Signals Inbox·July 20, 2026·AI Chips

Broadcom vs Marvell: who leads custom AI chips?

Broadcom leads custom AI chips today by a wide margin, with more revenue, deeper hyperscaler programs and greater control over the surrounding infrastructure, while Marvell is building the strongest challenge through AWS, Nvidia, optics and scale-up connectivity.

Q1Why are Broadcom and Marvell always compared in custom AI chips?

Broadcom and Marvell keep getting compared because they are the two large independent chip companies hyperscalers can call when they want a custom AI processor built on a leading-edge process node.

The customer usually brings the workload, part of the architecture and a long-term infrastructure plan. Broadcom or Marvell then helps turn that plan into silicon, packaging, boards and the connectivity around the processor. Both companies also sell the less glamorous pieces that decide whether the finished cluster works: Ethernet switches, SerDes, optical DSPs, retimers, PCIe or CXL components and network interfaces.

That overlap puts Broadcom and Marvell in front of the same unusually small group of hyperscalers and frontier AI labs. Each buyer can support a multi-generation program worth billions of dollars. Losing one design can leave a hole for years; winning one can reshape the supplier’s entire growth rate.

The rivalry has become much easier to see lately. Broadcom has added large new accelerator programs, while Marvell has deepened a major cloud relationship, joined Nvidia’s NVLink Fusion ecosystem and bought more optical and scale-up connectivity technology. They are chasing the same shift away from buying every AI accelerator as a standard merchant chip, but from very different starting positions.

Q2Why is the Broadcom vs Marvell race harder to score than it looks?

The Broadcom versus Marvell custom AI chip race is hard to call cleanly because the companies report different slices of the business and make money from different parts of the system.

Broadcom publishes “AI semiconductor revenue,” a bucket that includes custom accelerators and AI networking. Marvell publishes data-center revenue, which includes custom silicon, optics, switching, storage and other connectivity products. Neither company tells investors exactly how much quarterly revenue came from custom XPUs alone.

The timing creates more confusion. Broadcom already has mature programs generating billions of dollars, along with newer programs that are only beginning to ramp. Marvell is smaller today, but management expects a much steeper custom-silicon step-up in later fiscal years. Comparing current revenue favors Broadcom. Comparing percentage growth from a smaller base makes Marvell look closer.

We therefore separate four questions that often get mixed together: who ships more now, who is adding revenue faster, who has the firmer future orders, and who owns the technology around the processor. Broadcom can lead most of that race while Marvell still builds a serious position in optical and scale-up connectivity. The final verdict has to reflect where the money and production are today.

Q3How much bigger is Broadcom’s custom AI chip business than Marvell’s today?

Today, Broadcom’s custom AI chip operation is several times larger than Marvell’s entire data-center business, even though neither company gives us a perfect custom-XPU revenue number.

Broadcom’s latest quarterly results showed $10.8 billion of AI semiconductor revenue and $22.2 billion of total company revenue. Marvell reported $2.42 billion of total revenue in its latest quarter, including $1.83 billion from data centers.

The raw gap is striking. Broadcom’s AI semiconductor revenue was 4.5 times Marvell’s whole company and 5.9 times Marvell’s data-center division. Those ratios overstate the pure accelerator gap because Broadcom includes networking, but they also give Marvell generous treatment: Marvell’s data-center number includes optics, switches, storage and other products beyond custom processors.

Broadcom also generated $19.2 billion of AI semiconductor revenue across its two most recent reported quarters. Marvell produced about $8.2 billion across its full previous fiscal year. The periods and definitions differ, so this is not a market-share calculation. The order of magnitude is still clear. Broadcom is operating on a different commercial level today.

Latest reported Broadcom and Marvell revenue measures

Q4Does Broadcom’s AI revenue exaggerate its lead over Marvell?

Yes, Broadcom’s AI revenue makes the custom-chip gap look wider, but a cleaner comparison still leaves Broadcom far ahead of Marvell.

Broadcom said networking accounted for almost 40% of its latest quarterly AI semiconductor revenue. A rough split would leave around 60%, or about $6.5 billion, for custom accelerators and related compute silicon. Broadcom has not published that estimate, so we treat it as directional.

Even under that narrower calculation, Broadcom’s quarterly custom-compute revenue would be roughly 2.7 times Marvell’s total company revenue and about 3.5 times Marvell’s full data-center revenue. Marvell’s comparator still contains networking and optical products, so the exercise remains favorable to Marvell.

Counting networking is also reasonable in this market. A hyperscaler choosing Broadcom for a custom accelerator often buys Broadcom Ethernet, SerDes and optical components around the same processor. The customer is purchasing a working AI system, not an isolated die. We should separate the categories when measuring chip share, but removing the network from the competitive analysis would hide one of Broadcom’s strongest reasons for winning the processor design in the first place.

