Broadcom's Custom-Chip Boom Is Real. The Chokepoint It Owns Is the Network.

In early September, Broadcom reported the kind of quarter that most semiconductor companies only fantasize about. For its July quarter it booked $29.6 billion in revenue, up 86% from a year earlier, and roughly $16.7 billion of that came from AI chips — custom accelerators plus the networking that links them — nearly three times the year-ago figure. The stock fell anyway. A beat, a raised outlook, and a slide has become the pattern in 2026: the buy side keeps asking for more, and the price keeps giving it back.

Start there, not because the growth is in doubt, but because it frames the real question. Everyone can see that hyperscalers are buying custom silicon in volume. The useful question for an investor is narrower: in a supply chain this physical, which node actually decides how many chips ship, who owns it, and what has the market already paid to assume the answer?

The chip that undercuts the GPU maker

Broadcom doesn't sell these chips because hyperscalers like the logo. It sells them because the economics beat renting a GPU from Nvidia. The company's own cost math is the cleanest way to see it: a Broadcom custom XPU system runs roughly $12 billion per gigawatt of compute, versus about $18 billion for Nvidia's current Grace-Hopper generation and roughly $40 billion for the coming Vera-Rubin line. For a hyperscaler building several gigawatts a year, that gap moves operating costs by billions.

Custom ASICs are built to hardwire one workload, which makes them far more efficient at the task they were designed for — but that flexibility cost runs both ways. The deeper reason the big buyers want them is strategic. A hyperscaler does not want its entire AI build flowing through one supplier that sets the price and the allocation. Co-designing its own silicon in partnership with Broadcom is the way to keep that dependency in check, which is why the customer list reads like the who's-who of model builders: Google's TPUs since 2014, Meta's MTIA accelerators, OpenAI's newly unveiled Jalapeño chip, and Anthropic. Broadcom and Marvell together control roughly 95% of this custom ASIC co-design business, with Broadcom holding on the order of 70%.

The wave is real, and it is confirmed by booked commitments rather than slides. Management has said it has a line of sight to more than $100 billion in AI chip revenue in fiscal 2027, up from a 2026 target that was just raised to $58 billion. Anthropic is projected to become Broadcom's largest XPU customer in 2027, then OpenAI, and the three of them — plus Meta — are committing to gigawatt-scale buys extending into 2028.

The narrowest nodes aren't Broadcom's

Now apply the map that usually pays off. The design is Broadcom's. The physical capacity that turns that design into a shippable chip is not. Every XPU is built on TSMC's most advanced node, assembled inside TSMC's CoWoS advanced packaging, and wrapped in high-bandwidth memory from SK Hynix, Samsung, or Micron.

These are the narrowest nodes in the whole chain. Advanced packaging and HBM are the two parts of AI hardware that expand least quickly — you cannot pull a new packaging line or a new memory fab out of the air, and every AI accelerator from every vendor competes for the same slices of that capacity. And not one of those constraints sits on Broadcom's own balance sheet. Broadcom rents them.

That makes this a subtle kind of moat, and Hock Tan has been unusually explicit about it. He attributes the company's advantage not to chip design alone but to having secured HBM, wafers, and packaging capacity ahead of rivals. Translate the claim into plain terms: Broadcom's durable edge is that it is the one firm large and early enough to contract years of packaging and memory on behalf of its customers, turning a coordination of other people's chokepoints into a reserved seat at the front of the line. It is a real barrier — competitors like Marvell or upstarts don't carry the scale to lock the same capacity — but it is a coordination moat built on rented scarcity, not ownership of the scarce thing.

The chokepoint it actually owns sits in the network

And that is where the map turns in Broadcom's favor. The constraints that decide how many custom chips can ship live at TSMC and the memory makers. The genuinely ownable node in this chain is one layer over: the network that ties the chips together.

For every dollar a hyperscaler spends on accelerators, it typically spends another 40 to 60 cents on the interconnect fabric that connects them. As clusters grow to the point where thousands of accelerators must talk to each other, the network stops being an accessory and becomes the next thing that binds. Broadcom is the dominant supplier of that fabric — the Ethernet switching silicon (its Tomahawk and Jericho lines) that is displacing Nvidia's proprietary InfiniBand in favor of open Ethernet, the ultra-fast SerDes interconnects, and the co-packaged optics that connect ever-larger packages. In the quarter just reported, AI networking alone contributed about $4.5 billion of the $16.7 billion AI total, roughly a quarter of it.

This is the part of the story that is hardest to overstate and easiest to miss in a headline about custom chips. Broadcom is both the largest designer of the accelerators themselves and the owner of the interconnect chokepoint every competing accelerator also has to pass through. Multi-year swap costs and qualification gates protect that network position the way HBM contracts protect the XPU side — but unlike the packaging and memory, this is capacity and IP Broadcom controls directly.

What the price already assumes

The last question is whether the structure still beats the price. This is a $1.7 trillion company trading near 45 times trailing earnings, and it has entered the phase where the market no longer rewards good news — it punishes anything short of flawless. The shares fell roughly 12% after the June-quarter report despite a beat, and again after September's blowout, leaving them down about 13% over the past month.

Two pressures explain the skittishness, and both are real. Gross margin has slipped from roughly 77% to about 74% because custom ASICs carry more physical manufacturing content than Broadcom's older products — the rented packaging and memory pass through the income statement as cost. And revenue is concentrated: a handful of hyperscalers, at most six, drive the entire XPU business. The loss of a single major design relationship would register immediately.

The asymmetry the structure once offered has thinned. The dependency is confirmed — custom chips are a multi-year tailwind, and Broadcom owns the design talent and the network chokepoint that rides it. What's left is the recognition that the two scarcer inputs in the chain, packaging and memory, are rented, not owned, and that the stock already prices a near-flawless run. Broadcom is the counterweight to Nvidia that the custom-chip wave promised. The question for a new buyer is whether paying full price for the obvious winner still counts as finding the edge.