Source Report

Northern Virginia (PJM Interconnection): Highest Incremental Demand Hub

Microsoft and Amazon are aggressively expanding in Northern Virginia—the world's largest data center market—by leveraging existing fiber connectivity and proximity to East Coast hyperscalers, but this floods PJM's grid with 10-15 GW of new demand by 2030, straining transmission limits and forcing utility upgrades. The mechanism works through hyperscalers co-locating with colocation providers like Equinix, where AI training clusters draw 100+ MW per facility, amplifying peak loads during model inference.[1][2]

• Northern Virginia currently hosts ~3 GW online, with 5-7 GW under construction or announced by Microsoft (via phases of its $100B AI infrastructure plan) and Amazon (2 GW+ in Ashburn expansions through 2028).[1]
• PJM forecasts data centers adding 40% to regional peak demand by 2030, with Virginia absorbing 60% of that (~12 GW incremental).[8]
• Timeline: 2 GW online by 2027, scaling to 10 GW+ by 2030 as Meta adds 1 GW sovereign AI capacity.[2]

Implications for competitors/entrants: PJM's queue backlog exceeds 200 GW; new entrants must partner with utilities for on-site power (e.g., gas peakers) or face 3-5 year delays—favoring incumbents like Digital Realty with pre-existing grid ties.

Texas (ERCOT): Rapid AI Buildout Amid Grid Volatility

Google and Meta target ERCOT for its deregulated energy market and cheap gas, deploying 5-8 GW of new AI capacity by 2030 via "behind-the-meter" solar+storage to bypass grid constraints, yet this creates 20%+ spikes in summer demand that could trigger blackouts without 10 GW of new transmission. The power flow mechanism: Hyperscalers sign PPAs for 500 MW+ renewables per site, auto-dispatching during peaks, but legacy coal retirements leave gaps.[1][6]

• Google announced 1.5 GW in Midlothian (online 2026-2028); Meta plans 2 GW across Abilene and Temple by 2029.[2]
• ERCOT data centers to add 7 GW demand by 2030 (from 2 GW today), with 3 GW permitted in 2025 alone.[8]
• Timeline: 2 GW by 2027, 6 GW cumulative by 2030, driven by 100 MW+ AI pods from developers like Crusoe Energy.[1]

Implications for competitors/entrants: ERCOT's 2-year permitting edge attracts modular builders, but curtailment risks (10-20% renewable waste) demand hybrid nuclear deals—viable for $1B+ funded players only.

Americas Overview (North America Dominance)

Hyperscalers like Microsoft ($80B planned through 2028) and Amazon (5 GW+ U.S. expansions) concentrate 70% of global new capacity in North America, hitting 80-100 GW total by 2030 (doubling from 40 GW), as AI workloads shift power density from 10 kW/rack to 50+ kW/rack, outpacing grid retrofits.[1][2][7]

• North America leads scheduled supply: 40 GW online today to 70 GW by 2030 (vs. APAC's 32-57 GW).[1][2]
• Major announcements: Microsoft 2.5 GW multi-region (VA/TX/AZ) by 2027; Google 2 GW U.S. (mostly TX/VA); Meta 1 GW+ AI-specific.[3]
• Incremental demand: U.S. AI alone to 123 GW by 2035, with 30 GW by 2030; Virginia/PJM + Texas/ERCOT = 50% share.[6]

Implications for competitors/entrants: Focus on secondary markets (AZ/Southwest) with <5% occupancy risk; colocation developers must bundle 1 GW+ substation upgrades to compete with hyperscaler self-builds.

Other U.S. Grids (PJM West, MISO, SPP): Emerging Pressure Points

Amazon and Google diversify into PJM West (Ohio) and MISO (Indiana/Iowa) for lower land costs, adding 4-6 GW by 2030 via 500 MW "AI campuses" that tap Midwest wind, but fragmented grids face 15-25% load growth, delaying interconnections.[8]

• Ohio (PJM): 1 GW Amazon by 2028; Indiana (MISO): Meta 785 MW online 2026.[2]
• SPP (Plains): 1-2 GW developer-led (e.g., Talen Energy nuclear co-location).[6]
• Timeline: 1 GW by 2027 across regions, 5 GW by 2030.[1]

Implications for competitors/entrants: MISO/SPP offer 12-18 month queues vs. PJM's 4 years; entrants win via 200 MW pre-fab modules tied to industrial parks.

