MRVL上调AI连接增长目标
公司直接上调FY2028营收目标，并提供FY2029定制业务与FY2031收入、利润率和现金流目标；问答对大型协议纳入程度及增长约束作了补充。
中文摘要
核心结论
Marvell在2026年10月6日投资者日把FY2028营收目标提高至约200亿美元，并提出FY2031营收700亿至900亿美元、非GAAP（非通用会计准则）每股收益超过30美元的长期目标。管理层将增长归因于数据中心互连、交换/存储及定制芯片；目标能否兑现取决于客户项目量产、资本开支和供应链执行。
重要性评级
5/5（高）：这是MRVL年度投资者日，公司上调近期开支周期内的营收目标，并给出FY2029定制业务和FY2031营收、利润率与现金流框架；问答还说明大额协议机会如何进入预测。
关键事实
- 管理层估算，CY2030 Marvell可服务市场（TAM）约4000亿美元，2025至2030年复合增速约45%；其采用的数据中心资本开支假设约3万亿美元，复合增速约35%，并预计周期后段放缓至约20%。这是公司市场测算，不代表已取得收入。
- FY2028总营收目标升至约200亿美元，其中数据中心约180亿美元；公司称较FY2026的82亿美元总营收增长显著。FY2029定制业务目标提高至超过120亿美元，管理层称从FY2028水平增长逾三倍。
- FY2031总营收目标为700亿至900亿美元，数据中心约675亿至875亿美元；中点800亿美元按业务分为互连375亿美元、交换与存储100亿美元、定制300亿美元、通信及其他25亿美元。目标区间对应公司所述总营收复合增速约55%至60%。
- 定制业务涵盖XPU（AI加速器）及XPU attach（围绕加速器的定制配套芯片），包括网络接口、内存扩展、存储、服务器管理和推理加速。管理层称FY2031业务将由XPU与配套芯片较均衡构成，且不依赖单一插槽；这是基于公司所称的设计项目和客户关系作出的预测。
- 公司称机柜内扩展（scale-up）光学业务当前收入接近零，预计下一财年开始显著放量；CEO在问答中称初期规模可达数亿美元。近封装光学（NPO）与共封装光学（CPO）将并行推进。公司同时称100T以太网交换芯片已有产品发布，51.2T交换业务预计下一财年超过10亿美元；这些是不同业务和进度，不能把交换收入目标当作光学收入。
- 管理层称Google的TPU相关定制协议勾勒出六年半累计约1200亿美元的机会，且与现有连接业务分开。CEO在问答中称FY2031定制业务模型只纳入近期大型协议机会的一部分，按其判断的较可能情形估算；1200亿美元是公司描述的机会规模，不能视为已签订单或确定收入。AI实验室相关业务被CEO称为潜在上行选项。
- CFO给出的FY2031目标为毛利率56%至59%（受产品组合影响）、营业利润率44%至46%、税率15%、自由现金流率超过36%，以及非GAAP每股收益超过30美元。公司计划长期平均返还超过50%的自由现金流，同时继续投入增长并维持灵活资产负债表。CFO回顾FY2021至FY2026约80亿美元研发投入、110亿美元并购投入及约40亿美元股息和回购；演讲没有给出FY2031股数、稀释假设或未来回购金额。
作者观点与证据
这是Stock Analysis发布的投资者日演讲及问答转录，主要证据是Marvell管理层陈述。管理层称前十大云客户占数据中心资本开支比重将继续上升，公司已在多代产品上取得设计项目，并已为增长安排供应链产能；转录未附客户合同、订单明细或独立验证材料。问答中，CEO称公司预测假设资本开支逐步放缓，FY2031定制模型按“较可能”情形估算，并非把大型协议的最大潜力全部计入。关于土地、电力和数据中心机房容量，管理层称当前未见对已规划业务构成约束；这属于公司判断。
与相关标的的关系
内容直接涉及Marvell Technology（NASDAQ: MRVL）的业务组合、客户设计、产品路线图及多年财务目标。公司把互连、交换和配套定制芯片视为数据中心增长来源；演讲中的市场规模和收入预测仍是管理层展望。
时效性与限制
活动日期为10/06（未给出具体时刻）。财务目标、市场规模、客户项目进度和量产预期均来自公司演讲，不能当作已实现业绩。转录没有提供FY2031每股收益对应的股数计算、未来稀释或回购数量，也没有逐项说明已公布协议中的最低采购承诺。
后续跟踪
观察FY2028总营收及数据中心目标的兑现、FY2029定制业务的爬坡、scale-up光学从下一财年开始的收入和NPO/CPO项目量产进度；同时核对交换与存储、定制业务的收入拆分、资本开支变化、自由现金流返还执行，以及实际股数和稀释变化。
证据限制：Original text is extracted for authorized station-internal report reading under configured access protection.；第三方转录，公司目标未独立证实。全文仅内部归档，读者报告用简短转述和原链接。
原文
Marvell Technology (MRVL) Transcript: Investor Day 2026
Collapse
Marvell Technology, Inc. (MRVL)
NASDAQ: MRVL · Real-Time Price · USD
Full Chart Watchlist Alerts Compare
287.01
+15.76 (5.81%)
At close: Oct 6, 2026, 4:00 PM EDT
282.00
-5.01 (-1.75%)
Pre-market: Oct 7, 2026, 5:21 AM EDT
← View all transcripts
Investor Day 2026
Oct 6, 2026
Play Audio Download Transcript
Summary
FY28 revenue guidance rose to about $20B, while FY31 targets reach $70B-$90B revenue and >$30 non-GAAP EPS. Connectivity and custom/XPU attach underpin growth, supported by a roughly $400B CY30 TAM.
Operator
Please welcome to the stage, Marvell Chairman and Chief Executive Officer, Matt Murphy.
Matt Murphy
Chairman and CEO, Marvell Technology
Good morning, everybody. All right. All right, good morning, everybody. Yeah, there you go. I am fired up to be here today. Are you? Come on. All right, so welcome to Marvell's 2026 Investor Day. It is great to be back in New York and have so many of you with us today. I have great memories of the last time we were all here, which was about two and a half years ago. I remember having lunch afterwards and sitting with a bunch of you, and one of you spoke very fondly of this hotel. You had gotten married here. Another investor at lunch said to me, "You know, Matt, good call on the location at The Pierre. The Pierre is a bull market hotel." Look, I agree, from April 2024 to now, it has definitely been a bull market for semiconductors, hasn't it?
I am excited to be back here today to show you what the next phase of Marvell looks like. All right. Blasts from the past. Let us go back in time. This photo is from Marvell's first ever Investor Day in 2017, and that is me there on stage, and a number of you were there. In fact, I can recognize many of you just from the backs of your heads. You can probably find yourself. By the way, I am not sure who this young guy is on stage, but he is pretty handsome, isn't he? This is what 10 years as a public company CEO will do to you, okay, but I am still kicking.
I remember the first question I got at that Investor Day after we finished. Chris Rolland from Susquehanna, who I think is here today, stood up and very seriously asked me, first question, and said, "What company is this? Is this Marvell?" There was a reason he asked that question because, at that time, Marvell had never communicated ever to Wall Street what its mission or vision was or where the company was going. The company had gone public in 2000, by the way. Most people back then thought of Marvell as a consumer electronics company, right? Selling into smartphone, tablets, other digital devices. But look at what we were talking about back then. Look at my slide. The explosion of data and bandwidth demand, compute moving towards the data, end-to-end security, the need for power efficiency.
This was in March 2017, and we were already talking about the data center, data infrastructure, and where we believe the industry was going. Before Marvell, I spent 22 years at Maxim Integrated Products, which was one of the leading analog semiconductor companies. One of the great things about the analog business is you get exposure to virtually every end market and every major technology trend. The one thing I learned from that experience was that the market that you sell into is what matters. It is not just about having the best chip or the best part, the best widget, it is your end market exposure and then your long-term sustainable competitive advantage within that market. That is what matters. So actually, when I came to Marvell, I had a strong point of view about where the semiconductor industry was going and where I wanted to take the company.
That point of view I had was centered on data infrastructure in the cloud, even though at that time, most of Marvell's revenue was tied to consumer. In fact, back then, less than 10% of our revenue was actually in data center. That is what led us to develop Marvell's mission statement. Here it is. So we wrote this in January 2017, about six months after I became CEO. Nearly 10 years later, this mission has not changed. It is exactly the same words from January 2017. This might seem like motherhood and apple pie today. Make high performance chips that move data, store data, process data, compute networking storage, all those types of trends. But I can tell you, 10 years ago, this was unique. There was no category back then called data infrastructure. The whole semiconductor market was defined very differently back then.
I remember recently I was in a meeting with one of our senior fellows, who is a longtime Marvell employee. He was presenting to a big engineering group and he said, "I remember when Matt first presented this mission statement to us when we were sitting around at Marvell." I had already been at Marvell for 15 years, and I remember sitting there watching him and thinking, "I do not know what he is talking about." Then he stood up and he said, "Now, 10 years later, I see it. We have actually all the pieces now. Moving data, storing data, processing data, all the key elements. We did not have those back then, and we do now." So that was the vision.
Over the past decade, at each of our investor days, and we have had a bunch of them, we have continued to share our point of view about where we believe the industry was going. So let me show you a few examples. In 2018, eight years ago, we were talking about AI as a major driver of data center and how AI would evolve across both training and inference. It is a long time ago. This is pre-ChatGPT, pre the big AI wave. 2020, we were talking about the shift from application-specific workloads to data-centric workloads, with functions like networking, security, and storage increasingly moving into dedicated infrastructure silicon. 2021, we broadened that to the era of cloud-optimized silicon. This was five years ago. We said that compute, storage, networking, security, and optics were all becoming increasingly purpose-built.
And if you look at what we were describing back then, those were really the building blocks of what we call today XPU attach, before we even had a name for it. In 2024 when we were here in New York at the Pierre, we talked about how front-end and back-end networks were required for AI, and how optical connectivity would be a major driver of the next wave of AI infrastructure. Then finally last summer, in June, we showed you the breadth of our custom silicon capabilities, including the emerging opportunity we call XPU attach, and why we believe custom was going to be a major and much larger part of AI infrastructure. Throughout this journey, we've had a pretty consistent point of view where the industry was going, and we've invested aggressively ahead of these transitions.
But you can have a vision, you can see the shifts, but that is not enough. We had to build this company and the capability to capitalize on the opportunity. So what did we do? The mission in 2017 became the guiding principle for how me and the team allocated capital and architected Marvell over the last decade. In some areas, we invested in strong technology and expertise that we already had inside Marvell. We had a very talented team when I joined in engineering. In other areas, though, we pursued very strategic and, quite frankly, very ambitious M&A to add critical capabilities and domain expertise in a number of new important areas where we didn't have that capability. At the same time, though, we didn't just buy things. We narrowed our focus through divestitures, which created significant shareholder value.
We received premium multiples for these businesses that we sold, and it allowed us to continue to hone our focus on the data center. But I got to tell you, the real value of the acquisitions we made and those capabilities that we then gained, those came after those teams were inside Marvell and then part of our platform. We brought these teams together around one strategic direction and continued to invest aggressively in those businesses. And that allowed those teams to achieve far more than they ever could have on their own. So if you take a step back, we've invested roughly $20 billion bringing those capabilities into Marvell, decades and decades of expertise. We also invested another roughly $20 billion organically behind the platform. This was rocket fuel for the roadmaps and helped establish the leadership positions we have today and that you're going to hear about.
So that's roughly $40 billion of investment in the Marvell platform over the last 10 years. This has truly been a decade in the making. And the platform has actually arrived at the right moment. AI workloads are fundamentally changing the way that data moves through the network. We're moving beyond traditional training and reasoning models. We're moving beyond training into reasoning models, agentic AI, and inference now at scale. So those workloads are driving more and more data movement and diverse traffic patterns across AI infrastructure. And that's creating greater demand on connectivity. Efficient data movement is becoming a fundamental constraint on AI infrastructure. And this isn't just us saying it. This is what we're hearing from our customers. The world's largest hyperscalers are rethinking their network architectures as AI infrastructure scales. Across the industry, these changes are driving significantly more connectivity across the system.
Just take a moment and read it and look what they're saying. It's all about connectivity. If you look at the AI landscape overall today in semi companies, there are three very large companies whose businesses are primarily driven by compute. There's some networking in there, but compute is what they do really well. We've seen tremendous growth and value creation in the market from the compute wave that we've gone through. Then you have three large memory companies whose businesses are overwhelmingly concentrated at memory. That's what they do really well. Again, we've seen tremendous growth and value creation there as AI has driven the need for more memory, bandwidth, and capacity. But now we believe connectivity is the next defining unlock. At system scale, more and more data needs to move across every distance, and Marvell is fundamentally a connectivity-first company.
That gives us a very different position in the ecosystem. We work with the leading compute companies, the leading memory companies, and our customers, all the world's largest cloud operators. In many ways, we're enabling the entire industry, and Marvell is a partner to all. As AI infrastructure continues to scale, we have exposure across the ecosystem, regardless of which compute or memory architecture is deployed. Okay, let's put some numbers behind that. In FY 2026 last year, Marvell generated $8.2 billion in total revenue, 75% of that was in data center, and we break that down into the segments we're going to talk about today: interconnect, switch and storage, and custom. Most of our custom revenue last year was in compute, and some of our comms and other revenue is also in compute.
That gives you about 30% of total revenue landing in compute and 70% in connectivity. As we go forward, and as XPU attach ramps, more and more of the custom business will actually be part of our connectivity story as well. Coming back to data center overall, we'll walk you through these three categories today and show you how the technology is evolving, how the market is expanding, and what that means for Marvell. Then later, we'll come back to these businesses and look at the TAM, our growth expectations, and the long-term opportunity across each one. All right, so let's start with interconnect. As these network architectures evolve, connectivity has to scale across every distance in the system, from links between data centers to connections all the way inside the package. These aren't variations of the same problem.
Each distance represents a fundamentally different engineering challenge, requiring different technologies, different skill sets, and unique expertise. At longer distances, like scale across and scale out, the industry has already transitioned into optics. Marvell is a leader in those markets, and that's what's driving our interconnect revenue today. But as you move inside the rack and closer to the package, those connections today are still largely copper. In the middle is what we call the copper wall. But here's what's interesting, that wall is starting to move. As bandwidth continues to increase, the distance copper can reach gets shorter and shorter. That means connections inside the rack are starting to begin to transition to optics. That opens up an entirely new TAM for Marvell in scale-up interconnect. This is where technologies like near package optics or NPO or co-packaged optics, called CPO, are becoming increasingly important.
That transition plays directly into the optical capabilities we have been developing for years. You are going to hear much more about this today. In longer term, the same physics continues to apply. As bandwidth keeps increasing, the transitions to optics moves closer and closer to the compute. Ultimately, you can see and we see a world where optics extends all the way into the package. That is going to create another significant TAM opportunity for Marvell that we call scale in, and Xi is going to talk to you about that today. Let me move on to switching. We have established a strong position in scale-out switching with multiple customer wins and a business that will be well north of $1 billion next year. Now we are taking that same technology foundation into scale up, where accelerators require extremely high bandwidth, low latency connectivity. We are unique.
We are supporting all the major scale-up protocols, including NVLink, UALink, and Ethernet, giving customers flexibility across their different architectures. As bandwidth increases, scale up is becoming increasingly optical, which opens up an exciting new opportunity for Marvell, and you will hear much more about that today as well. Let us turn to storage and memory. Storage has been a core franchise since the earliest days of the company. We established leadership early on in hard drive and flash controllers, but over time, we have expanded that technology base across a much broader range of storage and memory technologies. Today, that spans HDD and NAND controllers, memory controllers and CXL, memory expansion, and best-in-class dense SRAM used in our custom products. What is important is that we have proven technology and decades of expertise across every layer in the storage and memory technology stack.
Increasingly, those same technologies are becoming the foundation for custom solutions that we design for hyperscale AI infrastructure. So customers can start with proven Marvell IP, work with us to optimize and customize it for their specific architecture. So what began as a leadership position in storage has become a much broader capability spanning storage and memory, and it is both in standard products as well as custom solutions. Let us go to custom. This has become a massive growth opportunity for Marvell. We are in this business because customers want access to Marvell's differentiated IP and ability to use that technology to realize their own unique architectures. This started with the XPU, but the opportunity today is actually much larger. As XPUs continue to grow, they pull through a whole new set of functions around them: networking, storage, memory, security, and inference acceleration.
