AI Infrastructure Capex: GPUs, Network, Storage, Facility Readiness and Software Deployment

Goldman Sachs Global Institute models $7.6T of AI infrastructure capex in its baseline scenario from 2026–2031. For enterprises and sovereign governments, there are five categories of capex to consider for AI infrastructure:

1. GPU (how much per GPU? how often to replace?)
2. Network (how much as rack density increases?)
3. Storage (how much to support workloads?)
4. Facility readiness (how much per MW?)
5. Software deployment (what does software deployment cost?)

How Much Is AI Infrastructure Capex?

Goldman Sachs models total AI infrastructure capex inclusive of compute, data centers, and power. There are several different scenarios in the model, and multiple drivers. The four most impactful assumptions in the model are:

1. How long AI silicon is economically useful (small changes to this assumption result in hundreds of billions of dollars in difference in the model output)
2. How expensive it is to build the next generation of data centers (increasing as power density and integration increases)
3. The mix of chips and architectures deployed (how inelastic is demand for compute)
4. How much supply-chain bottlenecks extend project timelines (power, labor, equipment, etc.), creating uncertainty about demand in more stressful scenarios

Two of these assumptions will be the basis for more detailed commentary on GPU and facility readiness capex below.

In terms of individual companies, Goldman Sachs notes that the top 5 US hyperscalers have provided guidance for 2026 combined capex of ~$660–$690B, nearly double the $380B for 2025 (via Futurum).

On a per-company basis, the 2026 guidance is ~$200B for Amazon, $175–$185B for Alphabet, tracking for at least $120B for Microsoft, $115–$135B for Meta, and ~$50B for Oracle.

GPU Capex (How Much per GPU?)

Goldman Sachs assumes $80.5k per GPU, including node costs, for a Rubin-class package modeled at 3,000W. This is a scenario assumption, not a universal GPU purchase price.

For GPU replacement cycles, Goldman Sachs notes that hyperscaler guidance has given a wide range of depreciation assumptions for materially similar hardware, from ~3 years to ~6+ years. These assumptions influence GPU replacement cycles and residual value.

Network Capex (How Much as Rack Density Increases?)

Network capex depends on GPU count, bandwidth requirements and fabric topology. Higher-density GPU racks often require more network bandwidth, but rack power consumption alone does not determine network costs.

Rack power consumption has jumped an order of magnitude in the last couple of years, with GB200 NVL72 racks consuming 130kW+ and requiring substantial cooling, whereas legacy enterprise racks may consume less than 10kW.

None of the three analyses reviewed from Goldman Sachs, Futurum, or build.inc breaks out capex for network as a discrete line item, so it's unclear what percentage or $/GPU to assume.

Storage Capex (How Much to Support Workloads?)

Similarly, none of the research from Goldman Sachs, Futurum, or build.inc breaks out capex for storage as a discrete line item.

To plan for storage capex, it is easiest to consider how much storage will be required to support the expected workloads.

Facility Readiness Capex (How Much per MW?)

Historically, to build out a cloud facility for hyperscale workloads, it was assumed that capex would be ~$10M/MW. Today, according to Goldman Sachs, capex for the next generation of AI data centers is more likely to be ~$15–$20M/MW due to the need for increased power density and integration. These upfront construction costs are one component of the broader AI infrastructure total cost of ownership.

In terms of the total capex buildout, build.inc notes that power infrastructure needs are now the first criteria by which developers are underwriting the buildout. Understanding the capital required for each phase, including transformers, switchgear, generators, UPS and cooling, is important for modeling financial exposure. However, the site's position in the grid connection queue is often the more immediate constraint.

Three factors affecting deployment:

1. Power: CenterPoint Energy reported a 700% increase in data center interconnection requests in its Houston-area territory, from approximately 1GW before summer 2024 to 8GW by October 2024.

1. Time: There is 12GW of data center capacity coming online in 2026 in the US, with around 5GW being "under construction" currently according to the US data center project tracker. 25% of those projects do not have a publicly disclosed power strategy yet.

1. Modular: We discuss modular in our total cost of ownership piece. Smaller modular data centers can provide 50–400kW per module, with manufacturing and delivery measured in weeks or months. Traditional brick-and-mortar construction typically takes considerably longer, but both approaches still require site preparation, power availability, integration and commissioning.

How Do These Factors Relate to Software Deployment as a Capex Consideration?

The customer ultimately pays for usable compute capacity. Delivering that capacity requires integrating multiple software components, including scheduling, orchestration, cluster filesystems and monitoring.

While our three analyses cited above do not include this component of the capex, it's worth noting that to make that capacity operational requires a plan to get these components online.

How Do These Considerations Impact the Approach to Planning for Capex?

The approach to capex planning depends on the project's primary constraint: power, GPU availability or deployment timeline.

If power is the limiting factor, start with available power capacity in MW. Given the cost per MW of a facility ($15–$20M per MW) and the timeline to get into the queue for power to a site, estimate the number of GPUs that can be purchased within the capex available for both the facility and the power to the site.

If chips are the limiting factor, start with the number of nodes being allocated. Given the baseline cost of a GPU (~$80.5k per GPU), scale the cost of the fabric appropriately to support the number of nodes required.

If time is the limiting factor, consider the manufacturing and delivery timelines of smaller modular facilities, measured in weeks or months, versus traditional brick-and-mortar construction, which typically takes considerably longer. Both approaches still depend on site readiness, power availability, integration and commissioning.

The project's capital structure also affects how these investments are funded. Depending on the AI infrastructure financing structure, companies may buy equipment outright, lease it or use partner capital.

Frequently Asked Questions About AI Capex

Does This Include the Cost to Operate? Or the Power Bill?

No. This is only the upfront cost to bring online a set of GPUs for use in running compute jobs. It includes the compute, the network, the storage and the data center (construction) and deployment. To see the cost to operate, or the economics of bringing these online and then using them, see our analysis of operating expenses (OPEX).

Why Is There a Discrepancy Between the $7.6T and the $690B?

The $660–$690B is the total capex for 2026 across the five named companies. The $7.6T is a scenario model of what it would cost to bring online all the compute, data center and power over the next six years, based on four underlying assumptions noted above.

How Can zFINANCE Help Finance Capex for AI?

We have built a service to help companies finance their AI journey. Whether it is to own, lease, or leverage partner capital, this category of capex is something we can help with. More on GPU financing here.