Everyone loves talking about Nvidia (NVDA) when it comes to AI chips. And for good reason: it practically owns the market at this point.
But there's a significant shift happening in data centers, and it's worth paying attention to.
More Top Stocks Daily: Go behind Wall Street’s hottest headlines with Barchart’s Active Investor newsletter.
Custom application-specific integrated circuits (ASICs) are growing explosively. The market is projected to expand at a 27% compound annual rate, ballooning to $118 billion by 2033.
And two companies are leading that charge: Broadcom (AVGO) and Marvell (MRVL).
Neither one gets the same headline ink as Nvidia. But both have quietly built businesses that hyperscalers are leaning on more and more, and both stand to benefit as that shift plays out.
Here's what's driving the transition toward companies like Broadcom and Marvell, and what it means for investors weighing exposure to the next leg of the AI infrastructure trade.
Why AI Chip Demand Is Shifting Toward Custom ASICs
The demand for ASICs traces back to the same reason Nvidia saw runaway success in crypto mining and AI: CUDA.
CUDA, which used to mean Compute Unified Device Architecture but is now being used as a standalone term, is a platform that essentially unlocks a GPU’s computing power for other uses. And since GPUs are better at doing similar calculations at the same time, they’ve become perfect for highly parallel workloads like crypto mining, AI training, and machine learning.
But you know what they say: jack of all trades; master of none. And even though this is a gross simplification of a complex subject, it still applies here.
Nvidia’s GPUs are incredibly powerful and flexible, but that power and flexibility are baked into the price. Vera Rubin, the company’s next-gen AI platform, is estimated to cost around $55,000 each. Meanwhile, smaller data centers might need hundreds to thousands of AI accelerators, while larger facilities might need hundreds of thousands.
So, as a rough estimate, if we say that a mid-sized data center needs 30,000 Vera Rubin chips, that means the company running that site needs to fork over roughly $1.65 billion. And that’s just for the chips; it doesn’t include the structure, cooling, and other components of a data center.
So you can understand why hyperscalers would grab any opportunity to cut down costs.
And that’s precisely why ASICs are seeing real market demand today.
As their name suggests, ASICs are designed to work for a specific application rather than trying to handle virtually every potential AI workload. That means they strip away much of the flexibility and general-purpose functionality that makes Nvidia’s GPUs so darn expensive, and instead dedicate the chip’s power to a specific calculation or task.
Of course, you’d first need to know what the chips are supposed to do to design them around that particular task.
Think of it like calling a plumber. A general handyman brings his whole toolbox because he doesn't know what he'll find. But if you tell him upfront it's a leaky pipe under the sink, he shows up with just the wrench and fitting he needs. Faster, cheaper, and no wasted gear.
That's a GPU versus an ASIC. Nvidia's chips are the whole toolbox, ready to tackle any job. ASICs are like the streamlined plumber who already knows it's the kitchen sink. If the custom silicon manufacturer knows what the chip will be used for, they can maximize its performance for that goal.
Now, ASICs are typically cheaper than NVIDIA AI GPUs. But the real savings for hyperscalers buying in volume go beyond the purchase price.
An ASIC’s specialization almost always translates to better performance per watt, lower ongoing operational costs, and higher throughput - the amount of work an AI does overall. Why? Because it doesn’t need any extra modules or circuitry for tasks that it won’t be doing in the first place. Instead, the silicon can be dedicated purely to the specific calculation or task it’s designed to do.
Now imagine that two-pronged savings, and multiply it across however many ASICs a hyperscaler needs to run its data centers. That could easily translate to hundreds of millions, or even billions, of dollars saved every year.
With all of those positives, you’d think that hyperscalers would be clamoring to replace all their golden Nvidia chips with their own custom silicon.
But there’s actually a catch. Designing and producing custom chips can be restrictively expensive.
Why ASICs Can't Fully Replace GPUs in AI Data Centers
Unlike Nvidia, which already has development and production down to a science and is just improving each iteration of their AI platforms, ASIC development needs to start from scratch - or at least, a lot closer to square one.
Designing a custom chip requires specialized engineers, architecture work, verification, software development, testing, and eventually manufacturing. Because the chip is built around a specific workload, getting the design wrong can be incredibly expensive.
You can’t simply make a few software tweaks and turn an AI ASIC into something completely different. If you design it for one task, then that's the task it will perform.
So, yes, custom chip development is a massive capital investment that not many companies are willing - or able - to make.
But for those who can shoulder the upfront costs in exchange for potentially longer-term savings, ASICs are the best choice.
And that’s where Broadcom and Marvell come in.