All of these functions are attached to the XPU, and increasingly, we are seeing them all move to custom solutions. That is what we call XPU attach. With the XPU, you are typically talking about a traditional, full custom ASIC model. XPU attach is a little different. These opportunities are more IP-led and semi-custom in nature. That difference between full custom and semi-custom is actually important, so I am going to take a minute just to talk about it. In many cases, Marvell works with the customer on the architecture and takes the lead on the front-end design with the proven Marvell IP set at its core. That IP comes from technologies that we have already developed, shipped at scale, and proven in high volume manufacturing. So it is proven. That model is gaining significant traction with customers.
We have already announced multiple large strategic agreements around these opportunities, and today we are shipping custom silicon to all four U.S. hyperscalers. We have grown our custom business essentially from scratch over the last five years, and last year it was about a $1.5 billion business. Along the way, we have set some pretty aggressive targets, including on this stage back in April 2024. Last year, we told you we expected custom revenue to more than double in FY 2028, putting us on a path to more than $10 billion in FY 2029. Today, that business is on track against that plan. Based on the opportunity we are seeing, we are raising our FY 2029 outlook for this business. We now expect custom revenue in FY 2029 to more than triple from FY 2028, reaching more than $12 billion and putting us well above our prior $10 billion target.
This update to our expectation is primarily driven by the rapidly expanding XPU attach opportunity we see in front of us, including the strategic agreement that we announced in August. The XPU attach opportunity continues to expand every day, and more importantly, this is not a single customer and it is not a single socket. We are seeing a broad set of product wins across multiple customers and multiple programs, spanning all the areas I just mentioned. That is what gives us confidence to raise the outlook from more than doubling to more than tripling. Let you digest that one for a second. All right, so let us take a step back and look at the market opportunity. By 2030, we expect total data center CapEx to reach roughly $3 trillion, growing at about a 35% compound annual rate, starting from calendar 2025 to calendar 2030.
Importantly, that assumes that the current rate of CapEx growth moderates over that period, moving closer to historical levels in the outer years. As you guys all know, we are growing faster right now than the 35%. So in our model, we expect some normal deceleration to a more normal base case in the outer years. Within that, accelerated compute represents roughly $1.6 trillion of TAM opportunity. Within that infrastructure, we see approximately $385 billion of data center TAM addressable by Marvell. When you add in the opportunity in the rest of our business, in communications and other, that represents another approximately $15 billion. So when you put it all together, we see an approximately $400 billion total addressable market all in for Marvell by 2030. That market that we are addressing is growing at a 45% compounded rate from 2025 to 2030.
It is growing faster than CapEx. To put that in perspective, just so everybody knows, there are a lot of big TAM numbers out there these days, but for us, the last time we laid out the TAM for the company, it was less than $100 billion. So this is like a 4x increase to our TAM opportunity that has occurred. That is how fast the market is moving for us. The growth though, which is a good thing, is not concentrated in just one part of the business. So let me dive deeper and show you what is underneath the $400 billion. All right, there are a lot of numbers up here, so let me give you a minute to absorb them, but I am going to start at the top and work my way down on the TAM.
In interconnect, we see the TAM growing at roughly 65% annually, reaching approximately $65 billion by 2030, as scale-out continues to grow and new opportunities emerge in scale-up and scale across. In switching and storage, we see the TAM growing at roughly 40% annually, reaching approximately $85 billion. This is driven by continued growth in scale-out switching and then the new emergence of scale-up switching. In custom, the market is expanding dramatically. We see the TAM growing at roughly 55% annually, reaching approximately $235 billion by 2030, driven by the growth in XPUs and then the layering in and emergence of XPU attach on top. In communications and other, as in approximately $15 billion. Put it all together, you get the $400 billion TAM I showed you, growing at 45% annually through 2030. Look, these are all very healthy growth rates.
But I want to point out, interconnect stands out at roughly 65% compounded growth over the period. As I said earlier, when you look at where the growth is coming from in the industry, it's connectivity. This breadth gives us the confidence in the overall growth potential of the company, even if the exact mix over time evolves and changes. Now let's look at what all of this means for our revenue outlook. In FY 2026, as I said, Marvell generated $8.2 billion in revenue, and this year we expect approximately $12 billion. That would be 47% year-over-year growth. Looks great. We're excited about that. On our August earnings call, we raised our outlook to approximately $18 billion. That represents 50% year-over-year growth, showing that even as the company gets significantly larger, our growth rate is actually accelerating. Today, we're raising that outlook again.
We expect approximately $20 billion in total company revenue in FY 2028, representing about 67% year-over-year growth. And then within that, we expect data center revenue to grow approximately 80%. When we raised our FY 2028 outlook in August, that increase was driven by interconnect, scale-up optics, and switching. The additional $2 billion of growth we're seeing for next year is being driven by the same trends. Scale-out optics continues to accelerate, scale-up optics continues to grow, and our switching portfolio is firing on all cylinders. So once again, it's connectivity leading the race. Okay, now I'm going to put this $20 billion in context for everybody. So you take a step back, when we were here in April 2024 on stage at the Pierre, the prior year, we had just printed $2.2 billion in data center revenue, which is our FY 2024.
At the time, we told you we believed we could grow our data center business to $15 billion by FY 2029. That was going to be 7x growth in five years, and I think people had a hard time wrapping their head around it. And I know, it was ambitious. I got it. It was ambitious. A year ago, in June, we had our custom AI event, and we basically said, "Look, the opportunity is developing faster than we thought." So at that time, we raised our FY 2029 data center revenue from $15 billion to approximately $18 billion. In other words, the target we had set a year earlier was already moving higher. Okay, now I'm going to show you where we are today. With the $20 billion in total company revenue we outlined for FY 2028, approximately $18 billion is going to come from data center.
That means we now expect to reach the $18 billion data center target we set for FY 2029 almost a full year early. That would represent roughly a 70% compounded annual growth for our data center business over that period from FY 2024 to FY 2028. Stepping back for a second. We've put some ambitious targets in front of you over the last few years, and the Marvell team has executed. We've gone from $2.2 billion of data center revenue in FY 2024, and we've got our plan to go to $18 billion next year. We're hitting our goals, and we're doing so earlier than planned. With that perspective in mind, let me now show you how we're thinking about the opportunity from here as we go forward.
For our purposes of our investor today, we're going to set our growth plan starting with FY 2026 last year and outline the opportunity through FY 2031. This matches the TAM framework I showed you earlier from CY 2025 to CY 2030. In FY 2026, of the $8.2 billion, $6.1 billion was data center. I showed you just now the plan to get to $18 billion in data center revenue in fiscal 2028, with $20 billion total company revenue. So during that period, FY 2026 to FY 2028, that's about 70% growth in our data center business compounded. If you remember, that's about the same growth that we were looking at from FY 2024 to FY 2028. So basically, we've been growing pretty consistently since calendar 2023 in our data center business. For FY 2031, we have a bottoms-up plan that I'm going to walk you through shortly. This is not just top-down.
I'm going to walk you through the bottoms-up in just a moment. But let's start at a high level. Looking forward from FY 2028, if you assume the $20 billion total revenue plan and you roll up by customer and product and design win and you layer it up, we see a path to continue growing our data center revenue at roughly 70% annually through FY 2031, despite the moderating CapEx assumptions. But look, it's a dynamic market. So today, when we talk about our revenue goals, we're going to give you ranges. At the top level, we're going to put a range starting at 55% at the low end. So taking that 55%-70% growth range, that implies data center revenue of approximately $67.5 billion to $87.5 billion in FY 2031.
Now, if you add in $2.5 billion for comms and other, you get a range of $70 billion to $90 billion in revenue for Marvell in FY 2031. I'll let you digest that for just a second. I understand. The absolute numbers are large, but the growth rates are at or below what we're already delivering today. With those numbers in mind, let's take a step back and look at the company as a whole. $8.2 billion in revenue last year in FY 2026. Plan to deliver $70 billion to $90 billion in five years, $80 billion at the midpoint. That would represent 10x growth over a five-year period and 55%-60% compounded annual growth over the same period. That's the growth framework behind the long-term outlook that we're setting today. These are very strong growth plans, but what gives us conviction is not the top-down view.
It's the bottoms-up plan, and we have a history of delivering at these growth rates. All of this is grounded in business we have today, our customer positions, and the opportunities that are already right in front of us. Let me now walk you through what underpins the bottoms up. Here we go. Okay, I'm going to walk you through one by one each business and what gives us conviction in our outlook. Let's start with interconnect. Scale-out has been an incredibly strong business for us, and in the last few years, that business has been roughly doubling each year. In this model, we actually assume that that growth rate moderates over time. At the same time, though, I want to say, keep in mind, we're just beginning our transition to 1.6T in volume in the market, and we have 3.2T ahead of us.
On top of that scale-out, which you're going to hear from Russ on, has a very strong future. In scale-across, we already have a very strong position with new use cases for that market that we're going to hear about. Scale-up optics, which is part of this, that's all in front of us. We have virtually no revenue in this area today, with strong growth beginning next year, and we already have significant traction. On interconnect, what we're assuming is actually pretty straightforward. Maintain our positions in scale-out and scale-across in a very high-growth TAM market, and then execute on the scale-up opportunity in front of us. That's what we've got to do. If we execute that plan, that gets us to $37.5 billion at the midpoint in FY 2031. Now let's look at switching and storage.
Storage is already a billion-dollar-plus business, and we assume it continues to grow at a double-digit rate. In scale-out switching, as I said earlier, we're going to be well north of a billion dollars next year with our 51.2T product just ramping in its first full year of production and our 100T product still in front of us. Based on the customers and programs we already have, we see scale-out switching becoming a multi-billion-dollar business over time. On top of that is an entirely new opportunity in scale-up switching. We already have multiple deep customer engagements in this area today, and as they materialize, those are going to provide another meaningful layer of growth in our switching business. When you put those pieces together, we see line of sight to approximately $10 billion in switching and storage revenue at the midpoint. Okay, now there's custom.
Custom actually might be the simplest of all of them. I just showed you that we expect this business will reach more than $12 billion in FY 2029. Our existing XPU programs are on track, and we're assuming that those programs ramp into that plan and continue to grow at a normal rate from there. On top of that, though, is the rapidly expanding XPU attach opportunity, supported by the design wins and our strategic customer agreements. Just to be clear, we've taken a very pragmatic view in judging the total opportunity covered by the large agreement we recently disclosed. If you start with more than $12 billion in FY 2029, the growth we're assuming from FY 2029 to FY 2031 is roughly in line with the market. We think that's a very reasonable assumption given the programs and the customer positions we have today.
That gets us to approximately $30 billion at the midpoint in FY 2031. By FY 2031, we expect custom, at that time, to be much more balanced between XPU and XPU attach sockets with no single socket concentration. Finally, in comms and other, we are assuming that we just maintain our position with very modest growth over the period. So that gets us to approximately $2.5 billion at the midpoint. So when you put all of that together, that's the bottoms-up case behind our $70 billion-$90 billion FY 2031 target. When you look at the assumptions behind each of these businesses, I think they're very reasonable. These are businesses that we have today, customer positions we already have, and opportunities that are in flight and in execution, and the opportunities are continuing to develop. We feel very good about our ability to get there. It's pretty exciting, isn't it?
It's exciting. Thank you. I'm excited. We're excited. You're going to hear about it today. Okay, so I'm going to wrap it up here and bring it all together for you. This is why invest in Marvell. I've given you a lot of big numbers today, so let me just take a moment and recap the ones I want you to remember. By calendar 2030, we see approximately $400 billion of total addressable market for Marvell. That's up 4X from the last time we updated you. In FY 2028, next year, we expect approximately $20 billion in total company revenue, up from the $18 billion outlook we gave you in August, driven by connectivity. In custom, we expect FY 2029 revenue to more than triple, reaching more than $12 billion, putting us well above our prior target.
Looking out to FY 2031, we see a path to $70 billion-$90 billion in annual revenue with $80 billion at the midpoint. As you'll see throughout the day, we're positioned as an ideal partner for all these hyperscale data center customers with deep trusted relationship and years and years of working together. They look today at Marvell as their primary connectivity partner, and we're deeply embedded across their entire architecture. So although the overall data center market by design is concentrated among these hyperscale customers, Marvell has built an incredibly diverse business with hundreds of products being delivered across this customer base and no single socket controlling our destiny. So when you see that kind of revenue growth, you're probably already doing the math. In fact, I see a lot of heads down in the audience on your PCs. Look up, stay with me. Hold on, we're almost there.
Because what you're doing, and you're probably wondering is, "Well, what's the earnings power for such a company of such scale?" I'm not going to give you that right now. I'm not going to do it. We have a new CFO at Marvell. He's outstanding, and his name is Dan Durn. He's been with us since June, but I've known him for a very long time. This is Dan's first Investor Day with Marvell, and I certainly don't want to steal his thunder. Dan's going to come back later and show you what this growth can mean for Marvell's earnings power and put it in the context of the overall financial model. Gross margin, operating margin, EPS, the whole nine yards. So for now, I've given everybody the revenue as a starting point. Okay.
Before I wrap up, I want to recognize the Marvell team that makes all of this possible. None of what I've shared with you today happens without the incredible work and commitment of our employees all around the world. For everyone at Marvell watching and listening today, thank you. I've been at Marvell for a little over 10 years now, and I've never been more excited and fired up about where this company is headed. We have an incredible team, the right technology, deep trusted relationships with the most important, significant customers in the industry, and relationships that we've built over many, many years. We have an opportunity we're looking at in front of us that is larger than anything we've ever seen before. I truly believe the best is still in front of us.
Thank you so much, everybody, for joining us today, and I'll talk to you guys later during the Q&A. Thank you.
Operator
Please welcome President and Chief Operating Officer, Chris Koopmans.
Chris Koopmans
President and COO, Marvell Technology
Good morning. It's great to be here with all of you today talking about this massive opportunity in front of us. This past summer also marked my 10-year anniversary at Marvell, and I've had a blast. I'm here to tell you that our best opportunity, our most exciting times are in front of us. We spent all these years building this platform and this company for this moment, and I'm going to talk about how we've been doing that and how we've been planning to scale this company. I'll talk about it along these three pillars. I'll start off talking about our customers, where are we focused, and how do we win. I'll talk about our core engineering and our systems and how are we set up to scale to deliver for our customers. Then I'll talk about the relationships in securing long-term capacity in the supply chain.
Let's start with our customers and where we see the overall TAM going. Last year at our Custom AI event, I showed that analysts were forecasting about $1 trillion in total data center CapEx by calendar 2028. That was showing basically a modest 20% growth CAGR from going forward after a really strong growth year in 2025. I remember when we put the $1 trillion up there, it kind of stuck in my mouth a little bit. $1 trillion? That's a lot of money. Is this a little bit crazy? Then you fast-forward a year and a half, and it wasn't crazy at all, actually. We ended 2025 at over $700 billion, and now we're hitting over $1 trillion this year already. Clearly, this is a fast-moving market and one that's hard to forecast, and it's growing faster than anybody expected.
As Matt said, analysts are now forecasting over $3 trillion by calendar 2030. That's after a much faster 2025, 2026, and even 2027, we're still expecting CapEx to moderate over time to about 20% after that. This is the CapEx envelope that all of the numbers we've shared with you today, okay, the total market, the $400 billion TAM, all the revenue targets, that's the assumption, the total data center CapEx. If you look inside that data center CapEx, the top 10 or so hyperscale customers are actually driving over 60% of that last year. As large as they are, they're actually expected to grow their spend even faster than the overall TAM, reaching more than 70% out in calendar 2030. That's astounding. If you do the math, that means that these top 10 companies are 200-plus billion-dollar TAMs each.
If you look within that, the top four U.S. hyperscalers alone were more than half of total data center CapEx last year. So you can see that any customer strategy in this market really needs to be hyper-focused on those companies. It's not just the largest economic opportunity. These are the technology leaders in the world, and so the focus that we put on them will pay dividends throughout the rest of our customer base. In fact, we have architected this company to be the perfect partner for the hyperscalers. They're large and sophisticated operators. They've been designing and deploying complex infrastructure for decades. They actually define what it means to operate at scale. They've seen in the real world how these workloads interact with the underlying architecture, and they have a point of view for how to optimize that architecture for the future and evolving workloads.
What they need is a broad portfolio of connectivity technologies to stitch it all together. At every hop and every distance, Marvell has a full portfolio of capabilities. Our teams are able to sit together and make bandwidth, reach, power, and manufacturing scale trade-off decisions and literally co-architect future cloud data centers. So at Marvell, we are laser-focused on these top 10 hyperscalers. In fact, within that top 10, we can hit the vast majority of the goals that Matt laid out with the top four U.S. hyperscalers alone. We aren't counting on some major new customer win or a new home-run design win. We need to focus on this finite set of customers and execute. So we've been doing this for a decade now, and our typical customer journey starts off with an off-the-shelf system, a branded storage switch optical module with Marvell silicon inside.
