In 2024, the California Energy Commission added projected data center growth to electricity demand forecasts.

TL;DR

• AI data centers are being reviewed through grid, cooling, and water operations, not only total size.
• This matters because delays often come from interconnection, permitting, water conditions, and local review.
• Readers should examine interconnection terms, cooling design, water agreements, and power management before judging expansion plans.

Example: A region plans new compute capacity during a dry season. The main question becomes grid access, cooling design, and water tracking.

The key point is practical. The main issue is not a broad claim about planetary harm. The issue is when, where, and how the grid can absorb demand. The water issue is similar. Official documents focus on cooling methods, drought-period operations, and tracking duties.

TL;DR

• The central issue is that AI data center resource use should be analyzed through operational metrics.
• Those metrics include electricity demand, grid interconnection, cooling methods, and water tracking obligations.
• Readers should first examine power conditions, cooling design, water agreements, and power management software.

Current situation

Official materials show that electricity is no longer only an abstract concern. In 2024, the California Energy Commission included projected data center growth in demand forecasting. To judge grid capacity, it helps to inspect utility interconnection assumptions and forecast models. That is more useful than only saying AI demand is growing. Data centers now appear as a variable in regional infrastructure planning.

The water issue is also reviewed through operating conditions. Linn County, Iowa requires water-use agreements for data centers. It also sets conditions for tracking, reporting, information sharing, and regulatory compliance. Review does not stop at whether a facility uses a lot of water. It also examines how use is recorded, reduced, and monitored. Project documents from the City of Pittsburg, California place environmental review materials under public review. Large data centers are now local permitting projects.

Analysis

From a decision-making view, the bottleneck may extend beyond chips. Power, cooling, and administrative procedures can shape project outcomes. If a region has grid interconnection capacity and clear water conditions, expansion review can become simpler. If demand forecasts are uncertain and water agreements are strict, timelines can become less stable. That is why compute constraints should also be framed as infrastructure readiness. The question is whether generation, substations, and cooling operations are prepared.

Official documents also identify technical responses. These include direct-to-chip liquid cooling. They also include hybrid air and liquid cooling designs. Airflow optimization appears as hot aisle and cold aisle separation and containment. On the power side, proposed measures include PUE improvement. They also include less mechanical cooling energy use, workload-specific power profiles, dynamic power allocation, and automated power redistribution.

These techniques have limits. Their existence does not mean every region can resolve grid or water conflicts. Supply chains, sites, permitting, and local opposition can remain difficult. Software alone may not solve them. Based on the present materials, uniform project comparisons are also difficult. That includes electricity rate changes, transmission reliability effects, and water savings by cooling method.

Practical application

When organizations review AI infrastructure plans, throughput under a power limit is often a better first question. That can be more useful than asking only how many GPUs fit. A practical framework has three parts. First, review utility interconnection and substation conditions. Second, review the cooling method and water-tracking system. Third, review whether power management software can reduce idle power. The power limit is a contractual condition. Throughput is an operational issue.

Even under the same power limit, outcomes can differ across operating policies. Teams may use shared policies or workload-specific power profiles. Training, inference, and batch jobs do not have identical power curves. A useful first step is measuring which tasks use the most power and when.

Checklist for Today:

• If a review document omits interconnection, cooling, or water-tracking items, request their inclusion.
• Cluster operations teams should test workload-specific power profiles and dynamic power allocation.
• For regional projects, check environmental review documents and water-use agreements before focusing on investment size.

FAQ

Q. Is the biggest resource issue for AI data centers electricity or water?
Reducing the issue to one category can hide important details. Official documents treat electricity demand and grid interconnection as major concerns. They review water together with cooling methods, drought-period operations, and tracking requirements.

Q. If cooling technology improves, will local conflicts also be resolved?
Not necessarily. Liquid cooling and airflow optimization can improve facility efficiency. They may not resolve interconnection delays, permitting issues, or community acceptance concerns.

Q. What metrics should companies examine first right now?
A combined view is more useful than a single metric. In practice, review the power limit, interconnection conditions, cooling design, water-tracking system, and power management software use.

Conclusion

It is more accurate to treat AI data center resource burden as an infrastructure operations issue. The evidence includes the 2024 forecasting change in California. It also includes facilities above 30 megawatts. It includes sites from 10 acres to more than 100 acres. Water-use agreements also matter. These details help identify likely bottlenecks. The next step is to examine how each project contracts for and manages power and water.

Further Reading

• AI Resource Roundup (24h) - 2026-07-02
• Bug Reproduction Tests as Signals for Code Agents
• DART-VLN Improves Discrete VLN Without Retraining at Test Time
• Dynamic 3D Reconstruction from Monocular Video with Generative Priors
• Interpreting RAG Retrieval With Sparse Autoencoder Features

References

• Data Centers | California Energy Commission - energy.ca.gov
• Data Centers | Missouri Department of Natural Resources - dnr.mo.gov
• Data Centers in Unincorporated Linn County | Linn County, IA - Official Website - linncountyiowa.gov
• AVAIO Data Center Project | City of Pittsburg - pittsburgca.gov
• Power and Thermals — NVIDIA GB200 NVL Multi-Node Tuning Guide - docs.nvidia.com
• Cooling and Airflow Optimization — NVIDIA DGX SuperPOD: Data Center Design Featuring NVIDIA DGX H100 Systems - docs.nvidia.com
• Domain Power Service Docs - docs.nvidia.com
• Domain Power Service Docs - docs.nvidia.com