EIA AEO2026: data-center servers reach 7% of commercial electricity in 2025 and 22%–33% of commercial-building use by 2050.
What happened: EIA Today in Energy (19 May 2026; Courtney Sourmehi) unpacks Annual Energy Outlook 2026 Counterfactual Baseline and High Electricity Demand cases for data-center server electricity inside the commercial building stock. AEO2026 released 8 April 2026; most cases consider laws and regulations only as of December 2025. Share lock: servers alone accounted for an estimated 7% of commercial-sector electricity in 2025, growing to 22%–33% of commercial-building electricity use by 2050 across cases. Census lock: by 2050, server electricity reaches 446–818 billion kWh; the high end is High Electricity Demand (faster server power-draw and installed-stock growth). Standalone data centers (other-buildings category) reach 581 BkWh of server consumption in 2050 in that high case. Intensity lock: commercial electricity intensity exceeds the 2003 historical high of 14.9 kWh per square foot for the first time in 2031–2032. Data-center floorspace cooling assumed up to 2.9× as energy-intensive as non-data-center floorspace; holding population and weather fixed, High-case space-cooling electricity is 84 BkWh higher than Counterfactual Baseline in 2050. Load-shape lock: server end-use load is essentially flat across hours. Counterfactual Baseline assumes that after 2040, average annual operational power draw falls 10% every three years above historical efficiency trends; High Electricity Demand makes no such extra-efficiency assumption and assumes AI servers a larger share of installed stock. This is a commercial-building long-term server-share scenario range — not yesterday’s EPA G20 fit-for-purpose reuse deal, not Pacific Institute facility withdrawals 270 gal–3.9 million gal/day / typical 74% consumed, not AESO 1,200 MW, not NYISO 50 MW, not ISO-NE not-yet-at-scale / 16 November, not MISO’s 120-day path, not SPP’s 90-day CHILL/HILLGA path, not CAISO EL26-71, not PJM ride-through after nearly 4,000 MW, not ERCOT Batch Zero, not IEA observed-2025 +17% / 485→950, not LBNL 649 TWh / ~11.8%, and not EIA September STEO 4,368 BkWh.
Why it matters: After diplomatic water-reuse governance that named data centers, the environment beat is how EIA’s AEO now splits servers out of commercial computing and shows their long-run share of the commercial building stock — scenario architecture, not another interconnection clock or facility gallons inventory.
Dallas Fed: a 10 pp higher automatable-task share tracks a 1.7 pp drop in Texas employment for new four-year graduates.
What happened: Federal Reserve Bank of Dallas Economics note by Samuel Dodini and Tucker Smith, AI plays a role in weak labor market for college graduates (22 September 2026), links Texas four-year public-university graduates to THECB / Texas Workforce Commission earnings and major-level exposure from pre-ChatGPT Lightcast field-of-study mentions plus Anthropic Economic Index / O*NET task automation. Displacement lock: after late-2022 ChatGPT, a 10-percentage-point higher share of automatable tasks is associated with a 1.7-percentage-point relative decline in the probability of finding employment in Texas within a year of graduating (vs less-exposed majors; pre-period difference stable vs May 2021 graduates). Wage lock: among those who found Texas work, first-year earnings for more-exposed majors fell about 5% from 2021 to 2024 relative to less-exposed majors. Adaptation lock: May 2024 four-year graduates from majors with 10 pp higher exposure were 1.4 percentage points more likely to enroll in graduate school within a year; two-thirds of 2024 computer-science returners chose a computer-science graduate program. New master’s recipients in more-exposed fields saw similar relative earnings declines as four-year graduates — limited returns to formal upskilling within exposed fields. Enrollment lock: for a 10-percentage-point difference in major automatable-task shares, undergraduate enrollment at Texas four-year institutions declined 4.8% between fall 2024 and fall 2025. Most-exposed majors include computer science, computer engineering, and languages; least-exposed include nursing, education, and psychology. Limits locked: authors’ views, not the Dallas Fed or Federal Reserve System; Texas administrative records; out-of-state employment outside the earnings frame; no national layoff headcount. This is a Texas public-university graduate / major-exposure event study — not yesterday’s Cleveland Fed WP 26-24 (+1 SD → 3.1 pp job-ad AI mentions), not Dallas Fed 1 September Lightcast postings −5% / ~−8%, not ILO China enterprise package 21 / 1,591 / 56% / 47% / 39%, not Atlanta Fed Workforce Currents Georgia skill map, not FRBSF bank AI-hiring, not NY Fed LSE vacancy work, not Board FEDS coder-employment paper, and not a mass-layoff census.