However, we do need to make a distinction between the two - beyond the most obvious size disparity, of course. We need to look at their businesses around the custom silicon opportunity to get a glimpse of their potential ceiling.
Broadcom's Business: Is It True Diversification?
So, let’s start with Broadcom, because this is arguably the more diversified company of the two. And when I say diversified, I don’t just mean that it has more products. Broadcom operates two major businesses: semiconductor solutions and infrastructure software, with the latter growing significantly after its acquisition of VMware.
Now, obviously, the custom chips business belongs to the semiconductor segment, and has been the company’s major growth driver.
Today, Broadcom holds somewhere between 60% and 80% of the AI ASIC market. Not surprising, given its partnerships with Alphabet (GOOGL), Meta (META), and OpenAI.
President and CEO Hock Tan said in its Q3 FY26 financials that “Demand for our custom AI accelerators and networking continues to be very strong. Q3 AI semiconductor revenue of $16.7 billion grew 221% year-over-year, and 54% quarter-over-quarter.”
That’s $16.7 billion out of $20 billion total revenue from the Semiconductor Solutions segment. Understandably, management is quite optimistic about its silicon business. According to Tan, “Q4 consolidated revenue growth is forecasted to increase 93% year-over-year” on the back of that AI chip growth.
Now, on the face of it, it’s easy to dismiss Broadcom as just a custom chip provider, especially with these numbers. But the business is actually more than that.
As part of the Semiconductor Solutions segment, Broadcom has built a broad portfolio around the physical infrastructure that keeps data centers and communications networks running.
Screenshot courtesy of broadcom.com As you can guess, that positions Broadcom to benefit from the AI infrastructure buildout from multiple angles. It supplies the custom accelerators that hyperscalers are increasingly adopting, while also providing the networking switches, optical components, connectivity chips, and other infrastructure needed to connect those accelerators.
And that’s the layer everyone seems to forget. Sure, custom AI silicon gets the headlines, but connectivity is just as vital as the processors running these data centers.
That's the semiconductor side. VMware is the other half of the business.
Broadcom completed its acquisition of VMware in late 2023, bringing the virtualization and cloud infrastructure software company into its portfolio. And while VMware doesn't have the same direct connection to the AI chip boom, it gives Broadcom exposure to a very different part of enterprise IT.
Broadcom has also been reshaping the VMware business around a subscription-based model, focusing its product portfolio on its core VMware Cloud Foundation platform. That creates a recurring revenue stream with a very different profile from Broadcom’s semiconductor business.
Screenshot courtesy of vmware.com Well, I say different, which is conceptually true - but VMware actually has an “in” on the AI buildout.
The platform’s virtualization and cloud management capabilities can essentially provide users with multiple virtual machines within a single physical computing unit, helping meet the growing demand for on-site data centers and private cloud environments.
And since everyone seems to be using AI for their workloads, VMware can also benefit as businesses upgrade and expand the infrastructure needed to run them.
But that’s the broad-stroke application. Recently, Broadcom launched VMware AI Factory, a platform designed to help enterprises deploy and manage AI workloads across their private and hybrid cloud environments. That puts VMware much closer to the AI infrastructure conversation, giving Broadcom another avenue to capture spending as businesses bring AI into their own data centers.
Screenshot courtesy of broadcom.com So there’s definite value in Broadcom’s software play. Now let’s see how big it is compared to the rest of the business.
How Valuable Is VMware to Broadcom?
As of Q3 FY26, the Infrastructure Software segment contributed 30%, or $8.7 billion, to revenue. That’s a pretty significant chunk of Broadcom’s business, especially when you consider that the segment tends to be more predictable than semiconductors.
Software revenue is largely recurring, thanks to VMware's subscription-based model, which gives Broadcom a steadier cash flow between semiconductor cycles.
Screenshot courtesy of broadcom.com Broadcom has also been putting considerable effort into making that business more profitable. Since acquiring VMware, the company has cut costs, streamlined the product portfolio, and pushed customers toward subscriptions. That has made VMware a much more important contributor to Broadcom’s overall financial profile.
Now, a 70/30 split doesn’t seem like an even division of revenue, but the segments used to be closer. In fact, in Q3 FY25, the split was 57% semiconductor and 43% infrastructure software. Both segments grew year over year; the semiconductor segment just outgrew software.
Screenshot courtesy of broadcom.com In other words, AI is rapidly changing the composition of Broadcom’s business, but that doesn’t make the software side irrelevant.
So, when we talk about Broadcom’s AI opportunity, it’s important not to look at custom silicon in isolation. The semiconductor business may be driving the explosive growth, but VMware provides a large, recurring, and, more importantly, AI-adjacent software business that lets Broadcom benefit from the data center buildout in another way.