This is how these hyperscalers get experience operating networks with Marvell chips. The next step is often to build a custom system with Marvell's standard silicon inside, optimized for their overall network design. Next, they will often ask for a custom feature, or maybe they want to put some of their IP in an otherwise standard chip, adding their secret sauce and adapting it to their own network topology. The last step is a full custom engagement, where all the IP that they've been using in those standard products for all those years is brought to market in a fully custom piece of silicon designed for that customer. We are not selling custom silicon to customers who are not also buying other parts of our platform. We are not an ASIC vendor. Give us a call if you want a custom chip.
We are not a back-end services company with layout engineers hanging around. We are not a manufacturing company. Call us if you need some wafers. We are a connectivity leader building deep partnerships with the largest hyperscalers in the world. In the end, this is a full platform engagement, merchant all the way through custom. Actually, I took a look at that fiscal 2031 forecast that Matt was talking about, and I broke it down by product and by customer, and we are talking about hundreds of products across this hyperscale base, just the top 10. If I look into the top four, more than 35 sockets, millions of dollars each to deliver that forecast. This is an incredibly diverse hyperscale data center engagement. At Marvell, we are really good at building these strategic engagements.
We have actually publicly disclosed strategic agreements with three of the five largest companies in the world by market cap. Each of these agreements is totally unique and tailored to the individual situation, built on top of years of collaboration. Our first announcement was with AWS in 2024. It was a five-year agreement covering custom AI products and a broad swath of networking and connectivity products, including switching, data center interconnect, and optics. Last year, we added our co-packaged optics technology to the agreement after our Celestial AI acquisition. Earlier this year, we announced a new strategic agreement with Nvidia. This again, was a very broad and very strategic agreement covering optics and photonics, scale-up networking with NVLink Fusion and AI RAM.
Most recently, the announcement was with Google, and this outlined a total of $120 billion opportunity over six and a half years for custom products attached to the TPU ecosystem. That is in addition and separate from all the connectivity business that we do together today. All these partnerships are multi-year, multi-product, multi-generational in nature, and they are all on track, in execution, and delivering value for both parties. These agreements are a natural evolution of our partnership. Usually, when we have already been working together for years at or nearing $1 billion a year in business together with a rapidly expanding design win funnel. That is typically when these sorts of discussions start. How do we move towards a win-win partnership, aligning interests on both sides? We are very strategic with how we approach these agreements.
Now let's talk about how strategically we approached scaling our engineering systems and organization in the company. We have been preparing for this growth for a long time. We've been consistently investing in R&D ahead of our revenue growth. Just take an example. Look at FY 2024, where Matt mentioned that we had done $2.2 billion in data center revenue. We spent $1.5 billion in R&D that same year, the vast majority of which was into the data center end market, which means we were investing well north of 50% of our revenue back into R&D to drive Marvell's growth. It's worked. If you take last quarter, fiscal Q2 of 2027, and you annualize it, we quadrupled our revenue in data center to $8.7 billion. So we are consistently investing heavily in our R&D to drive Marvell's growth.
We're making these investments in strategic R&D centers around the world. We're not just hiring from our competitors. These are very unique skill sets, very hard to hire people. So we're doing it right. We're building a long-term talent pipeline at Marvell. We're partnering with the local universities in each of those countries where we have these R&D centers set up. We're partnering with specific professors, and we're funding their research in things like optics and analog mixed signal design. We're also helping to influence the curriculum to make sure that the talent comes out with the right skills that Marvell needs. We're funding scholarships. More than 200 students are funded by Marvell scholarships every single year across more than 50 universities around the world. That brings us to our internship program, which we started a decade ago, and we've grown it every year.
It's now more than 500 students annually, the vast majority of which convert to full-time Marvell employees over time. Put it all together, Marvell has become a premier destination for talent, a company people actively seek out and want to come and work for. Recruiting talent is only half the story. When they come to Marvell, we arm them with the best toolset in the semiconductor industry. We have embraced and adopted AI tools throughout our design flows. We've deployed an ever-increasing pool of agents for common tasks, automating design, simulate, fix loops, and we've built homegrown tools such as the Marvell Design Assistant. This is a proprietary, homegrown tool that we built in-house with our own data and our own decades of experience designing chips. We launched it over two years ago, and it has grown dramatically since. It now becomes the foundation of Marvell's AI native workflow.
The results are dramatic. We now see engineers with five years of experience delivering the results that it used to take an engineer with 15 years experience on a traditional tool set. It's not just making engineers more productive sooner that's happening. We're actually accelerating and shrinking the overall design cycle. We're shrinking the time from concept to shipping silicon, allowing us to iterate more quickly on designs, find issues early, and adapt quickly. This is really important in the age of ever-shrinking hyperscale design cycles. None of this would be possible if we just kept doing it the way we always have, and that includes moving to the cloud. One of the advantages of working with the largest cloud operators in the world is we see how differentiated and performant their technology stack is.
So we went all in early in announcing our first deal with AWS in 2023 to move our EDA into the cloud. We invested ahead of the curve when it wasn't the obvious choice. In fact, my IT team told me at the time, "Chris, this is going to be more expensive if we do it this way." But I knew where Marvell was going wasn't going to scale without the cloud. So we worked closely with our entire tool set, all of our EDA vendors, and our cloud to optimize the complete flow and move it all into the cloud. Today, more than 80% of our total compute infrastructure is in the cloud. So you might be wondering, how does this matter, and does this actually help? Let me give you an example. Last year, we won a design win that we didn't have server provision for.
Within 60 days, we had 30,000 new cores and over a petabyte of data provisioned and engineers productive and working in the cloud. 90 days later, that was up to 70,000 cores. I can't even imagine how this would have been possible if I had people running around looking for servers and data center space to deploy it today. Completely impossible. By the way, that same chip, we were able to tape out in less than nine months. It's in the fab today, and it'll be delivered to our customers next year. Now, let's talk about our supply chain, a critical part of Marvell's scaling journey. Everything I said about focusing on fewer customers plays out 10 times in the supply chain. It's all about trust. These companies are buying buildings, hiring engineers, and buying equipment for product that are going to be delivered five years from now.
So they need to trust you. They need to know you're going to be there. They need to know your demand is real. So our strategy is to focus on a small number of critical suppliers who are the best in the world at what they do. These relationships often go back to the founders of these companies. When I took over operations in early 2021, Matt and I went on a sales campaign. We could see where all this was going, so we built relationships deep into the supply chain, not just foundry, but substrates, packaging, you name it. We needed a complete set of relationships all the way through. Then we told them our story. Where do all these chips fit? Why are they critical? Why should they bet on infrastructure at a time that they were seeing unprecedented demand from stay-at-home driven consumer end markets?
We showed them how diverse our platform is, how when anyone wins, Marvell wins. Because these suppliers are very sophisticated. They're very experienced in this market. They've been burned many times, and they're really wary of the one-ship moonshot customer. So these trusted relationships are also built on a track record of doing what we said we were going to do year in and year out. So we made the decision again to give our suppliers a five-year forecast. Five years. Like who does that? But that's what you have to do today. You can't just show up and say, "How many wafers can I get next year?" You needed to tell them five years ago how many wafers you were going to need. So in order to do that, we invested in a world-class, data-driven business intelligence platform. I have every piece of data at my fingertips today.
Every customer's forecast. What are the products? How's it evolved over the last eight quarters? What are all the pieces in the supply chain? How much capacity have we reserved? This isn't just an MRP feat. At Marvell, we have architectural insights across the stack. We know what's going on. So we build this forecast, I review it with Matt and the board every year, and we give these suppliers an early look. We've been doing it now for five years. So they can see how we've performed over that period. That fiscal 2031 forecast of $70 billion to $90 billion that Matt put up, that might have been a surprise to some of you, not to our suppliers.
Matt and I have visited these CEOs, shown them our forecasts, asked them if they could get us there, shook hands, and we are now hand in hand to deliver those results. The very last step is a capital investment for capacity reservation. This isn't where you start. You can't just show up with a checkbook and say, "How much capacity can I get?" They'll just tell you to go away. You needed to have built the trust, shown them that you have a track record. Then, yes, we are investing ahead of our growth to reserve the capacity. This is to cement the relationship and put skin in the game for our forecast. This is how you build a supply chain that can grow at the rates we have been growing and will continue to grow at.
As you can see, at Marvell, we are not short-term thinkers. We are building for the long term. We have deep, trusted relationships with the largest customers in the world. We've built account pipeline and tools to feed our growth and keep the company scaling. We have a supply chain that we've built based on trust with the critical suppliers who are the best in the world at what they do. My team and I are ready to deliver that $70 billion to $90 billion that Matt laid out. We've grown the company 4x in the last six years. We got all the demand and the supply I need to hit the $20 billion next year. We can certainly 4x from that with the wind at our back. We don't need any new home run design wins or any new sockets. We have what we need today.
We need to execute, deliver them to market, and put them into production. I'm here to tell you that my team and I are going to deliver that forecast. Thank you.
Operator
Please welcome, President, Data Center Group, Sandeep Bharathi.
Sandeep Bharathi
President of Data Center Group, Marvell Technology
Good morning, everyone. I am very excited and delighted to be here today. Every dollar that gets spent in the next five years by the AI hyperscale data centers and our customers will run into the same four constraints or walls, bandwidth, latency, power or energy efficiency, and data center reach or the distance data have to travel. Before I get into that, let me tell you, for the next 15 minutes, why Marvell is best positioned to solve those barriers and those challenges. But I wanted to give you my origin story. Eight years ago, Matt invited me to dinner, exactly eight years ago, and he laid out his vision for how he is going to transform Marvell to be a data infrastructure company. This was even before AI.
He told me and threw me a gauntlet at how I have an opportunity to transform the product strategy, technology, execution, and build a team to take it to the next level. I signed up to the challenge, and that is when my journey at Marvell began in February of 2019. In 2020, if we go back to the Investor Day, I presented a three-pronged pillar based on process, architecture, and design to transform Marvell to be the leading silicon provider for data infrastructure. Let us go through this. On the process side, we had to pivot because Marvell was historically a process fast follower. We were on technologies one generations or two generations behind, but implementing products such that it was already energy efficient, but a fast follower.
It had architecture to do Wi-Fi and storage products, about 200 square millimeters, but we had to pivot the team and the technologies to do reticle size shifts. Then came design and IP. Analog mixed signal IP that needed to be for the data center infrastructure market had to be high speed, high bandwidth SerDes. We already had 28G and 56G, but that had to transition to higher data rates, more like 100G, 200G at advanced nodes. Let us see where we ended in 2026, and I will go from the top. Process technology. This was a very important challenge. Matt wanted me to get a team to the next node. We were on 12 nm and 16 nm, and this was a company betting moment where we skipped 7 nm and went to 5 nm. This is very difficult to do.
Not only did we do this successfully in the years 2020 and 2021, we continued with leading-edge process and first to node with all the IP shuttles with our foundry partners at 3 nm, at 2 nm, and we are already on 1.4 nm or 14 angstroms. Not only that, we have transitioned from less than 50% A0 to production. What does A0 to production mean? A0 to production means you need to get it right the first time. Measure twice, cut once. But we measure many times and the masks get cut once, so we deliver time to market to the customer. Next is from an architecture standpoint, we wanted to deliver reticle-size dies, the ones that Will and Dave were going to talk to you about. 200 square millimeters and the reticle size die is about 856 to 858.
This is defined by the lithography limits and process technology that we can do today. I am here to say we have industry's first monolithic 100T switch. Nobody has that. Dave will talk all about it. We have about greater than 10 products at 600 square millimeters and beyond, shipping in volume and production. Will is going to talk about his XPU and XPU attach, where we have accomplished that. So we have completely transitioned from not only on process technology, but also architecture. When you take a look at these large dies, why is it hard? It is hard because let us say you have 856 square millimeters.
You may get maybe 30 or 40 dies, and you need to get it first time right because you need to have the methodology and the team to execute flawlessly to ensure that they account for different densities, signal integrity, and many other laws of physics in the silicon to optimize these architectures and get it first time right. Now let us talk about design. We were best in class IP portfolio, but fast follower, but we had to invest significantly in our technology to get leading-edge IP at 28G, 56G, 100G, 200G, and 400G SerDes. Just in ECOC and OFC, we have demonstrated in silicon. This is just not theory. It is actually done in multiple process technologies, and we have industry's best-in-class SerDes at all these data rates. So we have delivered and proven at every step, and this is foundational for Marvell.
Now, let us talk about what is AI tokenomics. In the building of the data centers, you get into the same barriers that I talked about: bandwidth, latency, energy efficiency or power, and distance or reach. What everybody is optimizing for total cost of ownership is really about delivering the maximum tokens per dollar per megawatt, but you also want to have the performance or the interactivity of the user. Whenever a query is served, you need responses immediately on the tokens, whether it is on research on LLM or even other applications that people are interacting with LLMs. So what happens here is AI is no longer a silicon problem. It is a system architecture and a data center connectivity challenge that Matt talked about.
Now, in order to increase performance, we need to take a look at all the constraints, and let me talk about the walls that you will run into, starting with the first one, the memory wall. The best way to think about it is an example of the human brain. The plasticity of the human brain gives immediate access to memory with the compute. It is so intertwined. Let's say for short-term memory, you will remember what you had for breakfast today, or from long-term memory, you will know what you studied in your universities. So in an ideal world, you want infinite memory, uniform latency, and a very high bandwidth interaction. Now, how is that achieved in the real world? In the real world, the memory wall is because of the hierarchy of memory that exists in the compute architectures and system architectures.
XPU has a distance chasm to get to the HBM memory, or when you have a pool of accelerators or XPUs, they need to access not only HBM, but pools of DRAM. So memory wall is a real bottleneck for performance for AI infrastructure scaling. Is Marvell positioned to take advantage of that? As memory wall as a shared resource, we have multiple technologies that Will is going to talk about in his section. We have controllers that can spread and reduce the latency for memory access. Compute express link technologies in our Structera family for memory expansion and pooling. Again, bringing memory as close to the compute as possible. Photonics interconnects for memory disaggregation, and Dave will talk to you all about those applications.
Not only that, inferencing accelerators and the performance needs very high density, highly capable SRAMs very close to the silicon, and Will will go into further details about that. So re-architecting systems using all these technologies is what Marvell is excellent at. Bringing memory closer to scale up fabric from all the history that we have had with these products. Now, let's take a look at the next wall. It is the reach or the copper wall. Matt went into considerable detail on this, but what I am here to tell you is the copper wall is extending all the way from clusters of AI compute to coming to rack and tray. The reason is, today, if you want to transfer or transmission a large amount of data movement, the shoreline of silicon and the tray is physically bound.
In order to get maximum amount of data transmission with the highest bandwidth, you need to get optics into the system, into the scale up rack, because today, scale up represents greater than 85% of the data center traffic. So optics, scale up optics is here and now. So solving the scale up bottleneck is the biggest gain we can do. Dave is going to talk about the scale of optics and across scale out and scale across, Xi and Russ are going to go into further more detail. Now, let's take a look at how Marvell is positioned to bring scale up optics into the system. We are integrating silicon photonics onto the silicon, onto the substrate to enable the scale up optics revolution.
If you just take a look at the entire gamut of technologies we have at our disposal, we were already leaders in different modulation schemes. Optical modulation schemes are very important for getting the reach bandwidth distance and the energy efficiency that every application needs. It is very different. It is not a one-size-fits-all. You have Mach-Zehnder modulation for distances greater than 20 km+ . Not only that, for within the silicon, if you want different modulation for highest energy efficiency, you have microring modulators, different modulation scheme. Then Dave is going to talk about the Photonic Fabric and the electroabsorption modulation, different modulation scheme. Scale and interconnect have newer micro emitter technologies like micro LED and micro-VCSELs. Xi is going to talk all about it. What is this actually telling you? What it is telling you is data movement challenges for those constraints require multiple technologies.
This is the core bet Marvell is making, and we have the industry's largest optical talent. Chris went into how we have enabled talent and building these teams, and this is just a decade in the making. We have the most brilliant optical engineers making all these technologies happen at scale, proven in silicon. Now, let us take a look at the common technology foundation that we are enabling across all these products. I have already talked about how photonics is a key piece of our Marvell competence. Not only that, you have advanced node and process, analog mixed signal, advanced packaging to solve every reach. Because ultimately, to solve those constraints, it is not just about compute, it is about connectivity. Because if the connectivity is not solved, then intelligence or access to compute means it is stranded compute, unused intelligence, not usable intelligence.
Let us take a look at what are we doing on the process technology. It is going beyond CMOS. Obviously, we talk about nanometers, but nanometers itself doesn't solve the problem. It is the table stakes to play in the market. However, we also have advances in silicon germanium, indium phosphide to implement and produce highly capable TIAs and drivers. Xi is going to talk all about how this is a key core part of his portfolio. Silicon photonics. This is where scale across is being enabled, including distances from 2 km to 1,000 km. Russ is going to talk all about it. Advanced material science research. Thin film lithium niobate. We work with advanced foundries across the board to bring technologies of the future, and this is here and now. Five years later, we are taking a look at what is going to be for high bandwidth, energy-efficient applications.
