We're adding a speculative stock to our Bullpen that can benefit from the AI data center boom

Alphabet unveiled its eighth-generation TPU lineup, including the TPU 8t for training and TPU 8i for inference, with the company saying the 8t delivers nearly 3x the compute performance per pod versus the prior generation. The note also highlighted Morgan Stanley's view that agentic AI could lift industry CPU spending to $82.5B in 2030 in a base case, from about $25B today, a constructive read-through for Alphabet, Amazon, and Arm. Separately, BWX Technologies was added to the Bullpen as a speculative nuclear power name tied to AI data-center power demand, while markets also tracked higher oil prices, the Iran ceasefire, and a heavy earnings calendar.

Analysis

The market is re-pricing the AI capex trade from pure GPU intensity toward the broader power-and-control stack. If the compute buildout keeps compounding, the next bottleneck is not semiconductors alone but grid access, backup generation, and nuclear-heavy baseload capacity; that shifts incremental value from the obvious leaders toward less crowded infrastructure names with embedded optionality. Alphabet’s custom silicon push also reinforces a structural advantage: the more workloads migrate to proprietary accelerators and Arm-based CPUs, the more cloud economics favor vertically integrated hyperscalers over rent-paying AI consumers.

Broadcom’s exposure is interesting because the TPU narrative is no longer just about one chip win; it signals that hyperscalers are willing to keep funding custom silicon as a strategic cost lever. That is constructive for AVGO, but the bigger second-order read-through is for Arm: custom CPUs may scale faster than investor models assume if agentic workloads are CPU-heavy and inference is latency constrained. The risk is that enthusiasm for “AI infrastructure” becomes too broad and starts pricing in every adjacent supplier, so near-term upside will likely be concentrated in names with demonstrable design-win visibility rather than generic AI beneficiaries.

BWXT is the more asymmetric but more dangerous expression of the power thesis. Nuclear supply chain scarcity can create real pricing power over a multi-year horizon, but the public market may be extrapolating SMR economics ahead of bankable deployment timelines; that creates a meaningful gap between narrative and earnings conversion. If power scarcity remains acute through 2025-2026, BWXT, GEV, and select nuclear suppliers can rerate further, but any delay in utility procurement or policy support would compress multiple fast because the stocks already embed long-duration optionality.

The key contrarian point: the current move is less about war-risk hedging and more about duration re-asserting itself in AI. If oil stabilizes while rates drift lower, capital-intensive infrastructure with credible cash-flow visibility should outperform high-beta software and cyclicals, but if earnings this week disappoint, the market could quickly rotate back to quality balance sheets and away from speculative capex enablers.

Key Decisions for Investors

• Long GOOGL vs. a basket of AI software names over 1-3 months: use the TPU launch as a signal that hyperscaler vertical integration is widening, with upside if cloud margin expansion starts to show through in next earnings print.
• Add AVGO on pullbacks for a 3-6 month hold: custom accelerator exposure remains underappreciated, but size modestly because the stock now requires continued design-win confirmation rather than just sector beta.
• Express the CPU/agentic AI thesis via long AMZN and/or GOOGL, and pair against a short in lower-quality enterprise software over 2-4 months: benefit from capex migration toward infrastructure while software monetization lags the compute spend cycle.
• Speculative starter long BWXT only on weakness, sized small: 6-18 month upside exists if SMR deployment milestones accelerate, but the rich multiple means any schedule slip could cut 15-25% quickly.
• If GEV holds post-earnings strength, consider a paired long GEV / short industrials proxy (e.g., XLI beta) for 1-2 quarters: the trade captures power-buildout scarcity while limiting macro cyclicality.