AVGO Q3 2026 Earnings Analysis
Broadcom delivered record Q3 revenue of $29.6B (+86% YoY), led by $16.7B AI semiconductor revenue (+221% YoY); Q4 guidance: $34.8B (+93% YoY) with FY2027-2028 AI guidance of $115B and $230B.
Key Metrics
Points clés
- AI semiconductor revenue surged to $16.7B, up 221% YoY, with XPU shipments up 3.5x YoY representing 73% of AI revenue.
- Broadcom signed long-term supply agreement with Google and guides FY2027 AI revenue to $115B, doubling from FY2026.
- Q4 guidance of $34.8B revenue (+93% YoY) with $21.7B AI revenue (+236% YoY) and 66% operating margin.
Écouter sur
Disponible en
Transcript
// Full episode scriptBefore we get into it, quick reminder: this podcast is AI-generated content for educational and entertainment purposes only. Nothing we discuss should be considered investment advice. Always do your own research and consult a qualified financial advisor before making any investment decisions.
And it's a good quarter to have that disclaimer front and center, because Hock Tan basically stood up and said "we're just getting started," and then backed it up with a $230 billion revenue forecast for 2028.
Let's start with the headline numbers. Q3 revenue came in at $29.6 billion, up 86% year-on-year. Operating income hit $20.1 billion, up 92%, with a record 68% operating margin. And free cash flow? $13.7 billion, 46% of revenue.
Those margins are the story for me. Gross margin actually dipped slightly, down to about 75%, because AI chips carry more memory content and lower margins than the rest of the business. But operating margin still climbed because revenue is growing so much faster than expenses. That's operating leverage doing exactly what it's supposed to do.
Right, and the AI piece specifically—AI semiconductor revenue was $16.7 billion for the quarter, more than tripling year-over-year. That's now 56% of total revenue, up from 49% just last quarter.
This company has essentially transformed into an AI infrastructure company that happens to also sell broadband chips and VMware software.
Speaking of which, let's talk customers, because this is where it gets interesting. Broadcom has six custom AI chip customers, but Hock Tan really zeroed in on four: Google, Anthropic, OpenAI, and Meta.
The Google relationship is the elder statesman here—a decade of TPU development, and they just signed a long-term deal for "multi-tens of billions of dollars" of TPUs annually. They shipped the new Ironwood TPU v7 this quarter and are already ramping the next-gen v8i.
But the real headline is Anthropic. Tan said Anthropic is on track to become Broadcom's largest XPU customer in 2027, deploying 5 gigawatts of TPU v8i next year and then 10 more gigawatts in 2028.
And OpenAI isn't far behind—their custom chip, nicknamed Jalapeño, is already reportedly outperforming Nvidia's Grace Blackwell in inference workloads at, according to Tan, less than half the cost. OpenAI's on pace for over 5 gigawatts by 2028, making them Broadcom's second-largest customer.
That "half the cost" line is the thesis of the whole call, honestly. Custom silicon, co-designed for a specific model's workload, beats a general-purpose GPU on performance and cost. That's Broadcom's pitch to the market.
Now let's talk about the number that got everyone's attention: guidance. Q4 AI revenue guided to $21.7 billion, up 236% year-on-year. Full fiscal 2026 AI revenue now expected at $58 billion.
And then they went further than usual—giving multi-year guidance. AI revenue is expected to double to $115 billion in fiscal 2027, then double again to $230 billion in fiscal 2028.
That's an unusually bold move for a company to lock in two years out. And CFO Amie Thuener was clear they don't plan to update it quarterly—so this is a stake in the ground, not a rolling estimate.
On the Q&A, analysts pushed hard on whether that's really achievable. Stacy Rasgon from Bernstein did some math on gigawatts versus dollars, and Tan clarified something important: not all the gigawatt capacity they've outlined will necessarily be "deployed" in that window—some of it depends on data centers, power, and shells being physically ready.
That land-power-shell constraint came up again and again in the Q&A. Multiple analysts asked about bottlenecks, and Tan was refreshingly candid that construction timelines, not just chip supply, could gate how fast this all plays out.
There was also a really interesting exchange about financing. Broadcom set up something called the "AI XPV Platform" with Apollo and Blackstone—essentially a vehicle to help Anthropic and OpenAI, as two labs still burning cash, fund massive infrastructure buildouts. They've already closed a $35 billion tranche for Anthropic's first gigawatt.
Tan's framing was memorable—he compared it to helping "two geniuses in Outer Mongolia get to college." Broadcom isn't directly lending the money; they're bringing in third-party financial partners and offering modest residual value guarantees, which they characterize as low-risk given how profitable they expect these labs to become.
That's worth flagging for listeners—it's a financially creative structure, and while management framed the risk as low, it's also new enough that it's worth watching how it evolves, especially since one analyst pointed out the first tranche already carried up to $29 billion in potential exposure.
On the non-AI side, things were much quieter—non-AI semiconductor revenue was $4.2 billion, up just 5%, and infrastructure software, largely VMware, was $8.8 billion, up 29%, with management expecting that to roughly stabilize going forward.
They did highlight a new product, VMware Private AI Cloud, aimed at enterprises that want to run AI workloads on-premises rather than in the public cloud—an interesting angle on how the AI boom is trickling into their more traditional software business.
So what does this all mean going forward? The bull case here is extremely clear: Broadcom has essentially locked itself into becoming the custom silicon supplier for the handful of companies building frontier AI models, and management is projecting that business to nearly quadruple over the next two years.
The risk case is just as clear, though—concentration. Six customers, with two of them, Anthropic and OpenAI, still not profitable enterprises in the traditional sense, and heavily dependent on continued fundraising and infrastructure buildout actually happening on schedule.
Before we sign off, one more required note: everything discussed is AI-generated analysis for educational purposes. Past performance doesn't guarantee future results. Please do your own due diligence.
Broadcom's next earnings call is scheduled for December 9th, when we'll get the official Q4 close-out and hopefully some more color on how that XPV financing platform is evolving.
Should be a fascinating one to watch. Until then, this has been Beta Finch.
Thanks for listening, and we'll catch you next time.