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. Since the January 2026 edition of this paper, every headline constraint has tightened: the US interconnection queue has grown from 2,300 to roughly 2,600 GW; water has moved from regulatory barrier to measured, disclosed and litigated resource; and community opposition has progressed from sentiment to blocked capital, with US$98 billion of projects delayed or stopped in a single 2025 review window.
The thesis is unchanged and strengthened: scaling AI without breaking the planet is not only possible, it is becoming a competitive advantage. The organisations that lead the next phase will understand power systems, water constraints, grids and communities as deeply as they understand models and code. The new evidence of this edition shows the cost of the alternative: capital stranded in queues, permits revoked, and a regulatory wave (300+ bills across 30+ US states in 2026 alone) arriving faster than most operators' planning cycles.
All headline metrics refreshed to June 2026. Water section substantially expanded with measured consumption data, disclosure legislation, and permitting case law. Grid section updated for PJM's reformed Cycle 1 intake and ERCOT's 410 GW large-load queue. Social-licence section updated with blocked-capital figures. The orbital escape-hatch analysis, previously drafted as a V4.0 section, has been expanded into a full companion paper: The Orbital Mirage (Paper II).
In five months, the constraint set has not eased anywhere. It has tightened everywhere it is measured.
| Metric | Jan 2026 edition | June 2026 | Direction |
|---|---|---|---|
| US grid interconnection queue | 2,300 GW | ~2,600 GW | Worsened |
| US data-centre share of electricity | 4%, heading to 12% by 2030 | 12% now projected by 2028 (LBNL) | Accelerated |
| PJM connection reality | 8+ year timelines | Cycle 1: 220 GW applied (Apr 2026) vs ~3 GW connected in all of 2025 | Worsened |
| ERCOT large-load queue | Not tracked | 410 GW, of which 87% data centres; 198 GW applied in Q1 2026 alone | New constraint |
| Texas data-centre water | 399B gal projected by 2030 | 50B+ gal measured in 2024; 2030 projection unchanged | Baseline confirmed |
| Blocked or delayed capital | Anecdotal | US$98B (Mar–Jun 2025); 25+ projects cancelled in 2025 | Quantified |
Critical markets under strain. Northern Virginia: interconnection delays with grid moratoriums emerging, and data centres already consuming roughly 40% of Virginia's total electricity (2024). Texas: a large-load queue equal to ERCOT's entire current peak demand, with utility-imposed transparency requirements on duplicate requests. Phoenix: extreme water stress compounding a thin reserve margin. Singapore: 6% reserve margin, extreme water stress and a tropical cooling climate, a triple constraint directly relevant to APAC operators.
In practice, electricity access remains the schedule driver, and the queue is no longer a pipeline but a dam: PJM has signed 103 GW of interconnection agreements since 2020, yet only 23 GW have entered service, with 74% of studied capacity withdrawing at some point. Wholesale power costs in PJM jumped 54% in a single year, prompting a federally mandated auction cap.
January's edition reported water as a permitting barrier. By June, it is a measured resource with federal disclosure legislation, enforced discharge rules, and revoked permits. This is the fastest-moving constraint in the paper.
US data centres directly consumed a measured 17.4 billion gallons in 2023 (EPA / LBNL), with projections of 38–73 billion gallons by 2028, and hyperscale facilities expected to account for half. Texas facilities alone consumed more than 50 billion gallons in 2024, validating the trajectory toward the 399-billion-gallon 2030 projection carried in earlier editions. Globally, the IEA forecasts data-centre water withdrawals exceeding 1,200 billion litres annually by 2030. Operator disclosures confirm the curve: Google's consumption rose from 4.3 billion gallons (2021) to 6.1 billion gallons (2024), with its largest campus consuming 2.8 million gallons per day.
Four developments since January change the operating environment. EPA discharge rules took full effect in 2026: NPDES permits are now required for cooling-tower blowdown, with limits on temperature, total dissolved solids, pH, PFAS and biocides. Federal disclosure legislation arrived: the Data Center Water and Energy Disclosure Act (introduced March 2026) would mandate reporting of both energy and water consumption. Permits are being revoked, not merely delayed: Google's Chile permit was partially revoked, and groundwater disputes continue in Arizona. State legislatures have engaged at volume: more than 300 data-centre bills across 30+ states in 2026, spanning moratoriums to disclosure mandates, with construction moratoriums plausible in multiple states by year-end.
