April 02, 2026

A Virginia appeals court just put a hard stop on one of the largest contested data-centre rezonings in the US. The decision blocks the Prince William Digital Gateway, a 2,000+ acre plan for 37 data centres and 14 substations that opponents said could ultimately pull 3+ GW. In a market where “just add power” has become the defining constraint, this is a reminder that the tightest bottleneck isn’t always a transformer — it’s permission.

The Big Stories

The Virginia Court of Appeals’ move to block the Prince William Digital Gateway halts a development that would have rewritten the scale map in Northern Virginia: 37 data centres, 14 substations, and a land footprint measured in thousands of acres. The Piedmont Environmental Council’s filings pointed to park, water-quality, and historic-resource impacts, plus the practical implication of 3+ GW of demand and Dominion Energy substation actions not landing until late 2025. Why it matters: the ruling underlines that the next wave of mega-campuses faces a two-front fight — grid timelines and local legitimacy — and either one can kill velocity.

On the demand side, the numbers are getting surreal. OpenAI raised $122 billion at an $852 billion valuation explicitly to expand compute capacity, cloud partnerships, and data-centre infrastructure, while also expanding a credit facility to $4.7 billion. The company says its platform is processing more than 15 billion tokens per minute and is spreading infrastructure across Microsoft, Oracle, AWS, CoreWeave, and Google Cloud with multiple silicon suppliers including Nvidia. Why it matters: this is capital formation at a scale that can overwhelm regional power planning — and it reinforces the “multi-vendor, multi-cloud” posture as the default for anyone trying to buy their way out of compute scarcity.

Europe’s homegrown AI ambitions are also shifting from rhetoric to steel-and-concrete. Mistral AI’s infrastructure push includes a Paris data centre planned with 13,800 NVIDIA GB300 GPUs, $830 million in debt, and a stated ambition to build 200 MW across Europe by end-2027, alongside a partnership with Nvidia and an acquisition of Koyeb. Why it matters: Mistral is acting less like a model shop and more like an operator with a balance sheet — which raises the competitive bar for European peers who hoped to rent their way to relevance.

Semiconductor capacity — not just data-centre capacity — is being pulled back under tighter corporate control. Intel will repurchase Apollo’s 49% stake in Fab 34 in Leixlip, Ireland for $14.2 billion, unwinding a 2024 minority sale that raised $11.2 billion. Intel will fund the buyback with cash plus about $6.5 billion of new debt, framing it as regaining full control of EUV manufacturing capacity tied to AI infrastructure and European chip sovereignty. Why it matters: if you believe AI infrastructure is a national-security-adjacent supply chain, this is Intel aligning capital structure with that reality — and paying up to do it.

India continues to be the most straightforward “follow the megawatts” story in the public market-adjacent ecosystem. Bharti Airtel’s Nxtra Data is raising $1 billion at an implied ~$3.1 billion valuation, with Alpha Wave Global, Carlyle, Anchorage Capital, and Airtel itself participating. Nxtra plans to expand its network (14 core sites and 120+ edge facilities) and build AI-focused campuses as it targets a jump from ~300 MW to 1 GW in the next few years. Why it matters: the scale target is the headline, but the subtext is that India’s demand narrative is now strong enough to attract growth capital specifically for AI campuses, not just generic colocation.

Behind the Headlines

Europe’s AI problem is increasingly an electricity problem — and policy people are starting to say it out loud. An ECFR paper argues the EU needs a “fast energy” programme to speed permitting and grid buildout while deploying “all available power technologies,” from nuclear and CCGTs to batteries and renewables, explicitly to keep AI electricity affordable and reliable (Europe must launch a ‘fast energy’ programme to power AI). It name-checks US-scale investment narratives (including “Stargate” at ~$500bn) and warns Europe will fall behind the US and China without urgency. The useful framing here is that “AI competitiveness” is being recast as “execution speed on power infrastructure,” which is a very different political fight than subsidies for chips or models.

The grid constraint is also spawning a new category: software designed to make data centres behave more like flexible industrial loads. Emerald AI’s $25M raise backs its “Conductor” platform to align data-centre energy use with grid capacity, with a cap-table that includes Energy Impact Partners plus strategic names across power equipment and cloud ecosystems, and an advisory board including NVIDIA and National Grid. The deeper point: interconnection queues and curtailment risk are turning “when you run” into a first-order design variable — especially for AI workloads that can be scheduled, paused, or shifted if the economics work.

Then there’s the long-shot-but-not-crazy siting idea: put the data centre on water. MOL and Hitachi’s MoU to develop floating data centres targets first operational units by 2027, with MOL handling vessel conversion and maritime ops and Hitachi leading technical design and IT infrastructure, including feasibility work using a second-hand vessel. This isn’t about novelty; it’s about escaping the land-use and grid-connection politics that are getting sharper in top-tier metros. If the concept survives the economics (power delivery, cooling systems, insurance, and uptime expectations), it could become a niche option for latency-tolerant workloads where “permitting certainty” is worth paying for.