Marvell Technology just put a number on the part of the AI chip market that Nvidia does not control. At its Investor Day this week, the company said its custom-silicon business for AI data centers is on track to reach roughly $30 billion in annual revenue by fiscal 2031, up from a prior path that already had it clearing $12 billion in fiscal 2029. For a company that spent years as a mid-tier networking-chip supplier, that is a startling reframe: Marvell is no longer pitching itself as an Nvidia alternative. It is pitching itself as the company hyperscalers call when they want to build AI chips of their own.
The timing matters. Nvidia still sets the price and the pace for the AI-chip market, as its continued share buybacks and record capital returns to shareholders make clear. But the $30 billion figure, confirmed in Marvell’s own investor materials and reported by outlets including TradingView and AD HOC NEWS, is the clearest signal yet that custom silicon, not merchant GPUs, is where a growing share of AI infrastructure spending is headed. This is a news analysis of what Marvell actually said, what it means for Nvidia, and where the real uncertainty in that $30 billion projection sits.
What Marvell Told Investors This Week
Marvell Technology, Inc. used its Investor Day to lay out an expanded version of the custom-silicon pitch it has been building for several years. The company said it is already shipping custom silicon to all four major U.S. hyperscalers, a detail executives used to argue that Marvell’s design business is no longer a bet on one or two customers but a diversified franchise spanning the biggest cloud buyers in the country. The company did not name all four hyperscalers in the material made public, so that detail should be read as a company claim rather than a confirmed customer list.
Marvell Technology CEO Matt Murphy framed the outlook in direct financial terms during the Investor Day, telling attendees: “We have increased our fiscal 2029 revenue outlook from more than $10 billion to more than $12 billion, and we are on a path to approximately $30 billion by fiscal 2031 at the midpoint of our long-term target range.” That is a company-stated target, not a locked-in contract, and it covers the custom business specifically rather than Marvell’s total revenue.
Marvell’s custom-silicon story sits alongside its separate, longer-running partnership with Nvidia. In March 2026, the two companies announced a strategic partnership built around Nvidia’s NVLink Fusion platform, under which Marvell supplies custom XPUs and networking silicon built to work with Nvidia’s architecture rather than against it. That detail is worth sitting with, because it undercuts the simplest version of the “Nvidia weak spot” framing: Marvell’s biggest growth story is partly a growth story for Nvidia’s own ecosystem, not a replacement for it.
Breaking Down the $30 Billion: XPUs vs. “XPU Attach”
The headline number obscures an important structural detail. Marvell’s own framing is that the projected fiscal 2031 custom business will be roughly balanced between two categories: XPU products themselves, meaning the custom AI accelerator chips Marvell designs for specific hyperscaler customers, and what the company calls “XPU attach,” the networking, memory, storage, and connectivity silicon that sits around those accelerators inside a server rack.
Murphy described the split at Investor Day: “On the magnitude of the opportunity, when you go out to 2030 calendar, fiscal 2031, and you’re looking at the $30 billion, what I said is, and I think the mix, we still don’t know exactly, but think of it as largely balanced between the existing XPU programs we have coming and then XPU attach.” That “we still don’t know exactly” qualifier is doing real work. It means the $30 billion figure is a planning target built on an expected mix, not a revenue breakdown Marvell has already locked in with signed contracts for every dollar.
This XPU-attach category is the quieter half of the custom-silicon trade. Every custom AI accelerator a hyperscaler designs still needs high-speed interconnects, memory controllers, and storage silicon to actually function inside a cluster, and Marvell’s argument is that it can win that attach revenue even on racks where it did not design the core accelerator. That is a meaningfully different, and arguably more defensible, growth thesis than simply hoping to out-design Nvidia’s own chips.
The Nvidia-Marvell Partnership and NVLink Fusion
NVLink Fusion is Nvidia’s program for letting outside silicon, including hyperscaler-designed custom chips and third-party accelerators, plug into Nvidia’s high-speed interconnect fabric rather than being locked out of it. Marvell’s role under the March 2026 partnership is to provide custom XPUs and networking components built to be compatible with that architecture, according to the companies’ joint announcement.
That arrangement is the single most important piece of context missing from the “Nvidia’s biggest weak spot” framing that has circulated since the Investor Day. A direct report on the custom-silicon trend has explicitly noted that Marvell’s growth does not directly threaten Nvidia, because Marvell maintains a strategic partnership with the company rather than competing head-on for the same sockets. Marvell is not trying to replace the Nvidia GPUs inside a hyperscaler’s AI cluster. It is trying to sell the networking and custom-compute silicon that sits next to, and increasingly plugs into, those same Nvidia systems.
