IEA: data-centre electricity jumped 17% in 2025 — bottlenecks, not just 2030 paths, now define the AI power story.
What happened: The International Energy Agency published a news brief on Key Questions on Energy and AI, building on its earlier Energy and AI work. From the page: electricity demand from data centres soared by 17% in 2025, and AI-focused sites climbed faster still — both well ahead of 3% global electricity-demand growth. Capex: driven by data-centre investment, five large technology companies spent more than $400 billion in 2025 and are set for a further 75% increase in 2026. Bottlenecks named: gas turbines, transformers, advanced chips and IT components, plus planning and regulatory systems holding up grid connections. Response layer: tech accounted for about 40% of corporate renewable power-purchase agreements signed in 2025; conditional offtake between data-centre operators and small modular reactor projects grew from 25 GW at end-2024 to 45 GW today. With slow grid connections, developers are advancing onsite natural-gas generation — largely in the United States — and IEA satellite tracking shows many of those projects remain early-stage. AI loads swing hard enough that onsite batteries are becoming critical, and with the right incentives could make sites a grid asset. Affordability line on the page: demand growth does not necessarily raise prices if policy and infrastructure keep up, but large, concentrated, fast-scaling data centres still create special challenges. This is an observed-2025 growth, bottleneck, and offtake layer — not a restamp of earlier IEA 2030 TWh paths, not Electricity 2026 Demand or Grids chapter queue totals, not LBNL’s national-lab TWh share, not ERCOT Batch Zero queue megawatts, and not yesterday’s PJM large-load ride-through event report.
Why it matters: After a stretch of 2030 forecasts, queue counts, and ride-through design, the environment beat is how fast AI load already moved in 2025 — and how supply chains, grid connections, PPAs, SMRs, and onsite gas-plus-batteries are the near-term scramble.
Minneapolis Fed: the fight is inside the job — automating records work can raise average wages and still hurt the people whose skill was the records.
What happened: The Federal Reserve Bank of Minneapolis For All magazine featured Freund and Mann’s Institute working paper Job Transformation, Specialization, and the Labor Market Effects of AI. Frame from the magazine: occupations are bundles of tasks; workers have bundles of skills. The authors use LLMs to organize about 20,000 O*NET occupation-specific tasks into 38 task clusters. Using other economists’ research, they estimate “processing and analyzing records” is the task most likely to be automated by large language models (examples named: financial analysts; information and record analysts). Counter-intuitive wage path: if that task were completely automated, occupations that do a lot of it experience wage gains, on average, because automation frees time for remaining tasks — customer-facing coordination, communication, and negotiation rise in significance. Workers strong at information processing but weaker on those other tasks are likelier to switch jobs and see wages fall; colleagues less specialized in records are likelier to stay and see wages rise; biggest winners include workers who switch in because they are good at customer service and coordination. Working-paper abstract: moderate exposure benefits workers on average but high exposure harms them, with large dispersion within occupations; the return to social skills rises and that to analytical skills falls; low-earners gain more than high-earners. Job transformation drives the results. This is a general-equilibrium model plus magazine feature — authors’ framework, not official FOMC policy, not Boston Fed CPP 26-8 household fear survey, not Cleveland Fed WP 26-22 complementarity, not Kansas City Fed industry-contribution bulletin, not St. Louis Fed occupation-and-task adoption, not Dallas Fed Lightcast posting declines, and not a 2026 layoff census.
Why it matters: After a run of survey fear, industry-contribution, and posting-decline prints, the jobs beat is within-occupation transformation — who wins inside the same job title when the records task disappears.
NIST puts an AI agent on the National Vulnerability Database — and keeps the modernization comment clock open through 13 October.