The technology spread remains roughly a 100:1 difference in water intensity: evaporative cooling at 1.8–2.6 L/kWh; immersion at 0.05–0.10; best-in-class geothermal and dry cooling at 0.02 (Meta Prineville). What changed is the consequence of ignoring the spread: siting in a water-stressed county with evaporative cooling is no longer a cost decision but a permitting and litigation exposure. Water Usage Effectiveness has become a board-level metric.
Two-thirds of new hyperscale campuses since 2022 sit in high or extreme water-stress counties. That is not a siting strategy; it is a class-action queue. The A+ Pathway's first pillar, appropriate siting, is now also the cheapest legal defence available to an operator.
The A+ Pathway is not a single technology. It is a system-level approach built on five reinforcing pillars, unchanged since V2 because the physics has not changed.
| Facility | Approach | PUE | g/kWh CO₂e | L/kWh water |
|---|---|---|---|---|
| Google Mayes County | AI-controlled HVAC + 24/7 wind PPA | 1.06 | 155 | 0.2 |
| Meta Prineville | Geothermal + dry cooling | 1.09 | 0 | 0.02 |
0–6 months: publish an open carbon ledger v1.0 with real-time CO₂e per job at under 2% error. 6–18 months: commission 5 MW immersion edge pods at PUE below 1.05 and water below 0.2 L/kWh. 18–36 months: secure 24/7 clean PPAs covering 50% of load with hourly matching above 90%. 36–60 months: deploy an SMR-anchored 1 GW campus at carbon intensity below 50 g/kWh, noting that IAEA timelines still place commercial SMR deployment in the early 2030s.
We are in an AI arms race. The unasked question remains: racing toward what?
The workload structure has inverted: training dominated compute in 2020–2022 (70–80%); by 2024–2026 inference dominates (60–70%) and grows at roughly 122% CAGR. Enterprise budget allocation (BCG, 2025) places support functions at 38%, operations 23%, marketing and sales 20%, R&D 13%, and climate-critical applications below 3%. Consumer scale compounds it: 800 million weekly ChatGPT users, with each query consuming roughly ten times the electricity of a traditional search (2.9 vs 0.3 watt-hours).
The carbon math has not improved with efficiency, because demand growth absorbs every gain within roughly eighteen months: carbon per unit of compute falls while absolute emissions rise. This is Jevons' paradox operating at silicon speed, and it is the strongest argument that allocation, not only efficiency, must enter the governance conversation. AI could mitigate 5–10% of global emissions by 2030, but only if the compute is allocated to let it. We are currently allocating the majority of inference to engagement, content generation and advertising optimisation.
January's edition argued opposition was economically rational. June's evidence shows it is economically effective.
The bill impacts that anchored the January analysis have continued to materialise: PJM capacity costs flowing through to residential bills ($18/month Maryland, $16/month Ohio), with projected increases of 8% on average across the US by 2030 and 25% or more in the highest-demand markets. Virginia illustrates the end state: nearly 600 facilities consuming on the order of 40% of state electricity, with more than 100 additional projects proposed.
The strategic conclusion strengthens with each data point: transparent engagement is now an approval prerequisite, and community impact has a price discoverable in blocked capital. Operators who treat social licence as communications rather than engineering are funding the case studies for the next 300 bills.
| Constraint | June 2026 status | Impact |
|---|---|---|
| Grid interconnection | ~2,600 GW queued; PJM 220 GW applied vs ~3 GW connected (2025); ERCOT 410 GW large-load queue | Primary schedule driver |
| Water | Measured baselines; NPDES enforcement; disclosure legislation; permit revocations | Permitting and litigation exposure |
| AI power growth | 12% of US electricity by 2028 (LBNL), accelerated from 2030 | Consumer bill increases |
| Social licence | US$98B blocked or delayed; 300+ bills in 30+ states | Capital-at-risk, quantified |
V3.1 (January 2026): terrestrial constraint set, A+ Pathway, misallocation analysis. V4.0 (May 2026, unpublished): added an orbital-horizon section. V4.1 (June 2026, this edition): orbital analysis expanded and moved to its own companion paper; all terrestrial data refreshed; water section rebuilt around measured baselines and the 2026 regulatory wave.