It is plasmonics. That is also being invested in today so that we can have products three to five years out. This is about technology enablement even before products are released three to five years out. We are fully integrated, process-agnostic, foundry-ready, and future-proof. Now, analog mixed signal is the heart of the data movement challenge. Why? Because high bandwidth and low latency requires SerDes. And SerDes is a very critical technology. If you take a look at what we have today, we have demonstrated and shipped SerDes high speed, whether it is optical or electrical, at 100G, 200G, and 400G has already been enabled in products which will be shipping in volume next year. That is 8x growth in data rate in six years. And SerDes is just not the only technology that is important for delivering on our products.
For XPU and XPU attach, because they are more than eight or 10 dies inside the package, you need highly capable die-to-die interfaces, and you might have heard picojoules per bit as a currency for high energy efficiency. We have advanced die-to-die interfaces to enable the communication and the data movement between these dies. Bringing it all together at the architecture level, you use these SerDes so that there is important constraints to be understood at the system level or the tray level or the package level when we have to connect XPU attached devices to the XPU. That requires advanced system-level simulation and technologies that we already have built and enabled and shipped in volume. Marvell SerDes leadership and advanced mixed signal is in here now proven in volume for every reach.
Let's take a look at the most important technology that's also in not only compute, but also high density switches. It's our packaging innovation. I talked about many of this in our June 2025 event. In the past, you had 2D packages, single dies, very simple to do. Then it went to 2.5D, wherein you have multiple dies on chip-on-wafer-on-substrate or any other silicon interposer material. But now where we are seeing this move to is integrating die-on-die, meaning bringing advanced memory on a logical die or multiple logical dies on a package, greater than 8, possibly 10 to 12. All the innovations that are important in packaging is critical because these substrates are greater than 100 millimeter on the side, going to 150 millimeters on the side. Some of the samples Dave is going to show you.
Not only that, because of the copper walls, because optics scaling comes into the package, we also innovate on co-packaged optics. So it's full integration and co-design to scale and volume. These are the capabilities that you need in every single product that we ship. I have talked about a multitude of technologies. It's easy to say, "Well, we have CMOS." It's easy for a company to say, "Well, we are excellent at silicon germanium," or, "We are excellent at silicon photonics." What I'm here to tell you is it requires just not one technology. It requires advances in packaging, all the capabilities in foundry, analog mixed signal, all under one roof. That's how we are well-positioned to capitalize on the opportunities that Matt talked about. This is foundational, and this is technology at its very best and bleeding edge.
Industry-leading core technology across all these domains means it's process and packaging, IP and architecture, and system-level solutions. You will hear the details from all my colleagues, Will, Dave, Russ, and Xi. It's proven execution and deployed at scale. To put it all together, let me start by saying tokenomics is just physics. I know when it's just physics. Physics doesn't negotiate with strategy. The company that solves the physics problem of bandwidth, latency, reach, and power is the one that is best positioned to go after the opportunity that Matt outlined today. I am here to tell you that the team that we have built at Marvell is exceptional at it. Marvell is trusted and undisputed leading solution provider. Are you all fired up?
Speaker 5
Yes.
Sandeep Bharathi
President of Data Center Group, Marvell Technology
Okay, thank you.
Operator
Ladies and gentlemen, at this time, we would like to pause for a 15-minute intermission. Kindly return to your seats within 15 minutes. Thank you. Our program will resume shortly. Please be seated as soon as possible. Once again, our program will resume shortly. Please be seated as soon as possible. Our program will resume shortly. Please be seated as soon as possible. Once again, our program will resume shortly. Please be seated as soon as possible. Please welcome Senior Vice President and General Manager, Connectivity Business, Xi Wang.
Xi Wang
SVP and General Manager of Connectivity Business, Marvell Technology
Good morning, everyone. Welcome from the break. Matt said Marvell is a connectivity-first company. I am thrilled to be here to talk about connectivity first. Just like Matt and Chris, this year marks my 10-year anniversary at Marvell. Of those 10 years, I spent five years at Inphi before the acquisition. Today, I want to show you what we actually mean when we say Marvell is a connectivity-first company, and why that position is becoming more critical as AI infrastructure scales. Let us start with what connectivity-first looks like in practice. Marvell has the broadest end-to-end connectivity portfolio. We organize our portfolio into four categories: scale across, scale out, scale up, and scale in. Together, they cover every distance that AI signal has to travel. That is everything from 1,000 km between data centers down to 10 centimeters inside the package. At each of these distances, Marvell has a solution.
Coherent module and DSPs for the longest reach. PAM, TIAs, drivers, and AECs for data center. ACC, NPO, and CPO for the rack. Today, I am also going to share more about what we are doing with optical IO chiplets using next-generation light source as we bring optics inside the package. No other companies cover this entire stack. Our competitors may try a piece of it, but Marvell offers all of it. That is the breadth, exactly, will let us support our customer's architecture with whatever implementation they go with next. In my presentation, I will focus on scale out as well as scale in. Russ will talk about scale across, and Dave will talk about scale up. Before we talk about the product and the business, I want to take a moment to show you the complete connectivity technology portfolio we have built over the past 30 years.
As a connectivity-first company, we have placed dedicated focus into being the first to solve tomorrow's AI infrastructure's challenges. We built the best franchise in high-speed analog, with expertise spanning across all compound semiconductors. We developed our analog IP that is tailored to the specific requirement interfacing optics. We built our DSP portfolio with advanced CMOS from the very beginning. We developed deep expertise in DSP algorithms, supporting both PAM and coherent modulations, and also with a variety of FEC options. We also developed photonic in-house, as well as forging partnerships with optical component vendors in the ecosystem. We work with a wide range of optics from silicon photonics to micro-LEDs, micro-VCSELs, and plasmonics. The result is a technology portfolio that spans every single aspect of connectivity. Not by chance, not because of one product. It is 30 years in the making.
I am very excited to be in this market. Personally, I earned my PhD in EE and Optics with a focus on plasmonics. I also spent years on high-speed analog with IoT, LEDs, and VCSELs. I developed data comm solutions before I joined Inphi Marvell. It's exciting to be able to bring all this technology together to support the market we are in today. There is one aspect of this enabling technology portfolio that does not get enough attention, the TIA and the driver. If you take a look inside the optical module, high-speed analog sits right at the boundary between the optical domain and the electrical domain. The job of the TIA and the driver is to convert electrical signal and optical signal and back again. The performance is everything here. Noise, linearity, and power. If you get it wrong, the whole link suffers.
You also have to do it at high speed, generation after generation, using circuit innovation, because as you know, analog does not scale with process. This is not a single product category. Each generation, we support a wide range of optics, which have different requirements. The market is also going to expand to support emerging form factors like NPOs and the CPOs. This is a component category where Marvell has built quietly with industry-leading performance, and it is foundational to everything else we do in optical interconnect. Marvell has a long history here. On the product side, we have led the industry since the 40G generation. That is where Inphi got its start. Today, we are at 1.6T generation, shipping in high volume, leading our competitor by two years and counting.
On the technology side, we have invested across multiple process nodes, across a wide range of compound semiconductor, developed together with our foundry partners. Because different optics, again, require different requirements on the bandwidth, output swing, and noise characteristics. That multi-foundry, multi-process leadership, combined with our capacity investment, is what keeps us ahead. All of that technology, developed fully in-house, support our flagship DSP portfolio. We have been leading since the inception of PAM4 with a fully co-designed DSP driver TIA solution. We have been leading at 200G, 400G, 800G, and 1.6T. 800G is the workhorse of AI today. It is what is powering all the AI models that you are using right now. There is a strong, continued demand from the market on our 800G solutions. This is what help us to drive this business to multi-billion dollar annually. We are leading at 1.6T.
First to sample, first to 3 nm, first to 3 nm in volume shipments. Marvell is the powerhouse here. We are in the middle of the fastest run of 800G and 1.6T. Interconnect is critical to AI, and the way we engage with the customer has evolved right along with it. 10 years ago, it is very simple. You got a spec from standard body, and you build to it. However, early on, we realized that model was not sustainable for the market, where it is heading. Marvell established a multi-generational co-design model, working directly with the cloud customer, system integrators, and optical component vendors in the ecosystem to build it together years ahead of deployment. The result is a solution that is aligned to the customer's architectural needs from day one and ready to run when the moment is needed.
This is even more important in today's AI era, where GPU and XPU refreshes occur every 12 to 18 months. Fast-forward to today, and our approach to 1.6T reflects just how much AI workload has diversified. Training, inference, networking, security. Rather than one solution, we are executing a multiple optimized 1.6T product in parallel, each purposely built. Ara being the foundation, Ara T optimized for power, for introducing LRO optics, Ara X for link robustness, Ara M for security with built-in MACsec, and Petra reflects our network development. Hyperscaler customer need this multiple optimized solution so they can have the optionality and the flexibility to deploy their network and compute. Our market has outgrown a one-size-fits-all approach. Long gone are the days that one product for each generation is sufficient to support this market. We expanded our portfolio to support it.
We have the parallel engineering resources to deliver all of them to our customers' timeline. We invest, and we scale. Looking ahead, the jump to 3.2T brings a new set of demands. Even as hyperscaler are deploying 1.6T, we are already working on the next step. Double the bandwidth at much larger scale, more robust, more power efficient, and still in the pluggable transceiver form factor our customer know how to deploy. We started working on this requirement more than two years ago. Marvell was first to show a live demonstration of a 400G per lane PAM4 in 3 nm at OFC 2025. By working with our customer, we learned something important. 3 nm alone was not the optimal answer, and PAM4 modulation has its limit. It is very simple physics. PAM4 has propelled the industry for more than a decade.
At 400G per lane, the reach maxes out around 500 m to 1 km. Most of skill out interconnect today runs 300 m to 500 m, but larger deployment require reach up to 2 km, and that calls for a more advanced modulation scheme. Our answer is a coherent-lite, where Marvell is first. It bridges PAM4 optics with traditional long-range coherent technologies, with a balanced approach to power and cost, giving customer the best of both worlds without forcing a trade-off. We do not just talk about it, we show it. Three weeks ago at ECOC, European Conference of Optical Communication, Marvell demonstrated industry's first 2 nm technology live. Low power, 2 nm, but 3.2T solution with 400G per lane PAM. 3.2T coherent-lite for data center and campus reach with balanced power and cost. We started this journey years ago.
As a matter of fact, we are in volume production and shipment of our first generation 1.6T coherent-lite with open optics for campus application. We are bringing coherent-lite into data center, and our 2 nm approach reflects our long-term understanding of the technology and the market, and the uniqueness of Marvell to support both PAM4 and coherent-lite. This is the same platform leadership story expanded. We build our DSP, TIA, and driver technology with our own leadership IP. It is what carried us from 200G through 1.6T. Where we are first to ship 3 nm in volume, at 3.2T, we are leading with 2 nm coherent-lite and a 2 nm 400G PAM. The roadmap continues on to 6.4T. We have the technology depth and breadth, and we develop it years ahead of when the market need it.
Scale out is a large revenue driver for Marvell interconnect business today, and it continues to be strong in the future. Now, let me talk about what we are doing to bring optics into the package, which is what we call scale in. Why is this important? As bandwidth continue to scale, interface between chip to chip and chip to memory will become the bottleneck. At the same time, the die size are increasingly getting bigger. So not only the speed is getting faster, it also need to transmit longer, in some cases, meters. Traditional interface shown on the left-hand side is built upon electrical SerDes. Electrical back-to-back SerDes has been a workhorse, but it does not scale to what is coming next. This could meet the bandwidth requirement by increasing the lane speed over time. However, the distance has its limitations.
Together with our customers, we identify a solution to it, designed to replace traditional SerDes for all future generations. We take next generation light source technology, micro-VCSELs, and micro-VCSELs, and integrate them directly into the package on the silicon with optical SerDes. This breakthrough is not the light source or the laser itself. It is the integration, the system-level design, and advanced packaging that make it all work, all based on high volume component that is ready to scale for an on-chip interconnect. This level of integration takes extraordinary engineering. We are already working with customers who are looking to bring this solution into their system. Over time, this could be yet another multi-billion dollar business for Marvell. So let me bring this together. Marvell has built a strong position across three things: technology, deep customer engagement, and scale. These are not three separate threads. They are one connected advantages.
In scale out, this is our most established business and existing leadership position, with new designs coming, continue to drive growth. Our DSP revenue has already gone into multi-billion annually, and our analog business is rapidly headed towards billion-dollar plus annually. We see strong growth continuing in scale out demand with data center, the data rate double generation to generation, accompanied by an increased content from Marvell. In addition, we are seeing new opportunities opening up for scale up with our TIA and drivers. There is a significant revenue ramping next year. As a result, I expect the DSP and analog business that I manage to remain the largest contributor to an overall interconnect revenue, which in aggregate, we expect to reach approximately $37 billion in FY 2031 at the midpoint of our forecast.
Across all three, the investment we are making today in technology, customer co-design, and manufacturing scale are what let us keep us leading at every hop, every distance, and every generation. Thank you.
Operator
Please welcome Senior Vice President and General Manager, Data Center Interconnect Business, Russ Esmacher.
Russ Esmacher
SVP and General Manager of Data Center Interconnect Business, Marvell Technology
Good morning, everyone. I am Russ Esmacher, and I lead the data center interconnect business at Marvell. It is a connectivity first-led company. Simply put, my business interconnects data centers with the highest bandwidth and the highest speeds over optical fiber. There has been a lot of 10th anniversaries from Marvell today, from Chris, Matt, Xi. This is also my 10th anniversary, my 10th month anniversary here at Marvell. But fear not, I see a lot of familiar faces out there. I have been in this industry for almost 30 years. Started with my graduate work in electro-optics and high-speed optics, and then progressing to some of the largest IP and optical OEMs in the market, including Cisco, Nokia, and most recently, Infinera. I could not be more excited today to be at Marvell and have the privilege to lead the most experienced, coherent, pluggable team in the market.
Let us hop right in. Marvell has the connections for AI infrastructure at all distances. We have heard Xi describe our advance and our leadership in scale-out and in scale-in, and Dave will cover next scale-up. I am here to talk to you today about scale-across. This market is very dynamic and data center interconnect is changing, so let us jump into scale-across. The DCI market has a lot of really positive tailwinds today. First, power efficiency. We heard Sandeep mention this in his talk. There has been a transition over the last 10 years from boxes to pluggable to interconnect data centers. This is nothing new to Marvell. Partnering closely with our hyperscaler customers, we pioneered and invented this market a decade ago. Data center densification. We are all very familiar. Data centers are trying to find more power, more location. This is densifying the network full of data centers.
What does this mean? More fibers, more connectivity, and it is actually shrinking the distance between the data centers, providing more opportunities for the application of coherent pluggables. Another very strong tailwind. Third, line system density. Millions of kilometers of fibers are going in annually into the network. These need to be lit by optical line systems. There has been tremendous innovation here in the last two or three years that are allowing this type of high density to accept all the bandwidth generated from coherent pluggables. Another third strong driver for coherent pluggables in the market. As strong as these three tailwinds are, there is a fourth one that has come on very strong in the last two years, and this is scale-across. This is redefining the data center interconnect market.
This is the direct connection of AI clusters to each other from distances from maybe 10 km on a campus all the way up to 1,000 km. However, the bandwidth here is not the same as traditional DCI just connecting data centers. It is 15x the bandwidth. We are going to jump into this and the impact this is having on the pace of innovation required to deliver such a massive amount of bandwidth in such a short period of time. Again, very strong tailwinds into this market. As I mentioned earlier, Marvell led the pluggable transition a decade ago, moving from boxes to smaller boxes, moving into pluggables. This move from boxes to pluggables provides a 75% TCO reduction for hyperscalers, and each generation of pluggable provides a 50% power reduction per generation. As we can see here, this is significant power savings and power efficiency.
But we can also look, the time frames are increasing. The pace of innovation is pulling the DSPs forward, and this requires an extremely strong technical and deep customer partnership to deliver these. We used to have four years between generations. It is now down to two. We used to be able to have one DSP for each generation, now it is multiple. Let us jump a little bit further into scale-across in the AI clusters. Again, 15x the bandwidth. What this does to the market is very profound. Not only is it more bandwidth, it drives the unit capacity of pluggables way up. Let me give you an example. This year, the total addressable market in units for coherent pluggables is right under 1 million. Versus Xi's business, this is a relatively small unit, but for coherent optics, this is a very large number.