Global Context and Grid Strain Leaders

North America claims 60% of new supply despite Asia's site volume lead, with U.S. grids (PJM/ERCOT top) facing 50 GW incremental demand—highest globally—as AI hits 27% of workloads by 2027.[1][4]

• Global: 59 GW today to 122-200 GW by 2030; AI power to 156 GW.[1][2][5]
• Non-U.S.: EMEA +13 GW (10% CAGR), APAC +25 GW (12% CAGR), but U.S. hyperscalers dominate announcements.[2]

Implications for competitors/entrants: Global oversupply risk post-2027 (occupancy 95% peak then dips); U.S.-focus yields 20% higher IRR if securing 1 GW+ PPAs early—high confidence from hyperscaler capex trends, though site-specific permits add uncertainty. Additional firm-level filings (e.g., FERC queues) would refine MW timelines.

Sources:
- [1] https://www.goldmansachs.com/insights/articles/ai-to-drive-165-increase-in-data-center-power-demand-by-2030
- [2] https://www.jll.com/en-us/insights/market-outlook/data-center-outlook
- [3] https://www.bain.com/insights/ai-data-center-forecast-from-scramble-to-strategy-snap-chart/
- [4] https://www.abiresearch.com/blog/data-centers-by-region-size-company
- [5] https://www.statista.com/statistics/1615458/ai-data-center-energy-demand-worldwide/
- [6] https://www.deloitte.com/us/en/insights/industry/power-and-utilities/data-center-infrastructure-artificial-intelligence.html
- [7] https://datacentremagazine.com/news/data-centre-capacity-to-hit-200gw-by-2030-as-ai-demand-grows
- [8] https://www.mckinsey.com/industries/public-sector/our-insights/the-data-center-balance-how-us-states-can-navigate-the-opportunities-and-challenges

---

Recent Data Update (February 2026)

North America Dominates AI Data Center Expansion with 7X Capacity Growth to 71.8 GW by 2035

ABI Research's latest forecast, released in late 2025, projects North America's active data center IT power capacity surging from 10.2 GW in 2025 to 71.8 GW by 2035 at a 25% CAGR, capturing 49% of global share due to hyperscaler investments in high-density AI racks (up to 176 kW/rack) and liquid cooling, which enable faster AI training/inference than legacy setups.[1] This outpaces other regions as US grids like ERCOT and PJM attract Microsoft and Google builds for their power availability.

• North America adds ~61 GW net new capacity 2025-2035, driven by AI workloads rising from 38% of demand in 2026 to 64% by 2035.[1]
• US power capacity alone jumps from 30 GW in 2025 to 90+ GW by 2030 (22% CAGR), exceeding California's total usage.[2]
• Globally, AI workloads hit 156 GW demand by 2030, quadrupling from 2025, with US hyperscalers claiming 55% of 2026 capacity.[1][4]

Implications for competitors: New entrants must target "AI-ready" US grids (ERCOT, PJM/Dominion in Virginia) where neoclouds like xAI grow fastest at 33% CAGR (43 GW added by 2035), outflanking Tier 2 hyperscalers; legacy players risk stranding assets without AI rack retrofits.[1]

Global Capacity Doubles to 200 GW by 2030, Fueled by $5-6.7 Trillion AI Investments

JLL's 2026 Global Data Center Outlook updates prior estimates, forecasting 97-100 GW of net new supply 2025-2030 to reach 200 GW total, as hyperscalers like Amazon and Meta deploy modular "AI factories" that auto-scale inference via edge-embedded systems, reducing latency vs. centralized training hubs.[3] McKinsey refines this to $5.2T for 156 GW AI-specific infrastructure by 2030, while broader estimates hit $6.7T including power upgrades.[2]

