// data centers

Source: Financial Times (paywall)

The Financial Times reports that intensive tech industry lobbying has pushed the EU toward diluting climate standards for gas-powered data centers in draft proposals. The exemptions undercut regulatory coherence as the bloc enforces the Digital Services Act and Green Taxonomy elsewhere. The move exposes tension between Europe's climate commitments and its accommodation of computational demands for AI training and cloud services—a precedent that invites similar carve-out requests from other industries citing critical infrastructure needs.

Source: Chartbook

While U.S. renewable energy capacity has expanded dramatically, AI model training is consuming electricity at a 40% annual efficiency gain rate—meaning chip makers are capturing productivity improvements faster than the grid can source them from wind and solar. Data center developers now compete directly with decarbonization goals for transmission capacity and water resources, particularly in water-scarce regions where both cooling and renewable generation depend on the same scarce input. Efficiency gains in AI chips no longer translate to reduced energy demand when training duration is simultaneously increasing 25% annually.

Source: Semianalysis

AI infrastructure operators are rapidly building private power generation and storage behind their own meters rather than relying on already-strained regional grids, with projections suggesting 40GW+ of capacity could be self-hosted by 2028. This fragments energy infrastructure and creates a structural decoupling where hyperscalers effectively become their own utilities, controlling generation, distribution, and consumption without grid arbitrage or oversight. The shift creates clear winners and losers. Companies with capital for solar, nuclear, and battery clusters gain energy independence. Utilities and regions lose datacenter tax revenue and grid stability contributions. It also exposes how quickly supply-constrained infrastructure becomes privatized when the public system can't adapt—a template that may apply to other critical systems when centralized capacity fails to scale.

Source: Works in Progress

Data center power delivery has become the hard constraint on AI expansion, not semiconductor manufacturing. Abilene's $20 billion Lantana power project typifies how utilities and grid infrastructure now limit GPU cluster placement. The bottleneck has shifted from Silicon Valley's chip design cycle to Texas utility politics and transmission line permitting, where 18-month Environmental Impact Statements matter more than TSMC's fab capacity. Cloud giants are scrambling to secure nuclear power contracts. Grid operators, not OpenAI, effectively control the pace of model training.