By calendar 2030, our FY 2031, the demand for coherent pluggables between data centers grows to over 3 million units. This is an extremely steep ramp. However, this is perfectly aligned to Marvell's strength as a semiconductor powerhouse and an at-scale optics provider. We are ready to meet this steep ramp and continue to lead and innovate for our customers. This is also pulling forward the pace of innovation with regards to speed. It feels like we just got to 1.6T because we just did. But now 3.2T is required. We must continue to invest and keep the pace of innovation rolling. Marvell is also unique in another regard. We have two distinct paths to market. Number one, our coherent module business, and we have been in this for a decade.
This allows us to partner very closely with hyperscalers to deliver these modules based on performance and power and speed and security to them. However, there is a very broad market, and we are the leading merchant coherent DSP provider in the market. This allows us to address all of the accelerating demand, whether it's from hyperscalers, neoscalers, or carriers, between these two very distinct and unique paths. Now, let's look inside a little bit deeper into the coherent module. I will tell you, these modules, they look innocent enough. They're very small. Fit in your pocket. I will tell you, these run at the highest bandwidth, highest speeds across every technology Marvell provides today. All the advancements that Sandeep mentioned, that Xi mentioned are inside of these modules.
Marvell is a powerhouse here because everything you see in blue is designed from the ground up by Marvell. Now let's talk a little about that laser that's in there. We co-design and co-develop this laser with the leading laser manufacturers in the world, who can provide us high quality, high reliability, and highest volume. Now we're also very fortunate, and I feel privileged to lead the largest coherent team in the market. We have the largest and most experienced DSP team in the world. They've been working on this since 2006, so they're in their 20th year now. Second, leveraging Marvell's powerhouse advanced SerDes that Sandeep walked us through. Third, we have the best, fastest, and deepest converters that convert the analog to digital and digital analog back into the DSP.
Last, but certainly not least, the most experienced and deepest photonics team to do the high package integration of the optics to the electronics. They've been working on this also for more than 20 years. Now putting all these technologies together is crucial. I will tell you, one of the most difficult pieces with this high speed, high bandwidth optics is the high volume manufacturability, and this is where Marvell excels. I've been in lots of different optics companies, and I can say with high confidence, this is a superpower for Marvell that allows us to deliver units at volume to meet that steep ramp. Today, Marvell has more than 1 million coherent modules in service over 15 billion field hours. We've been in this business since we started it. We continue to innovate and lead today.
For me, I look at all these core competencies, deep technical, this was a very compelling reason I feel very privileged to lead this team and have joined Marvell. Now, multi-generational scale across requires leadership from a long time. We got into this business at 100G ZR about 10 years ago. We then innovated with our customers and brought 400G ZR to market. Then we have 800G ZR. We partnered with our customers, and they said, "We would like you to please put encryption and higher performance in 800G. We need this now." So we built another DSP for them that had these new features inside of it. As you can see, as we go to 1.6T, multiple DSPs are required, whether it's for campus, metro, long haul. These are optimized for performance, power, and distance. 3.2T is under development as we speak.
This requires not only more DSPs, but much like 1.6T, new materials. As Sandeep mentioned, 1.6T required thin film lithium niobate. 30 years later, I am excited to work on lithium niobate again. Plasmonics is here. This requires higher bandwidth, lower power, shorter modulators, and we are making the investments in this technology to ensure that it is ready when needed for 3.2T and will also be first with 14 Angstrom DSPs. Marvell is the leader at 2 nm technology, and we did not do one, two new DSPs, new modules. We demonstrated these publicly for the first time two weeks ago at the European Conference of Optical Communications in Spain. I will say in our industry, this is rare to be able to deliver two DSPs at the same time.
I would contend it has been the first time it has been done in the industry that on leading node in the market, two simultaneous coherent DSPs were built, brought up, demonstrated with our technology and our partners as a leading merchant at the same time. This is a pace of innovation that our customers are counting on Marvell to continue to deliver, and I am very, very proud of the team who executed this flawlessly with the first tape out on the first shuttle in January. But like Xi said, demonstrating it is what is important. Having modules, having little samples, it is great. Showing it to our customers on time, first time right is what they count on us for. This is the live demonstrations in our booth and in our partners' booth publicly at the European conference.
Again, thank you to the team who was able to deliver this first for our customers. Now bringing it all together. In summary, I could not be more excited about the tailwinds in this business and the team we have assembled over decades to continue to deliver for our customers. Again, we have very strong tailwind. We see our TAM expanding rapidly, both in units and revenue. In addition, we are shipping to the top five hyperscalers today. We are deeply embedded with designs with them and making sure we deliver the right technology at the right time with the right features. As a result, we are on track to drive revenue from business from roughly $500 million in fiscal 2026 to safely over $1 billion in fiscal 2028 with a clear path to multiple billions in the subsequent financial years in the strategic period, FY 2029 through 2031.
I would like to thank our customers, our partners, and the Marvell team for continuing to deliver innovation at scale and on time for our customers. Thank you.
Operator
Please welcome Executive Vice President and General Manager, Data Center Networking Business, Dave Lazovsky.
Dave Lazovsky
EVP and General Manager of Data Center Networking Business, Marvell Technology
Good morning. I am Dave Lazovsky, the EVP and GM of the Data Center Networking group. I joined Marvell in February of this year with the acquisition of Celestial AI. I was formerly the founder and CEO of Celestial AI. This is my first Marvell Investor Day, and I could not be happier to be here. I am excited about the opportunity here before us. I am thrilled by the magnitude of the opportunity we have got both in scale-up switching and in optics. I feel that Marvell is layering strength on strength with our inorganic growth strategy. As you heard from Matt, we have invested as a company over $40 billion over the course of the last 10 years in innovative technologies and in leading teams. My business group is the beneficiary in large part for a lot of those investments.
I am excited about the position that we are in and could not be happier to be a part of Marvell right now. Let us dive into the data center networking business group. The group is comprised of three different business units, cloud optics that you just heard from Russ, that includes pluggable data center transceivers and DCI, and then switching and integrated optics. Switching includes both scale-up and scale-out, and integrated optics is NPO and CPO. I am going to focus my talk on these two areas of integrated optics and switching. Let us start with switching. As Matt mentioned, we have an established and growing business in Ethernet switches based on our Teralynx platform. We expect that business driven in large part by the ramp right now in 51T switches to exceed north of $1 billion and significantly so, next fiscal year.
In addition to that, what we are most excited about is a new flagship product that we have recently released, which is based on much of the technology that Sandeep introduced in his talk earlier today. This is our Teralynx T100 product, and it introduces multiple leading-edge disruptive technology capabilities, including it is the first and only monolithic full reticle limited die that is operating at a full 100 Tb per second of bandwidth. How we have achieved this is, as Sandy pointed out, Marvell has the world's leading 200G and 400G SerDes. We have 512 200G SerDes in a single reticle limited die in this device. Why does that matter? Well, it matters because if it is a uniform monolithic device, that means that the latency of the device is significantly lower than a multi-die, multi-chip module package. Our latency is leading edge.
We're on the order of 20% or more lower latency and, in fact, lower power than any leading competitor in this space. Why that matters is because we can use this device uniquely to serve both scale out applications for conventional Ethernet as well as scale up this emerging market that is taking off right now. We're thrilled about the opportunity for the 100T switch, and we expect the 100T switch as a product in and of itself to layer on incremental multiple billions of dollars of annual revenue in the coming years. So a huge opportunity from this switch alone.
But right behind that, we have a series of additional Ethernet-based switch solutions that include the 200T switch, includes the 400T switch, and we have ESUN, Ethernet scale up network solutions that include an E115, 115T Ethernet optimized for scale up, similarly 230 and 460 in the pipeline. So a broad suite of Ethernet-based switches and ESUN-based solutions for scale up and scale out. In addition to that, Marvell is developing for our customers UALink-based switches. We're building now and will be releasing next year a UALink U115 switch that is being deployed at one of the industry's leading hyperscalers. Right behind that, we have the U230 and a U460 switch. The industry's broadest portfolio of scale up and scale out switch technologies we have within Marvell. Excited about this opportunity, a growth engine for the company going forward.
The opportunity extends also to not just UALink-based protocols and Ethernet or ESUN-based protocols, but as Matt mentioned, we have partnered earlier this year with NVIDIA, and we now offer NVLink-based protocol solutions for our customers. So we have the ability to craft solutions for all of our customers for scale up and scale out to meet their requirements on an application-by-application basis and on a customer-by-customer basis. So let's discuss. What I'll do is I'll go into much more detail on scale up. You heard from Xi and you heard from Russ, the scale across, scale out, and scale in markets. I'm going to dive into scale up in detail, not just talking about the products and technologies, but about the application drivers at the artificial intelligence workload level that are driving requirements for the next wave of data center infrastructure, specifically scale up.
We feel that scale up is going to drive a tremendous opportunity for the company. It's an inflection point, and it's an inflection point in particular with this transition from copper to optical interconnectivity that Marvell is so well-positioned to lead. So let's begin with the workloads. The workloads that we've been addressing as an industry over the course of the last several years began with the ability to train multi-trillion parameter models. This is hard, right? So it puts a lot of pressure on all of the resources, on compute, on memory, a lot of network intensity. But over the course of the last couple of years, two years ago, specifically, there was a new wave of inference-based models, which are reasoning or inference time compute.
And what that did is it introduced a set of requirements for higher capacity bandwidth, higher capacity, high bandwidth memory, and it introduced the need for scale-up networks to increase the amount of memory semantic interconnectivity within the scale-up domain, XPU to XPU connectivity. Those workloads were driven by OpenAI o3 and o1, for example, and DeepSeek-R1 that was released a couple of years ago. Then, more recently, you see this push toward mixture of experts and longer context-based workloads. This was the result of a very important innovation that took place in June of 2023, about three years ago, at Berkeley. At UC Berkeley, a group introduced an open-source inference engine called vLLM, and one of the important innovations inside of vLLM was KV cache. And KV cache has now proliferated rapidly through the industry, and it's enabling much more efficient utilization of memory.
I'm going to spend some time in this talk talking about the impact that KV cache is having on the data center infrastructure that we're building with all of our hyperscale customers. But all of these innovations have come together and a shift at the same time has taken place, a shift from a training-centric environment for data center infrastructure to an increasingly inference-centric infrastructure. Today, about 60% of total data center infrastructure is inference. And as you all know, inference is the return on that ROI equation. Training is the investment. We're seeing not just an increase in inference, we are seeing an increase in efficient inference. So what that has done, in turn, is drive a significant increase in the operating profits of the cloud service providers. Three years ago, there were some cloud service providers that had operating profits in mid-single digits.
Today, almost all of them are seeing operating margins that are approaching 40%. The math is working. The flywheel is spinning. Those profits are being reinvested in data center infrastructure. The $1.6 trillion of CapEx that you just heard from Chris that's being invested this year, going to $3 trillion as we exit the decade. It's the largest infrastructure investment in the history of humanity, and the fact is it's justified because these are profitable workloads. So that investment is taking place across compute, across memory, across networking, and across connectivity. The market is coming to us. We're extraordinarily well-positioned for where we sit right now at Marvell. So let's dive into specifically what's happening within reasoning models and why we care about what this group of researchers at UC Berkeley did three years ago.
But we don't really care so much about that as much as we do about the utilization, the efficient utilization of memory. What we care about is how the model is building the context token by token. So again, this group introduced the concept of KV cache. So what is KV cache? KV cache is the storing of key and value matrices rather than requiring the need to go recompute that stuff. So you have the information available in high-speed memory so that it makes the interaction in reasoning models rapid, right? So you don't have to wait for recompute. Humans are inherently impatient, right? So the problem, though, is that the amount of KV cache is not unlimited. So it puts a lot of intensity on memory capacity, on memory bandwidth, and it puts a lot of intensity on the network latency.
But at some point, you exceed the amount of available KV cache in the system, and at that point, one of two things happens. You either need to reload from a different location in the memory hierarchy, those key and value parameters, or you have to go back and recompute them. And in both cases, you have to make room in the KV cache for the new key-value indices, and you have to eject or evict information to make room for it, which can potentially result in hallucinations. So in all cases, this reloading and recomputing results in inefficiencies. It drives higher memory latency, it drives a need for memory and compute transactions, overall, a negative impact on the efficiency of these systems. So ideally, we have systems that have, ideally, infinite amount of KV cache, high-speed memory that's available that can be addressed by any XPU.
The problem is it is not infinite. The KV cache is finite. So how we're addressing this issue is at the system level. We're addressing it first with maximizing the amount of high-speed SRAM on the chip. We then increase the amount of high bandwidth memory in the package. So if you look, for example, at state-of-the-art GPUs today, Rubin from NVIDIA, Rubin has roughly 300 GB of memory capacity in the package. The problem is that that 300 GB is orders of magnitude too small to deal with today's advanced AI models. So that has driven NVIDIA and the rest of the hyperscalers to build systems that interconnect multiple XPUs together in a memory semantic scale-up domain. In the case of NVIDIA, they use NVLink and NVSwitch. Right now, there's 72 interconnected XPUs inside of that rack. The problem is that the memory requirements are insatiable.
They're continuing to grow, and they're spilling over what will fit in a rack. The industry needs multi-rack scale-up domains growing to 512, 576, and beyond. So how do you get there? And what are the challenges, right? So some of the challenges are, today, 100% of all interconnected XPUs inside of all racks are copper, all of it. That is both the challenge and the opportunity for Marvell. So why is that? So why not just take that copper and extend it to a multi-rack domain? Well, the problem is that the data rates they're running at, which are 200, 224 Gb per second, they're going, by the way, to 448 Gb per second. But at 200 Gb per second, the reach that you've got is 2.5 m. It forces you inside the rack, right?
The only way to overcome this limitation within the scale-up domain is optics, rather than fighting the laws of physics, knowing that moving information over copper electronically, you're dealing with electrons, and unfortunately, electrons have mass. The beautiful thing about optics is that they have no resting mass, right? So it gives us the ability to leverage physics, use it to our advantage instead of trying to fight the laws of physics. It gives us the ability to increase fully connected scale-up domains with optics. Larger numbers of directly interconnected XPUs sharing the same memory with extremely high bandwidth and with extremely low latency. But the bar is high. It's very different than a scale-out network.
If you think about it, the XPU on the left here has to be able to operate, do a load store transaction, and address the XPU on the right, the memory in that XPU, and pull it across the network. Ideally, it is operating as if it's the same performance in terms of bandwidth and latency as the memory on the package. That is challenging, right? Because HBM has very high bandwidth. It's tens of terabytes per second. Right now, HBM4 has up to 16 TB per second of bandwidth, and you do not want the network to be the bottleneck for a memory transaction. You also want to be able to minimize latency. Again, the latency for these applications and reasoning translates directly to revenue for our customers. The latency inside of a package, less than 100 ns.
Latency going over a scale-out network, if you had to use a scale-out network for this, would be hundreds of times greater than the local memory latency. So what you really need is the latency down around 200 ns. Really, really fast memory transactions, also with low power. Sandeep pointed out earlier that the amount of data traffic going across scale-up networks is 85% of the total data center data traffic, 85% of it. Only 15% is scale-out. What that means is that if you can reduce power here, even a few picojoules per bit, you can repurpose that energy that's being used to move information and use it to compute information where tokens can be generated. Tokens aren't generated moving information, right? So the options that we've got for addressing this set of challenges are, unfortunately, not pluggables.
As you just heard from Xi, the leading-edge pluggables are going from 1.6 to 3.2. It's an order of magnitude lower in terms of bandwidth. The power's not there. The latency's not there. They're not designed for these applications. This application requires integrated optics. It requires NPO and CPO. The good thing about this circumstance, though, for Marvell, is that the market's just come to us because we have a full suite, as Xi pointed out, of optical technologies that we have built over the course of the last 25 years with the best team on the planet building these technologies. We bucket them in three categories. They're the category of working with module providers in the way that we have with pluggables, with supporting and providing these module providers in the overall optics ecosystem with drivers and TIAs, which we're doing today.
Then we have two categories of fully integrated, what we call light engines. One of those categories we call SPATs, which are 200 gigabit per second going to 400 gigabit per second per lane in optics. The other we call Flat Pipe. Flat Pipe optics are designed specifically for scale-up networking. Much lower latency, much lower power. We have two flavors of Flat Pipe. One is based on industry standards that are emerging now, OCI MSA, big set of acronyms that stands for Optical Compute Interconnect Multi-Supplier Agreement. That's basically the industry converging around ESUN-based optical scale-up networks, and they've landed there. So it's a set of optics that the industry has standardized on or is working on standardizing on. We have that as an option.
