Broadcom's Custom AI Chips Challenge AMD's GPU Growth Story

Competing AI Strategies Collide

AMD and Broadcom both delivered robust AI-driven quarters, but their business models put them on a collision course. AMD sells general-purpose Instinct GPUs and EPYC CPUs to any datacenter buyer, while Broadcom co-designs custom accelerators exclusively for a handful of hyperscale customers. That difference creates a structural challenge for AMD's growth thesis, according to analysis first reported by 24/7 Wall St.

AMD's Q2 fiscal 2026 revenue reached $11.54 billion, up 50.1% year over year, with its Data Center segment hitting $6.72 billion and growing 107%. CEO Lisa Su indicated the Data Center segment is expected to more than double again in fiscal 2027, with Helios rack shipments beginning later in Q3 2026.

Broadcom operates at a different scale. The company's Q3 fiscal 2026 revenue totaled $29.59 billion, up 85.5% year over year. AI semiconductor revenue alone accounted for $16.70 billion, growing 221% year over year and 54% sequentially. CEO Hock Tan guided Q4 AI revenue to $21.7 billion and outlined a fiscal 2027 AI outlook approaching $115 billion, with supply already secured. Broadcom's Q3 operating margin of 67.9% significantly exceeds AMD's 27% non-GAAP operating margin.

The Customer Overlap Problem

The strategic tension centers on shared customers. OpenAI, Meta, and Anthropic—three of AMD's marquee AI accounts—are simultaneously building custom accelerators with Broadcom. Tan has argued that a co-developed XPU "will outperform any GPU" and can operate at "half the cost of a GPU."

If these frontier AI labs increasingly shift training and inference workloads to their own custom silicon, AMD's addressable market at those accounts contracts even as it ships MI450 GPUs. AMD's product roadmap includes Instinct MI350 and MI450 chips plus Helios rack systems, while Broadcom delivers custom XPUs and Tomahawk Ethernet fabric.

Both companies face distinct risks. AMD must navigate export controls, memory supply constraints, and the maturity of its ROCm software ecosystem. Broadcom contends with customer concentration and significant debt levels.

Why It Matters

The divergence between merchant and custom silicon strategies will determine how AI infrastructure spending flows over the next several years. If hyperscalers conclude that custom accelerators deliver better performance per dollar for their specific workloads, the general-purpose GPU market AMD is targeting could shrink faster than the overall AI chip market grows. Broadcom's $13.66 billion free cash flow in Q3 and pre-sold 2028 capacity suggest customers are already making that bet at scale.

AMD shares have climbed 195.18% over one year and 135.63% over six months, while Broadcom has declined 14.44% in the past month. The 2027 ramp will reveal whether AMD's estimated $1.4 trillion AI accelerator total addressable market by 2030 holds once Broadcom is delivering tens of billions in custom chips annually to Google and other hyperscalers.

These details were first reported by 24/7 Wall St.