AI Infrastructure Rally Expands: Cisco, Broadcom Lead as Memory Constraints Drive Networking Capex

The dominant narrative in tech earnings this week has shifted from GPU concentration to infrastructure breadth. Within two days of each other in late April, the CEOs of Microsoft, Meta, Google, Amazon and Apple all flagged the same critical bottleneck: memory constraints that are not temporary. This observation has triggered a re-evaluation of the entire AI capex stack, moving investor focus upstream and downstream from Nvidia's dominance.

Cisco delivered the most concrete signal. The networking giant highlighted that AI buildout is widening into switches, optics and scale-across networking infrastructure, exactly the layers that enable training clusters at hyperscale. Broadcom shares rallied on the view that co-packaging demand and custom silicon requirements will sustain elevated capex for years, not just through the current GPU cycle. Meanwhile, Micron remains priced at 7x earnings despite memory being explicitly called out as the binding constraint by five of the largest tech companies in the world.

Meta's $21 billion deal with CoreWeave underscores this shift. The expanded partnership reveals how quickly AI infrastructure demand is scaling beyond model training into long-term inference capacity. Meta, Google and Amazon are building redundancy and geographic spread, not betting on a single vendor. This architectural shift favours diversified semiconductor suppliers, networking hardware specialists and infrastructure-as-a-service players over concentrated GPU makers.

The risk to the narrative is execution. While memory constraints are real, whether that translates to Micron rerating depends on demand stickiness beyond 2026. If AI capex moderates, Micron at 7x earnings could compress back to historical multiples. Additionally, some argue that custom silicon and architectural shifts (e.g. Google's claimed 6x memory efficiency gains with TurboQuant) could negate traditional memory demand. Investors are hedging by rotating into AVGO and semiconductor systems integrators (SMCI) rather than going all-in on memory plays.