3 More AI Infrastructure Plays Beyond the Big Names

Major technology firms' calls to slow AI development raise the prospect that future demand for AI training could shift. However, multi-year AI infrastructure projects continue to proliferate rapidly, supported by the lengthy timelines required to secure power and construct facilities, while AI applications expand into additional industries.

Analysis

The key investable distinction is between a potential slowdown in frontier-model training spend and the much stickier, already-contracted buildout of power, networking, cooling, and data-center capacity. A moderation in AI capex expectations would likely compress high-duration compute beneficiaries first, while suppliers tied to physical project milestones retain revenue visibility for the next 12-24 months. The bottleneck remains grid interconnection and power availability, making electrical equipment and generation exposure more defensible than broad AI software or semiconductor-beta exposure.

Second-order beneficiaries include Eaton (ETN), Vertiv (VRT), GE Vernova (GEV), Hubbell (HUBB), Quanta Services (PWR), and data-center REIT Digital Realty (DLR); each participates in infrastructure whose lead times limit near-term supply response. Conversely, an AI-training pause would be most damaging to the marginal demand assumptions embedded in Nvidia (NVDA), AMD (AMD), Super Micro Computer (SMCI), and highly levered data-center developers, particularly where capacity is speculative rather than pre-leased. The critical question is whether hyperscalers slow commitments before equipment orders convert into backlog, not whether public AI rhetoric turns more cautious.

Near term, this is not a standalone catalyst: sentiment around AI capex is too headline-sensitive and the article provides no evidence of canceled projects. Over 1-3 months, watch hyperscaler capex guidance, announced power procurement, and lead-time commentary from ETN/VRT/GEV; reductions in utility-load forecasts or data-center lease precommitments would be the first meaningful falsification. Over 6-18 months, grid constraints can turn into a margin issue if customers demand fixed-price delivery while labor, transformers, and generation equipment remain scarce.

Consensus may be too binary: slower model-training intensity does not necessarily mean lower data-center infrastructure demand, because inference, enterprise deployment, and redundancy requirements can shift the workload mix rather than eliminate power demand. The more likely de-rating risk is for companies priced on unconstrained GPU-unit growth, while the physical-infrastructure complex can remain supported—but only if backlog quality and pricing discipline hold.

Key Decisions for Investors

• Prefer a 6-12 month pair trade: long ETN or GEV versus short SMCI, sized modestly. It expresses durable power-equipment scarcity against the more cyclical, valuation-sensitive portion of AI infrastructure; reassess if ETN/GEV report backlog cancellations or SMCI demonstrates sustained gross-margin expansion with accelerating order visibility.
• Maintain or build VRT exposure only on pullbacks rather than chase momentum. Its upside case depends on conversion of cooling backlog into revenue over the next 12-24 months, but a 15-20% drawdown is plausible if hyperscaler capex guidance softens; use quarterly bookings, backlog, and pricing as the thesis checkpoints.
• Watch PWR and HUBB as lower-beta beneficiaries of interconnection and grid-hardening spend. Initiate only after confirmation that utility capital plans and large-load interconnection queues remain intact; a reduction in utility load-growth forecasts or delayed transmission awards would invalidate the demand-duration thesis.
• Avoid adding broad long exposure to NVDA/AMD solely on continued data-center construction. Require evidence that incremental facilities are being equipped on schedule through hyperscaler capex guidance and GPU lead-time data; a training-spend slowdown can hit unit-growth expectations well before physical projects are canceled.
• Set a catalyst calendar for the next hyperscaler earnings cycle: aggregate capex guidance, power-purchase commitments, and comments on training versus inference mix matter more than generalized AI caution. A coordinated downward revision in 2027 capex plans would favor reducing VRT/ETN/GEV exposure despite their near-term backlog protection.