Q1 FY2027 showed Marvell's custom AI chase (bespoke chips designed for individual hyperscaler buyers) from both sides: AI bookings lifted guidance 35%, free cash flow doubled, reported earnings fell about 80.
Marvell Technology has put nearly everything on being custom AI silicon's permanent No. 2, and the market has priced that bet accordingly. The first public accounting of how the bet is going, through the Q1 FY2027 earnings window, shows a company spending hard to win a slot it has not yet secured. GAAP net income collapsed about 80.4% even as Marvell reported $2.42 billion in revenue, up 27.6% year over year. Free cash flow more than doubled to $483 million. The numbers force a question the bull case does not want asked: whether Marvell's target of more than $10 billion in custom-chip revenue by fiscal 2029 is a durable moat in the AI build-out, or whether the few hyperscalers underwriting it are already hedging.
Custom silicon is the part of the AI stack that lives between the merchant accelerator and the data-center wall. Where NVIDIA sells off-the-shelf GPUs to every buyer, Broadcom and Marvell design bespoke chips for one hyperscaler customer at a time, plus the analog, mixed-signal, and photonic plumbing that moves data around AI clusters. Marvell's data-center business is now 76% of total revenue, up from a grab bag of storage, networking, and automotive parts a few years ago. The playbook belongs to Broadcom, and Broadcom trades at roughly $1.71 trillion in market cap to Marvell's $214.58 billion. Even a thin slice of that custom-AI pie has been the year's bull story.
The cost of claiming that slice is now in the GAAP line. Q1 FY2027 was the first earnings period where the custom-silicon thesis has been large enough to bend the income statement, and the bend reads as investment, not harvest. Marvell flagged more than 50 new custom AI design opportunities across more than 10 customers, and CEO Matt Murphy told investors the company is "significantly raising Marvell's revenue outlook for both fiscal 2027 and fiscal 2028" on the strength of AI bookings. Guidance points to roughly 35% year-over-year growth next quarter. The forward picture is bright. The trailing one shows what it costs to make it bright.
Custom silicon revenue is lumpy and concentrated among a handful of hyperscalers. Any one of those buyers can dual-source to Broadcom on the next process node, or move work in-house. On the same Q1 call, Murphy acknowledged that customers "may be pursuing multiple paths" on XPU supply, the industry's term for custom accelerators built around one buyer's workloads. That line is the bear case speaking in the bull case's own voice. A $10 billion revenue target concentrated in fewer than ten buyers is not the same product as $10 billion sold across an open market. The difference between those two is what the GAAP collapse is funding.
Marvell's distinct AI bets are easy to conflate. The custom-chip pipeline is the headline revenue bet. The optical interconnect acquisition of Celestial AI, paid for in part with the $2.5 billion Marvell received from selling its automotive Ethernet unit to Infineon, is a separate photonic bet on the chips and lasers that pull data off accelerators and move it across racks. XConn, also acquired out of the Infineon proceeds, is a smaller bet on chiplet packaging. Marvell joined NVIDIA's NVLink Fusion ecosystem, a partnership that monetizes every AI rack layer except the merchant GPU itself. Custom silicon, optical interconnect, chiplet packaging, and the NVLink partnership are four bets a reader could mistake for one. Marvell's earnings show what each one costs.
The next two quarters will reveal whether the $10 billion FY2029 target lands in Marvell's reported revenue, or stays a Marvell/Reuters claim the income statement refuses to confirm. The 80.4% GAAP collapse is what the company is paying to compete for a slot it has not yet locked in, with customers who have not yet stopped shopping that slot to Broadcom. The earnings so far only show what the chase costs. They do not yet show it works.