Broadcom's Custom AI Chips Are Quietly Squeezing AMD's Growth Story

AMD just posted one of its strongest quarters ever, with data center revenue more than doubling. But a rival chipmaker growing even faster on the exact same customer list is starting to raise real questions about how much of the AI chip market AMD can actually keep.

AMD's fiscal second-quarter revenue hit $11.54 billion, up 50.1% year-over-year, with data center revenue at $6.72 billion, up 107%, according to 24/7 Wall St's breakdown of the quarter. CEO Lisa Su said data center revenue more than doubled year-over-year and expects the segment to more than double again in 2027.

That's a genuinely strong number on its own. The problem is Broadcom's numbers from the same stretch are on an entirely different scale. Broadcom's fiscal third-quarter revenue reached $29.59 billion, up 85.5% year-over-year, with AI semiconductor revenue alone hitting $16.70 billion, up 221% year-over-year and 54% sequentially.

Broadcom's AI semiconductor business had already jumped 79% year-over-year to $15 billion the quarter before that, with AI-specific revenue up 143% to $10.8 billion, and AI bookings topping $30 billion, comfortably ahead of shipments, per Yahoo Finance's earlier coverage of the segment. Demand visibility, in other words, looks locked in well past the next couple of quarters.

"Data center revenue more than doubled year-over-year." — Lisa Su, Chair and CEO, Advanced Micro Devices

Here's the uncomfortable overlap for AMD investors: OpenAI, Anthropic and Meta are all AMD customers, and all three are simultaneously co-developing custom AI accelerators with Broadcom. CEO Hock Tan has argued that a co-designed chip can outperform any general-purpose GPU and run at roughly half the cost, a pitch that's clearly landing with hyperscalers.

AMD sells merchant GPUs and CPUs to any data center willing to buy them, which is a scalable model but not a sticky one. Broadcom instead embeds itself inside a handful of hyperscalers by co-designing custom silicon, which is harder to dislodge once it's built into a customer's roadmap. Broadcom's operating margin of 67.9% last quarter versus AMD's 27% non-GAAP margin reflects just how different those two business models really are.

Broadcom also guided fourth-quarter AI revenue to $21.7 billion, up 236% year-over-year, and sketched a fiscal 2027 AI outlook near $115 billion with supply already secured, according to 24/7 Wall St. That's the kind of forward visibility that makes Wall Street comfortable paying a premium multiple, and it's a big reason Broadcom's stock has held up even through a broader semiconductor pullback.

None of this means AMD's growth story is broken. Helios rack shipments are just starting in the current quarter, and AMD's Instinct MI350 and MI450 lines are still winning real deployments across Microsoft, OpenAI and Meta. The company's addressable market is expanding fast enough that both AMD and Broadcom can grow briskly without directly colliding everywhere.

The risk for AMD longer-term is concentration. If Broadcom's custom XPUs keep taking a bigger share of AI compute budgets at OpenAI, Anthropic and Meta specifically, that caps how much of the AI buildout AMD can capture even as the overall market keeps expanding.

For investors watching both names, AMD still looks like the more straightforward way to play broad AI GPU demand, especially with Helios shipments ramping and a cheaper relative valuation than Nvidia. Broadcom is the better pick for investors who want exposure to the custom-silicon trend and don't mind paying up for a business with structurally higher margins and 15 consecutive years of dividend increases.

Watch AMD's next earnings call closely for any commentary on Helios adoption rates and whether hyperscaler capex commitments are shifting toward custom silicon. That's the tell that would confirm whether Broadcom's structural advantage is actually taking share, or just growing alongside AMD in a market big enough for both.

Bottom line: AMD's growth is real, but Broadcom's custom-chip model is proving harder to compete against than the headline revenue numbers alone suggest.