US Grid Crisis + EU Enforcement Collapse = Europe's AI Infrastructure Advantage

The US Energy Arithmetic: Supply Can't Match Demand

The US power grid crisis is now quantified with precision. PJM Interconnection, serving 65 million people across primary AI data center corridors, fell 6.6 GW below reliability requirements in its 2027-2028 capacity auction. Capacity prices hit a record $333.44/MW-day. The fundamental arithmetic is brutal: data centers add 5-7 GW annually while new generation delivers only 2-3 GW. Grid interconnection queues average 5 years.

BloombergNEF projects US data center power demand reaching 106 GW by 2035, with 31 GW from the PJM region alone. Gartner projects 40% of AI data centers will be power-constrained by 2027. Lawrence Berkeley National Laboratory estimates US data center demand reaching 325-580 TWh by 2028 (up from 176 TWh in 2023) — a near-tripling in 5 years.

This energy constraint creates involuntary geographic fragmentation of AI compute. Training runs — which are latency-insensitive but power-intensive — are the first workloads to migrate. Norway's abundant hydropower (cheap, renewable, and available without 5-year interconnection queues) makes it the most attractive training destination outside the US.

Nscale Capitalizes on Dual Structural Failures

Nscale's $2 billion Series C (March 9, 2026) raised the company at a $14.6 billion valuation. The company has accumulated over $4 billion in 12 months to build GPU infrastructure in Norway, with Stargate Norway targeting 100,000 NVIDIA GPUs. Microsoft is already contracted for 200,000 GPUs across multiple Nscale facilities.

The investor composition — NVIDIA, Citadel, Jane Street, Dell — signals that hardware, compute-intensive finance, and sovereign infrastructure are converging on the same thesis. Board additions of Sheryl Sandberg (former Facebook COO) and Nick Clegg (former Meta Global Affairs, deep EU AI Act expertise) position the company for both IPO (18-24 months) and EU regulatory navigation.

The bitcoin mining advantage is underappreciated. Nscale was spun out of Arkon Energy — a bitcoin mining infrastructure company. The same power-dense, cooling-optimized facilities that mine bitcoin can host GPUs. This means the physical infrastructure buildout is faster and cheaper than greenfield data center construction, giving Nscale a 12-18 month head start over competitors building from scratch.

EU Enforcement Collapse Creates a Buildout Window

Only 8 of 27 EU member states have designated enforcement authorities — a deadline that was August 2, 2025 and is now 7+ months overdue. The EU has pushed high-risk deadlines to December 2027 (standalone) and August 2028 (embedded).

This creates a paradoxical opportunity: European AI infrastructure can be built and operationalized before the regulatory framework that governs it is enforceable. Companies building in Europe today face minimal regulatory overhead while gaining the 'EU-sovereign' positioning that will be valuable when enforcement materializes.

The residential electricity impact compounds the political dimension. US retail electricity prices have risen 42% since 2019, outpacing CPI by 13 percentage points. Goldman Sachs projects data center demand will add 0.1% to core inflation in both 2026 and 2027. Average household bills could rise $15-25/month. This creates political backlash against AI data center expansion in US communities — further pushing compute investment to jurisdictions where power is abundant and where AI data centers are politically welcomed (Nordic countries, where they represent economic development).

The Strategic Convergence: Three Tailwinds

Nscale occupies the intersection of three tailwinds:

• US grid constraints push compute demand to Norway: The 6.6 GW shortfall and 5-year interconnection queues make US-based training expansion economically irrational for anyone with access to Norwegian hydropower.
• EU enforcement delays reduce near-term compliance burden: Building capacity between now and December 2027 means Nscale achieves scale before regulatory compliance costs spike.
• EU sovereignty concerns guarantee long-term demand: The political desire for European-independent AI infrastructure will sustain demand for European-based compute once enforcement materializes.

Nick Clegg's board appointment (former Meta Global Affairs head, deep EU AI Act expertise) signals that Nscale is positioning for the world where regulation eventually arrives — but building capacity in the window where it has not.

What This Means for ML Engineers

Training workload strategy: If you have large training runs planned for 2026-2027, evaluate European cloud providers (Nscale, CoreWeave EU) for cost-per-FLOP advantages driven by lower energy costs. Norwegian hydropower offers 30-50% cost savings versus US grid rates, and the gap will widen as US electricity prices continue rising.

Sovereign infrastructure consideration: Teams with EU data residency or compliance requirements gain dual benefit: sovereign infrastructure positioning + lower training costs. This is a rare alignment where regulatory requirement and economic advantage point in the same direction.

Inference workload geography: Inference workloads remain latency-constrained to user-proximate locations. The energy advantage for training does not apply to real-time inference, which must be deployed closer to end users. Plan for geographic split: training in energy-rich regions (Norway), inference in user-proximate regions (AWS/Azure edge, local clouds).

Long-term hedging: Monitor US nuclear power investments (Microsoft's Three Mile Island restart, Oklo's NVIDIA partnership) as potential countermeasures. If US nuclear deployment accelerates beyond current timelines, the energy cost advantage for European infrastructure may erode in 3-5 years.