AI Capex Creates New Investment Cycle: NVIDIA & 2 Other Stocks to Buy

AI infrastructure spending is accelerating, with the five largest global technology companies projected to invest more than $1 trillion during 2025-2026; Alphabet lifted 2026 capex guidance to $195-$205 billion and Microsoft plans to expand data-center capacity to roughly 38 GW by 2032. NVIDIA's fiscal Q2 2027 Data Center revenue rose 117% year over year and it generated $69.9 billion of first-half free cash flow, while Arista's Q2 2026 revenue grew 37.7% and Vertiv's sales increased 24% with adjusted operating margin up 410 bps to 22.6%. The article identifies NVIDIA, Arista and Vertiv as beneficiaries across AI compute, networking, and power/cooling infrastructure, while flagging cash conversion and returns on invested capital as key measures of capex-cycle durability.

Analysis

The investable implication is a mix shift from compute silicon toward the physical constraints of AI deployment. NVDA remains the highest-beta read-through to accelerator orders, but incremental cluster spend increasingly shifts to networking, power distribution and liquid cooling as rack density rises; this favors ANET, VRT, Eaton (ETN), nVent (NVT), Hubbell (HUBB) and data-center REIT/developer proxies with secured power. The key second-order bottleneck is utility interconnection rather than GPU availability: delayed grid capacity can defer server revenue recognition while leaving equipment backlogs intact.

Near term (days to 1-3 months), the market will likely reward suppliers that convert backlog into raised 2027 revenue and margin guidance, not broad capex announcements. VRT has the clearest operating-leverage setup, but also the greatest risk that investors extrapolate peak pricing and expedited-installation margins; ANET's risk is customer concentration and Ethernet-versus-proprietary-fabric share assumptions. For NVDA, the relevant falsifier is not demand commentary but a material deterioration in hyperscaler depreciation-life assumptions, utilization disclosures, or orders shifting from purchased capacity to internally designed accelerators.

The consensus is overly focused on whether aggregate AI capex is sustainable and underweights timing mismatch: hyperscalers can sustain spend while returns remain weak, compressing their own FCF without immediately impairing component suppliers' booked revenue. That makes a relative-value structure preferable to outright long mega-cap platforms. Over 6-18 months, constrained electrical infrastructure should command a scarcity premium, but policy-driven power-rate increases, permitting delays, or a sharp fall in inference pricing could turn the buildout from a supply-constrained cycle into a returns-constrained one.

Key Decisions for Investors

• Initiate a 3-6 month long VRT / short GOOG pair, sized beta-neutral: VRT is more directly exposed to the power-density bottleneck, while GOOG bears the FCF and monetization burden of incremental capex. Review after each company reports; exit if VRT backlog conversion or gross-margin guidance weakens, or if GOOG demonstrates accelerating AI revenue sufficient to stabilize FCF expectations.
• Add ANET on post-earnings weakness rather than chase strength; target a 6-12 month hold into evidence of 1.6T orders and AI-cluster revenue mix expansion. Use a 10-15% position risk limit because a single large-cloud customer digestion cycle can overwhelm otherwise strong industry demand.
• Maintain NVDA exposure through a call-spread rather than unhedged equity for the next two earnings cycles: buy 6-9 month at-the-money calls and sell 20-25% out-of-the-money calls to retain upside to raised supply/revenue guidance while limiting valuation-contraction risk. Close if hyperscaler capex guidance turns down or NVDA's data-center gross-margin guide falls materially.
• Build a diversified electrical-infrastructure basket—long ETN, NVT and HUBB—over 6-18 months as a lower-multiple complement to VRT. The thesis is falsified by utility interconnection lead times shortening materially or by project cancellations showing that announced data-center capacity lacks contracted power and financing.
• Avoid adding broad long exposure to GOOG and MSFT solely on capex intensity. Set an alert for quarterly AI revenue disclosure, cloud margin, depreciation expense and FCF conversion; absent visible monetization, additional spending is more likely to be a multiple headwind than an equity catalyst.