Thursday, September 24, 2026 · Daily edition
NYISO’s 50 MW compute loads, bank AI hiring and SME credit, and a one-day NIST comment clock
Today’s ledger follows New York’s stakeholder design for computational large loads, a San Francisco Fed letter linking bank AI hiring to lending mix, NIST’s draft security gap analysis for AI data centers with comments closing tomorrow, an emergency-triage study on stigmatizing clinical language, and Caribbean readiness work at UNESCO’s global AI ethics forum.
NYISO large-load deck: Computational Large Loads at 50 MW+, Firm Withdrawal Service, and six-figure deposits
What happened. A New York Independent System Operator stakeholder PDF dated 14 September 2026 treats purpose-built computing demand as a Computational Large Load design problem. The deck sets the threshold at a peak load of 50 MW or greater and names Firm Withdrawal Service as one tool to blunt large-load impacts, with a refundable application fee tied to the Large Load request. Deposit ladders in the same slides run by size class — roughly $150,000 to $200,000 in the 50 MW bands, $150,000 to $250,000 for 200–499 MW, and $200,000 to $300,000 for 500 MW and above — as study and facility-entry style deposits, not a queue megawatt census. Limits are plain: this is stakeholder material, not a filed tariff order and not an interconnection queue count. It is a New York process layer for computational large loads, separate from New England’s not-yet-at-scale November clock, Midwest speed paths, or national terawatt-hour forecasts.
What to watch. After New England said the loads have not arrived at scale, the environment beat is how New York is defining the threshold, withdrawal service, and deposit ladders in stakeholder materials — process design, not another national power path.
Read the NYISO large-load stakeholder PDF →
San Francisco Fed: bank AI job postings hit 6.8% — and high-AI banks show a smaller SME loan share
What happened. Federal Reserve Bank of San Francisco economists Avaradi and Halbersleben match bank Call Reports to Lightcast job postings to measure in-house AI intensity across 1,006 banks covering more than 87% of banking-system assets. By end-2025, AI-related postings were 6.80% of banking-industry job ads, up from under 1% in 2015 and above the all-firm average near 2.7%. The split is sharp by size: about 1.2% at small banks, 4.5% at medium banks, and 8.9% at large banks, with 84% of small banks never posting an AI-related job. Over 2021–2025, banks with above-average AI posting share also show higher average returns and a lower small-business and farm loan share — about 12% versus 21% in the full sample. The authors treat the patterns as correlations, not causation, and note the proxy can overstate gaps if smaller banks buy AI through vendors. It is a staff Economic Letter on hiring intensity and portfolio mix — not a mass-layoff census and not FOMC policy.
What to watch. After national vacancy event studies, the jobs beat is how banks’ own AI hiring lines up with returns and a tilt away from soft-information SME credit — portfolio composition, not pink slips.
Read the San Francisco Fed Economic Letter →
NIST SP 800-239 draft: AI data-center security versus classic HPC — comments close tomorrow
What happened. NIST opened public comment on the initial public draft of Special Publication 800-239, a threat and security gap analysis for purpose-built AI infrastructure used in training, inference, and applications. The draft contrasts AI data centers with traditional high-performance computing across architecture, hardware, software stacks, workflows, and storage, then sketches possible safeguards. Comments remain open through 25 September 2026 — tomorrow from this edition’s date. Status is an initial public draft, not a Binding Operational Directive and not a statute. It sits beside, not inside, the finalized identity-token work and the still-open AI 200-2 and NVD modernization clocks.
What to watch. After token hardening and agentic-adoption guides, the policy beat is the closing window on how AI campuses differ from classic HPC threat models — a draft gap analysis, not finished control law.
Read the NIST SP 800-239 comment notice →
npj Digital Medicine: “frequent ED user” language shifted LLM triage against deteriorating patients
What happened. Yun, Ohuabunwa, Agboh, Courtney, Jamieson and colleagues ran a controlled, outcome-grounded experiment at two academic emergency departments. They paired notes for patients who deteriorated within six hours with severity-matched patients who did not, then asked three open-weight models — Gemma, Qwen, and DeepSeek — to score the notes before and after inserting one demographic, social, or stigma-related attribute across eighteen conditions and more than 220,000 comparisons. The stigmatizing “frequent emergency-department user” formulation produced significant harmful reprioritization in every model-dataset cell, with shifts up to about 9.4%, and beat its neutral counterpart within pairs in every cell. Psychiatric history also moved scores; race, language, and insurance did not show consistent harmful shifts. Limits are plain: simulation on matched notes, not a live bedside trial and not a clinical authorization.
What to watch. After discovery-workflow headlines, the health beat is how the wording of clinical notes can move open-weight triage models against patients who later got worse — input formulation as a safety control, not a green light to deploy.
Read the npj Digital Medicine triage study →
UNESCO Caribbean voices: RAM readiness work and a CARICOM-backed AI ethics roadmap
What happened. UNESCO’s recap of the 4th Global Forum on the Ethics of AI (14–17 September 2026 in Saudi Arabia) highlights Caribbean contributions from The Bahamas, Saint Vincent and the Grenadines, Suriname, and Trinidad and Tobago, plus AI Ethics Action Cohort training named for Curaçao, Suriname, and Trinidad and Tobago. The practical capacity tool is UNESCO’s Readiness Assessment Methodology. Suriname opened that process with a multi-stakeholder session of more than one hundred participants from dozens of institutions, following paths already taken by Antigua and Barbuda, Curaçao, Jamaica, and Trinidad and Tobago. On governance, the piece points to recent CARICOM trade-and-economic endorsement of a UNESCO Caribbean AI Policy Roadmap framed around human-centered, inclusive rules. Limits are plain: regional ethics-forum reporting, not a randomized trial and not a learning-outcomes print.
What to watch. After higher-ed uptake-versus-strategy censuses, the education beat is how Caribbean states are carrying readiness assessments and a regionally endorsed roadmap into the global ethics forum — capacity building, not another test-score story.
Read the UNESCO Caribbean voices article →
Also in today’s ledger
• A 50 MW computational threshold is not a national TWh path — and not New England’s still-empty November clock.
• Bank AI hiring intensity is not a pink-slip count — and a one-day NIST draft clock is not finished control law.
• Note-wording triage shifts are not a bedside authorization — and Caribbean RAM work is not a learning-outcomes print.
|