Remember when semiconductor stocks were all about smartphones and PCs? Yeah, those days feel almost quaint now.
Today, if you’re not riding the artificial intelligence wave, you’re basically treading water in this market. And few companies have positioned themselves as perfectly in the path of this tsunami as Broadcom.
When the bell rings this afternoon, all eyes will be on one number that matters more than any other: how much money Broadcom is actually making from custom AI chips. Not the networking stuff (though that’s massive too), but the bespoke silicon it’s building for the biggest players in artificial intelligence.
The Quiet Giant That’s Suddenly Impossible to Ignore
Let me put this in perspective. A year ago, most retail investors couldn’t have told you what Broadcom’s custom ASIC business was doing. Today? It’s arguably the most important growth driver in the entire semiconductor space outside of Nvidia itself.
The stock’s up 75% this year for a reason. The market has finally woken up to something industry insiders have known for years: when hyperscalers like Google, Meta, and now OpenAI want to reduce their dependence on any single chip supplier (looking at you, Nvidia), there’s really only one company they can turn to at scale.
That company is Broadcom.
What Wall Street Actually Expects Tonight
The numbers themselves are almost secondary at this point, but let’s run through them anyway because they’re pretty staggering.
Analysts are looking for $1.86 in adjusted earnings per share on $17.49 billion in revenue. That would represent about 25% year-over-year growth from the $14.05 billion they did in the same quarter last year.
But here’s what everyone actually cares about: guidance. The Street wants to hear something around $18.27 billion for the current quarter, with EPS closer to $1.95. More importantly, they want color on the AI revenue trajectory into 2026.
We expect investors to focus on FY26 AI revenue guidance, Google and OpenAI revenue contributions, and gross margin trajectory given the steep ramp of custom XPUs.
– Goldman Sachs analyst note, November 2025
Translation: Tell us how big this gets next year, and please don’t scare us about margins collapsing as these complex custom chips ramp up.
The Google Relationship: More Than Just TPUs
Let’s talk about Google for a minute, because this relationship is the foundation everything else is built on.
During the quarter we’re about to see results for, Google dropped a bombshell that barely registered with most investors: Gemini 3, their most advanced AI model yet, was trained entirely on Google’s own TPU chips. Not Nvidia GPUs. TPUs. Which, of course, are designed and manufactured by… you guessed it.
This isn’t new—Google has been using TPUs for years—but the scale and sophistication keeps ratcheting up. Each generation requires more complex, higher-performance chips. And Broadcom keeps delivering them.
What’s fascinating is how this relationship has evolved. It started with basic accelerators for specific workloads. Now we’re talking about entire system-on-chip solutions that integrate compute, networking, and memory in ways that are specifically optimized for Google’s infrastructure.
In my view, this is the template for every other hyperscaler. Once you prove you can deliver at Google’s scale—with their insane requirements for reliability, power efficiency, and performance—everyone else wants to talk.
OpenAI Just Changed Everything
And then there’s OpenAI.
When the news broke in October that OpenAI would begin deploying custom chips developed with Broadcom starting in 2026, I think most people missed the significance. This isn’t some side project. This is OpenAI—the company that arguably kicked off the entire AI revolution—deciding that the future of their infrastructure cannot be 100% dependent on Nvidia.
Think about that for a second.
The cost structure implications alone are mind-boggling. Industry sources suggest these custom solutions could be 30-50% cheaper than equivalent Nvidia-based systems at scale. When you’re spending tens of billions on compute, that kind of savings isn’t just nice—it’s existential.
- Lower capex for the same performance
- Reduced dependency on a single supplier
- Custom optimizations for specific workloads
- Better power efficiency (critical as data centers hit power constraints)
- More predictable supply chain
Every single one of these points is a direct threat to Nvidia’s stranglehold on AI training and inference.
The Networking Moat Nobody Talks About
While everyone obsesses over the custom chip business, there’s another part of Broadcom’s AI story that I think is actually more defensible long-term: networking.
Here’s something most investors don’t appreciate: in these massive AI training clusters, the cost of the GPUs (or TPUs, or custom XPUs) is actually only about 40-50% of the total system cost. The rest? Networking, memory, storage, power delivery.
Broadcom dominates the high-end networking space. Their Tomahawk and Jericho switches are basically the default choice for anyone building at scale. When you combine that with their optical components business, they have this beautiful full-stack offering that nobody else can match.
It’s like being the company that sells both the engines and the transmission for Formula 1 cars. Sure, someone could try to piece together a competitive solution from multiple vendors, but good luck matching the performance and reliability of an integrated stack.
The Margin Question Everyone’s Afraid to Ask
Now, let’s address the elephant in the room: gross margins.
Custom chip development is expensive. Really expensive. These aren’t off-the-shelf parts—these are multi-year, billion-dollar development programs with NRE (non-recurring engineering) costs that would make most companies blanch.
The bears argue that as this business scales, margins will get crushed. The bulls say these are essentially annuity-like contracts with massive upfront payments that more than offset the development costs.
I’ve been following this space for years, and my take is that the bulls are right—but with a caveat. The initial ramps can be messy. You often see margin compression for a quarter or two as new programs come online. But once you’re in production at scale? These become some of the highest-margin products in the entire semiconductor industry.
Why? Because the customer is locked in. They’ve invested years and billions in software optimization for your specific architecture. Switching costs are astronomical.
Where This Could Go From Here
Let’s play this out. If Broadcom is already doing, say, $10-12 billion annualized from AI-related revenue (and that’s probably conservative), and this grows at 50-70% for the next several years…
We’re talking about a company that could be generating $30-40 billion in AI revenue alone by 2028-2029. With operating margins north of 50%. On top of their existing infrastructure and networking businesses.
At that point, you’re looking at a company generating $20+ billion in free cash flow annually. With a market cap that could legitimately challenge the Magnificent 7.
Crazy? Maybe. But six years ago, nobody thought Nvidia would be worth $3 trillion either.
The AI infrastructure buildout is still in its early innings. We’re not even through the first wave of training cluster deployments, and already companies are planning the next generation that will be 5-10x larger.
Every single one of those clusters needs chips. And networking. And optical components.
And there’s really only one company that can credibly supply all of it at scale, with proven ability to deliver the most demanding custom silicon on earth.
Tonight’s earnings call won’t answer all these questions. But it will give us the next data point in what I believe is one of the most compelling growth stories in technology today.
Sometimes the best investments aren’t the ones everyone is talking about constantly. Sometimes they’re the ones quietly executing while the spotlight shines elsewhere.
Broadcom might just be that investment right now.
Disclosure: The author holds a long position in AVGO shares.
