Citadel: Compute Power Scarcity Remains the Biggest Bottleneck for AI, Hyperscale Cloud Providers May Emerge as Ultimate Winners

The competitive landscape of the AI industry is accelerating in its differentiation, and the key variable determining the winners is not model capabilities, but the availability of physical computing power.

The latest analysis from Citadel Securities points out that the demand for computing power continues to outpace the rate of supply growth. The truly scarce asset is not the chips in warehouses, but computing power that is powered on and can be put to immediate use.

This structural bottleneck is unlikely to be resolved in the short term—grid access and project approval cycles are often measured in years, and even massive capital expenditure plans cannot close the gap quickly. This means that the current cash flows of hyperscale cloud providers may underestimate the true profit potential of their existing infrastructure.

From market signals, the rental prices for computing power—including older generation GPUs—remain firm, indicating that current capacity is still being fully absorbed and the tight supply-demand situation has not been significantly alleviated by large-scale capital investments. Meanwhile, the rise of AI agents is amplifying computing power consumption—a single human command can trigger dozens or even hundreds of model calls, increasing the computing resources needed for a single task and further supporting demand.

The Essence of the Computing Power Bottleneck: More Than Just Chips

Citadel Securities analyst Nohshad Shah pointed out that "computing power" is not synonymous with GPUs alone. It is an aggregate of GPUs, electricity, data center space, memory, network, cooling systems, and professional operational capabilities. What is truly scarce is 'powered-on, readily available computing power,' not hardware sitting idle in warehouses.

Forward curves for computing power show that although supply is increasing, demand is growing even faster. Because grid access and project approval take years, even the largest capital expenditure plans can't fill this gap in the short term. The constraint currently facing AI is not a lack of users, but a lack of computing power that can actually be turned on and operated.

It is worth noting that a large amount of current computing capacity is tied to contracts signed for 2024–2025, when demand intensity was not yet fully apparent. As these contracts gradually expire, existing infrastructure is expected to be repriced amid rising utilization rates, while the cost base remains relatively fixed, further unlocking operating leverage. The analysis suggests that the market currently tends to account for capital expenditure costs right away, but may underestimate the longer-term upside in earnings potential.

In addition, a decline in token prices does not necessarily mean a negative impact. Cheaper intelligence will make more applications economically viable, and, combined with AI agents' multiplier effect on computing consumption, overall demand for computing power could continue to expand. The key variable is demand elasticity: if usage growth outpaces the decline in unit price, lower prices could broaden the market rather than shrink it; the true bear case would be token prices falling with no substantial increase in usage.

Industry Divergence Intensifies as Hyperscale Cloud Providers Hold a Dual Advantage

Citadel Securities' analysis outlines an increasingly polarized structure in the AI industry: on one end are cutting-edge labs such as OpenAI and Anthropic, whose product is intelligence itself; on the other are diversified hyperscale cloud providers like Google and Microsoft, who can monetize by covering the full stack regardless of which model ultimately wins.

Within this framework, leading-edge models will focus on high-value tasks such as planning, reasoning, programming, and orchestration, earning a premium on a relatively small share of tokens; cheaper or open-source models will handle highly concurrent execution-level work. Frontier labs may retain pricing power for high-end tasks, but hyperscale cloud providers and inference service providers will benefit from both tiers—because any kind of workload requires chips, memory, network, and electricity.

The core logic is: leading-edge models do the planning, cheaper models handle execution, and those with 'powered-on computing power' will reap economic value from both sides. For investors, at a time when the commercialization path for AI remains uncertain, owning the physical computing infrastructure as a hyperscale cloud provider may offer the most certain risk-return profile in this race.