I keep hitting the buy button on Broadcom because the company sells nearly everything inside a hyperscale AI cluster except the GPU. Every time I examine how Google, Meta, OpenAI, and Anthropic build compute, Broadcom’s name appears on the custom accelerator, Ethernet switch, DSP, optics, and fabric. That is why my cost basis keeps climbing.
The thesis: Broadcom (NASDAQ:AVGO | AVGO Price Prediction) captures the hyperscaler AI capital budget that does not flow to NVIDIA (NASDAQ:NVDA). Custom silicon (XPUs/ASICs) and AI networking interconnects are the two other legs of the data center stool, and Hock Tan’s team owns both.
The Receipts Behind the Conviction
AI semiconductor revenue jumped from $5.20 billion in Q3 FY25 to $10.80 billion in Q2 FY26, a 143% year-over-year increase. Q3 FY26 guidance calls for $16.0 billion, up over 200% YoY, with management guiding to $56 billion for full-year FY26 and in excess of $100 billion by FY27. Hock Tan stated that “Demand for XPUs and networking is simply insatiable,” and the $30 billion-plus in Q2 AI bookings backs the claim.
Q2 FY26 free cash flow was $10.26 billion, 46% of revenue, up 60.07% year-over-year. Adjusted EBITDA landed at 69% of revenue. Net income grew 87.51% to $9.31 billion. That margin structure funds both the $10 billion buyback authorized through December 2026 and the dividend, raised 10% last December to $0.65 per quarter, the 15th consecutive annual increase. This is a compounder’s dividend record on an AI growth chassis.
Networking made up almost 40% of AI revenue in Q2. Broadcom’s Tomahawk 6 is the industry’s only 100-terabit Ethernet switch, with a 200-terabit successor taping out this quarter. Anthropic committed to over 1 gigawatt of TPU compute in 2026 and another 5 gigawatts starting 2027. OpenAI is on the hook for 1.3 gigawatts in 2027 inside a 10-gigawatt by 2029 deal. Meta signed for 3 gigawatts of MTIA XPUs through 2028. These are multi-year, multi-generational commitments spanning several product generations.
Why Not Just Buy NVIDIA
I already own NVIDIA. The exposure NVIDIA does not provide is to customers building alternatives to it. Google’s TPU roadmap, Meta’s MTIA, OpenAI’s custom silicon, and Anthropic’s compute all route through Broadcom. When a hyperscaler wants to control cost and power per token, it hires Hock Tan’s team to co-design the chip and buys the switch fabric to connect it. That revenue stream grew 143% last quarter while consolidated revenue climbed 47.9%.
The Risk
Customer concentration is real. A handful of hyperscalers drive the AI number, and if one pulls a program or delays shipments, a quarter can move. The Q2 call disclosed $6 billion in orders from two additional core customers shipping late 2026, broadening the base. Total liabilities remain elevated from the VMware deal, and semiconductor cycles are cycles. The bookings, multi-year gigawatt commitments, and 46% free cash flow margin are doing the talking.
Why the Buy Button Stays Active
At $418.16, the stock trades at roughly 65x trailing and about 21x forward earnings, with analyst targets averaging $527.88. I am buying because a company generating 69% EBITDA margins on an AI business heading from $56 billion this year toward $100 billion next year is the rarest compounder: one whose customers plan their power grids around its ship dates.
Contact [email protected] for any questions or corrections.