Why it matters: After a U.S. job-ad AI-mention exposure paper, the jobs beat is how major-level automation exposure shows up in Texas new-graduate employment, first-year pay, grad-school return, and enrollment — administrative event-study evidence, not pink slips or ad-language elasticities.
GAO-26-107681: OMB AI guidance addresses only 2 of 10 expert privacy challenges — two recommendations still open.
What happened: GAO product page and HTML report Artificial Intelligence: OMB Action Needed to Address Privacy-Related Gaps in Federal Guidance (GAO-26-107681; letter date 26 March 2026; contact Marisol Cruz Cain). Finding lock: an expert panel identified privacy risks (including revealing sensitive information in raw datasets) and 10 selected challenges. OMB government-wide AI guidance addresses two (federal-workforce skills; accelerating/scaling AI with privacy protections) and does not fully address the remaining eight. Guidance also does not specify the types of known privacy-related risks agencies should consider when setting AI privacy policy. Recommendations lock: two, both to OMB, both Open on the fetched product page; OMB did not comment. Rec 1: specify examples of known privacy-related risks. Rec 2: facilitate information-sharing or issue guidance covering eight listed gaps (model evaluation/audit of sensitive models; separating sensitive data; internal-development norms; privacy performance metrics; public consent comprehension; PETs/tools; AI-specific PIAs / PII notice; privacy–performance tradeoffs). Forums named: Chief AI Officer Council or Federal Privacy Council. Status lock: congressional oversight of OMB guidance gaps, not a Binding Operational Directive and not a statute. This is an OMB privacy-guidance gap report — not yesterday’s NIST AI 200-3 ARIA planning manual, not AI 200-2 (comments still due 6 October), not SP 1353 (comments still through 15 October), not SP 800-239, not IR 8587, not EU AI Board ninth-meeting recap, and not a restatement of day-57 EU Article 50 duties (calendar only).
Why it matters: After NIST’s ARIA evaluation-planning manual, the policy beat is GAO telling OMB its government-wide AI guidance still leaves most expert privacy challenges uncovered — oversight of guidance gaps, not another evaluation-design manual or comment-clock restamp.
IARC EUcanAI: a 4-year EIC Pathfinder project starts 1 September 2026 to build AI for brain-tumour diagnosis and treatment.
What happened: IARC news (10 September 2026) announces EUcanAI (European Unified Cancer diagnostics and treatment AI agent). Official start 1 September 2026; preparatory work from 25 June. Led by Heidelberg University Hospital with IARC and 11 other institutions in Germany, Austria, and Denmark. Funded 4 years under EIC Pathfinder / Horizon Europe. Role lock: IARC will provide the WHO Classification of Tumours of the central nervous system and the framework to design, train, and implement AI modules. Goal lock: networked AI specialized for patient consultation, diagnostics, and therapy planning. Limits locked: project launch, not an RCT, not a sensitivity/specificity table, and not a marketing-authorization count. This is a brain-tumour AI consortium start — not yesterday’s FDA >1,600 authorized AI-enabled devices, not Nature/Anthropic wet-lab ~950-agent viral-genome hunt, not WHO/Europe Knowledge Community 22 pages, not npj triage-stigma experiment, not Nature Co-Scientist >700, and not Virtual Biotech multi-thousand-agent discovery.
Why it matters: After an authorization-count beat for U.S. AI devices, the health beat is a multi-country Pathfinder project standing up AI modules against the WHO CNS tumour classification — research infrastructure launch, not clearance volume or an unreviewed wet-lab headline.
UNESCO Digital Learning Week: ministers push education as a common good; 58 countries already supported on digital/AI competencies.