Industry analyst Jacob Bourne of eMarketer captured this dynamic in comments reported by Reuters on Nvidia’s investment in Marvell: “So Nvidia can maintain its dominant position while also expanding the scope and utility of the AI semiconductor sector.” (Reuters) That is a cooperative-expansion read of the relationship, not a disruption read, and it is the more evidence-backed interpretation of what is actually happening between the two companies.
Why Hyperscalers Want Their Own AI Chips
The underlying demand driver for Marvell’s custom business is not a secret: major cloud providers want more control over AI-data-center workloads than merchant GPUs alone give them. Buying a custom-designed accelerator, tuned for a specific internal workload like ad-ranking inference or large-model training, can let a hyperscaler cut power draw per unit of compute, avoid paying a merchant-chip margin, and reduce dependence on any single supplier’s roadmap and allocation decisions.
That logic is already visible elsewhere in the market. Google’s in-house TPU line has reportedly priced its newest generation well below comparable Nvidia silicon, a gap covered in detail elsewhere on this site, while Amazon has structured multibillion-dollar arrangements around its own custom accelerator roadmap, including the leaseback deal tied to its Nvidia chip holdings. Marvell does not build its own branded AI chip the way Google or Amazon do. Instead, it positions itself as the design-and-manufacturing partner that lets any hyperscaler build a TPU-style or Trainium-style chip of its own, which is precisely why Marvell describes its addressable opportunity as spanning all four major U.S. hyperscalers rather than just one captive customer.
Custom Silicon vs. Merchant GPUs: A Market Comparison
The table below lays out the structural differences between the custom-silicon model Marvell sells into and the merchant-GPU model Nvidia still dominates. These are qualitative distinctions drawn from how each approach is generally structured in the market, not hyperscaler-specific financial disclosures.
| Dimension | Merchant GPU Model (Nvidia-led) | Custom Silicon Model (Marvell-enabled) |
|---|---|---|
| Primary buyer relationship | Direct purchase of off-the-shelf accelerators | Design partnership for a customer-owned chip |
| Software ecosystem | CUDA and Nvidia’s mature developer tooling | Customer-controlled, often narrower workload fit |
| Supply allocation | Subject to Nvidia’s broader global allocation | Dedicated to the commissioning hyperscaler |
| Who captures the margin | Nvidia captures merchant-silicon margin | Hyperscaler captures silicon margin; Marvell earns design and manufacturing fees |
| Flexibility across workloads | General-purpose, broadly flexible | Optimized for a specific, known workload |
| Interconnect strategy | Nvidia’s own NVLink fabric | Increasingly compatible via NVLink Fusion |
Note the last row. Even hyperscalers pursuing custom silicon are not necessarily walking away from Nvidia’s interconnect standards, they are asking for compatibility with them, which is exactly the gap NVLink Fusion and the Marvell partnership are built to close.
Marvell’s Revenue Trajectory: The Confirmed Numbers
Here is what Marvell has actually confirmed about its custom-silicon revenue path, drawn directly from Investor Day materials, versus what remains projection rather than booked revenue.
| Fiscal Year | Custom-Silicon Revenue Figure | Status |
|---|---|---|
| Fiscal 2029 | More than $12 billion (raised from a prior “more than $10 billion” outlook) | Company-stated revenue outlook |
| Fiscal 2031 | Approximately $30 billion | Midpoint of a long-term target range; not realized revenue |
| Revenue mix at FY2031 target | “Largely balanced” between XPU products and XPU attach | Company characterization; exact split undetermined, per Murphy |
The gap between fiscal 2029’s confirmed-outlook figure and fiscal 2031’s target range is the part investors will be watching most closely over the next two years. An “approximately $30 billion” target at the “midpoint” of a range implies Marvell itself is working with a band of outcomes, not a single locked number, and the company has been explicit that it does not yet know the precise mix driving that figure.
How Broadcom Fits Into the Custom-Silicon Picture
Marvell is not the only chip designer chasing hyperscaler custom-silicon budgets. Broadcom has built the larger of the two major independent custom-AI-chip franchises, supplying design work tied to some of the same category of hyperscaler customers Marvell is now courting, and Broadcom’s own AI chip revenue forecasts have drawn scrutiny and regulatory attention this year. The competitive dynamic between the two companies is less a two-horse race for the same socket and more a split: different hyperscalers tend to concentrate their custom-silicon design work with one partner or the other, based on existing relationships and prior design wins.
What Marvell’s Investor Day numbers do confirm is that the custom-silicon category as a whole, across both major suppliers, is now large enough to be treated as a distinct line item by Wall Street analysts rather than a footnote to the merchant-GPU story. AMD, meanwhile, has pursued a hybrid approach, selling its own merchant AI accelerators while also forecasting that AI chip demand across the industry will outpace available supply well into 2028, a supply-constrained backdrop that benefits every credible chip designer, custom or merchant, simultaneously.
