The AI Bottleneck Most Investors Are Missing but Shouldn't

The article argues that AI’s next major bottleneck is communication, not computation, creating a potential growth opportunity for optical networking and photonics companies. It highlights Broadcom, Marvell Technology, Lumentum, Coherent, and POET Technologies as potential beneficiaries of rising AI infrastructure spending. The piece is largely thematic and forward-looking rather than event-driven, so near-term market impact is likely modest.

Analysis

The market is still underwriting AI as a compute story, but the bigger marginal dollar may migrate to interconnect as cluster size scales nonlinearly. Once GPU count per pod rises, the hidden tax is not just capex; it is latency, power density, and the opportunity cost of chips sitting idle while data moves. That creates a second-order winner set in high-speed optical transport, where revenue can compound even if unit GPU demand normalizes, because every incremental training step requires more network dollar content per compute dollar.

This is most constructive for the picks-and-shovels names with the broadest attach points into AI racks and backplanes: AVGO and MRVL for platform exposure, and COHR/LITE for optical component leverage. POET is the highest beta expression, but also the least forgiving if design wins slip, because its equity value depends on a narrower commercialization window and a faster conversion of technical promise into socket revenue. The more interesting angle is that this trend can expand the AI supply chain beyond the current GPU duopoly, redistributing bargaining power toward networking silicon and optical subsystems over the next 12-24 months.

The main contrarian risk is that investors are front-running a solution before the bottleneck is fully binding. Hyperscalers may extract more life from existing electrical architectures through software optimization, topology changes, and incremental packaging improvements, which would delay the optical refresh cycle by quarters rather than years. If AI capex pauses or shifts from training to inference, the urgency for expensive optical upgrades could fade quickly, pressuring the smaller names first and leaving only the diversified incumbents with durable upside.

Near term, this is a relative-value trade more than a broad thematic long: the market should reward companies with immediate content per AI rack rather than pure-play concept names. Over a 6-18 month horizon, the asymmetric setup is to own the winners with real revenue torque and fade the names priced for a TAM expansion that may take multiple procurement cycles to realize.

Key Decisions for Investors

• Long AVGO vs. short a basket of legacy network-equipment proxies over 3-6 months; AVGO has the best blend of AI content, pricing power, and lower execution risk if optical demand broadens.
• Long MRVL on pullbacks for a 6-12 month trade; use a 15-20% stop because the upside is tied to hyperscaler deployment timing, but the stock should rerate if interconnect becomes a bigger share of AI capex.
• Speculative long COHR or LITE as a momentum/second-order AI infrastructure trade, but size at half-risk versus AVGO/MRVL; the reward is multiple expansion if optical content inflects, while downside is sharp if demand proves cyclical.
• Avoid initiating large longs in POET until there is clearer evidence of repeatable commercial wins; if trading it, use call spreads rather than equity to define risk given binary adoption risk.
• For portfolio hedge: pair long AVGO/MRVL against short NVDA as a relative beneficiary trade if AI spend rotates from compute to connectivity; this works best if data-center capex grows but GPU enthusiasm cools over the next 1-2 quarters.