Q5Is Broadcom or Marvell growing faster in custom AI chips right now?

Right now, Broadcom is growing faster in custom AI chips. Marvell’s more dramatic acceleration still sits mainly in its forward guidance.

Broadcom’s AI semiconductor revenue grew 106% year over year in the previous quarter and 143% in the latest one. The company expects the next quarter to reach $16 billion, representing more than 200% growth from the year-earlier period. That is an unusually sharp acceleration from a business already measured in many billions of dollars.

Marvell’s latest quarter was strong in normal semiconductor terms: total revenue increased 28%, data-center revenue rose 27%, and the next-quarter midpoint implies 35% company growth. Marvell also raised its medium-term outlook and expects its custom business to more than double in fiscal 2028, helped by a new tier-one XPU and more than ten XPU-attach programs entering higher-volume production. Management now expects custom revenue to exceed $10 billion in fiscal 2029.

Broadcom’s acceleration is already visible in reported revenue. Marvell’s biggest jump remains a forecast, however credible the underlying design wins may be.

Q6Who has better custom AI customers and firmer future demand: Broadcom or Marvell?

Broadcom currently has better custom AI customers and firmer demand than Marvell.

Broadcom’s visible relationships include a long-term Google agreement covering future TPU generations through 2031, OpenAI’s planned 10-gigawatt accelerator rollout and Meta’s MTIA program. Meta’s initial commitment exceeds one gigawatt and is meant to grow into a multi-gigawatt deployment across several processor generations. OpenAI plans to deploy Broadcom-based accelerator and Ethernet systems through 2029. Anthropic has separately agreed to access roughly 3.5 gigawatts of next-generation TPU-based capacity through Broadcom beginning in 2027.

Marvell’s best disclosed account is AWS. Their five-year, multi-generation agreement covers custom AI products, optical DSPs, active electrical cable DSPs, PCIe retimers, data-center interconnect modules and Ethernet switches. Marvell also says it works with all four top hyperscalers and has built a pipeline of more than 50 custom opportunities across more than ten customers, with an estimated $75 billion of lifetime potential.

Broadcom’s advantage is that more of the demand has already become orders and scheduled capacity. The latest earnings discussion put quarterly AI bookings above $30 billion and purchase orders at $6 billion. Marvell described exceptional AI bookings without publishing a comparable amount. A $75 billion opportunity pipeline can contain excellent programs, lost bids and projects that crawl along for years, so we give more weight to purchase orders and scheduled gigawatt deployments.

Customer depth and future custom AI demand

Q7Who has more custom AI chips running in production: Broadcom or Marvell?

In production today, Broadcom has more custom AI chips running at meaningful scale than Marvell, although some of Broadcom’s newest programs still have to prove themselves outside the lab.

Google’s TPU relationship gives Broadcom years of experience across several chip generations, packaging cycles and cluster deployments. That history is hard to recreate quickly. It shows that Broadcom can keep a hyperscaler program alive after the first design, when cost reduction, yield, software compatibility and the next architecture become the real work.

The newer Broadcom programs sit at different stages. OpenAI’s Jalapeño engineering samples are already running machine-learning workloads at production target frequency and power, with deployment expected to begin later in 2026. Meta’s first 2nm MTIA generation with Broadcom belongs to a multi-year program, but the largest volume from that collaboration lies ahead.

Marvell already ships custom silicon in production, particularly through major cloud relationships, and its latest results show that data-center products are no longer an experimental side business. Still, Marvell’s expected leap depends on a new tier-one XPU and a collection of attach programs reaching larger volumes in fiscal 2028.

Broadcom has the wider production base and more repeat generations. Marvell’s upside depends more heavily on ramps that customers still have to complete.

Q8Who is better at turning a custom AI chip into a working cluster: Broadcom or Marvell?

When a custom AI chip has to become a working cluster, Broadcom currently takes the project further than Marvell.

The modern bottleneck goes far beyond designing the compute die. A customer needs advanced packaging, high-bandwidth memory, boards, rack integration, scale-up links, scale-out networking, power, cooling, manufacturing slots and enough financing to build the data center. One weak link can delay the whole program.

Broadcom’s OpenAI work covers chip implementation, boards, racks, high-performance networking and scalable production. The Meta partnership reaches across chip design, packaging and Ethernet. Broadcom has even joined Apollo and Blackstone in a platform designed to finance more than 20 gigawatts of XPU-based infrastructure, beginning with a $35 billion transaction supporting more than one gigawatt for Anthropic.