Does This Really Threaten Nvidia?
The honest answer, based on what has actually been confirmed, is: not directly, and not yet. The “Nvidia’s biggest AI weak spot” framing attached to this story is a headline characterization rather than a verified claim, and the closest thing to an on-record assessment of the relationship points the other way. Marvell’s custom business grows in large part through hyperscalers wanting more design control and better cost structure on specific workloads, not through those hyperscalers abandoning Nvidia GPUs for their general-purpose training and inference needs.
What the $30 billion figure does represent is a redistribution of where AI infrastructure dollars land. If hyperscalers spend a growing share of their chip budgets on custom silicon and the networking and memory components that attach to it, that is dollars that might otherwise have gone entirely to merchant GPU purchases. Nvidia’s own response, including its participation in NVLink Fusion and its direct investment in Marvell’s partnership, suggests the company sees this less as a threat to be fought and more as an adjacent market to be integrated with, so it can still capture a share of the interconnect and ecosystem spending even on racks built around someone else’s custom accelerator.
Historical Context: From Networking Chips to AI Silicon
Marvell’s current pitch did not appear overnight. The company built its original business on networking and storage controller chips, the unglamorous silicon that moves data between servers and drives, long before “custom AI silicon” was a Wall Street buzzphrase. That networking heritage is exactly why Marvell is positioned to sell “XPU attach” products today: interconnect and storage silicon is the same category of chip the company has designed for two decades, now repackaged for AI racks instead of enterprise data centers.
The broader custom-silicon trend itself traces back further, to hyperscalers’ earlier moves into application-specific chips for non-AI workloads, like video transcoding and network switching, before AI training became the dominant driver of data-center capital spending. What is different in 2026 is scale: the dollar figures now being discussed for custom AI silicon, across Marvell, Broadcom, and the hyperscalers’ own internal chip programs, are an order of magnitude larger than anything the custom-ASIC business generated in prior cycles. The memory-chip shortage running alongside this buildout, which has pushed DRAM pricing past the cost of advanced logic silicon, is itself a byproduct of how much custom and merchant AI silicon the industry is now trying to build at once.
What Analysts and Marvell’s CEO Are Saying
Beyond the headline revenue figures, Murphy used the Investor Day to make a broader infrastructure claim about Marvell’s role in the AI buildout: “As Xi said, every model so far that’s ever been created has been trained using Marvell connectivity as an example.” (Marvell Investor Day 2026 transcript) That is a sweeping statement about Marvell’s networking silicon being present somewhere in the training infrastructure of essentially every major AI model to date, and it should be read as a company claim about its ubiquity in the supply chain rather than a measured market-share statistic.
On where the next leg of growth comes from, Murphy left the door open to expanding beyond Marvell’s current four-hyperscaler base: “The upside could be as they look at more custom or other types of solutions, that’s an opportunity for Marvell as well.” (Marvell Investor Day 2026 transcript) Combined with Bourne’s read on the Nvidia relationship cited above, the analyst and executive commentary around this story points to expansion and integration rather than disruption as the dominant theme.
What Hyperscalers Gain — and Give Up
Commissioning custom silicon through a partner like Marvell is not free of trade-offs for the hyperscalers buying into it. The upside is real: a chip tuned to one company’s specific training or inference workload can beat a general-purpose GPU on performance-per-watt for that narrow job, and owning the design means the hyperscaler is not waiting in line behind every other Nvidia customer for allocation during a supply crunch.
The downside is development risk and software lock-in of a different kind. A custom chip that underperforms its design target, or arrives a generation late relative to Nvidia’s release cadence, can leave a hyperscaler stuck running an inferior accelerator across a fleet it already committed billions of dollars to building. That risk is part of why even the most aggressive custom-silicon programs, including Google’s TPU line and Amazon’s in-house accelerators, have continued buying large volumes of Nvidia GPUs alongside their custom fleets rather than replacing them outright, a dual-track approach that server deals like HPE’s recent AI infrastructure agreement also reflect on the merchant-silicon side.
The Unknowns in Marvell’s Forecast
Three specific gaps in what has been disclosed are worth flagging plainly, because they are the difference between a confirmed business and a confirmed target.
- The identities of all four “major U.S. hyperscalers” Marvell says it ships custom silicon to have not been fully disclosed in the available Investor Day material.
- The approximately $30 billion fiscal 2031 figure is explicitly a company expectation at the midpoint of a target range, not revenue that has already been booked or contracted.
