Networking for AI

The question

Is the networking content per AI cluster -- switches, optics, DSPs, and retimers -- going to grow faster than the accelerator count, and are the vendors selling into that layer priced for steady-state enterprise demand rather than AI scale-up?

Base rates

The reference class is component layers inside a computing platform shift. In the PC cycle, memory content per unit rose roughly 3x faster than unit volume for a decade. In the smartphone cycle, RF content per handset rose roughly 4x faster than handset volume between 2010 and 2016. The base rate for a bandwidth-adjacent layer compounding well through a compute cycle is the 70th-80th percentile of tech subsectors. The base rate for optical component vendors specifically is less flattering -- closer to the 45th percentile -- because historical telecom demand is lumpy and inventory-driven. The question is whether the AI data-center mix shifts the reference class from telecom cyclicals to hyperscaler royalty.

Consensus forward numbers currently imply networking revenue per AI dollar of roughly 8-10%, consistent with pre-AI data-center builds. Hyperscaler technical disclosures on recent clusters imply the figure is closer to 15-20% and rising as cluster size grows. That is the gap we are pricing.

Why the consensus view is wrong (or incomplete)

The sell-side models AI networking as a linear function of accelerator count. It is not. Training bandwidth requirements scale approximately with model parameter count and with the square of cluster size for all-to-all collective operations. As clusters grow from 10k to 100k accelerators, the network fabric's share of the bill of materials does not stay constant -- it rises, because the fabric has to carry more per-accelerator bandwidth AND more hops. The causal mechanism is geometric: an accelerator count going up 10x requires switch port count going up 10x AND per-port bandwidth going up roughly 2x per generation. That produces a compounding effect the linear model misses.

The second mechanism is the shift from copper to optical inside the rack. As per-port speeds cross 800G and head toward 1.6T, copper reach collapses; optical transceivers and linear-drive interconnect become mandatory rather than optional. That is a content-per-box step-up, not a volume story.

Position construction

Two 20% anchors: ANET for the merchant-silicon Ethernet switch franchise and MRVL for the custom DSP and electro-optics IP that sits inside nearly every 800G and 1.6T transceiver. CIEN at ~18% captures the DCI (data-center interconnect) layer that links geographically distributed clusters -- a segment whose growth rate is bound to cluster federation, not enterprise refresh. FN (~11%) is the contract optical manufacturer with a capital-efficient model and a long-tenured NVDA transceiver relationship; it converts the volume story into a gross-margin-stable earnings stream.

LITE (~9%) and MTSI (~8%) are the photonic component and III-V laser franchises -- narrower moats but pure bandwidth-generation exposure. QRVO (~6.5%) is a hedge: mixed RF and infrastructure, with a lower AI beta but a cheaper entry multiple. VIAV, CALX, and EXTR (~7.2% combined) are smaller tail positions covering test/measurement, access, and enterprise campus switching that benefit secondarily from the buildout. The tail is kept small because the AI sensitivity is lower and valuation is not as asymmetric.

Asymmetric payoff

If networking content per AI dollar expands from ~10% to ~15% over three years while AI capex grows 15% annually, weighted book revenue compounds roughly 25-30%. With flat multiples the return is 20-28% annualized. If the networking-share thesis fails and revenue tracks accelerator count linearly, revenue growth drops to ~12% and multiples compress; the bear case returns -20 to -30% cumulatively. The right tail -- 1.6T transitions pull in by a year and optical attach rates surprise -- is plus 60-80% on a 2-year horizon.

At a 60% base case, 25% bear, 15% bull, expected value is roughly +14 to +18% annualized. The payoff is asymmetric because the bear requires a specific physical claim (copper reach does not collapse) that is already being falsified in lab disclosures.

Three things that would change our mind

1. A credible co-packaged optics transition inside the accelerator package that collapses external transceiver demand by more than 30% within 18 months.
2. Arista or Marvell reporting two consecutive quarters of AI-segment revenue growth below 15% year-over-year, indicating content-per-cluster is not expanding.
3. Hyperscaler disclosures indicating a decisive move to proprietary scale-up fabrics (UALink-equivalent captive stacks) that displace merchant Ethernet in training clusters.

What we are explicitly NOT betting on

We are not buying accelerator vendors or foundry capacity -- those live in the picks-and-shovels book and have different risk factors. We are also not buying the legacy telecom carrier equipment complex; the historical base rate for carrier capex is poor and the AI signal there is second-derivative. The networking thesis requires only that per-cluster bandwidth keeps outrunning per-cluster accelerator count. That is a narrower, more testable claim, and the book is sized accordingly.