Last issue was about what the labs shipped. This one is about what it costs — and who pays.
Most AI coverage stops at the announcement. But the interesting part of 2026 is what's landing downstream: on your power bill, in the job market for 22-year-olds, and in the small print under the videos in your feed.
In today's issue:
⚡ Data centres are showing up on household electricity bills
📉 Stanford: young workers in AI-exposed jobs are 19% behind
🏷️ Europe now requires AI content to identify itself
🌍 Abu Dhabi gave away a 375B-parameter model — training data included
📌 Quick hits
Reading time: 5 minutes.
⚡ The AI boom is on your power bill now
For three years, AI's energy use was an abstraction. It has stopped being one.
THE BREAKDOWN:
US residential electricity prices rose 7.1% in 2025 — more than double inflation. Some states saw over 20%.
In areas dense with data centres, prices have jumped as much as 267% over five years.
Data centre energy demand is projected to nearly double from 80 GW (2025) to 150 GW by 2028. A single hyperscale facility draws around 100 megawatts — enough for 100,000 homes. Meta's Hyperion site in Louisiana is planned at 5 GW, roughly three times what all of New Orleans uses.
In Virginia — nearly 600 data centres, 100+ more coming — data centres consumed almost 40% of the state's electricity in 2024. 75% of Virginia voters now blame them for rising bills.
Water matters too: large facilities can use up to 5 million gallons a day for cooling, and two-thirds of data centres built since 2022 sit in water-stressed regions.
WHY IT MATTERS: Every "AI is free" product is running on infrastructure someone pays for. Increasingly, part of that someone is the household on the same grid. This is the fight that decides how fast AI can actually scale — not benchmarks, but whether local regulators let the next campus connect.
Worth watching: the rules on who pays for grid upgrades — the utility's ratepayers or the data centre — are being rewritten state by state right now. That's the whole ballgame.
📎 Source: Consumer Reports
📉 Stanford: the entry-level gap widened to 19%
Erik Brynjolfsson's team at the Stanford Digital Economy Lab updated their "Canaries in the Coal Mine" study on August 12, 2026, using ADP payroll data through mid-2026.
THE BREAKDOWN:
Workers aged 22–25 in AI-exposed occupations now sit 19% behind similar-aged peers in less-exposed jobs. That gap was 15% in July 2025.
Experienced workers show no comparable gap. The effect is concentrated at the entry level.
The pattern: jobs built on codified knowledge — the stuff written down in manuals and documentation — are shrinking for young people. Jobs built on tacit knowledge — judgement built through practice — are holding or growing.
The gap persists after controlling for interest rates, remote work and tech-sector employment.
The researchers' own caveat, which most coverage skips: these are "descriptive patterns, not causal estimates," and they explicitly say no single study is definitive.
WHY IT MATTERS: The AI-and-jobs debate is usually all-or-nothing — mass unemployment or nothing to see here. The data suggests something narrower and more useful: AI is eating the first rung of specific ladders. If your work is mostly applying documented procedure, that's the exposure. If it's judgement, relationships and knowing what the manual doesn't say, that's the moat.
📎 Source: Stanford Digital Economy Lab
🏷️ In Europe, AI content now has to say it's AI
Since August 2, 2026, the European Commission and national authorities have been enforcing the EU AI Act's transparency rules.
THE BREAKDOWN — what changes in practice:
Chatbots must disclose they're not human.
Deepfakes must be labelled.
AI-generated or altered content must carry machine-readable markers so tools can detect it.
The accompanying Code of Practice gets specific: a persistent icon plus an opening disclaimer for video, a fixed visible icon for images, audible disclaimers for audio deepfakes, and lighter-touch disclosure for creative work so it doesn't wreck the piece.
Over 180 organisations have signed that Code.
WHY IT MATTERS: This is the first serious attempt to make "is this real?" a question with a technical answer rather than a vibe. Two open problems: labels only help if platforms display them, and a world where AI content is labelled can quietly make unlabelled fakes more convincing. Watch for the first enforcement action — that's when we learn what this actually means.
If you make content: if any of your audience is in the EU, this applies to you, not just to Big Tech.
📎 Sources: European Commission · Tech Policy Press
🌍 Abu Dhabi gave away a 375B model — with the training data
On September 3, the Institute of Foundation Models (part of MBZUAI in Abu Dhabi) released K2 Horizon: six models from 0.9B to 375B parameters, all under Apache 2.0.
THE BREAKDOWN:
Released: weights, source code, the complete training data, and the training recipe. That last part is rare — most "open" models publish weights only.
The fleet spans a 0.9B model that runs on modest hardware up to a 375B flagship with 23B active parameters.
IFM claims state-of-the-art performance at the small end (0.9B, 3.7B, 7B) and around a 3x speedup from a diffusion-distillation technique.
Founder Eric Xing: "Open source is much more than open weights... Every model in the fleet ships with its training data, recipe and evaluations."
WHY IT MATTERS: Two things at once. First, if you can inspect the training data, you can actually audit a model — for bias, for contamination, for what it was fed. Almost nobody lets you do that. Second, the geography: the most open frontier-scale release this year came from the Gulf, not California. The map of who builds AI is being redrawn, and Abu Dhabi is buying influence with generosity rather than secrecy.
📎 Sources: IFM · BigDATAwire
📌 Quick hits
The picks-and-shovels round. Gimlet Labs raised $300M at a $3B valuation, led by Andreessen Horowitz with Arm and Microsoft's M12 participating. The pitch: software that spreads AI workloads across different chip architectures, so you're not locked to one vendor. Telling detail — it raised an $80M Series A earlier this year. Source
Follow-up from last issue: OpenAI's GPT-6 Astra is rolling out to Plus, Pro and Enterprise users this week, and Nvidia's $12.9B Hugging Face deal now goes to regulators. Both worth watching for what happens next rather than what was announced.
🧠 What this actually means for you
Audit your own exposure honestly. Not "will AI take my job" but: how much of what you're paid for is written down somewhere? That's the part under pressure. The judgement part isn't.
If you hire, the entry level is your problem to solve. The Stanford data says juniors are getting squeezed out of exactly the roles that used to train them. Whoever figures out how to build experience without the old first rung has a real advantage.
Label your AI content now. Not because the EU will fine you tomorrow, but because "we tell you when it's AI" is about to be a trust signal — and retrofitting it later looks like getting caught.
💬 One question
Have you noticed AI showing up in your actual costs — power bill, subscriptions, insurance, anything?
Hit reply. We're collecting real examples for a future issue, and the specific ones are always better than the studies.
That's it for today. If someone forwarded you this, you can subscribe below.
— The artificialintelligence.ai team