Marvell has excellent pieces of the same puzzle, especially optical DSPs, CXL, PCIe, retimers and scale-up connectivity. Marvell’s AWS agreement also shows that the company can sell several components into one customer architecture. We have seen it coordinate the entire route from a central XPU to a gigawatt-scale deployment much less often.

Winning the chip design is only the first half of the job now. Broadcom has assembled more of the engineering, networking, supply and capital needed for the second half.

Q9Who leads AI networking and optical connectivity: Broadcom or Marvell?

Commercially, Broadcom leads AI networking today. Marvell is the more credible challenger in optical and mixed-architecture connectivity.

Broadcom says networking represented almost 40% of its latest quarterly AI semiconductor revenue. Our calculation puts that contribution around $4.3 billion for one quarter. Broadcom can supply Tomahawk and Jericho switching, Ethernet network interfaces, SerDes, PCIe products and optical components, often alongside the customer’s custom accelerator.

Marvell is much smaller, but its portfolio is becoming harder to dismiss. The company says 800G and 1.6T optics, 51.2-terabit Ethernet switches, scale-up optical products and data-center interconnect modules are driving its raised outlook. Marvell’s Ara 1.6T optical DSP is already shipping in mass volume, which gives the optical story more substance than a roadmap slide.

The two companies also offer customers different choices. Broadcom pushes an Ethernet-centered stack optimized from processor to network. Marvell supports Ethernet while letting custom XPUs plug into Nvidia’s NVLink Fusion environment, which suits cloud companies that want proprietary compute without rebuilding every layer around it.

Broadcom has the revenue, installed base and integrated system wins. Marvell has a credible route to gain share if optical links and heterogeneous scale-up fabrics capture more of the value inside each AI rack.

Q10Who launches custom AI chips faster: Broadcom or Marvell?

The fastest recent custom AI chip launch belongs to Broadcom, although OpenAI’s nine-month Jalapeño schedule still needs to be judged after large-scale deployment.

Broadcom and OpenAI said Jalapeño moved from design to production in nine months. OpenAI used its own models during parts of the development process, while Broadcom and Celestica helped industrialize the chip, boards, racks, networking and production system. Engineering samples are already running workloads at their target frequency and power.

Nine months is exceptionally fast for an advanced accelerator. Most custom programs take well over a year and often closer to two years before meaningful production. Marvell has described a roughly two-year path from a new XPU win to revenue, which fits the normal rhythm of architecture, implementation, tape-out, validation and customer qualification.

The claim still needs some proportion. Early internal tests cannot prove superior performance per watt across real customer workloads, and production samples tell us little about stability across thousands of racks. Broadcom has at least produced a concrete chip on an unusually compressed schedule instead of announcing another distant roadmap.

Marvell may match that speed on a future program, especially after expanding its design and connectivity platform. Broadcom owns the best recent proof.

Q11Whose custom AI chip strategy is harder to copy: Broadcom’s or Marvell’s?

Broadcom currently has the harder custom AI chip strategy to copy because one major XPU win can pull through several generations of compute, packaging and networking revenue.

A hyperscaler rarely changes its silicon partner casually. The customer and supplier spend years aligning the architecture, physical design, memory system, packaging, software and cluster network. Once the first chip works, the second and third generations can arrive faster and with less execution risk. Broadcom’s long Google relationship shows how durable that position can become.

Broadcom then adds more products around the processor. Ethernet switches, SerDes, optical components and PCIe connectivity increase the revenue attached to one customer program. They also give Broadcom more information about how the whole cluster behaves, which can improve the next XPU.

Marvell spreads its bets more widely. The company pursues central XPUs and the attach chips around them, including networking, memory, security and optical components. That approach lets Marvell earn money even when another supplier owns the main accelerator. It also reduces the damage from losing a single processor bid.

Broadcom’s lock-in has limits. During the latest earnings discussion, management acknowledged that a major custom-chip customer could diversify suppliers. A hyperscaler can keep Broadcom on its main program while handing a smaller generation or a surrounding component to another vendor. That is exactly the opening Marvell needs.

Marvell gains resilience from its breadth, but gives up some control. The central XPU partner shapes more of the architecture and usually earns more from each program. Attach sockets face more frequent competition, and customers often have more suppliers to choose from. Broadcom’s deeper position remains harder to break.

Q12Can Marvell’s Nvidia deal and acquisitions close the gap with Broadcom?

Marvell’s Nvidia partnership and recent acquisitions give the company a real path to narrow the custom AI gap, but most of the financial payoff still lies ahead.

