The $725 Billion AI Capex Cycle Has 3 Bottlenecks: Power, Memory, and Optical Bandwidth. 3 Stocks Poised to Win Big.

The article argues that AI infrastructure spending could reach $765 billion this year, with the biggest opportunities in power, memory, and optical networking. It highlights GE Vernova’s gas turbines and $163 billion backlog, Micron’s sold-out HBM through 2027, and Marvell’s new 102.4 Tbps switch as key beneficiaries of AI data center buildout. The piece is broadly constructive on these names, but it is primarily an investment thesis rather than fresh company-specific news.

Analysis

The real signal is not that AI demand is broadening, but that it is becoming bottleneck-driven: power, memory, and interconnect are now the gating variables for incremental capex. That tends to shift economics away from the obvious compute beneficiaries and toward the “picks-and-shovels” names with the highest switching costs and longest backlog visibility. The second-order effect is that every incremental AI deployment raises the floor for non-labor industrial capacity demand, which should support a multi-quarter rerating in select infrastructure suppliers before it fully shows up in end-market earnings.

GE Vernova looks like the cleanest scarcity trade because the constraint is physical, not software-driven. The key is not just turbine demand, but the emergence of private power as a financing workaround for data centers that cannot wait on utility interconnect queues; that creates a long-duration order book with very little substitution risk in the near term. The risk is execution and cycle timing: if utility capex, permitting, or grid upgrades accelerate faster than expected, some of this captive generation demand could be deferred 12-24 months, but that is unlikely to eliminate it.

Micron’s setup is less about spot memory pricing and more about contract quality. Strategic multi-year agreements should reduce the classic memory boom/bust discount, but the market may be underestimating how quickly HBM supply can be the bottleneck to AI server ramp if OEMs keep increasing memory-per-chip intensity. Marvell sits in the most underappreciated lane: as models scale, network throughput becomes a tax on every inference cluster, and the winners will be vendors that can sell full systems rather than component parts.

The consensus may be too complacent on the durability of these trends. If AI capex moderates, the highest-beta beneficiaries can de-rate quickly because current expectations already embed years of growth; if capex reaccelerates, the upside is still asymmetric because supply constraints are tangible and slow to clear. The best contrarian expression is not chasing the most obvious AI leaders, but owning the supply-constrained enablers while fading names whose upside depends on a smoother data-center buildout than reality allows.

Key Decisions for Investors

• Long GEV on a 3-6 month horizon; use pullbacks to build. Risk/reward is favorable because backlog and private-power demand can re-rate the multiple before revenue fully catches up.
• Buy MU on weakness rather than strength; treat it as a 6-12 month hold. The thesis improves if HBM sell-through remains sold-out into 2027, but position sizing should reflect memory-cycle volatility.
• Go long MRVL vs short AVGO as a tactical pair for 3-9 months. MRVL has more torque to optical networking growth and system-level content expansion, while AVGO is more crowded and better owned.
• Consider GEV/MU/MRVL basket long against a short of a broad AI capex proxy ETF or mega-cap AI spenders if available. This isolates the bottleneck winners versus the cash-burning capex generators.
• Use call spreads instead of outright longs on the highest-beta name in the group, especially MRVL, to express upside while capping valuation compression risk if AI sentiment cools over the next quarter.