What happened: NIST hosted an Information Technology Laboratory AI webinar on 17 September 2026 covering development of an AI agent enrichment workflow at the National Vulnerability Database. From the event page: NVD is NIST’s standards-based U.S. government vulnerability-management repository used across public and private sectors. Increasing scale and complexity of discovered vulnerabilities — plus rising use of AI tools to aid discovery and exploitation — make timely, actionable enrichment harder. NIST has begun work on an AI agentic workflow to help enrich vulnerability information; the webinar covered approach, architecture, implementation issues, and early results with the tool at NVD. Parallel clock: NVD-modernization Request for Information (Federal Register 2026-16371, published 12 August 2026) seeks comments on or before 13 October 2026, 11:59 p.m. Eastern, aimed at scalability, automation, interoperability, transparency, and utility. This is a webinar plus RFI modernization track — not a published accuracy benchmark, not NIST AI 200-2 (comments still due 6 October), not SP 800-239 AI data-center security draft, not NCCoE agent-identity blog, not yesterday’s FDA radiology CAD 510(k) denial, and not live EU Article 50 chatbot/mark duties (day 48 calendar only).
Why it matters: The policy beat is operational cybersecurity infrastructure — an AI agent on the government’s vulnerability ledger — separate from labeling statutes and still-open evaluation-framework dockets.
First-trimester machine learning beat existing early pregnancy risk tools across more than half a million pregnancies — and still misfit one country.
What happened: News-Medical reported on 18 September 2026 a Journal of Medical Internet Research multicenter model-development study, Machine Learning–Based First-Trimester Antenatal Risk Prediction for Adverse Maternal and Neonatal Outcomes. From the recap: models using information available in the first 14 weeks of pregnancy analyzed data from more than half a million pregnancies across Sweden, Chile, and Singapore and generally outperformed the early risk-assessment methods currently used in each setting. Discrimination: Sweden and Chile showed substantially better ability to separate higher- from lower-risk pregnancies; Singapore’s improvement was smaller but still statistically significant. Calibration and equity: one model did not perform equally well in every population — Swedish and Singaporean models showed reasonable agreement between predicted and observed risks; the Chilean model was less well calibrated. Social and demographic factors were among the most important predictors in some populations, information traditional assessments may overlook. Authors, via the recap: not intended to replace clinicians; decision-support for closer monitoring or earlier intervention. This is a retrospective ML risk-tool study via secondary journalism — not a bedside patient-outcome RCT, not yesterday’s Nature Medicine prospective-evidence comment, not I3LUNG, and not a claimed AUROC table locked here.
Why it matters: After a run of evidence-standard comments and simulation papers, the science beat is early-pregnancy risk stratification at population scale — with a clear warning that calibration does not travel automatically across countries.
World Bank: The Job I Hope For — free introductory AI training for Sub-Saharan African adults, register by 7 October.
What happened: The World Bank published the campaign brief The Job I Hope For, inviting young people across Africa to reflect on hoped-for jobs and receive free introductory AI training. Eligibility from the page: 18 years or older; citizens of named Sub-Saharan African countries; not directed at minors under 18. Clock: deadline to register 7 October 2026; three-part online sessions in English and French in October 2026. Course modules: Generative AI Fundamentals; How to Write Effective AI Prompts; AI Do’s and Don’ts — a beginner practical introduction. Registration also collects hoped-for job, three needs, and the role of AI in career ambitions; responses in English, French, or Portuguese. After training, aggregated Step-1 answers are to be published as a snapshot of youth aspirations; personal data deleted 6 months from the end of the campaign. Scale language on the brief: between now and 2050, more than 600 million young people will enter Africa’s labor market. This is a campaign / intro-course brief — not a learning-outcomes RCT, not World Bank Europe and Central Asia skills-divide coverage, not UNICEF Education Strategy 2026, not UNESCO ROSA / TECH SPARK Africa, and not a registrant headcount.
Why it matters: The education beat is access to beginner AI skills for Africa’s working-age youth — opportunity design, not another connected-campus pilot or PISA print.
