You know how this story is supposed to go, right? Nvidia owns AI chips. Everyone else is fighting for scraps. The GPU is the only thing that matters. Wall Street's told you this a hundred times and you've probably stopped questioning it.
Broadcom just proved the entire narrative is incomplete. While you were watching Jensen Huang's leather jacket collection evolve, Hock Tan was quietly racking up a $73 billion AI backlog — committed orders, with TSMC capacity locked through 2028. AVGO hit a $1.96 trillion market cap at $411 a share (+18.6% YTD). And the AI revenue? Q2 clocked in at $10.8 billion, more than doubled year-over-year, representing nearly half of total revenue. This isn't a side hustle. This is the main event.
Here's what you need to understand: Broadcom isn't competing with Nvidia. It's building the chips that Nvidia doesn't — custom ASICs purpose-built for individual hyperscalers. Google's TPU v7 "Ironwood"? Broadcom. Meta's first 2-nanometer AI silicon? Broadcom. OpenAI's first-ever custom chip, tied to a 10-gigawatt infrastructure deal and targeting 2027? Broadcom. Anthropic dropped a $10 billion order. ByteDance and Fujitsu are in the door. Apple's name keeps surfacing. That's six confirmed hyperscaler customers — and the list is growing.
The custom ASIC co-design market is basically a duopoly: Broadcom and Marvell own roughly 95% of it. But here's the thing — Broadcom's not just a chip designer. They're the networking backbone of every AI cluster on the planet.
The Invisible Moat Nobody Talks About
GPUs get the glory, but switches do the actual work. Every GPU in an AI data center connects through Ethernet switch chips — and Broadcom owns roughly 70% of that market. Their Tomahawk 6, announced in March 2026, is the world's first 102.4 terabits-per-second switch chip. That's not a typo. One chip. A hundred trillion bits. Every second.
The datacenter Ethernet switch market grew 63% year-over-year in Q4 2025, per IDC. As AI clusters scale from thousands to millions of GPUs, the networking layer becomes more critical — and more valuable — than the compute itself. Broadcom collects a toll on every single connection.
The Math That Changes Everything
Q2 AI revenue: $10.8 billion. Q3 guidance: AI revenue triples to roughly $16 billion. And Hock Tan just told Wall Street that FY2027 AI chip revenue is heading to $100 billion or more. At a trailing P/E of 68, you might flinch — until you see the forward P/E is 21.22. That's not expensive. That's a growth stock priced like a value play because the market still hasn't internalized how big this gets.
The $73 billion AI backlog isn't analyst speculation — it's committed orders with TSMC capacity locked through 2028. These chips are already spoken for. The revenue is a matter of when, not if.
Nvidia's the face of the AI buildout. But Broadcom is the spine — the custom silicon and networking infrastructure that makes every AI workload physically possible. And with $73 billion in committed orders, we're not guessing about demand. We're just waiting for the revenue to land.
— The Signal
