Broadcom's AI revenue is growing at 221%, a figure The Globe and Mail highlighted this week while making a case that custom chips could become a bigger business than GPUs. That's the headline number. And it lands at a moment when hyperscalers are quietly rerouting billions of dollars away from merchant graphics processors and toward silicon designed for their own workloads.

The Globe and Mail piece frames the argument in its title: "Broadcom's AI Revenue Is Growing at 221%. Here's Why Custom Chips Could Be a Bigger Business Than GPUs." The claim rests on a structural shift. Cloud operators that once bought Nvidia's GPUs by the rack now co-develop application-specific accelerators with Broadcom's chip design teams, then pay for the silicon, the packaging and the networking that surrounds it.

Why 221% matters more than the absolute number

Growth rates flatter small bases. But 221% growth in AI revenue is the kind of trajectory that turns a side business into a pillar. The Globe and Mail report positions Broadcom's custom accelerator franchise — the design work it does for major cloud customers, paired with its Tomahawk and Jericho switching families — as the engine behind that expansion.

The economics are straightforward. A hyperscaler that commissions its own accelerator owns the architecture, tunes it to its own models, and escapes the margin its GPU supplier would otherwise capture. Broadcom collects design revenue plus per-unit silicon. Both sides win, and the GPU vendor loses share at the largest accounts.

What does the custom-chip argument actually claim?

The Globe and Mail's thesis, stated plainly: custom silicon could become a bigger business than GPUs. That's a bold framing. GPUs currently dominate AI training, and Nvidia's shipments remain the industry's reference volume. But the custom-chip camp has momentum on its side:

• 221% AI revenue growth at Broadcom, per the report's headline figure
• Deepening co-design relationships between Broadcom and its cloud customers
• Custom accelerators paired with high-margin networking silicon, which Broadcom also supplies
• A purchasing shift among hyperscalers toward parts optimized for their own workloads rather than merchant hardware

The report's logic is that AI inference, not training, will dominate future compute demand. Inference rewards purpose-built silicon. It penalizes general-purpose overhead. If inference volumes scale the way cloud operators expect, custom accelerators built for those specific workloads capture the majority of the value.

The competitive picture

Broadcom isn't arguing this alone. Its custom-silicon business competes with Marvell, with internal design efforts at Google and Amazon, and with the gravitational pull of Nvidia's CUDA software ecosystem. The GPU maker's moat remains real; developers know its tools, and its interconnect strategy ties clusters together in ways that are hard to replicate.

But the 221% figure says the moat is leaking at the top of the market. The biggest buyers — the companies spending the most on AI infrastructure — are exactly the customers with the scale and engineering depth to go custom. Broadcom sits on the profitable side of that trade.

For investors and supply-chain watchers, the piece reads as a re-rating argument: value the AI boom not only through GPU shipments but through the design-and-silicon pipeline that serves customers who will never buy a merchant GPU again. The Globe and Mail puts the growth number first for a reason.

If Broadcom's custom accelerator revenue keeps compounding anywhere near that reported pace, the question shifts from whether custom chips can rival GPUs to how quickly the hyperscaler transition already underway leaves general-purpose silicon with a shrinking slice of AI compute.