The other is the technology we developed at Celestial AI, the Photonic Fabric, which is fully optimized, again, ideally for scale-up networks where we've got strong customer demand. We put checks in the win column as it relates to customer demand across all three of these categories. We're winning in delivering drivers and TIAs to module providers today. We have won a major socket recently, and we're winning multiple sockets in Fast Pipe optics, and we have won sockets, including with the world's leading hyperscalers, in Flat Pipe as well. Let's double-click on this and take a look at the level of complexity of these devices. What's in here? In the case of the Fast Pipe light engines, here's our 200 and 400G. You see we're able to leverage things like our optical DSPs and the range of optical DSPs that we've got that are fully retimed or transmit retimed optics.
If you blow open the light engine, you can see a lot of the innovation that Marvell has, again, has developed over the course of the last 25 years in advanced drivers and TIAs, including silicon germanium. We have, again, tapping into the know-how that we've got in silicon photonics with Mach-Zehnder interferometers and all of the optics and packaging and scale that you heard from Russ and Xi, leveraging all of that in an NPO form factor. So a large opportunity, Fast Pipe optics, excited about that opportunity. On the Flat Pipe side, very similar, other than the fact that this technology has been designed from the ground up for this application, scale up. What we've done is the light engine is far more integrated. It's a deeper level of integration in the optics.
All of the discrete components are now fully integrated in an advanced ASIC. This is the drivers, the TIAs, the controllers, and most importantly, analog SerDes. These systems are designed to eliminate the need for digital signal processors. We're using analog equalization, and by eliminating the DSP, you take the largest power contributor and the largest area contributor out of the equation. Game-changing as it relates to performance, game-changing as it relates to latency, the ultimate optimal solution. That Flat Pipe is applicable for both. I'll jump back because it's important. It's applicable for both MRM-based silicon photonics as well as EAM or electroabsorption modulator-based photonics sharing the same control electronics, the same advanced ASICs. This is where we've brought the teams together from all that horsepower or power that's come from Marvell over the last 25 years with our team from Celestial AI.
Now, form factor. Some customers are interested in NPO-based form factor, some are interested in CPO. From our standpoint, the answer is yes. We're agnostic. If a customer's operational priorities are field replaceability of the optics or worry about blast radius of a failure or if they want reconfigurability, use an NPO. Plug it into the board, and if there's a concern, you swap it out, not a problem. For those customers, on the other hand, that are looking for the ultimate in performance, the lowest latency, the lowest power, we're offering today and we have customers deploying co-packaged optics. This is the deepest level of integration, which provides the most superior performance for scale-up network solutions. We have developed both, and we are in the process of deploying both.
Marvell is in a position where we don't have to force our customers to make a choice. We can embrace them, and in fact, that's how we do it. If you think about how all of this technology comes together and how we work with customers, our customers are operating with us as an extension of their own teams. We're engaging with them to understand what their requirements are for architecting and developing the right scale-up network across the entire link, from the XPU interface, in some cases going from copper to the right optical I/O, determining what is the right physical layer, what optics do they want to run, what is the right logical layer, what protocol do they want to run to the switch, all of it. And these things are highly integrated and highly interdependent.
Most of these programs that we have with the largest hyperscalers in the world have started years ago. We're working with them to develop solutions for scale-up networking that takes time to get developed and implemented and deployed. And it enables us to design and optimize for leap performance, including power latency, bandwidth. It enables us to accelerate time to market and time to first token generation with them, and it also helps them de-risk the operation of these devices by leveraging things like telemetry, so they can see the stability and the resiliency of their links and employ things like security to ensure that in the world of agentic AI, their links are fully secure. These are examples of how we're working together using the leading-edge technology that Marvell has brought together to address these rapidly evolving architectures in scale-up networking.
Now, what I'm about to show you is how we are adding new capabilities and tools to the toolbox to help our customers disrupt this, meaning help them craft system architectures that can provide next-level capabilities for dealing with things like KV cache sizes, the ability to decouple memory scaling, high speed, high capacity, high bandwidth memory scaling from compute. What you're looking at right here is what we call the Photonic Fabric memory module. These modules, 8 to 16 of them, are integrated in a Photonic Fabric memory appliance. And what it does is it provides, the first time really in history, the ability to decouple memory scaling from compute. The only sockets that are currently available for HBM today reside 1.5 mm away from an XPU or a GPU.
What that means is that our customers are forced to use their GPUs as the world's most expensive memory controller. This technology changes that. It changes fundamentally the economics of memory, and we're excited about where we stand with this. This is a product that we taped out earlier this year. It's being brought up now, and we'll be sampling with multiple customers as we move into next year. Very excited about this opportunity. In closing, I'd like you to take away three primary things from what I covered. The first is that we have an established scale-out switching business. All of our business in switching today, all of it, is scale out. Now, that business is growing. It's doubling year-over-year.
This year it will more than double again year-over-year next to well over a billion-dollar annual run rate, and that will grow to a multi-billion dollar annual run rate in scale-out switching alone. In addition to that, we are leading and enabling this new emerging market segment of scale-up switching, and there we have significant customer traction. We expect scale up to layer in an additional multi-billion dollar annual recurring revenue opportunity for Marvell. In addition to that, we have scale-up optics. What is interesting about scale-up optics is the fact that there is significant Marvell content. There is significant content for a few reasons. One, because the bandwidths are so high, the other because you saw the level of content in these highly integrated devices like the Fast Pipe and the Flat Pipe light engines, a lot of content.
If you think about it in terms of the dollar content for scale-up optics per XPU, you should think about this in the context of on the order of thousands of dollars of scale-up optics per XPU. It is large valuable technology, but it is bigger than that. The optics are required on both sides of the link, and our customers in large part are looking for comprehensive end-to-end solutions. The optics are required on both the XPU side as well as the switch side of these links. We are excited about the magnitude of what the scale-up optics market segment brings to Marvell and the fact that this market has just come to us. With that, we are in the right place at the right time with the right technology, and we have got the right trusted team to execute. With that, I would like to say thank you.
Operator
Please welcome Executive Vice President and General Manager, Custom Cloud Solutions Business, Will Chu.
Will Chu
EVP and General Manager of Custom Cloud Solutions Business, Marvell Technology
All right. Good morning. My name is Will Chu. I have been with Marvell for nine years, and I lead our custom business. At last year's Custom AI Investor event, I talked about how Marvell was becoming a leader in custom solutions. I thought my team was pretty busy then. It turns out we were just getting started. Since then, our traction and opportunity have grown dramatically. 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. Today, I am super excited to show you what is driving that growth. Okay. Let me start with how we built the custom business from the ground up. Very early on, we saw that hyperscalers were going to need significantly more custom solutions.
We recognized that Marvell already had the core IP to address this market and world-class engineering teams with 30 years of experience in storage and custom ASICs, 25 years in networking, compute, and security. We also had decades of experience taking the core IP from our merchant products to create semi-custom and full custom solutions. My mission has been to integrate these teams with our core IPs and build a custom solutions leader. We are focused on XPUs and a diverse set of XPU attach solutions and on customers that leverages Marvell's entire technology platform. Our strategy is working. At last year's Custom AI Investor event, I talked about the 18 sockets we had already won. Since then, we've continued with many more designs and with multiple multi-generational wins across the top four hyperscalers, along with early wins with many others.
That continued momentum is what's driving our strong revenue growth. Let's start with XPU. Hyperscalers were initially interested in custom XPUs, so that's where we started. Our initial XPUs were already complex devices, one to two reticle size dies with integrated HBM using 2.5D packaging technology. Today, our XPU business is on track. We've already shipped in high volume, and we're executing very well across all our programs. Looking forward, XPU complexity will continue to increase as hyperscalers push for more performance. You see more compute dies, more HBM, and standalone copper I/O chiplets all integrated using 3.5D packaging. By the n+2 generation, we expect even more computes and HBMs, and that these XPUs will approach the copper wall and need to integrate optical I/O. This is why at our custom AI investor event last year, we highlighted the critical technologies you see here.
Our deep investment in connectivity IP is another big differentiator and positions us extremely well as XPU architectures continue to scale. Hyperscalers will need a partner like Marvell that already has a track record of taking complex XPUs into high volume production and leading NPO and CPO technology. As Dave highlighted, we are focused on customers that leverage our complete end-to-end portfolio of XPUs with our I/O chiplets, optics, and networking. That combination of connectivity leadership and XPU experience sets Marvell up to be a big winner as AI architectures scale. Our close partnership with Nvidia further expands that platform. Nvidia and Marvell are jointly engaged with customers on NVLink Fusion solutions, combining Marvell's custom XPUs and Nvidia's scale-up connectivity and rack scale infrastructure. Our opportunity extends beyond the XPU, and there are a number of exciting XPU attach opportunities I'd like to discuss next.
The XPU opportunity really evolved from the traditional data center market. The market had a well-established ecosystem of merchant solutions using standard interfaces like PCIe to support x86 CPUs. Marvell participated in that market for many years. We developed networking, processing, security, storage, and memory products and shipped millions of units to the hyperscalers. As the industry transitioned to AI, hyperscalers initially leveraged those merchant solutions. As XPUs became more complex, they needed increasingly optimized solutions around the XPU. We saw this trend early and proactively engaged our hyperscaler customers to address these new requirements. Marvell was a natural partner. We had the underlying IP, product expertise, customer credibility, and the experience of turning our merchant IP into semi-custom solutions. What makes XPU attach very attractive for Marvell is that these sockets are very sticky. They can persist across multiple generations of XPU.
They heavily leverage Marvell's IP and expertise, and they extend our reach across the entire infrastructure for AI, even with non-Marvell XPUs. Let me walk you through a few of them, starting with NICs. In a traditional data center, a vast majority of the market were merchant NICs. Foundational NICs are the most prevalent. There are smart NICs with integrated CPUs, and IO Cs are typically 100 gig or less. Marvell had shipped NICs for decades and millions of parts into production. Marvell invented the smart NIC back in 2005 and has been in production for over 20 years now. There are a few merchant NICs on the market today, but all hyperscalers have or are moving to custom because standard NICs lack either the optimization or the performance they need. Hyperscalers are looking for a specific mix of I/O speed, bandwidth, and configuration.
They need optimized data paths, accelerators, as well as custom networking protocols and integrated CPUs dimensions just right for their applications. All of these requirements tie directly to how efficiently data moves across our network to support their specific workloads. Marvell is a trusted partner for hyperscalers to solve these challenges. We have leading SerDes IP at 100 gig, 200 gig, 400 gig. Decades of experience with NICs, high-performance data path compute, and acceleration. We also have the capability to design very complex custom solutions. For example, our latest NIC is greater than 400 millimeters square to give you a sense. We are doing really well here. We have design wins with multiple hyperscalers across multiple generations that are ramping or in execution. Now, let's take a look at memory.
In traditional data centers, workloads were compute-bound, not memory-bound, and having memory directly attached to the CPU was generally sufficient CXL was originally developed for traditional data center applications, not AI. We at Marvell had early investments in CXL, but this market developed slowly. AI changed everything. Workloads like inference and KV cache are memory hungry and latency sensitive, driving higher memory capacity and bandwidth for both CPUs and XPUs. This surging memory demand, of course, has created a memory shortage. Very early on, we saw the memory wall challenge and worked closely with a leading hyperscaler to define a differentiated solution, leveraging our early investments in CXL. Today, Marvell is a leader in memory expansion and near memory compute with the broadest portfolio of high-performance solutions in 5 nm. Our early design wins are ramping at multiple hyperscalers and with multiple multi-generational programs in execution.
Our Structera X products support DDR5 and DDR4 for memory recycling, industry-leading memory capacity and compression to extend the capacity further. Our Structera A products bring compute closer to memory to accelerate memory-intensive workloads like LLMs and offload the host CPU. We also have a broad ecosystem across memory, CPU, and GPU providers. This is a memory add-in card developed with a leading memory vendor, and the opportunity is expanding beyond CPUs. XPUs need more memory capacity and bandwidth as well. We're now developing custom memory expanders that connect directly to the XPU at very high speeds using our high-speed SerDes IP, giving the XPUs direct attach to expanded memory. This creates a very exciting new growth vector for our memory expansion business. With that, I'll move to storage. Historically, the storage controller market had many merchant solutions, often bundled by the flash vendors with their NAND.
There are also merchant accelerators that offload storage and data management functions from the CPU. Again, Marvell brings decades of storage experience with millions of Bravera SSD controllers and accelerators shipped across multiple generations. Looking ahead, hyperscalers want to control their storage firmware and software so they can leverage their investment across multiple generations of SSDs and multiple NAND suppliers without being tied to a specific drive vendor. AI workloads also rely on flash now to continuously feed data to their XPUs, making storage offload performance increasingly important. This plays directly to our storage heritage. We leverage our merchant technology to build a leading portfolio of semi-custom 5-nm SSD controllers and storage accelerators. For hyperscalers developing do-it-yourself SSDs, our multi-vendor controller delivers high performance and custom features while giving them full control of their software.
For storage accelerators, we optimized our merchant technology for an initial hyperscaler and have already secured design wins beyond that initial engagement. Our accelerators increase storage throughput and provide confidential compute and virtualization offload, freeing up the CPU. I also see storage requirements continuing to increase, and they are roughly doubling generation over generation. This is creating a strong tailwind for our custom storage business. I would like to discuss a couple emerging XPU attach opportunities that we have never talked about before. The first is a new class of AI infrastructure management silicon. We have talked before about NICs, memory, and storage. Today, we are introducing an entirely new XPU attach category driven by a fundamental change in AI infrastructure. In traditional data centers, BMCs monitor the health of a relatively simple server, typically one to two CPUs.
These servers have approximately 150 sensors generating relatively small amounts of data, so the processing and IO requirements are pretty modest. The impact of a failing server are also relatively contained, typically one to two CPUs. AI changes this dramatically. AI servers are some of the most complex servers ever created, with more than 1,500 sensors and generating hundreds of megabytes of data per day. Just managing this complexity requires 1,000X more performance, and the impact of failure is much, much greater. A single XPU failure can disrupt the work of thousands of XPUs, and traditional BMCs weren't designed for this. AI infrastructure needs an entirely new level of performance, processing, and IO performance. At Marvell, we recognized this challenge early and proactively engaged with multiple hyperscalers to develop a new infrastructure management solution.
They needed a much more powerful multi-core processor to monitor their servers and switches across their AI infrastructure. We already had deep relationships with these customers and a long history of delivering infrastructure class processors. So we partnered closely with them and developed an optimized solution leveraging our Octeon technology, proven 5-nm IP, and experience in high reliable mission-critical applications. The Marvell semi-custom processor delivers the 1,000X increase in performance needed for these AI infrastructure management workloads. Today, we have programs with multiple hyperscalers creating an attractive XPU attach growth opportunity ahead of us. Now let's move on to another emerging XPU attach opportunity. The next opportunity is inference acceleration, and this is an exciting emerging category that could significantly expand the XPU TAM. Inference, as you all know, is growing rapidly, and architectures are evolving to deliver significantly more performance.
One of those changes is the use of specialized inference accelerators to attach to the XPU to offload specific workloads. These are highly complex, large devices that require the most advanced process nodes. This represents another significant silicon opportunity around the XPU, and Marvell was already investing in the technologies needed to address it. That includes our high-speed SerDes and die-to-die connectivity with a path to optics, our SRAM technology, which can deliver 17 times more bandwidth at two-thirds lower standby power than standard solutions. Advanced 3.5D packaging technology, integrating 3D stacked memory and our logic die to enable even more performance. Together, these technologies allow us to optimize inference accelerators for performance, bandwidth, and power. We believe inference accelerators can become a significant new XPU attach market, and we're already engaged with multiple customers. Marvell has the core technologies to win here. Okay, so let's summarize.
We're engaged with all the major hyperscalers. Since last year's custom AI investor event, we've deepened our customer relationships and have won additional multi-generational designs. All of our XPU programs are on track, and XPU attach has become a big multiplier to our opportunity. Standard products are moving to high performance and more complex semi-custom solutions for NICs, memory, and storage. New attach categories are emerging, including AI infrastructure management and inference acceleration. As AI performance scales and XPU complexity grows, we see higher attach rates and higher ASPs, significantly expanding our opportunity. You can imagine one to two sockets attaching to each XPU with ASPs around $1,000. You guys are going to write all that down. Do the math and you can see how each of these attach categories can represent a billion opportunity, with some reaching several billions.
As these programs ramp, we're building a large and increasingly balanced custom business across both XPU and XPU attach. Okay, I'm very pleased with how quickly our custom business is growing and diversifying. As XPU attach ramps, it will drive greater balance between XPU and XPU attach. We are on track for our custom revenue to more than triple from fiscal 2028 to more than $12 billion in fiscal 2029, well above our prior fiscal 2029 target of more than $10 billion. Looking further ahead, I am personally very confident in our ability to grow custom to approximately $30 billion in fiscal 2031 at the midpoint of our long-term target range. My confidence is driven by the momentum I discussed today across all the tier one hyperscalers in both XPU and a diverse set of XPU attach programs.
