Broadcom: Riding The XPU Wave

Broadcom's XPU shipments rose approximately 3.5x year over year and accounted for nearly 75% of AI revenue in Q3 FY26. Its co-designed XPUs, including Google's TPU v8i and OpenAI's Jalapeno, are positioned as lower-cost alternatives to traditional GPUs for inference. Broadcom is also pursuing AI and private-cloud demand through its AI XPV Platform and VMware AI Factory, alongside partnerships including Apollo, Blackstone, OpenAI, and Anthropic.

Analysis

The useful signal is not shipment growth by itself but the possibility that custom silicon shifts inference economics: if lower-cost-per-inference XPUs expand deployment, total compute demand could rise even as some GPU workloads are displaced. That supports Broadcom’s design-and-networking position, but shipment counts do not establish revenue, gross profit, or durable customer economics; verify AI revenue conversion, backlog, and customer concentration in reported results.

The key second-order risk is buyer power. Google’s in-house TPU roadmap gives it an alternative and negotiating leverage; large model developers may likewise seek multiple silicon sources. Broadcom can win design activity while still facing concentrated programs, pricing pressure, and lumpy ramps. Nvidia is exposed to inference substitution, but remains a beneficiary if total AI infrastructure spending expands or GPU-heavy training remains dominant. Advanced foundry and packaging capacity could benefit if custom-chip volumes scale, though capacity and allocation data are needed to underwrite that spillover.

Days: optimistic positioning may already price in the shipment narrative, so avoid chasing. Over 1–3 months, focus on reported AI revenue, backlog/conversion, and customer or platform disclosures. Over 6–18 months, the structural question is whether XPUs broaden beyond a few large programs and sustain attractive economics. The thesis weakens if Broadcom’s AI revenue or forward commentary fails to convert shipment momentum, or if Nvidia’s results show inference demand expanding without meaningful GPU substitution.

Key Decisions for Investors

• Prefer a staged, relative-value expression over an outright momentum buy: consider a modest long Broadcom / short Nvidia position after confirming the next earnings update supports AI revenue conversion. Keep it beta-aware; both names remain exposed to AI-capex reversals.
• Do not underwrite the claimed XPU cost advantage as Broadcom margin expansion until reported segment economics or customer disclosures support it. Track AI revenue, backlog, design-win concentration, and conversion timing.
• Treat Google as both a potential ecosystem beneficiary and a bargaining-power risk to Broadcom; reassess if Google signals greater internal silicon sourcing or a change in TPU deployment plans.
• Falsification: reduce or exit the relative-value thesis if Broadcom’s AI revenue/backlog trajectory disappoints while Nvidia reports sustained broad-based inference demand, or if custom-chip ramps are delayed.