Share with your CTO
Broadcom is making a credible case that custom silicon for hyperscalers is the most valuable seat at the AI infrastructure table. After a broad chip selloff pushed the PHLX Semiconductor Index into bear-market territory, a head-to-head comparison of AMD and Broadcom surfaces a telling gap: Broadcom guided for 200%-plus year-over-year AI chip revenue growth in Q3, reaching $16 billion, while generating free cash flow of $10.3 billion in a single quarter. AMD’s data center revenue grew 57% in Q1, with Q2 guidance implying acceleration, but at roughly twice Broadcom’s earnings multiple.
What this means for your business
The chip layer your infrastructure bets sit on is not a neutral commodity decision anymore. If your organization is deep into hyperscaler-hosted AI workloads on AWS, Google Cloud, or Microsoft Azure, Broadcom’s custom accelerator dominance is already embedded in your stack whether your team has noticed or not. AMD’s EPYC server processors and Instinct AI accelerators matter most if you’re running on-premises GPU clusters or actively evaluating alternatives to Nvidia for model training and inference inside your own data centers.
The structural story here is the divergence between two AI chip strategies that used to look roughly equivalent. Broadcom builds custom AI accelerators (chips designed to a specific hyperscaler’s exact specifications, rather than general-purpose GPUs sold to anyone) for a concentrated set of cloud giants, and that concentration is currently producing explosive economics. AMD competes in the merchant GPU market, selling general-purpose accelerators to a broader customer base. The merchant model offers wider distribution but slower adoption curves, and AMD’s MI-series chips are still in ramp mode against Nvidia’s entrenched CUDA software ecosystem, the developer toolchain that makes switching painful.
The valuation gap between the two stocks at roughly 32x versus 67x forward earnings is actually a signal worth reading from an enterprise planning perspective. Markets are pricing AMD as though the MI450 ramp will capture material share from Nvidia over the next two to three years. If your procurement team is evaluating AMD accelerators partly on the assumption that the vendor will keep investing aggressively in the platform, that assumption looks reasonably well-funded. But the scenario is not yet delivered, and Broadcom’s 46% free-cash-flow margin suggests a business already through the hardest part of proving its model. I’d revise this view if AMD’s MI450 customer pipeline converts at the rate Lisa Su’s commentary implies, but right now the momentum is Broadcom’s.
Concept deep-dive: Custom silicon vs. merchant silicon
Custom silicon means a chip designed from scratch for one buyer’s specific workload, the way a bespoke suit fits one person. Merchant silicon is the off-the-rack alternative: a general-purpose chip sold competitively to many buyers. Custom silicon yields better performance-per-watt for the specific task and locks in a supplier relationship, but it requires the buyer to commit enormous upfront engineering resources. The custom model is why Broadcom’s revenue is so concentrated in a few hyperscalers and why its growth can be so extreme in a single quarter when those customers scale up.
Based on reporting from AMD vs. Broadcom: The Better AI-Chip Stock to Buy After the Sell-Off, originally published 2026-07-20 05:22:00.
