SURVIVING
THE SURGE

Executive Summary

Artificial intelligence is scaling faster than the physical systems that support it. The central constraint on AI growth in 2026 is time-to-power: the ability to secure electricity, cooling, land, permits, and social licence fast enough to deploy responsibly.

Compute capacity, model complexity, and inference demand continue to accelerate, while power generation, grid connections, water availability, and permitting move at infrastructure speed. The US interconnection queue stands at approximately 2,600 GW as of early 2026, up from 2,300 GW in January. Berkeley Lab now projects data centres could consume 6.7 to 12 percent of total US electricity by 2028 — pulling the upper-bound demand case forward relative to earlier 2030 framing.

Water has moved from a regulatory barrier to a measured and increasingly politicised operating constraint. Community opposition has evolved from reputational friction to capital friction.

1The Scale Problem Has Accelerated

In five months, the constraint set has not materially eased in the main markets that matter. It has tightened where it is measured, and in some cases the pace of tightening appears faster than operator planning cycles.

Critical markets remain under strain. Northern Virginia illustrates the extreme case of electricity concentration and queue pressure. Texas shows the consequences of large-load demand moving faster than planning assumptions. Singapore is directly relevant to APAC operators: it combines a tight reserve margin, water stress, and a tropical cooling penalty inside a small, highly regulated system.

Electricity access remains the schedule driver. The stronger framing for V4.2: in leading AI infrastructure markets, queue depth, time-to-power, and grid integration friction are the real deployment bottlenecks — not compute availability, not model readiness.

2Water: From Barrier to Measured Constraint

January's edition treated water mainly as a permitting issue. June's evidence shows that was too narrow. Water is now a measurable operational variable, a planning constraint, a public-policy issue, and in some locations a potential litigation trigger.

2.1 The Measured Baseline

US data centres directly consumed 17.4 billion gallons in 2023, and published projections place 2028 direct consumption in a 38 to 73 billion gallon range. That difference matters. The baseline is measured. The forward number is a scenario envelope. They carry different epistemic weight and should never be written as though they are equivalent.

2.2 The Regulatory Wave

Cooling-water discharge, reuse, and water reporting are moving into closer regulatory scrutiny, and operators increasingly face permitting, disclosure, and compliance obligations that vary by jurisdiction. V4.2 avoids overstating regulatory scope: where bills have been introduced but not enacted, the paper says exactly that. Future compliance is framed as "would mandate," not "mandates."

The core claim is preserved and strengthened: water is becoming harder to externalise. The attack surface is removed. That is not a retreat — it is analytical discipline.

2.3 The Engineering Answer

Cooling choice, siting choice, and water context can produce order-of-magnitude differences in water intensity. Water Usage Effectiveness belongs in the boardroom because the constraint is now strategic, not merely technical. Evaporative cooling in a high water-stress county is not just an operating decision — it is a permitting, reputation, and potentially legal-risk decision.

2.4 Figure 1 — Optimisation vs. Depletion

Figure 1. Efficiency gains inside the data centre do not automatically translate into lower physical impact. Without constraints on siting, water, and allocation, efficiency can reduce unit cost while increasing total system pressure. An optimisation narrative that ignores allocation and physical context does not eliminate impact — it displaces it.

3The A+ Pathway

The A+ Pathway remains one of the most useful parts of the paper. It works because it treats the problem as a system rather than a technology shopping list.

3.1 Five Pillars

3.2 Benchmarks

Benchmark values are from public operator disclosures. They are indicative of engineering direction, not strict like-for-like comparisons across identical reporting boundaries.

3.3 Implementation Roadmap

4The Misallocation Crisis and the Carbon Math

Most AI infrastructure commentary asks how to supply more compute. This paper asks what the compute is for. That remains the right question.

The workload structure has inverted: training dominated compute in 2020–2022; by 2024–2026 inference dominates at 60–70% and grows at roughly 122% CAGR. Enterprise budget allocation (BCG, 2025) places support functions at 38%, operations at 23%, marketing and sales at 20%, R&D at 13%, and climate-critical applications below 3%.

Efficiency gains reduce energy intensity per unit of compute, but observed demand growth can absorb those gains and increase total system load — especially in inference-heavy markets. V4.2 frames this as an interpretive risk, not a mechanically proven law for every workload. The Jevons dynamic is a serious structural warning, not a certainty.

The enterprise allocation argument is portfolio governance: if scarce infrastructure is being allocated to low-social-value workloads, then compute allocation becomes an infrastructure policy question — not merely a product-market outcome. That is the conversation the industry has not yet held.

5Social Licence: From Sentiment to Capital Friction

This section is right on the substance and stronger with tighter method. V4.2 fixes the method.

Opposition now has measurable economic weight. Delays, local resistance, moratorium proposals, tax debates, and utility bill impacts have moved beyond public-relations nuisance. They are now part of project bankability and delivery risk. But headline figures need explicit methodology before they can anchor a serious argument.

5.1 Constraint Summary — June 2026

The strategic conclusion strengthens with each data point: transparent engagement is now an approval prerequisite. Operators who treat social licence as communications rather than engineering are funding the case studies for the next legislative wave.

6What Changed: January to June 2026

Data-Source Reference Pack