As you heard from Matt, we don't need new wins to achieve our target. We are in full control of our destiny. We have the customers, the team, the technology platform, and the programs in place to deliver on the tremendous custom opportunity ahead of us. Like I said at the beginning, we're going to be pretty busy. We're just getting started. Thank you.
Operator
Please welcome Executive Vice President and Chief Financial Officer, Dan Durn.
Dan Durn
EVP and CFO, Marvell Technology
Hi, everyone. Good morning. Thank you all for coming today. I am Dan Durn, Chief Financial Officer, and I have been at Marvell a little more than three months. From a career perspective, I got my start in the military. I am absolutely honored to have worn the uniform of my country for a dozen years. Subsequent to my time in the military, I spent a little more than two and a half decades in and around the semiconductor industry. First as an investment banker with Goldman Sachs in their merger department covering semiconductor companies. Then in private equity investing in semiconductor companies. Now in the last 15 years in operating roles as a CFO leading finance organizations. Before diving into the financial model, I want to share how excited I am about the opportunity we have in front of us. Marvell, we are at a unique inflection.
We believe we are on the cusp of increasing the magnitude of our revenue by a factor of 10, and that is going from $8 billion to $80 billion of revenue at the midpoint of our long-term model and doing it in a very short period of time. It is all predicated on a vision. One that was laid out 10 years ago by Matt and the team, and it has not changed. A decade later, it is the exact same vision. So what has changed? We have taken that vision, that North Star for the company, and a point of view of where our end markets are going, and we have backed it up. We have backed it up with conviction and a decade of investment. We have invested tens of billions of dollars to bring this vision to life, and now we are purpose-built. We are purpose-built for this moment.
We are intersecting the largest wave of infrastructure investment in human history, and we are doing it with critically enabling IP and foundational technology. It is IP and technology that directly addresses, strikes right at the heart of the single biggest rate limiter in data center performance today. That is connectivity. How do you efficiently move, store, process, and secure data? How do you do it at an absolutely massive scale and power efficiently? We, Marvell, we are built for this moment, and it has been a decade in the making. As CFO, I take comfort in the underlying drivers of growth. You heard it from Matt and the team today. The drivers of growth, they are diverse. They are diverse across products, customers, technologies, and multiple generations of design wins. We have got many shots on goal, and no mega socket is needed to deliver on our long-term targets.
I can't think of a better time to be at Marvell. With that, let's dive into the financials. Before looking forward, let's do a quick recap of how Marvell has performed relative to its long-term targets that we put forward a handful of years ago at our last Investor Day. I think it's important context setting and forms the foundation of our journey ahead. At the time, in 2021, we laid out three key priorities and Marvell has delivered on each. Revenue. We delivered a 23% compound annual growth rate from FY 2021 to FY 2026, exceeding the high end of our target long-term revenue growth of 15%-20%. On earnings, we delivered strong operating leverage. Our non-GAAP operating income growing at a 32% compound annual growth rate, nicely outgrowing revenue over this time period. Finally, capital allocation. We've invested heavily in growth.
$8 billion of organic R&D investment, and we've complemented it with $11 billion of inorganic investments. All while returning about $4 billion to shareholders via dividends and share repurchase. To recap, we've invested in growth. We've returned capital to shareholders, all while maintaining a strong and flexible balance sheet. With that behind us, I know you're here to focus on the next five years. Let's jump into that. What have you heard today? First, we see a huge revenue opportunity in the next five years, and it's largely coming from our focus on connectivity and compute for our hyperscaler customers. We've got a massive $400 billion TAM in calendar year 2030, and that's a greater than 4X increase from the last TAM we provided.
It's growing at a 45% compound annual growth rate, even when factoring in a moderation in data center CapEx, which we have growing at a 35% CAGR over this same time period. So we window back out. Data center CapEx, 35% CAGR. Our opportunity, our TAM, 45% CAGR. Our revenue growth, 55%-60% CAGR. Second, we purpose-built a vast array of connectivity and custom technologies to address the massive TAM that we see. That includes our leadership positions in DSP and optics, our analog mixed signal capabilities, our silicon and packaging technologies. These capabilities are foundational to our success. They're truly differentiated, and they're truly world-class. Third, combination of this massive TAM, our key differentiated technologies, that positions us to accelerate our revenue growth in the years ahead. Now let's put some numbers around this.
Going forward, we remain focused on the same three metrics: driving strong revenue growth, delivering earnings leverage, growing cash flow. I'll dive deeper into each of these in the subsequent slides, but I just want to share a few quick highlights. Revenue, we expect the revenue CAGR to be roughly 55%-60%, which is significantly faster than the growth of data center CapEx and our opportunity, our TAM. On earnings, we expect to grow earnings meaningfully faster than revenue and deliver operating leverage and do it at scale. Finally, we expect to generate strong and growing cash flow that gives us positive optionality for capital allocation in the years ahead. Now, let's double-click into each of these individual pieces of the three focus areas we just talked about. Matt and Chris showed some of this earlier today.
They highlighted in a forecast $400 billion TAM for Marvell in calendar year 2030. As you know, that is the equivalent of our fiscal year 2031. Within that opportunity, we believe Marvell can grow our revenue from $8 billion last year to $80 billion of revenue at the midpoint of our long-term model in FY 2031. We expect the communications and other business to grow at the same low single-digit rate that we've discussed in the past. So no change. As you can see, our data center business, this is the primary driver of the growth with a 65% CAGR at the midpoint of our long-term model in the coming years. Let's dig a bit deeper into the data center.
Consistent with the data center splits that we've highlighted in the past, we see strong growth across all three pillars of our data center business in the years ahead. Interconnect, custom, switching, and storage. For interconnect, with the move to optical connectivity supercharging growth, we forecast a revenue CAGR of about 65% at the midpoint, driven by scale out, scale up, and scale across. For switching and storage, we see a revenue CAGR of about 45%, with our switching business expected to grow even faster. That's driven by scale out and scale up. Finally, custom. We see a revenue CAGR of approximately 80% at the midpoint, and that's driven by both strong acceleration of our XPU business and strong acceleration of our differentiated and sticky portfolio of XPU attach products.
Taking a step back, the key takeaway, all three pillars of our data center business have got strong growth in the years ahead. We have three large, diversified, high growth engines that are powering our data center business. Next question. How does this revenue growth translate into earnings? As we've said for many years, Marvell's gross margin, that's mix dependent. Custom silicon carrying a relatively lower gross margin, but has an attractive operating margin due to customer funded NRE and scale of the business. Merchant silicon, on the other hand, has a relatively higher gross margin, but also has a higher R&D profile. The net of this dynamic, both custom and merchant yield attractive operating margins despite the variations that they have from a gross margin perspective.
We expect to grow operating expenses at roughly half the rate of revenue growth, and that will deliver nice operating leverage as well as a sizable upside to the operating margin in the years to come. Netting this out, we remain committed to growing earnings at a rate that's nicely above our revenue growth between now and FY 2031. Based on this combination of significant revenue growth and strong operating discipline and execution, we're going to deliver a structural increase in the cash flow generative capabilities of Marvell. Our priorities for this significant increase in cash generation, we're going to remain oriented towards growth. We're going to continue to fuel the growth of company, and that's going to be both through organic investments and inorganic investments.
And we are committed to returning greater than 50% of our free cash flow on average over time, and that is all while maintaining a strong and flexible balance sheet. So let us summarize all of this in a single table. Our new long-term target model, FY 2031. We expect to deliver revenue of $70 billion-$90 billion, and that is driven by a diverse set of customers, products, multi-generational design wins, and that is across all three pillars of the data center business. We expect a gross margin of 56%-59%, with the range being mix dependent, and that is consistent with what we have shared in the past. It is also important to keep in mind this mix dependence, it acts somewhat like a hedge. The low end of the gross margin range, it likely requires a significant upside to the base case revenue forecast for our custom silicon business.
Conversely, if connectivity revenue nicely outperforms, our gross margin will likely tend towards the upper end of the 56%-59% range. So the mix dependence acts as a bit of a hedge. We expect an operating margin of 44%-46% as OpEx grows at roughly half the rate of revenue. And taking a step back, putting that in context, our new operating margin target is six points higher than the prior long-term target we put forward a handful of years ago. This significant increase in profitability is a result of the increased scale and differentiated positioning we have as a company. We expect a tax rate of 15% and a free cash flow margin of greater than 36% over time.
And I know that you guys all took a stab at what the earnings power of the company is a couple of hours ago when Matt previewed and laid some breadcrumbs. So to demystify it with these assumptions baked in, confirm that you guys got it right, which I am sure you did, here is the math. Combination of these metrics imply a non-GAAP EPS of greater than $30 a share in FY 2031 at the midpoint. So when I window back out and I think about the earnings power of this company and how to frame that, I think of 30 in 30. Greater than $30 of EPS in calendar 2030, which again, is the equivalent of our FY 2031. So greater than $30 of EPS in calendar 2030, 30 in 30. So let me bring you back to a slide from the end of Matt's presentation. Why invest in Marvell?
You heard about the technology, you heard from the team. We talked about the large market opportunity. There are five numbers I want you to walk away with from this room. $400 billion TAM calendar year 2030, massive and growing. $20 billion revenue guidance for FY 2028, $2 billion higher than our prior guidance. $12 billion-plus custom revenue target FY 2029, up from $10 billion-plus. $70 billion-$90 billion revenue target FY 2031, built bottoms up with a conservative view on data center CapEx. Last, 30 in 30. Earnings power, long-term model, greater than $30 of non-GAAP EPS in FY 2031, 30 in 30. We at Marvell, we have worked hard to earn the trust of our largest partners. Similarly, we appreciate the trust you, our investors, place in us. On behalf of the entire Marvell management team, I wanted to express appreciation and say thank you.
With that, let me welcome the rest of the team back up and let's kick off the Q&A. Thank you.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
The Q&A, but before you do that, I think Matt just wanted to kick it off with one comment.
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah. Anyway, it's great to see everybody. We'll do a Q&A, and then we'll have lunch. But before I do that, I actually want to give just a special recognition for a moment of a very important person here who has supported you guys tirelessly for really almost two decades, and that's Ashish Saran. Ashish joined me in 2018, and Marvell was not in the place that we are today. He was very brave to take this on with me. They say being a CEO is a lonely job, and it is. I think one of the few roles and people that understands that implicitly always is the head of investor relations. In my darkest times and deepest criticism and things just not going the way we wanted, because look, it hasn't been all up and to the right for the last 10 years.
Although if you look at it's pretty good, right? The 300, not bad. Along the way, Ashish, you and I had a roller coaster, and I just want to have everybody. Ashish, stand up, please. Give Ashish Saran a round of applause. He's retiring. He's retiring soon, but we're keeping him on the payroll for a little bit longer because we just can't seem to let him go. Anyway, with that, why don't we kick off, Ross, and we can moderate the Q&A.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Perfect. So raise your hand. Wait for the mic to get to you so those on the webcast can listen. I have been in your seat, so please listen to this last thing. One question. Not five part question, one question. All right, first up is Vivek.
Speaker 12
Great. Thank you so much for a great conference and this day. Matt, my question has to do with your platform opportunities. So the color that I would love to have is how many source are there? Then when you look at the forecast that you gave of $30 billion, when you said about the platform XPU attach, how does that align with the large opportunities coming with the major manufacturers where they can handle opportunities a lot higher than that number. So just how many source, how sticky these are, and then how many do you put that 13 into in context with a lot of opportunities that you're designing on that item?
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah, sure. That's a great question, Vivek. Thanks. I think first, the way to think of it is fundamentally, these programs that we're executing on are all proprietary sole source. It's Marvell IP and technology, especially when you get to the more complex products and on XPU attach. It actually starts with Marvell at the core, and in some cases, it's our entire design. It's what you call build to spec, which means we basically get a specification from the customer, we work with them, and we deliver it to them. 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.
I could see that tilting one way or the other. I think it depends on how some of these programs take off. But with respect to the XPU attach and the large strategic agreement that we announced in August, the way that we've got that modeled right now in fiscal 2031 is very pragmatic and judged. So don't think of this as the max case when you built your model on what the value of that warrant is. As usual, the way I do it is I'm not conservative, and I'm not aggressive. I'm pragmatic. I use most probable outcome, and we start from there. Then as we build over time, we adjust. We adjust our update. So that comprehends not the full value of the warrant, but some portion of it that in our view is a very appropriate judge view for this time.
So what that means is you have a very healthy set of other sockets coming in, because as Will talked about, we had 18 of these already won a year ago. We have layered in more, and they are across all the different hyperscalers, across all those different technologies. So it really will be a very diversified, rich business when you get to custom in 2030. And it is actually a great way to play custom because you are not rolling the dice on one giant socket and hoping and seeing what happens. We have developed quite a portfolio here, and I think that is going to serve us well as we grow through the next few years. Thanks for the question, Vivek. Next question.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Let me go to Tim.
Matt Murphy
Chairman and CEO, Marvell Technology
I will direct traffic on these. So ask the question, and I will farm it out.
Speaker 13
Hey, thanks. So Matt, one emerging theme is sort of powered shell and whether there is going to be enough powered shell for all these XPUs. And Jensen would say that there will be enough for his product, but there will not be enough for anybody else's. So I do not know. I know that you are being very conservative in your XPU model for fiscal 2031, but how do you think about that? When you are talking to hyperscale CEOs, is this something that is coming up?
Matt Murphy
Chairman and CEO, Marvell Technology
No, not in the context of our XPU programs or our custom business or quite frankly, anything to do with our current revenue. The way we have our CapEx, I think to set aside land, power, and shell, just take a step back. If you look at our model, which is 35% CapEx growth over our period from fiscal 2026 to fiscal 2031, that's not a linear CapEx. Right now it's growing 70% or more. So we anticipate at the tail end of this, it's probably growing more like 20%. So we think we've built in our model just a lot of cushion for the AI business to kind of digest over time. If CapEx does better, actually, we'll do better, but we don't see that as any kind of constraint or any kind of customer discussion at this time.
And remember, we're across all the fleets and all the different customers, and we sell to everybody, and we sell the systems, not only custom-based systems, but obviously our broad-based business connects to every GPU, XPU accelerator on the planet. So not an issue for us. Next question.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Let's go to Thomas.
Speaker 14
Hey, guys.
Matt Murphy
Chairman and CEO, Marvell Technology
Hello.
Speaker 14
Thanks for doing the day. Appreciate it. For the interconnect TAM, I was curious if you guys could break that down a little bit further into calendar year 2030. I think it was $38 billion. Is that more DSP share, more DSP volume, NPO, CPO? Any kind of breakdown that you could give us? Because that's obviously a really large number, and where you see the most growth.
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah. I'll give you some of the assumptions behind it. Maybe, Chris, you want to chime in after that. I think the way to think about it, just on the growth is that, and I think Xi did a great job. Where is Xi? Yeah. Making a very compelling case for why scale-out continues to perform. That's our largest business today. That's going to continue. So you kind of know what that model looks like today. I think you should assume that we at least can grow with the market. We don't intend to lose any market share there, just to be super clear. I think Russ did a great job outlining the scale across, which has already been growing, but with this sort of scale across use case.
With the data traffic increasing and the number of units going up, you should assume that we scale with that, and definitely hold our own. By the way, in both of those businesses, each of those generational changes that we go through, let's say it's from 800 gig to 1.6T to 3.2T, all of those come with ASP uplifts. So it's not just units, but it's also the dollars. I think you can make a model on that. You kind of know those numbers. Then you should just assume in interconnect that scale-up optics, which Dave talked about, that's going to layer in the rest. Obviously, that's going to grow at a much faster rate because effectively we have zero revenue there today. Multi hundred million starting next year. But that's going to ramp extremely hard. Then who knows?
There's a call option on scale in, as Xi talked about with some of this micro emitter technology. That's not really material at that point in FY 2031, but it could be. So that's kind of the bucketing I would think about today. I think we're not going to give a lot more precision than that. But Chris, is that about right?
Chris Koopmans
President and COO, Marvell Technology
Yep. That is right. You got it.
Matt Murphy
Chairman and CEO, Marvell Technology
Okay, cool. Thank you.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Next question. Why don't we go to Ben up here, please, Carmen.
Ben Reitzes
Analyst, Melius
Hey, thanks a lot. Ben Reitzes with Melius. Thanks for doing this, Matt. Hey, I was wondering in the target through FY 2031, is there anything in there for the AI labs and how are you working with them? Is that potential upside, and if so, which category would it be in? Would it be in customer attach or optics? And how are you thinking about your relationship with those folks?
