Key Takeaways
- • Compute has emerged as the critical bottleneck in AI infrastructure. Apollo Global Management's Robert Bittencourt characterizes computational capacity as the "spice" of 2026—an indispensable resource constrained by physical supply chains including accelerators, high-bandwidth memory, networking equipment, datacenters, and electricity generation.
- • AI workload intensity is accelerating faster than supply expansion. Demand for compute is growing more than twice as fast as Moore's law, with AI computational needs projected to reach approximately 100 gigawatts of new demand in the US by 2030, far outpacing infrastructure buildout.
- • The shortage is fundamentally a power and infrastructure problem, not merely a chip shortage. Data centers running AI workloads consume enormous amounts of electricity at rates straining power grids. Eight hyperscalers expect a 44% year-over-year increase in spending on AI data centers and computing resources, reaching $371 billion in 2025.
- • Pricing power is consolidating among scarce input providers. Nvidia compute time has risen approximately 48% in recent months, reflecting supply constraints. As demand exceeds supply, providers of advanced semiconductors, power generation, transmission capacity, and data center infrastructure are capturing significant margin expansion.
- • Market conditions are shifting from abundance to scarcity. Rather than an outright shortage, the market is transitioning into a supply-constrained environment where access and pricing power will increasingly favor infrastructure providers and chip manufacturers over AI developers.
Summary
The artificial intelligence infrastructure market is experiencing a fundamental supply-demand imbalance centered on computational capacity. According to Apollo Global Management's thematic investing analysis, compute—the computational capacity required to train and run AI models—has become the critical constraint across the industry. This shortage is not a temporary disruption but reflects a structural mismatch between accelerating AI adoption and the physical infrastructure required to support it.
The constraint operates across an interconnected supply chain spanning semiconductor accelerators, high-bandwidth memory, networking equipment, large-scale datacenters, and electricity generation and transmission. Critically, the bottleneck is not limited to chip supply; it is fundamentally a power infrastructure problem. Data centers supporting AI workloads consume electricity at rates that are actively straining electrical grids, creating real-world capacity constraints that extend beyond manufacturing timelines.
Demand acceleration is outpacing even aggressive supply expansion efforts. AI computational demand is growing at a rate exceeding twice the historical pace of Moore's law improvements, with projections indicating the US market alone will require approximately 100 gigawatts of new computing demand by 2030. This acceleration reflects not only increased user adoption but also a marked shift in AI workload characteristics—newer models and use cases are dramatically more compute-intensive than predecessors. Eight leading hyperscalers have committed to 44% year-over-year increases in capital expenditure for AI infrastructure, signaling an industry-wide recognition that supply-side constraints are binding.
The shortage is translating directly into pricing power for scarce input providers. Nvidia compute rental costs have risen approximately 48% in recent months, a metric reflecting the persistent imbalance between available capacity and demand. Rather than a complete shortage, the market is transitioning into a supply-constrained environment where pricing increasingly reflects scarcity premiums. This dynamic concentrates value among providers of advanced semiconductors, power generation and transmission capacity, and data center infrastructure, while raising operational costs for AI developers and downstream enterprises.
The constraint operates across an interconnected supply chain spanning semiconductor accelerators, high-bandwidth memory, networking equipment, large-scale datacenters, and electricity generation and transmission. Critically, the bottleneck is not limited to chip supply; it is fundamentally a power infrastructure problem. Data centers supporting AI workloads consume electricity at rates that are actively straining electrical grids, creating real-world capacity constraints that extend beyond manufacturing timelines.
Demand acceleration is outpacing even aggressive supply expansion efforts. AI computational demand is growing at a rate exceeding twice the historical pace of Moore's law improvements, with projections indicating the US market alone will require approximately 100 gigawatts of new computing demand by 2030. This acceleration reflects not only increased user adoption but also a marked shift in AI workload characteristics—newer models and use cases are dramatically more compute-intensive than predecessors. Eight leading hyperscalers have committed to 44% year-over-year increases in capital expenditure for AI infrastructure, signaling an industry-wide recognition that supply-side constraints are binding.
The shortage is translating directly into pricing power for scarce input providers. Nvidia compute rental costs have risen approximately 48% in recent months, a metric reflecting the persistent imbalance between available capacity and demand. Rather than a complete shortage, the market is transitioning into a supply-constrained environment where pricing increasingly reflects scarcity premiums. This dynamic concentrates value among providers of advanced semiconductors, power generation and transmission capacity, and data center infrastructure, while raising operational costs for AI developers and downstream enterprises.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.