. The bet is that hyperscalers increasingly want application-specific chips tuned to their own workloads, chips they own outright instead of renting compute built for everyone. Marvell supplies the pieces that turn a hyperscaler’s architecture into working silicon: ASIC design, high-speed SerDes links that move data between chips, HBM memory interfaces, advanced packaging, chiplet integration and access to leading-edge manufacturing nodes.
The company already works with all four of the largest hyperscalers and has disclosed three XPU design wins plus nine XPU-attach programs, according to its investor materials. Management told analysts it expects the custom business to grow more than 20% in fiscal 2027 and more than double again in fiscal 2028, with the real acceleration in fiscal 2029 when multiple design wins ramp at once. Chief executive Matt Murphy has gone further, saying there is “a lot of upside bias” in fiscal 2029 numbers and that custom silicon will likely come in “a lot larger” than the Street has modeled.
A market moving the target
The context favors the pitch. Nvidia released Vera this year, a fully redesigned central processor built for AI agents, and said it expects the CPU market to reach $200 billion by 2030. AMD predicted the whole CPU market would hit $220 billion in sales by 2030, up from a 2025 forecast of $60 billion. Futurum estimates $118 billion in total CPU sales in 2027, nearly double its own May forecast. Every one of those revisions points the same direction: the AI buildout is spreading from GPUs into the rest of the server, and custom silicon rides along.
Marvell’s rivals are not standing still. Broadcom remains the custom ASIC incumbent and dominates the merchant networking silicon Marvell also sells into. Cadence and Synopsys are pushing AI agents into chip design itself, with Synopsys and OpenAI announcing a joint model to operate EDA tools autonomously. Meanwhile the OpenAI-Synopsys work and Broadcom’s deepening ties to other hyperscalers mean the design-win pipeline Marvell counts on is contested at every socket.
What analysts will watch
Three things from today’s event will move the stock. First, whether Marvell raises the fiscal 2029 custom revenue target above $10 billion or merely reaffirms it with more timing detail. Second, any disclosure on the pace of Google order flow, since the warrant tranches unlock on actual purchases and the first ones should already be visible in backlog. Third, margin commentary, because custom ASIC work carries structurally lower gross margins than merchant products, and a $120 billion headline means less than it looks if the mix shift drags profitability down with it.
The deeper question is whether hyperscalers keep diversifying away from Nvidia at all. Google has its TPU program, Amazon has Trainium, Microsoft has Maia, and Meta has split its volume across vendors. Every dollar that moves to custom silicon is a dollar not spent on Nvidia’s 70%-plus margins, which is exactly why the GPU leader has started courting the same customers with its own reduced-margin offerings. Marvell’s pitch is that it wins either way, as the neutral supplier that the giants trust with their most sensitive designs.
Today’s session will not settle that. Design cycles in this business run years, and warrants only pay out if orders land on schedule. But the stock’s 160% run this year was a bet that the August headline converts into scheduled revenue, and after Tuesday, investors will finally have numbers to check that bet against instead of a press release.