Thursday, August 27, 2026 · Daily edition
IEA data-center demand, live AI Act enforcement, and BLS’s decade-ahead jobs map
Today’s AI footprint runs through metered power, a live European enforcement clock, and an official U.S. jobs projection that cools some office demand while still growing computer occupations. The IEA’s demand chapter puts global data-centre electricity on a path from about 415 TWh in 2024 to about 945 TWh by 2030, with accelerated AI servers doing most of the lifting. Day 25 of AI Act enforcement maps who polices what — and what is not yet due. BLS’s 2024–34 overview is the official spine, not a layoff census. Two Nature Medicine papers from August 19 add liver-CT guidance and a prospective ED decision-support study. UNESCO’s Digital Learning Week is 12 days out.
Data-center power use is on track to more than double by 2030 — mostly from AI servers
What happened. The International Energy Agency’s Energy and AI demand chapter puts global data-centre electricity at about 415 TWh in 2024 — roughly 1.5% of world electricity — after five years of about 12% annual growth. In the Base Case, consumption reaches about 945 TWh by 2030, just under 3% of global electricity, growing around 15% a year from 2024–2030, more than four times electricity growth from all other sectors. Accelerated servers (mainly AI) grow about 30% a year and account for almost half of the net increase; conventional servers about 9% a year and roughly 20% of the increment; other IT about 10%; cooling and other infrastructure about 20%. The chapter is a consumption outlook built partly on server-shipment projections, with Lift-Off, High Efficiency, and Headwinds sensitivities. It is not a plant inventory. A separate IEA news note says 2025 data-centre electricity jumped about 17% to around 485 TWh amid grid, equipment, and siting bottlenecks — a year marker, not a re-lock of the 415→945 path.
What to watch. The measurable record is metered load, interconnects that clear, and whether efficiency gains offset AI server growth — not a single global percentage treated as destiny. A “just-under-3%” 2030 share is still a large absolute jump on grids that build slower than halls.
Read the IEA demand chapter →
Day 25 of live AI Act enforcement: who polices what, and what is not yet due
What happened. The European Commission’s Digital Strategy enforcement-framework page maps the live architecture for the AI Act. Powers of the AI Office and national competent authorities apply from 2 August 2026 for prohibited practices already in force, general-purpose AI model duties, and transparency rules — chatbots must disclose they are AI; deepfakes must be labelled; synthetic content needs machine-readable marks. The AI Office covers GPAI model providers (including systemic-risk models), systems from the same provider or group as the underlying GPAI model, and systems inside designated very large online platforms or search engines under the DSA; the EDPS covers EU institutions; national authorities cover other systems. Sanction maxima listed for information only: prohibited practices up to €35 million or 7% of worldwide annual turnover; other breaches including GPAI obligations up to €15 million or 3%; failure to answer an AI-systems information request up to €7.5 million or 1%. The split clock continues: CSAM and non-consensual intimate-image prohibitions from 2 December 2026; standalone high-risk Annex III rules from 2 December 2027; high-risk systems embedded in regulated products from 2 August 2028. This is process architecture, not a measured drop in deepfakes.
What to watch. The measurable record is opened investigations, labelled products in the wild, and whether the 7% figure is wrongly applied to chatbot disclosure. Transparency is live; most high-risk product rules are not.
Read the Commission enforcement map →
BLS’s decade-ahead map: generative AI cools some office demand — and still grows the computer occupations
What happened. The Bureau of Labor Statistics Monthly Labor Review January 2026 overview of industry and occupational employment projections for 2024–34 (Rieley & Colato) is the official industry-plus-occupation spine. Total employment is projected to rise from 170.0 million to 175.2 million — +3.1%, or about 5.2 million jobs — much slower than the +13.0% gain over 2014–24. Growing adoption of AI, including generative AI, and resulting productivity gains are expected to dampen labor demand in sales, design, and administrative support. Four sectors are expected to lose jobs, with the bulk of losses in retail trade. Healthcare and social assistance is the fastest sector (+8.4%, roughly 2.0 million). Computer and mathematical occupations grow +10.1% / +545,600, with data scientists +33.5% / +82,500 and information security analysts +28.5% / +52,100 near the top. Related wage-and-salary industries: computing infrastructure / data processing / web hosting +20.3%; software publishers +19.3%; computer systems design +15.8%. Projections embed judgment about AI adoption pace; they are not a 2026 layoff census.
What to watch. The measurable record is whether hiring and openings match the dampened office path and the still-growing computer path — not slogans that “half of all jobs” vanish.
Read the BLS Monthly Labor Review overview →
Two Nature Medicine papers: liver CT guidance and an ED clinical-decision tool
What happened. Two peer-reviewed Nature Medicine papers share a 19 August 2026 version-of-record date. Zhang et al. describe LiON, a large-scale AI-guided liver malignancy diagnosis effort framed as a multicenter study and a single-arm trial. Extended-data captions say the system analyzes contrast-enhanced CTs, predicts patient-level malignancy scores, and segments eight types of liver lesions. That is not an RCT and not a mortality trial. Leibovitch, Ahituv et al. report a prospective evaluation of a large language model clinical decision support system in the emergency department at Rambam Health Care Campus / Technion, Haifa. Figure captions name a Wing A intervention allocation of n = 584; a propensity-score figure reports n = 1,133. Analysis code is public; patient data are controlled-access. Abstract result tables were not independently extracted for the site edition, so performance headlines stay off this email.
What to watch. The measurable record is prospective patient outcomes, activation rates, and independent replication — not publication-day buzz. Single-arm imaging and prospective CDSS can still matter clinically without being sold as autonomy.
Read the LiON liver paper →
UNESCO’s Digital Learning Week is 12 days out — facts, frictions, and “ghost learners”
What happened. UNESCO’s Digital Learning Week 2026 runs 8–11 September at Headquarters in Paris under the theme Education in the age of AI: Facts | Frictions | Frontiers. The institutional event page flags AI tutors reaching learners without prior personalized support; agentic systems beginning to act more autonomously; synthetic content reshaping how knowledge is produced and validated; and “synthetic research” and “ghost learners” as challenges to evidence, participation, and presence. Stated purpose: keep human agency, equity, and the public purpose of education at the centre. A related ministerial statement on sustaining education as a common good is slated around the week. This is a convening agenda, not a multi-country learning-outcomes RCT.
What to watch. The measurable record is what ministers actually commit to, which classroom policies change, and whether “ghost learner” and synthetic-evidence problems get operational definitions — not the conference branding alone.
Read the UNESCO week page →
Also in today’s ledger
• IEA news note: 2025 data-centre electricity ~+17% to around 485 TWh amid bottlenecks — a year marker, not a re-lock of the 415→945 Base Case (IEA).
• AI Act split clock: CSAM / intimate-image bans 2 Dec 2026; product-embedded high-risk 2 Aug 2028 (European Commission).
• Same BLS table’s expansion side: computing infrastructure +20.3%; software publishers +19.3%; systems design +15.8% (BLS MLR).
|