RockstarMarkets
All news
Topic hub

AI Capex: Who's Spending, Who's Earning, and What's at Risk

Tracking AI infrastructure capex — hyperscaler spend, data center buildouts, memory demand and the margin compression risk.

AI capex is the supply-side mirror of the AI demand story. MSFT, GOOGL, META and AMZN have collectively guided to over $250 billion in annualized capex, dwarfing the dot-com era's peak telco spend. The winners are clear (NVDA, AVGO, ASML, MU, the data center REITs), but the longer-term question is whether the spend earns ROIC above hurdle rates — or whether AI model commoditization crushes the margins of the hyperscalers themselves.

This hub aggregates the narrative coverage on capex announcements, earnings revisions, infrastructure bottlenecks (power, cooling, advanced packaging), and the margin compression debate that keeps coming back as foundation models become cheaper to train.

Latest coverage

Frequently asked

Why is AI capex so high in 2026?

Foundation-model scale, sovereign cloud build-outs, and competitive pressure between MSFT, GOOGL, META and AMZN are driving capex to roughly $250 billion annualized — about 4x the pre-AI baseline. Each hyperscaler is investing to avoid being relegated to a second-tier AI infrastructure provider.

Will AI capex pay back for hyperscalers?

The bull case: AI inference revenue ramps faster than capex depreciation, ROIC stays above WACC. The bear case: model commoditization compresses the value of inference, and capex investments earn below cost of capital. Both scenarios are live; the answer is asymmetric per quarter of inference revenue data.

Which beneficiaries get the highest share of the spend?

NVDA captures roughly 70% of AI GPU spend. AVGO is the leader in custom ASICs and high-performance networking. ASML supplies the EUV lithography essential to leading-edge chip manufacturing. MU and SK Hynix are the largest beneficiaries of HBM memory demand.

What is the margin compression risk?

If frontier-model performance plateaus and inference becomes commoditised, the gross margins on AI workloads collapse. Hyperscalers would be left with a depreciated asset base earning utility-grade returns rather than software-grade returns. Watch GOOGL and MSFT cloud gross margins in quarterly earnings for the first signs.