Nvidia invested $2 billion in Marvell and brought the company into NVLink Fusion. Marvell can now pair custom XPUs and scale-up networking with Nvidia CPUs, network interfaces, DPUs, switches and rack-scale systems. For a customer that wants specialized compute while keeping Nvidia’s broader ecosystem, Marvell becomes an easier choice.

The optical acquisitions target another pressure point. Celestial AI adds a photonic fabric for high-bandwidth, low-latency scale-up connectivity. Marvell expects Celestial AI to reach a $500 million annualized revenue run rate late in fiscal 2028 and $1 billion late in fiscal 2029. XConn adds CXL and PCIe switching, while Polariton strengthens Marvell’s high-speed optical component technology.

Marvell can narrow the gap without displacing Broadcom everywhere. A custom processor tied to NVLink, an optical fabric around another company’s XPU, or a CXL and PCIe socket inside a mixed system can each add meaningful revenue. Several wins across those layers can add up.

Timing and integration now decide whether the strategy pays off. Celestial AI and XConn entered Marvell’s accounts only from their acquisition dates, and the largest revenue targets begin later. Marvell has bought a strong collection of assets. Customers still have to adopt them as one platform rather than a set of separate components.

Q13Who has more financial room for custom AI chips: Broadcom or Marvell?

Financially, Broadcom has far more room than Marvell to fund custom AI chips, reserve supply and absorb the inevitable delays that come with large semiconductor programs.

Broadcom generated $10.26 billion of free cash flow in its latest quarter and ended the period with $19.63 billion in cash. Marvell generated a record $639 million of operating cash flow and held $3.84 billion in cash.

The cash-flow measures are different, so an exact multiple would mislead. The scale difference is still obvious. Broadcom can finance several leading-edge programs at once, absorb a failed tape-out, support customers through a delayed data-center build and commit capital to supply before the revenue arrives.

Marvell remains financially healthy, with Nvidia now a strategic shareholder, and the company has enough scale to buy Celestial AI and XConn while continuing heavy research spending. Its current plans are fully fundable.

Broadcom’s advantage appears when the projects become unusually large. A ten-gigawatt program or a multi-generation 2nm project can require years of engineering and supply commitments. Broadcom can take those bets without making the whole company depend on one successful ramp.

Q14Could Broadcom beat Marvell in custom AI chips but still lose the wider AI race?

Broadcom can lead Marvell in custom AI chips while Nvidia remains the dominant force in the broader market for AI compute.

Custom accelerators work best for a small number of customers with huge, predictable workloads. Those companies can justify the design cost because a purpose-built chip may lower power use, improve inference economics or reduce dependence on merchant GPUs. Most enterprises cannot support that development effort.

Nvidia still offers the broadest software ecosystem, a mature developer base and processors that can handle changing research and production workloads. Even the largest buyers are likely to mix custom XPUs with Nvidia systems rather than choose one architecture for every task. Marvell’s NVLink Fusion partnership openly embraces that mixed future.

The custom market also remains concentrated. A handful of cloud companies control most of the available programs, and each company can split work among suppliers or bring more design in-house. Broadcom may dominate the independent design-partner layer without capturing most AI accelerator spending overall.

So the claim stays narrow: Broadcom is winning the contest to become the main outside partner for hyperscalers building their own accelerators and Ethernet-based systems. Nvidia continues to set the pace in general-purpose AI compute.

Q15Who leads custom AI chips right now: Broadcom or Marvell?

Broadcom clearly leads Marvell in custom AI chips right now, with much greater production revenue, firmer future demand and a stronger ability to deliver the processor with the network and infrastructure around it.

The current commercial gap carries the most weight. Broadcom already runs an AI semiconductor business several times larger than Marvell’s entire data-center division. Broadcom is also accelerating faster from that larger base, while Marvell’s sharpest custom-silicon growth remains scheduled for later fiscal years.

The customer evidence also favors Broadcom: mature Google production, a large OpenAI program, multiple Meta generations and committed Anthropic capacity. Marvell has an excellent AWS relationship, wider pursuit activity and a valuable Nvidia alliance, but fewer publicly quantified orders.

The third deciding factor is system control. Broadcom can connect custom compute, Ethernet, optics and rack infrastructure inside one customer plan. Marvell’s optical and scale-up portfolio may become more important as data movement consumes a larger share of power and cost. For now, that possibility strengthens Marvell’s long-term case without changing the leader.

Broadcom now needs to turn its new gigawatt commitments into stable clusters and repeat orders without letting customer concentration or deployment delays damage the economics. Marvell needs its new tier-one XPU, attach programs and acquired optical platforms to produce several billion dollars of real revenue. A couple of successful ramps would make the contest much closer. Until then, Broadcom’s lead is decisive.

Broadcom vs Marvell custom AI chip verdict