- The exact split between XPU products and XPU attach within that figure is, in Murphy’s own words, still unknown even to Marvell’s management team.
None of that makes the $30 billion target implausible. Marvell has already demonstrated it can raise its own guidance, moving the fiscal 2029 figure from “more than $10 billion” to “more than $12 billion” in the same Investor Day presentation. But a target raised twice in public remarks is still a target, and five fiscal years is a long runway for a hyperscaler customer to shift budget, delay a program, or bring a design in-house.
What Happens Next: Five Predictions
Based on the trajectory Marvell has laid out and the broader custom-silicon market around it, here is how this is likely to play out over the next two to three years.
- Marvell will likely name additional hyperscaler or large-scale AI customers in future quarterly disclosures, as competitive pressure pushes the company to substantiate the “four major hyperscalers” claim with specifics.
- The XPU-attach category, not the headline XPU accelerator chips, will probably grow faster in the near term, since networking and memory silicon ships alongside every AI rack regardless of which company designed the core accelerator.
- Nvidia’s NVLink Fusion ecosystem will likely expand to include more custom-silicon partners beyond Marvell, making interconnect compatibility rather than raw accelerator competition the main battleground.
- Broadcom and Marvell will probably continue splitting hyperscaler design wins rather than directly outbidding each other for the same contracts, keeping the custom-silicon market a duopoly-plus-insourcing structure rather than a single winner-take-most race.
- Marvell’s fiscal 2031 target will likely be revised, up or down, at least once more before that fiscal year arrives, given how far out the projection sits and how explicitly the company has flagged the mix as undetermined.
Market Impact for Chip Investors and Buyers
For investors, the practical takeaway is that Marvell’s custom-silicon story is now large enough to move the stock independent of its legacy networking business, which is why Wall Street coverage has shifted to treating the $30 billion figure as a standalone data point. For enterprise buyers and cloud customers who do not design their own chips, the more relevant effect is indirect: as hyperscalers shift more capital into custom accelerator programs, the mix of hardware available inside public cloud regions, and potentially its pricing, will increasingly reflect each cloud provider’s own silicon choices rather than a uniform Nvidia-everywhere baseline.
That shift is already visible in how differently Google, Amazon, and Microsoft are pricing and positioning their respective accelerator options relative to Nvidia instances. A business running AI workloads in the cloud should expect that divergence to widen, not narrow, as more of Marvell’s and Broadcom’s custom-silicon pipeline reaches production over the next several fiscal years.
Frequently Asked Questions
What did Marvell actually announce about its AI chip business?
At its 2026 Investor Day, Marvell Technology said its custom-silicon business for AI data centers is projected to reach approximately $30 billion in annual revenue by fiscal 2031, up from a raised fiscal 2029 outlook of more than $12 billion.
Is Marvell competing directly against Nvidia?
Not primarily. Marvell’s custom-silicon business operates alongside a March 2026 strategic partnership with Nvidia built around the NVLink Fusion platform, under which Marvell’s custom chips and networking silicon are designed to be compatible with Nvidia’s interconnect architecture rather than to replace it.
What is “XPU attach” in Marvell’s forecast?
XPU attach refers to the networking, memory, storage, and other connectivity silicon that surrounds a custom AI accelerator inside a server rack. Marvell expects its fiscal 2031 custom revenue to be roughly balanced between XPU products themselves and this XPU-attach category.
Which hyperscalers is Marvell working with?
Marvell has said it ships custom silicon to all four major U.S. hyperscalers, but the company has not publicly named all four customers in its disclosed Investor Day materials, so specific identities remain unconfirmed.
Is the $30 billion figure already locked in as revenue?
No. Marvell describes it as the midpoint of a long-term target range for fiscal 2031, not revenue the company has already booked or contracted. Marvell’s own CEO has said the exact mix driving that number is still undetermined.
How does Marvell’s custom-chip business compare to Broadcom’s?
Both companies run large custom-AI-silicon design businesses serving overlapping categories of hyperscaler customers, but they have tended to win design contracts with different customers rather than competing head-to-head for the same programs, effectively splitting the custom-silicon market between them.
Does custom silicon threaten Nvidia’s market position?
Not directly, based on current reporting. Analysts covering the Nvidia-Marvell relationship have characterized it as expanding the overall AI semiconductor market rather than displacing Nvidia, partly because Nvidia itself has invested in and partnered with Marvell’s custom-silicon push.
Why are hyperscalers building custom AI chips instead of just buying Nvidia GPUs?
Custom silicon lets a hyperscaler tune a chip to its own specific workloads, potentially improving performance-per-watt and cost structure, while also reducing dependence on Nvidia’s allocation and roadmap for every unit of AI compute it needs.
