There are a lot of things that are in short supply in the IT sector these days, and one of them is teams of engineers who can design complex chips and get the successfully through the fabs and packagers and onto system boards where they can be embedded into systems. Broadcom and Marvell, rivals in the chip shepherding businesses, are seeing their businesses explode thanks to this relatively new line of business, and it will not be long before this utterly dominates their business.
Wall Street is so used to the “beat and raise” game that it freaked out when Broadcom only affirmed its guidance for this year and next, and just because Nvidia co-founder and chief executive officer Jensen Huang proclaimed in the wake of a partnership expansion with Marvell that this chip shepherd would be the next $1 trillion market cap chip company, Marvell got a pop when Broadcom got smacked across the nose. (Broadcom stock recovered a few days later when sense prevailed.)
The thing is, Broadcom is already approaching $2 trillion in market cap, and that will keep rising as its AI business grows both revenues and profits. That Broadcom is talking about $56 billion in fiscal 2026 and more than $100 billion in AI revenues in fiscal 2027 is nothing short of remarkable. It wasn’t that long ago when Nvidia itself was this “small.”
And, of course, given its optics datacenter switching businesses as well as its chip design business, Marvell will break $1 trillion in market cap. It is already more than a quarter of the way there as this story goes to press, and Huang has the inside information on exactly how much it is going to spend on Marvell components in the coming years. (Which we presume will be large.) So maybe Huang is channeling something. Broadcom should have between $95 billion to $100 billion in sales for its fiscal 2026, which ends in November this year, and Marvell is tracking to reach $11.5 billion in fiscal 2027 ended in February 2027 and $16.5 billion in fiscal 2028 ended the following February.
Both companies will be dominated by their AI datacenter businesses, which are a mix of custom CPU and AI processors, homegrown network ASICs, and electro-optical components. They are both utterly transformed from what they were only a few years ago, and there is no reason to believe that the hyperscalers and the cloud builders are not going to increasingly emphasize their homegrown AI accelerators as they have been doing with their homegrown CPUs and that somewhere north of half of the flops in their datacenters will eventually come from these homegrown and shepherded chips. It is so much less expensive to make their own devices – it costs 40 percent of the cost of an Nvidia or AMD GPU at retail prices, and they can rent it for 60 percent to 70 percent of the price they charge to rent GPUs, thus keeping the 20 points in their pockets to fuel more AI spending.
That’s what I would do, and that’s what you would do, given the same situation.
This rebalancing away from a single vendor to a mix of bought and designed compute engines is inevitable, so long as Broadcom and Marvell keep delivering on their promises to the hyperscalers, cloud builders, and now the AI model builders. And that means that Broadcom and, to a lesser extent, Marvell, are indeed the real counterbalances to Nvidia at the hyperscalers and the cloud builders – much more so than AMD. Something has to provide an alternative because economic substitution is an economic law, not just a good idea.
With that, let’s drill down in the Broadcom numbers and then follow suit with Marvell.
Broad Communications, Broad Computing
In its second quarter of fiscal 2026, which ended in the first week of May, Broadcom pulled in $22.19 billion, up 47.9 percent year on year and up 14.9 percent sequentially. Operating income was up nearly double that to 10.79 billion (rising 85.1 percent) and net income was $9.31 billion (up 87.5 percent).
Needing to preserve cash to keep this chip shepherding business going, Broadcom was hoarding its cash and not paying down the massive $67 billion in debt it took on to acquire VMware. The company ended up with $19.63 billion in cash and has $64.91 billion in debt. If you were Broadcom and you had Google and Anthropic as TPU customers and Meta Platforms ramping up its MTIA accelerators, you would ride the credit card hard and keep a lock on the savings account, too. Heaven only knows what HBM memory will cost down the road, and even if the costs get passed back to Google and Meta Platforms, you have to pay upfront for memory allocations as well as your place in line at Taiwan Semiconductor Manufacturing Co.
Broadcom has two major groups, one that makes chips and the other that peddles legacy (and profitable) systems software for the enterprise.
The Semiconductor Solutions group had a tad over $15 billion in sales, up 78.5 percent, with operating income of $9.31 billion, up 41.1 percent thanks in large part to sales of Tomahawk, Trident, and Jericho switch ASICs into the datacenter. The AI XPU business drives revenues, but we do not think it drives margins quite the same way. Our guess is that Broadcom is doing better than the OEMs and ODMs, who get single digit operating margins for AI servers, and maybe has gross margins in the range of 50 percent to 55 percent on the AI chippery and more than 65 percent for all of its other chips. And a lot of that margin in the AI column at Broadcom is most definitely not coming from AI XPUs, but rather its own AI chippery for networking.
It is better to be Broadcom than to be Dell and Hewlett Packard Enterprise, but then again, someone has to build the machines. . . .
The Infrastructure Software group had $7.18 billion in sales, up 8.8 percent, which is respectable given its legacy nature, and more importantly, Broadcom was able to increase operating income by 12.8 percent to $5.67 billion, which is a 79 percent operating margin. Chip operating margins were at a blended 62 percent in Q2 F2026, and that is with a lot of help from very pricey switch ASICs.
