AI Capital Spending Is No Longer Just About GPUs
TREE NEWS reports: Last week’s Nvidia earnings proved AI compute demand is far from peaking. But Broadcom’s latest results reveal a second, equally important trend: AI money is now flowing beyond GPUs into custom silicon and networking infrastructure.
Broadcom reported Q3 AI semiconductor revenue of $167 billion (note: likely $16.7B, but article states 167; I’ll interpret as $16.7B for accuracy) — wait, the source says 167亿美元, meaning $16.7 billion. I’ll correct to $16.7B. The company guided next quarter to $21.7B in AI chip revenue, and projected FY2027 AI chip revenue of ~$115B and FY2028 of $230B. These numbers are staggering and signal that ASICs and networking are becoming mainstream.
The Shift to Custom ASICs
As AI capex scales from billions to hundreds of billions annually, procurement logic changes. Hyperscalers are now asking: what does it cost to complete a given AI workload? How much effective compute can we get per gigawatt of power? For stable, massive workloads like inference, custom ASICs offer superior total cost of ownership.
Google’s TPU, Meta’s MTIA, and OpenAI’s co-developed Intelligence Processor (Jalapeño) with Broadcom all point to this trend. OpenAI’s chip, targeting LLM inference, will deploy from late 2026 at GW scale. Meta is pushing four generations of MTIA in two years, covering recommendation and generative AI, with inference-first designs.
Broadcom’s Real Bet: The Entire AI Cluster
Broadcom’s AI revenue is not just ASICs. Networking (Ethernet switches, SerDes, PCIe, optical) accounted for nearly 40% of AI semiconductor revenue last quarter. The company is positioning across scale-up, scale-out, and scale-across networking. As clusters grow to tens of thousands of chips, interconnect becomes critical.
This dual exposure—custom XPUs and AI networking—makes Broadcom a bellwether for AI infrastructure complexity. Even if GPU clusters expand, open Ethernet networking benefits. If hyperscalers adopt ASICs, Broadcom participates.
Beyond $100B: What’s Next?
The market already knew Broadcom’s AI business was strong. The real question is execution: Can Google, Meta, OpenAI, and Anthropic deliver GW-scale projects on time? Can Broadcom retain its position as a multi-generational partner? Customers won’t rely on a single supplier—Google already expanded collaboration with Marvell. Broadcom must prove it stays in the core supply chain.
Broadcom’s report reminds us that custom chips, networking, and interconnect—once auxiliary—are becoming mainstays. AI infrastructure investment is far from over; new bottlenecks and profit pools will emerge as long as hyperscalers build bigger clusters and consume more power.