Broadcom's Customer Concentration Risk
But you’ll notice that there’s a bunch of AI words in the last two sections, which can be a red flag for seasoned investors. After all, that suggests Broadcom’s business, while relatively diversified, still relies heavily on today's AI boom. And that’s a perfectly valid concern - which is why I looked into it.
In its Q2 FY26 10-Q report (because there's no 10-Q for the third quarter yet), the company reported that its top five end customers accounted for about 45% of total net revenue, up from roughly 40% in the comparable period a year earlier.
Screenshot courtesy of broadcom.com I say the more interesting number is its distributor concentration. One semiconductor distributor accounted for 42% of Broadcom’s total company revenue in Q2 FY’26, compared with 29% a year earlier. Now, that doesn’t mean one customer is responsible for 42% of Broadcom’s sales, since the distributor ultimately serves multiple end customers. Still, it represents a significant concentration in Broadcom’s sales channel.
And Broadcom specifically calls out the risk this creates for its custom AI accelerator business. Large AI customers can reduce, delay, or cancel orders, while also negotiating for lower prices or more favorable contract terms.
The company even notes that customers could eventually lease custom AI accelerators or entire AI racks rather than purchase them outright, which could affect Broadcom’s cash flow and overall risk profile.
So, despite having VMware and a broad semiconductor portfolio, Broadcom still has considerable exposure to a relatively small number of AI hyperscale customers.
Marvell's Data Center and Custom Chip Business
Marvell, on the other hand, is the newer player in the space, though it has gotten the attention of industry leaders. Jensen Huang once called it “the next trillion-dollar company.”
But does that statement hold up under scrutiny? Let’s see, starting with its segments. The company has two: Data Center and Communications and Other.
As you’d expect, Data Center is focused on the chips and infrastructure used inside data centers, including products for compute, networking, storage, optical connectivity, and, of course, custom ASICs.
Speaking of which, Marvell is estimated to hold 20-25% of custom ASIC market share, driven by key design wins with Amazon (AMZN) and Microsoft (MSFT).
Communications and Other covers Marvell’s remaining semiconductor businesses, including chips used in carrier networks, automotive applications, and consumer and enterprise markets.
So, a pretty simple delineation. Now let’s see how each segment contributes to Marvell’s top line.
Screenshot courtesy of marvell.com 
Screenshot courtesy of marvell.com As of its Q2 FY27 report, the Data Center segment has contributed 79%, or roughly $2.17 billion, of the total revenue, “driven by continued strong demand across our Data Center portfolio,” according to Chairman and CEO Matt Murphy. “AI-related bookings remain exceptionally robust, and we expect our revenue growth to accelerate further through the remainder of fiscal 2027.”
So it looks like the same story as Broadcom’s, right?
Well, broadly speaking, yes. Both companies are riding the AI wave. Their data center businesses are taking the spotlight and driving meaningful growth.
But unlike Broadcom, Marvell is much more concentrated around semiconductors and data infrastructure. There’s no VMware-like software business sitting alongside it, so the company’s fortunes are much more closely tied to the demand for chips, connectivity, and data center infrastructure.
The numbers make that concentration clear.
Marvell's Customer Concentration Risk
In its latest 10-Q, covering the quarter ended August 1, 2026, Marvell disclosed that one direct customer accounted for 16% of quarterly revenue, while one distributor represented a much larger 44%. The distributor figure, like Broadcom’s, doesn’t necessarily mean one end customer generated 44% of sales, but it still represents a significant concentration in Marvell’s sales channel.
The risk, however, becomes even more pronounced when I checked the company’s latest annual filing.
There, management stated that only 10 customers accounted for 82% of fiscal 2026 revenue.
Screenshot courtesy of marvell.com The report doesn’t explicitly say these customers are from its custom silicon business, but it doesn’t take a genius to connect customer concentration with data center revenue growth.
Now, it's hard to say the 82% figure from 10 clients is materially different from Broadcom’s 45% from its top five customers. After all, we’re looking at a different number of customers and different reporting methodologies, so no apples-to-apples here.
However, these figures still tell us something important. Marvell has a much smaller cushion between its biggest customers and the rest of its business. A delayed project, reduced order, or lost design from any of those 10 clients can have an outsized impact on its financial performance.
Meanwhile, Broadcom’s top five customers account for less than half of its revenue, meaning more than half comes from everyone else. And with its business more materially diversified than Marvell’s, that cushion looks more like an effective safety net.