What happened: Keep IEA as observed-2025 data-centre electricity +17% vs global 3%; five-firm capex >$400 billion and +75% in 2026; tech about 40% of 2025 corporate renewable PPAs; SMR conditional offtake 25 → 45 GW; onsite gas-plus-batteries as a swing-management response when grid connections lag. Keep separate from earlier IEA 2030 TWh paths, Electricity 2026 Demand/Grids chapters, LBNL TWh share, ERCOT Batch Zero, and PJM ride-through.
Why it matters: Near-term bottlenecks and offtake are a different evidence class than long-run electricity-share forecasts or single-event disconnection reports.
What happened: Keep Minneapolis Fed Freund–Mann as within-occupation specialization: 20,000 → 38 task clusters; records-task automation → average wage gains via remaining tasks; moderate exposure benefits on average, high exposure harms, with large within-occupation dispersion; social-skill returns up, analytical down; low-earners gain more than high-earners. Keep separate from Boston Fed CPP 26-8 fear survey and from Cleveland / Kansas City / St. Louis / Dallas prints.
Why it matters: Model-based job transformation is a different evidence class than household fear surveys or firm posting declines.
What happened: Keep NIST as AI-agent enrichment for the National Vulnerability Database (webinar 17 September) plus modernization RFI comments due 13 October 2026. Day-48 Article 50, NIST AI 200-2 still due 6 October, SP 800-239 still due 25 September, FTC personalized pricing still due 25 September, Canada ISED still open through 23 September, and FDA GenAI-device comments still due 19 October remain standing calendar only.
A first-trimester risk model is not a bedside trial.
What happened: Keep the JMIR/News-Medical study as early-pregnancy ML risk stratification across Sweden, Chile, and Singapore — more than half a million pregnancies; first 14 weeks; Chile less well calibrated. Standing clock only: WHO GI-AI4H Hangzhou ended 18 September with no communique locked here.
Why it matters: Population-scale early risk tools need country-specific calibration — not claimed outcome gains from a secondary recap alone.
What happened: Keep World Bank The Job I Hope For as free beginner genAI training for SSA adults 18+ (register by 7 October; October EN/FR sessions). Standing context only: UNESCO consultation comments due 15 October.
Why it matters: Access and aspiration design for working-age youth is the beat — measured multi-country learning gains remain unclaimed.
Five large tech firms spent more than $400 billion in 2025, set for +75% in 2026; tech ~40% of 2025 corporate renewable PPAs; SMR conditional offtake 25 → 45 GW.
Bottleneck and offtake response — not ERCOT Batch Zero; not PJM ride-through.
Slow grid connections push U.S. onsite natural-gas projects; IEA satellite tracking finds many still early-stage; onsite batteries critical for AI demand swings.
Near-term design response — not a metered annual TWh census.
Full automation of the records task raises average wages in high-records occupations via remaining tasks; moderate exposure benefits on average, high exposure harms; social-skill returns up, analytical down.
Within-occupation transformation — not Cleveland Fed WP 26-22; not Kansas City Fed contribution bulletin.
JMIR multicenter study (News-Medical 18 Sep): ML using first-14-week data across more than half a million pregnancies in Sweden, Chile, and Singapore generally beat existing early risk tools; Chile less well calibrated.
Retrospective risk tool — not a bedside RCT; not Nature Medicine prospective-evidence comment.
World Bank campaign: free three-part introductory AI training for Sub-Saharan African adults 18+; register by 7 October 2026; EN/FR sessions in October.
Intro-AI campaign — not World Bank ECA skills divide; not UNESCO ROSA / TECH SPARK.
Day-48 EU Article 50 calendar only; Canada ISED closes 23 September; FTC personalized pricing and NIST SP 800-239 still due 25 September; NIST AI 200-2 still due 6 October; NVD RFI still due 13 October; UNESCO consultation still due 15 October; FDA GenAI-device comments still due 19 October.
