Why AI Infrastructure Spending Is Surging Toward a $3 Trillion Milestone

AI Infrastructure Pushes Data Center Capex Forecast Above $3 Trillion

The forecast from Dell’Oro Group has nearly doubled since January 2026. For anyone running models in production, the headline is less about GPU supply than about the infrastructure bill attached to every accelerator: power, cooling, networking, storage, and the servers around the chip.

The practical issue is throughput. AI infrastructure is becoming a system-level bottleneck, not a component-shopping exercise.

The accelerator is only one line item

Data Center Knowledge reports that AI accelerators could represent about one-third of the projected $3 trillion in data center capex. That still leaves most of the spending elsewhere.

AI clusters require servers to host the accelerators, specialized networking to keep those systems fed, and storage for training and inference workloads. Higher-density deployments also require changes to power distribution, liquid cooling, containment, and the supporting architecture inside the facility.

That is the part many infrastructure plans quietly undercount. Buying more compute does not automatically produce more useful throughput if the rack cannot receive enough power, shed enough heat, or move data fast enough. An accelerator waiting on storage or network traffic is expensive idle capacity. An accelerator triggering an OOM or thermal limit is worse: capital expenditure with a broken SLA attached.

The Dell’Oro outlook assumes global data center power availability will rise above 200 GW by 2030. That assumption is doing significant work. Power is not a background utility in this model. It is a capacity constraint that determines where AI infrastructure can actually be deployed.

Hyperscaler scale changes the procurement math

The four largest US cloud providers are expected to account for about half of global data center capex, according to the report cited by Data Center Knowledge. Their purchasing power allows them to secure preferred pricing and capacity commitments through long-term supplier agreements.

That creates a predictable problem for smaller buyers. Component availability can tighten, lead times can extend, and prices can rise. Custom chips and proprietary architectures may lower costs at hyperscaler scale while putting pressure on server manufacturers and limiting the value of standardized procurement for enterprises.

AI-specialized cloud providers, including model developers and neoclouds, are projected to grow at an annualized rate of nearly 60%. At the same time, demand for general-purpose servers is expected to increase as inference, agentic AI, and storage workloads expand.

For engineering teams, this points to a less fashionable but more useful planning question: which workload deserves dedicated infrastructure, and which workload should remain rented? The report indicates that enterprises may initially prefer rented GPU capacity because it avoids a large upfront commitment while utilization and returns remain uncertain.

That is not just a finance decision. It is an operational one. Renting shifts some procurement, power, and hardware-refresh risk to the provider. It does not eliminate cost; it changes the cost curve and the failure boundary.

Hybrid deployment is the likely pressure valve

The forecast suggests a hybrid path. Stable, heavily utilized AI workloads could eventually move on-premises when ownership becomes cheaper, while variable or incremental demand remains in the cloud.

I would treat that as a utilization test, not a deployment ideology. A workload with steady demand may justify owned capacity, but only after the organization can account for the full stack: accelerators, host servers, networking, storage, power, cooling, and the staff required to keep it alive. Low utilization turns expensive hardware into stranded capacity. High utilization exposes every bottleneck in the facility.

Data Center Knowledge also reports that operators are being advised to add capacity in stages, identify where high-density computing is genuinely required, and integrate new cooling methods with existing infrastructure. That is the sensible sequence. Start with the power envelope and rack density. Then measure network and storage throughput. Only after that should the accelerator count become the headline metric.

The capex forecast matters because it confirms that AI infrastructure is no longer a narrow GPU procurement problem. The deployment surface is the data center. My rule remains simple: deploy owned capacity only when utilization and workload stability can support the complete operating cost. Keep bursty demand in the cloud. Otherwise, the cheaper accelerator will be the one that never gets purchased.