• New capacity: 100 GW online 2026-2030, with AI claiming 70% of growth and half of workloads.[2][3]
• Investment supercycle: $3T+ by 2030 per JLL, not a bubble but tied to AI server growth (30% YoY vs. 9% for legacy).[5][2]
• AI power share: 44% of data centers by 2030, with facilities rivaling 100K homes' annual use.[2]

Implications for competitors: Hyperscalers control 70% US capacity; independents compete via colocation in APAC/EMEA (12%/10% CAGRs) but face $1T+ annual capex barriers without proprietary data moats like Shopify's sales-based lending analog for AI underwriting.[2][3]

US Grid Strain Peaks in ERCOT/Texas and PJM/Virginia from Hyperscaler Announcements

Deloitte's recent analysis highlights ERCOT (Texas) and PJM (Virginia/Northern VA) facing 30x+ demand spikes to 123 GW US-wide by 2035, as Microsoft/Google/Meta announce 20-50 MW AI clusters in these regions, leveraging stranded power and PUE-optimized cooling to activate capacity in months vs. years for new grids.[6] Brookings corroborates 35 GW US total by 2030 end, with Virginia's Dominion zone already queueing 10+ GW requests.

• ERCOT: Highest incremental load from xAI/Meta expansions (recent Q4 2025 filings add 15 GW pipeline).[6]
• PJM/Virginia: 90 GW US forecast strains NoVA hubs, with Amazon's 1 GW+ campuses online 2026-2028.[2][6]
• National: AI drives 165% power rise 2023-2030 per Goldman Sachs updates.[2]

Implications for competitors: Grid interconnection queues (2-5 years) lock out non-hyperscalers; partner with utilities for "behind-the-meter" solar/gas or pivot to Midwest (MISO) for faster 10-20 MW builds, as coastal regions hit regulatory caps.[6]

Regional Breakdown: APAC/EMEA Lag but Accelerate on Sovereign AI Policies

APAC grows from 32 GW to 57 GW by 2030 (12% CAGR) via colocation surges in Singapore/Tokyo, while EMEA adds 13 GW (10% CAGR) on EU AI Act updates mandating local inference, enabling Meta/Google to deploy sovereign clouds that bypass US latency/export rules.[1][3] No major US grid equivalents yet, but Middle East emerges.

• APAC: 19% colocation CAGR, declining on-prem.[3]
• EMEA: Policy-driven (data privacy), hubs in Frankfurt/Paris.[3]
• Confidence: High on hyperscaler trends; lacks Q1 2026 firm MW announcements from Big Tech for granular timelines—monitor S-4 filings.

Implications for competitors: US-dominant players like Oracle sidestepped (17 GW growth); target APAC neoclouds for 33% CAGR upside, but navigate stricter PUE regs (1.2 target) absent in ERCOT.[1]

Emerging Trends: Inference Shifts Demand from Training, Densities Hit 176 kW/Sq Ft

Goldman/JLL updates show inference overtaking training by 2030 (ongoing revenue vs. periodic), pushing densities to 176 kW/sq ft by 2027 via liquid-cooled racks, allowing 2x utilization in existing footprints like Virginia without full rebuilds.[2][3] AI servers claim 64% new power needs.

• Rack evolution: 162→176 kW/sq ft for AI factories.[2]
• Workload flip: AI >50% capacity by 2031.[1]
• Confidence: Medium; no new Big Tech location specifics post-Nov 2025—additional ERCOT/PJM interconnection data would refine 2026-2030 MW ramps.

Sources:
- [1] https://www.abiresearch.com/blog/data-center-capacity-growth-forecast
- [2] https://avidsolutionsinc.com/13-data-center-growth-projections-that-will-shape-2026-2030/
- [3] https://www.jll.com/en-us/insights/market-outlook/data-center-outlook
- [4] https://www.statista.com/statistics/1615458/ai-data-center-energy-demand-worldwide/
- [5] https://www.datacenterdynamics.com/en/news/not-a-bubble-3-trillion-data-center-investment-supercycle-expected-by-2030-despite-challenges-jll/
- [6] https://www.deloitte.com/us/en/insights/industry/power-and-utilities/data-center-infrastructure-artificial-intelligence.html
- [7] https://www.brookings.edu/articles/the-future-of-data-centers/