AI Buildout Signals Shift Beyond GPU: Networking Gear, Optics, Inference Scaling

Cisco's latest results delivered an unexpected wake-up call to the market: AI capex buildout is no longer a GPU monopoly story. While NVIDIA dominates training chips, the networking layer, switches, optics, and routing, is becoming the next bottleneck. Cisco's strength in AI-adjacent networking gear signals that hyperscalers are scaling inference infrastructure, not just training clusters. This shift has downstream implications for entire supply chains.

Meta's $21 billion multi-year agreement with CoreWeave, announced this week, crystallizes the trend. CoreWeave is a specialist in inference capacity, not training. Meta's willingness to lock in long-term inference contracts suggests that training is becoming a solved problem; the real capital intensity now lies in sustained inference workloads. Microsoft, similarly, is deploying billions into AI infrastructure, but analyst debate centers on how quickly that capex converts to revenue and margin expansion. Most of the company's AI spending is still being capitalized on the balance sheet rather than expensed, masking profitability questions.

For NVIDIA, this creates a mixed picture. Training GPU demand remains strong enough to support consensus price targets, but the narrative is shifting from 'AI capex is endless' to 'AI capex is diversifying.' Networking, memory (Broadcom, Micron), and power management are gaining relative strength. Equities broadly are grinding higher on this broadening AI theme, but breadth is tightening. The top 10 mega-cap names (the so-called Magnificent 7 plus Tesla and Broadcom) are doing the heavy lifting; mid-cap and small-cap participation lags.

Skeptics worry that Meta and Microsoft's capex commitments reflect desperation to secure inference supply during a supply crunch, not evidence of durable demand. If inference capacity becomes abundant over 18 months, utilization rates could compress margins. Additionally, the shift from training to inference may actually reduce total dollar capex if inference workloads are less compute-intensive per dollar of economic value.