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah, I think the answer is yes, and I think the opportunity set is very broad for the labs as well as the current customers we had. As Chris pointed out, I think here's the good thing. In our model today, the core driver and engine of the revenue growth through the period that's currently forecasted is highly leveraged to the big four, but very diversified. Like I said in my opening, we're talking about selling 15, 20, 30 products each to some of these big companies. Very deeply embedded in the infrastructure. That growth of CapEx is robust, but then we see the top four taking more and more of that CapEx. At the same time, I think it's a real call option, some of the things that we're doing as well with the labs. But we would be embedded in that infrastructure no matter what.
As Xi said, every model so far that's ever been created has been trained using Marvell connectivity as an example. That's going to be broadly used across all of the model makers around the world. The upside could be as they look at more custom or other types of solutions, that's an opportunity for Marvell as well.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Why don't we go to Mark over there, please?
Mark Lipacis
Analyst, Evercore ISI
Great. Mark Lipacis, Evercore ISI. Thanks so much for the great presentations. Really appreciate it. Really appreciative of you guys taking a stab at what you think about your business in 2030. I think that's something we all try to do, and it's all appreciated. The transparency is really appreciated. At the same time, it's really hard to go out that far, and I think if you look at historically when you have high growth markets like this, your customers will often make a guess at their demand and give you a forecast that is more than they want so they get what they need. I'm wondering if you could just talk a little bit about how you think about managing that risk, and what's different this cycle versus last kind of super cycles. I'd love to hear Matt Murphy and Dan Durn, you've been around for a while.
If you have a perspective on that, I would love to hear that. Thank you.
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah, great. No, thanks, Mark. I will start off, and then I will let Dan comment. I think a couple things. And yeah, you are right, we have both been around the block. I am almost 33 years in this industry and kind of right at the front line. So I have definitely seen the cycles. I have touched the stove before. I am not saying I am perfect, but I understand. A couple things give us comfort on this. The first is that while we are framing this as a five-year, and we are, right? Because last fiscal year was the first full year to start the measurement period. You go out five years, that is pretty typical. But think about it. We have line of sight right now to $20 billion in revenue next year.
That is 2027. Our forecast is out through 2030. You have to assume and we feel really good about the next year. Okay? Then you have to go 2027 to 2028 to 2029 to 2030. That is it. So I am actually not calling the ball five years from today. I am calling the ball with a retrospective view of what our business has done and how it has performed. So actually, just to go from 2027 to 2030, I think we feel really good about that. Now look, we got it. We are all exposed if the world turns upside down and something happens, and I can give those caveats like anyone would always do in the semiconductor industry. Well, what if something big happens? But at the current moment, we do not see that.
These programs are completely on track. AI investment continues to be extremely robust. It has outperformed all of our expectations. When I go back to April 24 in the bull market hotel versus now, it has continued to rip. I think what is built in from a hedge point of view for us, one, we have CapEx moderating at the end of the period. The other thing is we have given you a range. A range of outcomes, which I think should give you a sense of where we can land. But if you just take the 20, even to the low end of the range, it is not. So I know these numbers are big, and I think everyone is going to try to catch up, but as Chris said, we have had our head wrapped around what this could look like for a while.
We got the supply chain lined up. So we have a lot of conviction on what we can go do, but certainly it is market dependent and certainly it is dynamic. But I think of it as a three-year kind of a pick, not a five-year. Dan, you want to add anything?
Dan Durn
EVP and CFO, Marvell Technology
I think that's well said. I don't have anything to add to that.
Matt Murphy
Chairman and CEO, Marvell Technology
Okay.
Dan Durn
EVP and CFO, Marvell Technology
Great explanation of where we are and captures our thinking.
Matt Murphy
Chairman and CEO, Marvell Technology
Cool.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Can we go to Tore, please?
Tore Svanberg
Analyst, Stifel
Thank you. Tore Svanberg from Stifel. Thank you for all the presentations. By the way, I love the 30 by 30. Simple. Thanks, Matt. I had a question on the XPU attach or the custom business. Based on my math, I think you are only targeting about 13% share. Given that XPU attach is so much networking related and there are not many custom players out there, why would not that number be bigger? Are you just basically giving us your guidance based on the bottoms up? Are you being conservative? Just wondering why it is not bigger than 13%. Thank you.
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah. Totally get the question, and I just want to take a second, just take a step back. The fact that a real question in the room is why is $30 billion so low? That is kind of the question. Because of the market. The market has really run away from everybody, I think. We sat here in April 2024, and we called out a $40 billion XPU custom TAM. It was a custom TAM, $40 billion, and there was this huge debate, well, is custom even going to happen? Could it be 20%? Remember all these debates? Now our view is it is $235 billion. Where we think about it is, I think we got a very realistic plan. You should assume on the XPU attach side, that size of the TAM has gotten way bigger than we thought, even from a year ago.
We said it was about $15 billion a year ago. It has gone up by multiples of that in terms of where we can land. We know because of our design win hit rate, our share is going to be very strong in XPU attach. So think of it as XPU attach, very strong, multitude of design wins, and on XPU, very selective, few key flagship programs carrying us very deep, very long term. That is kind of how you get that diversified mix over time. You are right, if you just do the simple math and you say $30 billion out of $235 billion, that is what you get. You get 13% share. Okay. So that is why I think it is actually not an unreasonable target because we are. By the way, there are shots on goal all over the place.
It is just about ultimately where we think we can add value and where our customers really want us to participate. That is kind of how we are approaching it. We are going to be there for them if they want us to do more. But at the current time, you look at Will's book of business, you look at the design wins we have, you just layer it through. I think we have a very compelling path to get to the 30 with a range on it.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Go over to right here, please.
Speaker 18
Hey, guys. Thanks for the day. I want to go back to the scale-up opportunity and maybe talk about how that's evolved since the Celestial acquisition. NPO is a big topic, kind of where you're seeing design activities there. I am just curious, if you threw in micro-VCSEL, is that a scale-up opportunity, and where are we at with that?
Matt Murphy
Chairman and CEO, Marvell Technology
Okay. I will do an intro. I am going to have Dave do part and Xi do part. Here is what has happened on scale-up, and it is kind of very exciting. In 2025, we were looking at the scale-up opportunity strategically as a massive inflection. I remember we did our strategic annual review. Xi actually stood up there. He showed this chart, and it was just this crazy hockey stick up on scale-up optics, like this is happening. We have a great team inside of Marvell developing scale-up optics. Dave showed you that, by the way. That light engine roadmap that was in the middle, that NPO, that team has been working on that for years. We said is, look, we think this market is going to evolve.
We decided to go out and go big and looked at what was out there and ultimately aligned, and I will have Dave come in a minute, aligned with Dave and his team to join us and create really this powerhouse optics team and franchise to go after this market. What has happened is the NPO part of it has really gotten a lot of legs. In parallel, we had gone off and won and gotten huge position and print position in all of these NPO modules that have been developed over the last few years with our SiGe products. Then you bring in Celestial. All of a sudden, we close Celestial, we get into 2026, we start talking roadmaps combined. We start showing our scale-up switching roadmap, and it is like out of the woodwork, right? All these designs that we had, one, came out of their latent state.
The designs were already won a year ago on NPO at the module vendors. Light engine's already been developed. Those are now landing in real customer design wins today. The way it's going to happen, Blaine, I think, is that NPO's going to ramp for sure, CPO still tracking, and it's not going to be one goes and then it just falls off, and then the other one takes off. It's not going to work that way. NPO is a train that's going to run. It's paving the way with most customers who are ultimately going to CPO. But we have our own CPO engagements where Dave said some customers want to go on some of their platform to the final step. We're going to drive that in parallel. So don't think of this as some sequential thing. Okay, here's an NPO cycle, here's a CPO cycle.
Even further than that, here's a copper cycle, here's an NPO cycle. It doesn't work like that. All these are going to exist in parallel, but that's sort of how we arrived at scale up. Dave, why don't you make a comment on your perspective, and then, Xi, you can cover the micro emitter opportunity a little bit more and how we think about where that fits in the application stack.
Dave Lazovsky
EVP and General Manager of Data Center Networking Business, Marvell Technology
Yeah, maybe a few comments on the acceleration. What we are seeing is an acceleration of optics based on the drivers that I just walked through. In fact, a half node earlier, we're seeing the industry being effectively forced based on the requirements for artificial intelligence needs to move toward whatever optics they can push into production as quickly as possible. In many cases, that is using the pluggable data center transceiver and the Fast Pipe light engines that I walked through, repackaging that into an NPO form factor, leveraging either coming from us directly or coming from our module providers. In both cases, for NPO in that form factor, we win. When our module partners win, we win. And when we're providing direct Fast Pipe optics, we win. The way to think about this market, though, is a little bit confusing.
I think there's a little bit of just general confusion around the fact that for us, it doesn't matter, right? So every one of those optical technologies that I walked through, we can provide in either an NPO or a CPO form factor. If our customers want Photonic Fabric in an NPO form factor, the answer is yes. It's relatively straightforward to restructure the form factor into exactly what they want. So one thing that we've done, as I'd mentioned, as soon as we integrated the Celestial team with the Marvell optics team, is we took the Flat Pipe teams, then we merged them quickly, right? So we have a singular analog team that's working on fully integrated optics, where we can leverage the whole suite of silicon photonics foundries, and we're working with all of them, including Si and SiN.
We can work across all the different modulator technologies, as I pointed out and as Xi pointed out. So we are completely agnostic. We start with the customer's requirements, and we work backwards from there.
Matt Murphy
Chairman and CEO, Marvell Technology
Great, Dave. Thank you. Xi, just maybe some comments on Micro VCSEL, Micro LED.
Xi Wang
SVP and General Manager of Connectivity Business, Marvell Technology
On this go out acquisition from a copper to optical, this is a once in a lifetime opportunity. I have seen the sheer volume and also the application use cases we have never seen before. So we developed all this technology in the past. We focus on the traditional laser source to enable this technology, as Matt and Dave mentioned. But we also start looking at the micro emitter technology, right? Because this market is so big, we are looking for diverse competition, not just the traditional laser sources, but the brand new type of laser sources. That can give you more flexibility to help the customer to deploy this unprecedented volume. That is where we start looking at the micro emitter technology, and that can combine with our experiences, optical experiences, and everything. We can combine them as an integrated solution to provide to the market.
That is how we are thinking about it, hosting holistically.
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah. So for scale in and scale up.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Time for two more questions.
Matt Murphy
Chairman and CEO, Marvell Technology
Great question.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Harlan. Back there, please, gentleman.
Harlan Sur
Analyst, JPMorgan
Yeah, great. Thanks for hosting the event. Harlan Sur, JP Morgan. The XPU attach strategy, I think, is a very smart strategy, as you guys mentioned. It really allows you to exploit the IP, whether that's storage, whether that's memory, whether that's networking. It really allows you to sort of exploit a lot of the portfolio of IP that the Marvell team has developed. I think relative to, let's say, a full-blown custom XPU, because XPU attach is more heavily biased towards existing Marvell IP, less contribution from your customer. First question is, does that reduce the cycle time, the velocity of bringing these products to market. In other words, relative to a full-blown XPU accelerator, can you bring XPU attach to the market design cycle-wise, like sooner relative to XPU full-blown accelerator. That's the first question.
Second question, because of maybe the faster time to market and because more Marvell IP per XPU attach, is the gross margin profile relatively richer versus a full-blown, full custom XPU accelerator?
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah. Hey, Harlan. Great question. Maybe I'll just cover it quickly. I think the first kind of framework is our customers, regardless of XPU or XPU attach, we got to move super fast. That being said, on XPU attach, because it is highly leveraged and we can actually show we have products in the market that are or maybe the customers already use that product as an example. We can spin very quick time to market adjustments and roadmap adjustments and really tailor these. That's a superpower we have, actually, is to be able to take some core we have and then customize it. So that's high leverage, which is great. But on the XPU side, say, if there's incremental things that need to be done, we can do those fast.
But if you're starting from brand new process, brand new package, certainly that would be a longer cycle time. Yes, we even said this at our Investor Day last year, you should expect that the XPU attached side is more gross margin rich than the other. What we said last year is primarily that some of which is our IP and differentiation, but it's also the volume. Although I'll just finally say what's interesting is, we started out with this idea of cloud-optimized silicon in 2021, and I think everybody was like, "What is that?" Even XPU attached last summer, just to be fair, I think it was, "Hey, is this a diversion because you're just not winning enough XPUs or something?" I'm like, "I don't think you understand. This is a big market.
This is really going to be a good business for us. It turns out it's a really good business for us, to your point. It's where the market is moving, and that's going to be a sizable portion of the custom revenue. But definitely the volumes, which I think we were thinking were going to be smaller, are not small. Will talked about this. You start thinking about number of accelerators, the kind of ASPs we're talking about, the attach rate, and these are meaningful opportunities for a company our size, and they layer in nicely as we head out towards the $80 billion. I actually really like the balance in the portfolio there. Thanks.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
All right. One last question. I guess it's kind of poetic. We'll have Chris Rolland do it.
Matt Murphy
Chairman and CEO, Marvell Technology
Chris Rolland. All right.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
There we go.
Matt Murphy
Chairman and CEO, Marvell Technology
You got to do it for me.
Chris Rolland
Analyst, Susquehanna
All right. Well, first of all, thank you for the shout-out, Matt. I was right. It was a very different, unrecognizable Marvell. The stock was probably low teens at the time, too. So congrats on your success. What was not so successful was the picture of my very large bald head.
Matt Murphy
Chairman and CEO, Marvell Technology
We'll circle it for the replay of that.
Chris Rolland
Analyst, Susquehanna
My question is actually coming back from ECOC. Scale-in, everyone's talking about it, the new buzzword. Matt, you mentioned it as well. People are talking about it as multiples of scale-up even. I think maybe we're having some definitional problems here just trying to nail this down. I think it's all about remote memory access via optics. This seems like it could be a potentially massive market that you guys are really well-positioned for. Matt, you said there wasn't anything in your 31 number regarding this. Maybe you can talk about the size of this market, where you think it can be over time, what engagements look like today, what timing looks like, because we were hearing maybe calendar 2029 would be the earliest scale-in opportunities, and anything else you could say about this market.
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah, no, I think it is a great way to wrap up the Q&A because it is long-term, what happens in the future. I think, look, we could give you a whole presentation on that right now. We actually have a point of view. We are spending R&D on this right now. We have multiple engagements with multiple customers in this area. I think it is actually going to be massive, but I think today is not the day. We just loaded you guys chock-full of all kinds of awesome information about how the core business alone can take us to multiples of where we are today. But we did not want to ignore it because I actually think, like I always do for you guys, I do think about this. What are we saying now that we can come back to in a few years and then update you on, right?
But I think just for the group, I think scale-in will be another massive, meaningful leg of growth to Marvell, which could start in the period. But certainly when you think about a future Investor Day and you roll years beyond that, I think that will be a key thematic part of the Marvell connectivity story, where you are adding an entirely new, basically fourth leg. Now remember, leg three has not even started, which is scale up, but we have already got leg four kind of in the chamber, and we are investing aggressively here. This is not just a couple of people running around doing an experiment. We have full-blown teams working on this today with key partners. So I think that will be the last question. Do you want me to say a few words and wrap it, Ross? Or do you want to say anything first?
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Go ahead, Matt. I will hit them with some logistics.
Matt Murphy
Chairman and CEO, Marvell Technology
Okay, perfect. He will handle the logistics. Really appreciate the time today for everyone to come out, and everybody on the webcast, investors, Marvell employees, customers, everybody watching, thanks so much for your interest in the company. It has been an incredible journey over the last 10 years. It has been an honor of a lifetime. When I walked into this company 10 years ago, I think none of us had any clue where it was going to go. I certainly did not. I had a point of view, though. I really did, and it is very cool to be here today and be up here with such an incredible group of people. You heard it in all the presentations today, the word team. The word team. And this is a fully integrated team. I have also got the rest of my team up here too.
I've got our CTO, our head of legal, head of M&A. Don't talk to him, by the way. Head of HR, finance, you name it. We're all here. And if you're not here and you're on the Marvell team, the whole Marvell team, by the way, is 9,000-plus strong. You're on the call. Thank all of you. Really, you guys did an amazing job today. I'm really proud of all of you, and thank you for your interest in the company. We're going to have a great lunch, and we're going to tell some stories and talk about Bull Market Hotel one more time. Ross, go ahead.
Ross Seymore
Senior Vice President of Investor Relations, Marvell Technology
Perfect. Thank you, Matt. Lunch straight across, past the stairs. The whole executive team will be at tables in there. No need to come up and mob everyone quite yet. Do that on the other side. Let us have a chance to get the microphones off. So head over there, start getting into lunch, and we'll join you momentarily. Thank you, everyone.
Matt Murphy
Chairman and CEO, Marvell Technology
Yeah, great. Thanks, everybody.