Hardware, data centers, managed services and compute capacity each capture value differently, and the connections between them shape how much of it lasts.
Chips absorb most of the capital in AI infrastructure, and the data centers, contracts and operations around them determine whether that capital earns a return.
AI infrastructure is one of the largest capital buildouts in modern industry. Goldman Sachs Research’s baseline estimate puts cumulative capital for compute, data centers and power at about $7.6 trillion between 2026 and 2031, with annual AI capital spending rising from roughly $765 billion in 2026 to $1.6 trillion in 2031. Numbers on that scale raise a practical question: where in the stack is value created, and where does it last?
The layers of the stack behave very differently, so each one is worth examining on its own before looking at how they connect.
Most AI Infrastructure Spending Goes to Hardware
Hardware is the largest line item. The Goldman Sachs baseline assigns about $5.1 trillion of the $7.6 trillion to compute, roughly two-thirds of the total, against $2.1 trillion for data centers and $358 billion for power. JLL puts the global average shell-and-core construction cost at roughly $11.3 million per megawatt in 2026, and tenant fit-out for electrical distribution, racks and cooling can add up to $25 million per megawatt more, according to Cora Advisors. Power, cooling and the building itself are substantial costs, and compute hardware still outweighs each of them.
Hardware is also the layer with the fastest-moving economics. Each new GPU generation raises performance and resets the value of what came before. Owning hardware creates value only while the hardware is used productively, and that window is measured in years.
Data Centers Capture Value Through Scarcity of Power and Density
The data center layer earns its value from what is hardest to replicate: energized power and the ability to support high-density AI racks. Modern AI racks draw many times the power that most existing facilities were designed to deliver. Facilities that can provide that density, with the power and cooling behind it, are scarce relative to demand.
Long lead times for power equipment and grid connections reinforce the scarcity of energized capacity. Lead times for high-voltage circuit breakers climbed to about 125 weeks in the second half of 2025, up from 77 weeks in 2023, according to Wood Mackenzie data reported by Reuters, and grid interconnection in some regions now takes years. Value in this layer comes from securing power and density before others, and from a physical asset that outlasts several hardware generations.
Compute Capacity Earns Value Through Utilization and Contract Structure
Compute capacity, the GPU hours sold to model developers, AI companies and enterprises, is where hardware and facilities turn into revenue. The economics depend heavily on utilization and contract structure. A cluster running near full capacity under a multi-year agreement produces predictable revenue. The same cluster sitting partly idle, or selling into a volatile spot market, can struggle to cover its costs.
Spot pricing moves in both directions, which is why the commercial layer favors long-term contracts. One-year H100 contract pricing rose almost 40 percent between October 2025 and March 2026 as on-demand capacity sold out, according to SemiAnalysis, and one GPU cloud provider raised its on-demand H100 price from $3.85 to $4.50 per GPU-hour in October 2026, according to Uvik Software’s GPU Price Index. Committed capacity avoids that volatility, so value shifts toward keeping capacity committed and productive across the hardware’s operational life.
Managed Services Create Value Over the Full Hardware Lifecycle
Managed services are the least visible layer and influence the economics of everything above them. A GPU cluster requires continuous attention to power, cooling, network orchestration and node health. In large training jobs, a failing node can interrupt work running across many GPUs, so operational quality directly affects how much revenue a cluster can produce.
The lifecycle extends beyond day-to-day operations. The Goldman Sachs baseline assumes a chip economic life of five years, according to Augment Market, which means refresh, migration, decommissioning and resale of retired equipment are all part of the economic picture. A well-run cluster also recovers value at the end of its life, in addition to what it earns along the way.
The Connections Between Layers Shape Returns
Viewed one layer at a time, AI infrastructure looks like a set of separate markets: hardware vendors, landlords, cloud providers and operators. In practice, the economics depend on how well those layers connect. Hardware that arrives before the facility is ready earns nothing. A facility without committed customers sits idle. Capacity without reliable operations loses revenue to downtime.
Each handoff between independent providers adds coordination time and risk. A reasonable view is that a meaningful share of value in AI infrastructure comes from shortening the gap between capital spent and revenue earned: coordinating capital, procurement, facilities and operations so that each layer is ready when the next one needs it.
A Framework for Evaluating AI Infrastructure Businesses
For investors, a useful way to assess any AI infrastructure business is to ask which layers the business participates in and how it manages the connections between them. How quickly does capital turn into active compute? How much of the capacity is committed, and for how long? What happens to the hardware at the end of its life? The answers reveal more about durable value than the size of a hardware order.
Vertical Data was built to shorten the distance between capital and active compute. Vertical Data’s three platforms bring GPU provisioning and managed infrastructure (VerticalData.io), structured financing (GPUFinancing.com) and data center development (VerticalEdgedc.com) together under one company, so fewer handoffs sit between capital spent and revenue earned.
