Custom AI Chips, Not Merchant GPUs, Seen as Nvidia’s Real Competitive Battleground

Debate over Nvidia’s biggest challenger in AI hardware often defaults to Advanced Micro Devices, whose Instinct accelerators and EPYC processors have made it the most visible merchant GPU alternative. But a growing argument suggests the more consequential competition comes from a different direction: custom silicon designed for hyperscale customers.

At the center of that argument is Broadcom, which develops application-specific AI chips, including XPUs built in partnership with large cloud operators. Rather than selling standardized GPUs into an open market, Broadcom’s custom-accelerator model ties its AI revenue to multi-year programs with a handful of very large buyers — precisely the customers with the scale and motivation to reduce dependence on Nvidia’s product roadmap and pricing.

The market backdrop underscores the scale gap that still separates the two. Nvidia’s market capitalization stands at roughly $5.56 trillion, compared with about $1.99 trillion for Broadcom and approximately $779.6 billion for AMD. Nvidia remains the dominant supplier of data center AI accelerators, while Broadcom’s Semiconductor Solutions segment spans networking connectivity alongside its custom chip business — a positioning that covers both the accelerators and the switching fabric that links them in large AI clusters.

Shares reflected a mixed session. Broadcom traded up 1.35% at $364.30, and AMD gained 2.49% to $514.51, while Nvidia slipped 0.73% to $223.67.

The custom-silicon thesis rests on a simple dynamic: hyperscalers collectively spend enormous sums on AI infrastructure, and even a partial shift of accelerator purchases from merchant GPUs to in-house designs co-developed with Broadcom would meaningfully alter the competitive mix. Broadcom has also flagged accelerating growth expectations for its AI semiconductor business in recent quarters, which is fueling attention on the category.

That said, the two approaches are not strictly substitutes. Merchant GPUs serve a broad customer base of enterprises, governments, and AI startups that lack the volume to justify custom chip development, while custom XPUs are viable mainly for a small set of giants with predictable, massive workloads. Nvidia, for its part, has been building out networking — including its NVLink and Spectrum-X lines — that competes directly with Broadcom’s connectivity franchise.

What to watch

• Broadcom’s upcoming quarterly results and any updated guidance on AI custom-chip revenue and its pipeline of hyperscaler XPU programs.
• Nvidia’s next earnings report and commentary on demand for its networking portfolio alongside GPU sales.
• Announcements of new or expanded custom accelerator deals between Broadcom and cloud providers.
• AMD’s data center GPU momentum, including customer wins and roadmap updates for its Instinct line.

Source: original release