The closing position of every edition stands, with more evidence behind it each time it is retested: AI does not fail because of lack of ambition. It fails when physical limits are ignored. Scaling AI without breaking the planet is not only possible; it is becoming the competitive advantage. Success belongs to operators who solve physics, not those who optimise spreadsheets, and, as Paper II now shows, not to those who believe physics ends at the Kármán line.
US DOE / LBNL interconnection and demand reports (2024–26) · queue ~2,600 GW early 2026; data centres 12% of US electricity by 2028.
PJM Interconnection (2025–26) · Cycle 1 intake 811 projects / 220 GW (Apr 2026); 103 GW agreements signed since 2020, 23 GW in service; 74% withdrawal; wholesale costs +54% in one year; capacity shortfall to 15 GW by 2030.
ERCOT / Ascend Analytics (2026) · 410 GW large-load queue, 87% data centres; 198 GW applied Q1 2026.
Goldman Sachs (2024); IEA Energy & AI (2024) · global data-centre electricity 415 TWh (2024) → 945 TWh (2030); US bills +25% in highest-demand markets by 2030.
EPA / Shehabi et al., LBNL (2024–25) · US direct consumption 17.4B gal (2023); 38–73B gal projected by 2028; hyperscale ≈ half.
Texas analyses (2025–26) · 50B+ gal consumed 2024; Texas Water Development Board 399B gal/yr projection by 2030; SB 7 active.
IEA (2025–26) · global data-centre withdrawals >1,200B litres/yr by 2030.
Operator disclosures (2025) · Google 6.1B gal consumed 2024 (4.3B in 2021); largest campus 2.8M gal/day; Equinix 1.2B gal consumed 2024.
Bloomberg (May 2025) · two-thirds of new US hyperscale campuses since 2022 in high/extreme water-stress counties.
Regulatory record (2026) · EPA NPDES cooling-tower discharge rules in full effect; Data Center Water and Energy Disclosure Act (Mar 2026); Google Chile partial permit revocation; Santa Clara recycled-water mandate; Johor 30% rejection rate.
LBNL (2022) · evaporative cooling 1.8–2.6 L/kWh. Submer / Asperitas / Icetope (2023–24) · immersion PUE 1.03–1.10, WUE 0.05–0.10. Meta Prineville (2024) · 0.02 L/kWh, zero carbon intensity. Google Mayes County · PUE 1.06, 155 g/kWh, 0.2 L/kWh.
IEA Net Zero Pathways (2023–24) · data-centre emissions 2–3% of global CO₂e by 2030. Google/BCG (2023) · AI mitigation potential 5–10% of global GHG by 2030.
Deloitte (2025); AllAboutAI (2025) · inference 60–70% of AI energy. Bloomberg Intelligence (2024); Bain (2025) · demand at 2× Moore's-law rate; global requirement toward 200 GW. IDC (2025) · AI infrastructure spend $82B in Q2 2025, +166% YoY.
BCG (2025) · enterprise allocation: support 38%, operations 23%, marketing/sales 20%, R&D 13%, climate-critical <3%. a16z (2025) · ChatGPT 800M weekly users. OpenAI (2025) · 2.9 vs 0.3 Wh per query.
Data Center Watch (2025) · US$98B blocked/delayed Mar–Jun 2025. Heatmap Pro (Jan 2026) · 25+ cancellations in 2025. Multistate (2026) · 300+ bills, 30+ states. PJM capacity auction (2025–26) · $9.3B increase; $18/mo MD, $16/mo OH. Bloomberg (2026) · Virginia data centres ≈ 40% of state consumption (2024).
Lira, H. (2026). The Orbital Mirage: Environmental Externalities of Space-Based AI Infrastructure and the Limits of Off-Planet Escape. HML Services Ltd, Paper II, June 2026.
© 2026 HML Services Ltd. All rights reserved. · hml-services.com · info@hml-services.com