Our best guess is that VMware drove $5.28 billion in sales, up 15 percent, with operating income of over $4 billion, up 32.6 percent. Yeah, that sounds crazy. But Hock Tan squeezes the hell out of every business, and VMware had a lot of fat and prices that were not what the market would bear before Broadcom acquired it. If former chief executive officer Pat Gelsinger had been as ruthless, he would still have a job. (Ditto for Gelsinger’s stint at Intel.)
Broadcom does not talk about its divisions anymore, but here is our stab at trying to figure out what might be happening down inside the chip group:
Yup, we think enterprise networking took a pause even as AI networking is skyrocketing. The hyperscalers and cloud builders are spending like crazy on 102.4 Tb/sec Tomahawk 6 chips, which has plain vanilla and co-packaged optics versions as well as Tomahawk 6 Ultra version aimed at scale up networks to lash together GPUs and XPUs. Don’t worry. That orange line is going to skyrocket once enterprises catch up, and they will.
That brings us to the AI cut at Broadcom. The company is explicit about the split, which is refreshing, even though it is giving fewer and fewer hints about what is happening within the AI segment in terms of XPUs versus optics versus racks versus other. Here is the AI versus non-AI split over the past several quarters at Broadcom:
The non-AI business is just trending along, and we think reflects the fact that companies big and small are shifting IT budget away from legacy workloads and towards new AI workloads.
Here is a more detailed table I put together showing AI versus non-AI with a further bifurcation of compute and networking:
The thing that is not in this table is the forecast for Q3 and Q4. If you do the math AI chip sales will be $16 billion in Q3 and around $21 billion in Q4 to reach that $56 billion. Broadcom will probably beat this, but maybe not hit as high as $60 billion, which is was talking about a few months ago. (Our guess? HBM allocations determine this number more than any other factor in a complex design and supply chain.)
Marvelling At The AI Shazaam
In two weeks, Marvell will be added to the S&P 500 index, knocking out the Campbell Soup Company. A sign of the times, indeed. But remember: You can’t eat AI, even if it can eat you.
Marvell’s AI business is about a tenth the size of Broadcom’s and it is a bit different in that it relies more heavily on AI networking than it does AI compute. At least for now. Amazon Web Services is the anchor customer for both AI compute and AI networking, but Nvidia is coming on strong in the latter category for various kinds of optics.
In the first quarter of fiscal 2027, which also ended in early May, Marvell had $2.42 billion in overall revenues, up 27.6 percent. Operating income was a mere $339 million, only 14 percent of revenues, and thanks to rising costs across the board, net income fell by more than an order of magnitude to $34.5 million. Obviously, Marvell does not have an enterprise software business to fall back on. But it does have a very fast-growing AI business and the prospects of – how do these companies put it? – accretive operating income even if there is operating margin dilution.
The only thing worse than selling AI gear to the hyperscalers and clouds is not selling it at all.
Marvell exited Q1 F2027 with $3.84 billion in cash (up by a factor of 4.3X year on year), and had $4.96 billion in debt. This is not bad, and certainly not as deep as the debts that Broadcom has. But Broadcom has a much larger and more profitable business, and as the third largest buyer of HBM memory in the world and a position near the front of the fab line at TSMC, it has great sway. Marvell is on the rise, though.
Like Broadcom, Marvell has lumped its businesses into two groups: Datacenter is the first, Communications & Other is the second. And the Datacenter group is now 4.5X larger than the rest of the company, and grew by 27.2 percent to $1.83 billion in fiscal Q1. The Communications & Other group kept pace, however, with $585.1 million in sales, up 28.7 percent year on year. This is the profitable part of the business, so growth is very important here to counterbalance the AI XPU business, which is not as profitable.
Marvell does not talk about its divisions anymore, also like Broadcom, and we have therefore also taken a stab at breaking sales down into the old divisions:
The numbers for Q4 F2026 and Q1 F2027 are hunches based on a hint here and there, nothing more. Take them for what you will.
What everyone wants to know is how the AI business is doing within Marvell, and here is our breakdown of the datacenter business, including AI and non-AI components:
I think Marvell did just shy of $500 million in AI XPU sales in Q1 F2027, about the same as in Q4 F2026 but still up 71 percent from the year ago quarter. (Dominated by AWS Trainium chips, of course.) The sale of electro-optical components for AI – components for high speed network transceivers thanks to the Inphi business – dominate the AI sales for Marvell. My best guess is that they accounted for $916 million in the first quarter, up 47.5 percent. There is some Teralynx switch ASIC revenue in there somewhere perhaps, or maybe that is classified in other datacenter revenues. (It is not clear how this is allocated.)
Here’s the data visually for those of you who like pictures:
In any event, if my model is correct, then that means other sales in the Datacenter group fell by 20.7 percent to $419 million.
