Have you ever watched a stock you love jump 3% the moment earnings hit, only to watch it completely reverse and drop hard twenty minutes later? That was Broadcom last night, and honestly, it felt like one of those movie moments where everyone in the theater gasps at the same time.
The numbers themselves were beautiful. The guidance was even better. Yet by the time the conference call Q&A started, the mood flipped. One analyst question about customers potentially bringing chip design in-house, one slightly nuanced answer from CEO Hock Tan, and boom — the after-hours selling began. Classic Wall Street overreaction or something worth worrying about? Let’s dig in.
A Quarter That Should Have Sent the Stock Soaring
Broadcom’s fiscal Q4 was, putting it mildly, ridiculous in the best way possible.
Revenue hit $18.02 billion — that’s 28% growth year-over-year and comfortably ahead of the $17.5 billion the Street expected. Adjusted EPS came in at $1.95, crushing the $1.86 consensus. EBITDA margins expanded massively, climbing to just over 67% of revenue. For a company this size to grow top and bottom line that fast while also expanding margins is borderline unfair.
The driver, of course, is AI. AI-related semiconductor revenue surged 74% to $6.5 billion in the quarter, beating the company’s own guidance from three months ago. Networking remains on fire as hyperscalers race to wire up clusters before they drop in the actual accelerators.
“Backlog for AI switches now exceeds $10 billion… when you add custom XPUs the total AI backlog sits above $73 billion.”
– CEO Hock Tan on the call
Seventy-three billion dollars of future AI revenue already in the bag. Let that sink in for a second.
The Custom Chip Business Just Keeps Getting Bigger
Here’s where it gets really fun. Remember last quarter when management casually mentioned a mystery fourth customer dropping a $10 billion order for custom accelerators? Last night they confirmed it’s Anthropic — and then dropped the bombshell that Anthropic doubled down with another $11 billion commitment for delivery in late 2026.
On top of that, a brand-new fifth customer signed an initial $1 billion deal, also for 2026 delivery. These aren’t small proof-of-concept orders. These are multi-billion-dollar, multi-year bets on Broadcom’s ability to design and manufacture exactly the silicon each customer needs.
Management also confirmed the chips going to Anthropic are the same Ironwood-generation XPUs (Broadcom’s name for custom ASICs) that power Google’s latest Gemini models. Other users reportedly include Apple, Cohere, and ByteDance. The scale, as Tan put it, “could be significant.” Understatement of the year.
Guidance That Should Have Been Celebrated
For the current quarter (fiscal Q1 2026), Broadcom guided to roughly $19.1 billion in revenue — almost a billion above consensus. They explicitly said AI semiconductor revenue will double year-over-year to $8.2 billion, driven by both custom XPUs and networking.
EBITDA margin guidance came in at ~67% of revenue, again ahead of expectations. The only “miss” was the infrastructure software segment guiding slightly below lofty hopes, but that’s largely VMware digestion noise and was more than offset everywhere else.
So again… why did the stock drop?
The Comment That Spooked Everyone
It all came down to the very first question in Q&A. An analyst asked whether hyperscale customers might eventually bring custom accelerator design in-house, reducing Broadcom’s moat over time.
Tan’s answer was long and thoughtful — maybe too thoughtful for a twitchy after-hours crowd looking for soundbites. He explained why purpose-built silicon will always beat general-purpose GPUs on efficiency, why the silicon road map moves so fast that customers are better off focusing on models rather than transistors, and then said this:
“I see this concept of customer tooling as an overblown hypothesis, which frankly, I don’t think will happen.”
Read in print, that sounds pretty dismissive of the risk. But some traders heard “he didn’t say it’s impossible” and immediately pictured Google, Meta, Amazon, and Microsoft all hiring thousands of chip designers tomorrow.
Look, I get it. We’ve seen this movie before. Amazon developed Graviton CPUs and Inferentia AI chips. Google has its TPU program (which, ironically, Broadcom co-designs). Tesla does Dojo. When you’re spending tens of billions on compute, the incentive to vertically integrate is real.
But here’s the counterpoint Tan was making, and I think he’s right: designing bleeding-edge silicon at scale is insanely hard, capital intensive, and the goalposts move every 18 months. Even if a hyperscaler wanted to go it alone, they’d be playing catch-up to Nvidia’s GPU road map on one side and Broadcom’s partnership-driven XPU road map on the other.
Most importantly, the order flow says customers are voting with their wallets — and right now they’re voting Broadcom, in $10–20 billion increments.
Segment Breakdown: Where the Growth Is Coming From
Let’s run through the actual numbers because they’re worth lingering on.
- Semiconductor Solutions (the big one): $11.07 billion, +34.5% yoy
- AI portion: $6.5 billion, +74% yoy
- Non-AI portion: $4.6 billion, +2% yoy (wireless actually strong thanks to iPhone 17 ramp)
- Infrastructure Software (mostly VMware): $6.9 billion, +19% yoy
- Bookings strength continues — total contract value booked hit $10.4 billion in Q4
The software side still has some integration work left, but the renewal rates and new bookings trends look solid. More importantly, the margin profile of that business (high-80s gross margins) is helping fund the overall margin expansion story.
Perhaps my favorite chart from the deck was the AI backlog conversion timeline. Management expects to recognize roughly $8.2 billion of the current $73 billion AI backlog in the current quarter alone. That’s real, near-term revenue, not some 2030 pipe dream.
So… Buy the Dip or Wait?
Here’s my take, and I’ve been following Broadcom for years: this drop last night was 90% tax-loss harvesting and profit-taking in a name that was up 75% YTD and trading at all-time highs in mid-December. Add in a nuanced CEO answer on a theoretical long-term risk, and you get exactly the kind of short-term overreaction growth investors dream about.
The fundamentals didn’t change in twenty minutes. The $73 billion AI backlog didn’t evaporate. Anthropic didn’t cancel their $21 billion in combined orders. Nvidia didn’t suddenly stop evolving GPUs, making custom silicon obsolete overnight.
If anything, the guide for AI revenue to double again next quarter reinforces that demand is accelerating, not topping.
Valuation-wise, Broadcom trades around 34× forward earnings today. Expensive? Sure. But for a company growing revenue 30%+, expanding margins, and sitting on multi-decade AI tailwinds with a fortress balance sheet and a growing dividend, it’s hardly nosebleed territory compared to some other “AI darlings” I could name.
I’m not pounding the table for new money at the absolute high, but if this weakness sticks around for a few sessions, I’ll be looking very closely. In my experience, the best time to add to best-in-class compounders is exactly when the crowd fixates on a risk five years out and ignores the rocket ship taking off right now.
Sometimes the market hands you a gift wrapped in fear. Last night felt like one of those nights.
We’ll see if today’s regular session brings more selling or if calmer heads prevail. Either way, the long-term story looks stronger than ever — and that’s what ultimately matters.
