Meta plans to spend $145 billion this year, more than every military budget except the U.S., China and Russia

Meta plans up to $145 billion of 2026 capex, a roughly 101% increase from $72.2 billion in 2025, driven by AI compute, data centers, cloud agreements and higher component costs. Total 2026 expenses are projected to reach as much as $169 billion as Meta recruits AI talent and expands infrastructure, while Zuckerberg expects AI to improve advertising performance and create new consumer and enterprise revenue streams. Alphabet, Amazon, Meta and Microsoft are collectively on track to spend about $730 billion on AI in 2026, up from prior estimates of $600 billion.

Analysis

The relevant equity transmission is not headline capex but the conversion of incremental infrastructure into ad pricing, engagement, and eventually external revenue. If roughly half of the incremental build is depreciated over 4-5 years, the resulting annual depreciation burden could approach $7-9B before associated power, networking, and labor costs; META needs sustained high-single-digit incremental ad-revenue growth or demonstrable new revenue to preserve operating-leverage expectations. A larger installed base also raises the probability that 2027 becomes a free-cash-flow reset year even if 2026 EPS remains supported by advertising.

Near-term beneficiaries should be the constrained layers of the stack rather than broad software: NVDA/AVGO for accelerators and custom silicon, ANET for high-speed networking, and VRT/ETN for power and thermal infrastructure. The non-obvious bottleneck is grid interconnection and electrical equipment availability: data-center construction commitments can pull forward orders for VRT and ETN while delaying usable compute capacity, making their backlog more valuable than nominal GPU demand over the next 12-18 months. CEG/VST benefit only where load contracts are both incremental and economically priced; headline data-center demand alone does not guarantee merchant-power upside.

Consensus appears to treat aggregate hyperscaler spend as uniformly bullish for AI suppliers and benign for platform margins. That misses competitive duplication: four buyers building similar capacity can create a 2027-28 compute surplus if enterprise AI monetization lags, pressuring cloud pricing and reducing returns on capital. META is uniquely exposed because its external-compute ambitions would move it from a high-margin ad platform into competition with MSFT, AMZN, and GOOG, where utilization and customer-acquisition economics are materially less attractive.

The next 1-3 month catalyst is whether META quantifies ad conversion/ROAS gains and holds expense guidance despite the buildout. Falsification for the margin-risk thesis would be a material acceleration in ad pricing or engagement accompanied by stable operating-margin guidance; confirmation would be another upward capex revision, rising depreciation guidance, or management deferring a timeline for enterprise AI revenue.

Key Decisions for Investors

• Overweight VRT and ETN for 6-18 months versus the broader AI basket: electrical distribution, cooling, and power-density constraints have stronger pricing power than commodity compute. Use a 15-20% relative drawdown versus SOXX as a risk trigger; the thesis weakens if backlog conversion slows or lead times normalize.
• Pair trade over the next 3-6 months: long NVDA or AVGO / short META in equal beta-adjusted notional. Suppliers recognize demand earlier, while META absorbs depreciation and execution risk; cover if META reports clear ad-productivity monetization sufficient to offset the implied $7-9B annual depreciation step-up.
• Avoid adding outright META solely on AI-spending enthusiasm until the next earnings release demonstrates measurable advertising uplift and no further capex-guide increase. A post-results entry is preferable if revenue growth accelerates while operating-margin expectations remain intact.
• Monitor ANET as a higher-beta confirmation trade: initiate only if hyperscaler networking commentary or order backlog supports continued 800G/AI-fabric deployment. The key risk is custom-networking adoption or data-center commissioning delays reducing switch attach rates.