What happened: UNESCO news covers Digital Learning Week 2026 at UNESCO Headquarters, Paris, 8–11 September 2026, under the theme facts / frictions / frontiers (AI safety and sovereignty, human learning, teacher agency, higher education, deliberative governance). Capacity lock: since 2024, UNESCO seminars on AI Competency Frameworks involved participants from more than 100 countries; UNESCO has directly supported 58 countries with frameworks, curricula, and quality-assured training on digital or AI competencies. Backdrop lock: 2023 Guidance for generative AI in education and research; 2024 AI competency frameworks for students and teachers; 2021 Recommendation on the Ethics of AI adopted by 194 Member States. Consultation lock: working document Sustaining education as a common good in the age of AI plus six background papers opened for public comment; input to be consolidated into policy briefs for the next Digital Learning Week in March 2027. Limits locked: ministerial / competency-support coverage, not a student-uptake census, not a learning-outcomes RCT, and not PISA (still unextracted). No minister headcount locked from the extracted HTML. This is a ministerial competency-support beat — not yesterday’s UNESCO MOOC ED-004 7/49/5 EN/ES/PT package, not EC/Eurydice nearly two-thirds system AI-strategy coverage, not UNESCO IITE/IFAP MIL webinar, not Caribbean RAM / CARICOM roadmap, and not IESALC’s LAC higher-ed census (200 / 19 / 87% / 26%).
Why it matters: After a teacher MOOC catalog item, the education beat is how many countries UNESCO has already helped build digital/AI competency frameworks — capacity-building reach, not another course package or system policy-coverage percent.
A commercial-server share path is not a facility gallons inventory.
What happened: Keep EIA AEO2026 as servers 7% of commercial electricity in 2025; 22%–33% of commercial-building use by 2050; 446–818 BkWh servers by 2050; standalone 581 BkWh in High; 14.9 kWh/ft² surpassed 2031–2032; cooling up to 2.9×; High-vs-Baseline cooling +84 BkWh in 2050; post-2040 Counterfactual 10%/3 years. Label long-term AEO scenario range (laws as of Dec 2025), not observed-2026 TWh. Keep separate from EPA G20 fit-for-purpose reuse, Pacific Institute 270 gal–3.9 million gal/day / 74% consumed, AESO 1,200 MW / BYOG, NYISO 50 MW / Firm Withdrawal, ISO-NE not-yet-at-scale / 16 November, MISO 120-day / ZGIA, SPP 90-day CHILL/HILLGA, CAISO EL26-71, PJM ride-through ~4,000 MW, ERCOT Batch Zero, IEA +17% 2025 / 485→950, LBNL 649 TWh / ~11.8%, and STEO 4,368 BkWh.
Why it matters: AEO commercial-server share paths are a different evidence class than diplomatic reuse frameworks, facility withdrawal inventories, or ISO connection-limit allocations.
A Texas graduate exposure paper is not a job-ad AI-mention elasticity.
What happened: Keep Dallas Fed 22 Sep 2026 as 10 pp higher automatable-task share → 1.7 pp relative drop in Texas employment within a year; ~5% relative first-year wage drop 2021–2024; May 2024 +1.4 pp grad-school; two-thirds of 2024 CS returners stayed in CS; undergrad enrollment −4.8% fall 2024–2025 per 10-pp gap. Label Texas administrative event-study, not CES and not a layoff census. Keep separate from Cleveland Fed WP 26-24 (+1 SD → 3.1 pp), Dallas Fed 1 Sep Lightcast −5%/−8%, ILO China 21 / 1,591 / 56% / 47% / 39%, Atlanta Fed Workforce Currents Georgia skill map, the San Francisco Fed bank-lending letter, NY Fed LSE, the Board FEDS coder-employment paper, Philadelphia Fed LIFE, Minneapolis Fed Freund–Mann, Boston Fed CPP 26-8, and St. Louis RPS shares.
Why it matters: New-graduate major-exposure event studies are a different evidence class than job-ad mention elasticities or firm interview packages.
An OMB privacy-guidance gap report is not an ARIA planning manual.
What happened: Keep GAO-26-107681 as 26 March 2026; 10 expert privacy challenges; OMB guidance covers 2, not 8; two open recommendations; OMB no comment. Day-57 EU Article 50 remains calendar only. Standing clocks only: NIST AI 200-2 still due 6 October; NVD modernization RFI still due 13 October; UNESCO consultation, SPP Large Load Q&A, and NIST SP 1353 still due 15 October; FDA GenAI-device comments still due 19 October; CAISO Board still 28 October; ISO-NE Section 205 still 16 November; EU AI Board tenth still 18 November.
A Pathfinder project launch is not an authorization count.