Two Different Ways to Invest in the AI Chip Boom
By now, the distinction between the two companies should be pretty clear. But I think there’s another way to look at it: Broadcom and Marvell offer investors different ways to participate in the same AI infrastructure cycle.
Broadcom gives you several sources of growth at once. Custom silicon and networking are benefiting directly from AI spending, while VMware adds a subscription-based software business with its own customer base and demand drivers. That means the company can benefit from AI without making AI the sole determinant of its future.
Marvell offers much more direct exposure. Its business is increasingly built around the chips and connectivity large-scale data centers need, particularly as hyperscalers develop their own custom silicon. If those customers continue increasing their infrastructure spending, Marvell stands to benefit significantly - and fast.
In fact, these differences have already shown up in each company’s stock performance.
Over the last five years, Broadcom (blue) has significantly outperformed both Marvell (black) and the S&P 500 (SPX) (orange).
MRVL vs. AVGO and SPX via Barchart. But if we zoom in to the last 52 weeks, Marvell comes out on top by a big margin, while Broadcom is just above the S&P 500.
MRVL vs. AVGO and SPX via Barchart. To me, that’s a useful illustration of the trade-off we’ve been talking about.
Broadcom’s broader business has historically delivered stronger long-term performance and more consistent growth, while Marvell’s more concentrated exposure can bring more momentum and volatility when the market is particularly enthusiastic about AI infrastructure.
I like to call this the “degree of sensitivity.” Broadcom has more insulation if one part of the technology market weakens - and we saw plenty of that this year. Despite those downsides, though, its performance has been relatively stable.
On the other hand, Marvell has less insulation, so its price reacts more strongly to broader short-term headwinds and tailwinds. However, it also offers stronger upside participation when the core AI market starts firing on all cylinders, as we can see in its price peaks.
But overall, neither structure is objectively better. It all boils down to preference: do you want a steadier compounder with multiple sources of growth, or a higher-volatility play that gives you more direct exposure to the AI infrastructure cycle?
Your risk appetite will dictate where you lean. But what does Wall Street say about these two?
What Analysts Are Saying About These Custom AI Chip Stocks
Interestingly, analysts are quite bullish on both AVGO and MRVL. But dive into the numbers, and you’ll see they tell a slightly different story.
AVGO analyst ratings via Barchart. A consensus among 42 analysts rates Broadcom a Strong Buy, with an average score of 4.69. The high target price is set at $715, translating to roughly 94% potential upside at the time of writing.
MRVL analyst ratings via Barchart. Meanwhile, a consensus among 36 analysts rates Marvell a similar Strong Buy, though with a lower average score of 4.53. The high target price at $400 implies around 77% upside potential.
So, Wall Street is giving both companies a strong vote of confidence, but Broadcom has slightly more upside. Of course, these are analyst estimates, not guarantees, and the highest target represents the most optimistic case.
However, it reinforces the broader point I’ve been trying to make: both companies have substantial AI opportunities ahead of them, but they offer investors two very different ways to capture them.
What Could Change This AI Infrastructure Outlook
So who wins? Broadcom or Marvell?
Well… I’m leaning toward Broadcom.
Both companies are “Buys” in my book. It doesn’t matter to what degree one is higher than the other.
What matters more is understanding the risks of owning either company. After all, the AI market will likely look very different in three to five years, and for long-term investors, it’s the long-term catalysts that matter more.
Like Broadcom, with its VMware segment, for one simple reason: real diversification. The point of that software platform was to give Broadcom a source of recurring revenue outside chip manufacturing and reduce its concentration risk. If that line of business continues to grow despite a broader semiconductor downturn, then it’ll be working just as intended.
Marvell has a somewhat different issue. It’s performing extremely well today, given that AI spending among the world’s top hyperscalers is like a runaway train with the brakes broken. However, when that spending inevitably slows, its extremely concentrated exposure to the AI market could drag it down to new lows.
So can Marvell broaden its customer base before AI spending slows down?
After all, selling ASICs during an AI boom is one thing. Turning that demand into a durable and diversified business is another thing entirely.
On the date of publication, Rick Orford had a position in: GOOGL , META , AMZN , MSFT . All information and data in this article is solely for informational purposes. For more information please view the Barchart Disclosure Policy here.
More news from Barchart
2 Reasons to Buy Super Micro Stock — And 1 Reason to Be Cautious Scott Bessent Thinks Norway May Trade Treasuries for Assets Like Fannie Mae. What That Actually Means for FNMA Stock. A Rare Quant Signal Just Flashed for Contrarian Options Traders in Home Depot Stock Why 1 Veteran Analyst Just Trimmed His Intel Stock Price Target