EIA: data-center servers hit 7% of commercial electricity in 2025, on a path to 22%–33% by 2050

What happened. EIA’s Today in Energy note by Courtney Sourmehi unpacks Annual Energy Outlook 2026 for data-center server electricity inside the commercial building stock. Servers alone took an estimated 7% of commercial-sector electricity in 2025. Across AEO cases they grow to 22%–33% of commercial-building electricity use by 2050, with server consumption in a 446–818 billion kWh range. The high case puts standalone data-center server use at 581 billion kWh in 2050. Commercial electricity intensity is projected to pass the 2003 high of 14.9 kWh per square foot in 2031–2032. AEO treats laws and regulations mostly as of December 2025. This is a long-run scenario path, not an observed 2026 megawatt clock and not a facility water inventory.

What to watch. After diplomatic reuse rules that named data centers, the environment beat is whether commercial-building electricity growth is really a server story — architecture of the load, not another interconnection queue or gallon total.

Read the EIA AEO2026 server-energy note →

Dallas Fed: higher automatable-task share tracks weaker Texas outcomes for new graduates

What happened. A Dallas Fed Economics note by Samuel Dodini and Tucker Smith (22 September 2026) links Texas four-year public-university graduates to state earnings records and major-level automation exposure. After ChatGPT, a 10-percentage-point higher share of automatable tasks is associated with a 1.7-percentage-point relative drop in the chance of Texas employment within a year of graduating, and about 5% lower first-year wages from 2021 to 2024 versus less-exposed majors. May 2024 graduates from more-exposed majors were 1.4 points more likely to re-enroll in graduate school. Undergraduate enrollment in more-exposed majors fell 4.8% from fall 2024 to fall 2025 for each 10-point exposure gap. Most-exposed majors include computer science and computer engineering; least-exposed include nursing, education, and psychology. This is a Texas administrative event study, not a national layoff census and not a job-ad language paper.

What to watch. After U.S. job-ad AI-mention elasticities, the jobs beat is whether major choice and new-graduate placement already bend when automation risk is high — employment, pay, grad school, and enrollment, not pink slips.

Read the Dallas Fed graduate-labor note →

GAO: OMB AI guidance covers only 2 of 10 expert privacy challenges

What happened. GAO report GAO-26-107681 (letter date 26 March 2026) says an expert panel identified ten selected privacy challenges for federal AI use. OMB’s government-wide AI guidance fully addresses two of them — workforce skills, and accelerating AI with privacy protections — and does not fully address the remaining eight. Guidance also does not specify the kinds of known privacy risks agencies should weigh when setting AI privacy policy. GAO made two recommendations to OMB; both were still open on the product page, and OMB did not comment. This is congressional oversight of guidance gaps, not a new evaluation-planning manual and not a live enforcement action.

What to watch. After NIST’s ARIA planning manual, the policy beat is whether OMB closes the privacy holes GAO named — examples of known risks, and information-sharing or guidance on the eight remaining challenges.

Read GAO-26-107681 →

IARC EUcanAI: four-year Pathfinder project starts for brain-tumour AI

What happened. IARC announced EUcanAI, a European Unified Cancer diagnostics and treatment AI agent, with an official start date of 1 September 2026. The four-year project is funded under EIC Pathfinder / Horizon Europe and led by Heidelberg University Hospital with IARC and 11 other institutions in Germany, Austria, and Denmark. IARC will provide the WHO Classification of Tumours of the central nervous system and the framework to design, train, and implement AI modules for consultation, diagnostics, and therapy planning. This is a project launch and research infrastructure build, not a completed trial and not a U.S. device-authorization count.

What to watch. After marketing-authorization tallies for AI devices, the health beat is whether multi-country Pathfinder work can put WHO CNS classification into real diagnostic and treatment AI modules — infrastructure first, outcomes later.

Read the IARC EUcanAI announcement →

UNESCO: 58 countries already supported on digital and AI competencies

What happened. At Digital Learning Week 2026 in Paris (8–11 September), education ministers pressed the case for education as a common good in the age of AI. UNESCO says that since 2024 its AI Competency Framework seminars have involved participants from more than 100 countries, and that it has directly supported 58 countries with frameworks, curricula, and quality-assured training on digital or AI competencies. The 2021 Recommendation on the Ethics of AI remains the backdrop, adopted by 194 Member States. A working document and six background papers are open for public comment, with briefs aimed at the next Digital Learning Week in March 2027. This is capacity-building reach, not a student-uptake census and not another teacher MOOC catalog entry.

What to watch. After free teacher-course packages, the education beat is how many countries already have help building digital and AI competency systems — frameworks and training capacity, not a single course page.

Read the UNESCO Digital Learning Week note →

Also in today’s ledger

• A commercial-server share path is a different evidence class from facility gallons and ISO interconnection clocks.

• A Texas graduate exposure paper is not a job-ad AI-mention elasticity — and an OMB privacy-gap report is not an ARIA planning manual.

• A Pathfinder project launch is not an authorization count — and 58-country competency support is not a teacher MOOC catalog.

Full ledger

This is the short version.

The complete source-linked ledger is on AI Footprint.

Open today’s full AI Footprint edition →

Open the dated September 28 edition →