Marvell has told Wall Street that its custom AI chip designs will exceed internal forecasts, driven by strong hyperscaler demand for Nvidia alternatives.
Marvell raised its own numbers twice in two months and still told investors Tuesday that the real upside hasn't been modeled yet.
On October 6, Marvell CEO Matt Murphy told New York investors something chip executives rarely do: that his own guidance was too conservative. At the company's Investor Day, Murphy outlined a custom silicon roadmap centered on partnerships with Google, Microsoft, and Amazon. He then told the room that even the newly raised fiscal 2028 revenue target of $18 billion—up from $16.5 billion just weeks earlier—understates what's coming once those programs ramp past 2029. Marvell's custom AI chip revenue reached roughly $1.5 billion in fiscal 2026, representing about 18% of total sales. Management now expects that share to grow more than 20% this fiscal year, double the year after, and reach $10 billion annually by fiscal 2029—a target Murphy himself called conservative on stage.
Google is the centerpiece. Marvell's custom silicon agreement with the search giant has been reported at roughly $120 billion in value. Murphy called it "game changing" on an earnings call, adding that AI chip revenue could end up "a lot larger" than current targets. That's an unusually direct statement for a CEO to make about his own forecast. Most executives pad guidance to beat it quietly next quarter. Murphy instead told investors the beat is already coming: Google contributes meaningfully through fiscal 2028, then becomes the dominant driver from fiscal 2029 onward, once the custom chips ship at volume.
Broadcom runs the same playbook, only bigger. Hock Tan's company raised its fiscal 2026 AI revenue guidance to around $58 billion and told investors it has "line of sight" to more than $100 billion in AI chip revenue by fiscal 2027. Broadcom's custom silicon backlog hit $73 billion over an 18-month window, with bookings alone topping $30 billion in a single quarter in June. Its customer list overlaps almost exactly with Marvell's ambitions: Google, Meta, OpenAI, and Anthropic, along with seven generations of co-designed Google TPUs dating back to 2014.
That history is the real risk beneath Marvell's pitch. Broadcom didn't win Google's TPU business this year—it won it over a decade of iteration. Switching an AI training fleet off a chip architecture that works is not something hyperscalers do lightly. Marvell is betting it can carve out separate, defensible lanes with Microsoft and Amazon chips where Broadcom isn't entrenched, rather than taking share directly from Broadcom's existing accounts. The Google deal is new territory for Marvell, not a switch from an incumbent.
Nvidia's Jensen Huang has publicly praised the broader custom silicon trend, and Marvell stock jumped on the endorsement earlier this year. That's an odd position: the company whose merchant GPUs custom ASICs exist to replace is cheering the architecture that competes with it. Huang sees hyperscaler compute diversification as inevitable, with room for both merchant GPUs and custom silicon to grow. Nvidia would rather appear comfortable with that reality than fight it publicly.
What Marvell actually needs to prove isn't that custom silicon works. Google, Amazon, and Meta have already shown that a well-designed ASIC beats a general-purpose GPU on cost per inference for specific workloads. What matters is whether Marvell can execute delivery on three simultaneous hyperscaler programs. There can be no repeat of the slip that cost it credibility after its August earnings report, when shares fell 10% despite record revenue because investors wanted firm timing on the Google ramp, not just bigger long-term numbers.
Marvell shares were up roughly 203% year to date heading into the investor day. That means expectations were already elevated before Murphy spoke. Raising the fiscal 2028 target to $18 billion and talking up fiscal 2029 doesn't give investors new information so much as a number to hold the company to. The next real test isn't this presentation. It's the quarter where Google's custom chips are supposed to start shipping at the promised volume, and whether the revenue shows up on schedule or slips again.