Why this comparison matters more than it first appears
This cohort covers Broadcom and Marvell individually elsewhere. This briefing puts them side by side directly, because together the two companies are estimated to enable roughly 80% of hyperscaler custom AI silicon [2] — meaning the balance of power between these two specific companies has more influence over which hyperscalers can field competitive custom accelerators than almost any other single relationship in the industry outside Nvidia itself.
The numbers, side by side
Broadcom’s reported AI backlog stands at $73 billion, against a stated $100 billion annual AI revenue target by 2027. Marvell’s pipeline is smaller in absolute terms — up to $11 billion in 2026 AI ASIC revenue, with 18 design-win sockets and a reported $75 billion in lifetime pipeline opportunity, a figure that, as this cohort’s Marvell briefing notes, describes addressable opportunity rather than signed revenue [4] [3].
Why “rivalry” is only partly the right word
The two companies compete for some of the same category of business — hyperscaler custom-silicon design contracts — but their customer rosters only partially overlap. Broadcom’s flagship relationship is Google’s seven-generation TPU partnership; Marvell’s are Amazon’s Trainium and Microsoft’s Maia. A hyperscaler generally commits to one design partner per major accelerator program rather than splitting the work, which means Broadcom and Marvell more often compete to win a new hyperscaler’s business than to displace each other from an existing, already-committed relationship. Analysts framing this as a “valuation showdown” are comparing two companies pursuing the same overall opportunity from largely separate customer footholds, not two vendors directly undercutting each other on the same active contract [1].
| Metric | Broadcom | Marvell |
|---|---|---|
| Flagship relationship | Google TPU (7 generations) | Amazon Trainium, Microsoft Maia |
| Reported backlog / pipeline | $73B backlog | $75B lifetime pipeline (unsigned opportunity) |
| Near-term revenue figure | $100B target by 2027 | Up to $11B in 2026 |
| Confirmed major customers | 6 | 2 flagship + smaller programs |
What would actually settle the comparison
A cleaner test than backlog size alone is design-win velocity going forward: which company adds the next new hyperscaler customer, and how large that customer’s committed capacity turns out to be once contracted rather than merely pipelined. Broadcom’s larger current backlog reflects a longer operating history in the space, anchored by the Google relationship dating to 2014; it does not necessarily predict which company wins the next hyperscaler considering a first custom-silicon program. Both companies’ next several quarters of new-customer announcements, more than either company’s current backlog total, are what will actually determine how this comparison looks by 2027.
Why investors treat this as a package trade rather than a pick-one bet
Multiple analysts frame this comparison explicitly as a “valuation showdown,” language that reveals something about how the market actually treats these two companies: less as mutually exclusive alternatives and more as a paired bet on the same underlying thesis — that hyperscaler custom silicon keeps taking share from merchant GPUs — expressed through two different vehicles with different risk profiles. Broadcom offers a larger, more established, more diversified revenue base with AI as a fast-growing but not sole component; Marvell offers a smaller, more AI- concentrated position with correspondingly higher torque to the same underlying trend, for better or worse. A reader trying to decide between the two as an investment case is, in an important sense, asking the wrong question — the more useful question is how much exposure to the custom- silicon thesis as a whole they want, and in what risk-and-diversification shape.
The scenario that would hurt both companies equally
It is worth naming explicitly the one scenario that would undermine both Broadcom’s and Marvell’s custom-silicon businesses simultaneously, since the head-to-head framing can obscure how correlated their fates actually are: a slowdown in hyperscaler capital expenditure broad enough to reduce demand for new custom-silicon programs across the board, regardless of which design partner a given hyperscaler favors. Both companies’ growth depends on the same underlying hyperscaler capex cycle this cohort’s market-and-capex research file documents in detail. In that sense, “Broadcom versus Marvell” is the less consequential question for either company’s near-term fortunes; “will the roughly $725 billion in 2026 hyperscaler capex keep growing at anything like its recent pace” is the more consequential one, and it applies to both companies equally.
