Apollo President Warns AI Infrastructure May Undermine Returns

Apollo Global Management president Jim Zelter warned that the AI spending boom may not translate into strong investor returns despite creating a massive demand surge for digital infrastructure. Zelter estimates US data centers alone could need $5 trillion to $6 trillion over the next five years, and he questioned whether the economic owners of that capacity will be able to harvest commensurate returns. "We've seen this many times in our last 30 years, whether it's cell phones or other technology uses. There's no doubt they're going to have a massive utility," Zelter said, adding that the critical question is who gets the returns.

The AI buildout shifts capital requirements across multiple layers of the stack and changes unit economics for technology companies. Key technical and capital facts to note:

• •Scale of capital: Apollo and industry briefs project trillions in required capital, with figures like $3 trillion through 2028 and higher global estimates to 2030 depending on source assumptions.
• •Asset profile: The market is moving from asset-light software models toward asset-heavy commitments for land, power, cooling, and long-lived data center campuses.
• •Power and density: Data center growth implies large additional grid loads; estimates show tens of gigawatts of new power demand, driving separate investments in generation and transmission.
• •Developer role: Hyperscalers will still rely on third-party developers and operators to secure sites and grid connections, which is why private owners are buying companies like Stream Data Centers.

This is not just an engineering problem; it is a finance and macro problem that will affect allocation of compute, resilience, and cost curves for AI teams. The shift to heavy capital intensity means:

• •Financing conditions and credit spreads matter more for data center economics than they did for cloud-native software.
• •A potential $1 trillion to $1.5 trillion funding gap between projected hyperscaler capex and expected external capital raises could force slower buildouts, higher lease rates, or tighter siting constraints.
• •Investors and operators will compete over locations with favorable power and permitting, creating regional bottlenecks that influence latency, redundancy, and geopolitics of AI deployments.

For machine learning teams, higher infrastructure costs and localized capacity constraints can change cost-per-inference and time-to-deployment calculations for large models. For capital allocators, the risk profile of infrastructure plays is different from software: long duration, operational complexity, and dependency on regulated utilities.

Apollo is not just sounding an alarm; the firm is deploying capital to capture the opportunity, including acquisitions like Stream Data Centers and emphasizing private credit and long-dated capital as solution sets. That dual stance-warning about returns while backing investments-signals conviction that careful structuring and developer-led models can be profitable, but also that execution and financing terms will be decisive.

Monitor actual capital deployment versus announced plans, utility permitting and transmission buildouts, and spreads on private credit and investment-grade debt tied to infrastructure developers. If the funding gap persists, expect higher prices for colo/lease capacity, slower hyperscaler expansion in constrained regions, and more bespoke financing structures for AI compute projects.