Perplexity's Agentic AI Revenue Surge Signals Inference Infra Bottleneck — Not Nvidia

If Nvidia's next-generation Blackwell architecture dramatically reduces the memory-bandwidth constraint per inference call (through on-chip memory expansion or architectural efficiency gains), the bottleneck shifts back to raw compute and the thesis fails. The observable signal would be Nvidia announcing HBM capacity per GPU chip significantly above current roadmap specs, or hyperscalers publicly stating they are pausing networking upgrades because GPU utilization is the binding limit.

The strongest bear case is that Nvidia's vertical integration strategy — investing directly in Perplexity — is specifically designed to lock in GPU demand by subsidizing a customer, which means the GPU itself remains the scarce and monetizable asset, not the surrounding infrastructure. If Nvidia successfully makes GPUs the irreplaceable bottleneck through software lock-in (CUDA ecosystem, NIM microservices), then commodity networking and memory suppliers face margin compression as hyperscalers play them against each other. The specific assumption most likely to be wrong: that memory bandwidth is the binding constraint at current agentic workload scales — if Perplexity's revenue growth is driven by breadth of users rather than depth of agentic chains, the inference profile may still be closer to chatbot than true multi-step agentic, and the HBM/networking bottleneck thesis is premature. The falsifying data point: if Micron or the networking names report flat AI-specific revenue in their next earnings while Nvidia's data-center segment accelerates, the bottleneck is confirmed to still sit at the GPU.

The 2000 fiber-optic buildout offers a precise analogy: Cisco and Lucent were the 'picks and shovels' of internet infrastructure, but the real bottleneck — and the biggest sustained winner — turned out to be the optical component suppliers (Corning, JDS Uniphase) whose glass and lasers were the true scarce input. The GPU-to-networking parallel is tight: just as bandwidth, not servers, was the true constraint in web-scale computing, memory bandwidth and interconnect, not raw FLOPS, are the true constraint in agentic inference. Corning's stock tripled from 1998-2000 on fiber demand before the bubble, while Cisco peaked first. The lesson: follow the actual physical bottleneck, not the branded hardware. Today, HBM3e supply is physically constrained (SK Hynix and Micron are the only qualified suppliers), and AI-cluster Ethernet switching is a two-player market. These are the 'Corning' equivalents of the current cycle.