What happened: Keep IARC EUcanAI as official start 1 September 2026; 4-year EIC Pathfinder / Horizon Europe; Heidelberg + IARC + 11 institutions in DE/AT/DK; WHO CNS tumour classification as training framework; patient consultation, diagnostics, and therapy-planning modules — project launch, not endpoints. Keep separate from FDA >1,600 authorized AI devices, Nature/Anthropic ~950-agent wet lab, WHO/Europe 22-page dialogue, npj triage-stigma experiment, Nature Co-Scientist more than 700 / 108 / 1, AISB/OpenFold3, and Virtual Biotech multi-thousand-agent discovery.
Why it matters: Multi-country research-infrastructure launches are a different evidence class than marketing-authorization lists or multi-agent hypothesis generation.
A 58-country competency-support print is not a teacher MOOC catalog.
What happened: Keep UNESCO Digital Learning Week as Paris 8–11 September 2026; more than 100 countries in competency seminars since 2024; 58 countries directly supported on digital/AI frameworks, curricula, and training; 194 Member States on 2021 ethics recommendation; consultation briefs targeted March 2027. Standing context only: UNESCO consultation comments due 15 October. Keep separate from UNESCO MOOC ED-004 7/49/5, EC/Eurydice nearly two-thirds, UNESCO IITE/IFAP MIL webinar, Caribbean RAM / CARICOM roadmap, IESALC 200 / 19 / 87% / 26%, algorithm-in-the-room think pieces, and PISA (still unextracted).
Why it matters: Country competency-support counts are a different evidence class than course catalogs or system policy overviews.
The full daily ledger keeps broader source-linked coverage organized by topic. Story dates are shown separately from the September 28 edition date.
September 28 · EIA AEO2026 servers
EIA Today in Energy (19 May 2026): AEO2026 servers 7% of commercial electricity in 2025; 22%–33% of commercial-building use by 2050; 446–818 BkWh servers by 2050; standalone 581 BkWh in High; laws as of Dec 2025.
Commercial-server share path — not EPA G20 reuse; not Pacific Institute gallons; not IEA 485→950; not LBNL 649 TWh.
Commercial electricity intensity exceeds 14.9 kWh/ft² (2003 high) in 2031–2032; data-center cooling up to 2.9× non-data-center; High-vs-Baseline cooling +84 BkWh in 2050; post-2040 Counterfactual extra efficiency 10%/3 years.
AEO intensity assumptions — not AESO 1,200 MW; not MISO 120-day; not STEO 4,368 BkWh.
Dallas Fed Economics (22 Sep 2026; Dodini–Smith): after ChatGPT, 10 pp higher automatable-task share → 1.7 pp relative drop in Texas employment within a year of graduating; ~5% relative first-year wage drop 2021–2024.
Texas graduate event study — not Cleveland Fed WP 26-24 3.1 pp; not ILO China 21/1,591; not FOMC.
Dallas FedSeptember 28 · Grad school and enrollment
May 2024: +1.4 pp grad-school enrollment per 10 pp exposure; two-thirds of 2024 CS returners stayed in CS; undergrad enrollment −4.8% fall 2024–2025 per 10-pp gap — limited returns to upskilling within exposed fields.
Adaptation and enrollment — not Dallas Fed 1 Sep Lightcast −5%/−8%; not Atlanta Fed 2.93%; not a national pink-slip census.
GAO-26-107681 (26 Mar 2026): expert panel 10 privacy challenges; OMB AI guidance addresses 2, not 8; does not specify known privacy-risk types; two open recommendations to OMB; OMB did not comment.
OMB guidance gap — not NIST AI 200-3; not AI 200-2; not SP 1353; not EU AI Board ninth.
Day-57 EU Article 50 calendar only; NIST AI 200-2 due 6 October; NVD RFI due 13 October; UNESCO consultation, SPP Q&A, and NIST SP 1353 due 15 October; FDA GenAI-device comments due 19 October; CAISO Board 28 October; ISO-NE Section 205 16 November; EU AI Board tenth 18 November.
UNESCO Digital Learning Week 2026 (Paris 8–11 Sep): more than 100 countries in competency seminars since 2024; 58 countries directly supported on digital/AI frameworks, curricula, and training; briefs targeted March 2027.
Competency-support reach — not MOOC ED-004 7/49/5; not EC/Eurydice nearly two-thirds; not IESALC 87%/26%; not PISA.
