Broadcom's Custom-AI Boom Runs on TSMC's Toll Road

Broadcom had the quarter that a custom-AI skeptic could not have invented. In the three months through early September it sold $16.7 billion of AI semiconductors—more than triple the year-ago figure—and the company now talks about a $230 billion custom-chip and networking business by fiscal 2028. And yet it fabbed none of it. Most of that production depends on one landlord for foundry and advanced packaging: Taiwan Semiconductor Manufacturing Co., though secondary packaging suppliers such as ASE, SPIL, and Amkor also take on overflow orders.

That is not a footnote to the Broadcom story. It is the structure of the trade. The custom-AI boom that bears Broadcom's name is physically routed through a single landlord, and the landlord is raising the rent at the exact moment the tenant is promising the biggest growth in its history. What that means for each stock is the useful question.

The boom, in one number

Broadcom doesn't sell the flashiest product in AI. It sells the other chips—the custom accelerators that Google, Meta, OpenAI, and Anthropic design with it to run their models more cheaply than an off-the-shelf Nvidia GPU, plus the Ethernet networking silicon that links them. In its fiscal third quarter, its AI semiconductor revenue was $16.7 billion, up 221% year over year. That was 54% sequential growth against total company revenue of $29.59 billion.

The scale is why the story is worth attention. Broadcom's management has guided Wall Street to expect AI chip-plus-networking revenue to grow from $2.6 billion in fiscal 2022 to $230 billion in fiscal 2028—a two-order-of-magnitude climb the company says is supported by secured wafer, memory, and packaging capacity, alongside a large reported order book. The reason the biggest buyers keep signing up: custom ASICs cost hyperscalers roughly 30–50% less in total cost of ownership than general-purpose GPUs on specific workloads.

Now hold that growth next to a hard physical fact.

Why TSMC is the landlord, not a supplier

An AI accelerator is not one chip. It is a stack: the compute die, the high-bandwidth memory beside it, and a silicon interposer that wires them together into one package. That final step—TSMC's CoWoS technique—is the narrowest node in the entire AI supply chain. Even after the compute die is fabricated perfectly on a leading-edge node, it sits useless until it passes through this packaging step, which in 2026 is sold out with lead times stretching past a year.

Broadcom is one of the two or three biggest buyers of that scarce capacity, demanding roughly 300,000 CoWoS wafers in 2026 and around 484,000 in 2027. Its multi-year revenue targets are only credible because it has locked in foundry and packaging capacity through 2028. But locking in capacity is precisely the language of a tenant, not an owner.

And the landlord behaves like one. With demand for advanced nodes and packaging outstripping supply, reports put TSMC's CoWoS packaging price hikes as high as 20%. The company has also told customers to expect a further 5–10% base price increase plus a surcharge on excess high-performance-computing orders beginning in early 2027. The company that owns every toll booth is rewriting the tolls.

This resolves into a genuine difference in economic role. Broadcom is the designer, and design carries real margin—about 65% gross on its AI chips. But it is a tenant in the physical plant. TSMC is the collector, charging a rent on every AI chip anyone sells, Broadcom's custom ASICs and Nvidia's GPUs and AMD's accelerators alike, and its pricing power is now a visible earnings lever rather than a distant risk.

What the tape already priced

The two stocks have sailed in opposite directions this year, and the market has effectively already cast the roles. Through early September, TSMC was up about 43% year to date with a 120-day gain near 32%, while Broadcom was up roughly 5% and had fallen around 13% over the trailing 20 days—the aftermath of a quarter that beat estimates but guided the next one slightly below a very high bar.

Valuation tells the same story from the other side. TSMC trades around 33 times trailing earnings and carries the broader, simpler thesis—a toll on every AI chip in a market where no one seriously competes with it at the leading edge. Broadcom trades nearer 45 times trailing earnings, because the market is paying for the design margin and the custom-ASIC share it is projected to hold.

Here is the discipline that matters. The Broadcom–TSMC dependency is confirmed and durable: Broadcom cannot realistically manufacture or package its own accelerators, TSMC cannot be replaced any time soon, and the relationship is a genuine chokepoint rather than a passing bottleneck. That is the map. It is not, by itself, a verdict on either stock.

The asymmetry the map once suggested has partly closed on both sides. TSMC's rent-raising is now discussed openly and much of its rerating is behind it, so the "hidden landlord" is no longer hidden. Broadcom's growth is enormous but its expectations are exacting: it must keep winning custom-ASIC programs, keep securing scarce HBM and packaging allocation as its customers multiply, and absorb TSMC's rising tolls within margins already under scrutiny. The structural relationship is real. Whether the price still offers the edge is a separate question—and the two are easier to confuse now that both names have already moved.

What is confirmed is the dependency itself. Custom AI chips, no matter whose name is on them, pass through the same gate. Owning the boom at the scale Broadcom promises means paying the landlord, and the landlord has decided the price is going up.