A frontier model has a body — silicon, electricity, water, copper, land. We start with an honest accounting of that cost, then ask the question it forces: who writes the rules, and who pays the bill?
Tyson sets the question, the vocabulary, the stakes. The slides that follow are the shared substrate — not a lecture.
The floor is yours. Ask anything about the ethics and the law of AI today — the seed prompts are only a starting point, and disagreement with them is fair game.
We tie it back to your own work — what actually changes in your own practice. Bring the harder question, not the safe one.
A frontier model is not weightless. Training and inference draw megawatts; cooling evaporates freshwater; the chips and the buildings carry carbon paid before the first token. Part 01 itemizes it.
Choosing a smaller model is an ethical-AI behavior question. Who is accountable for the grid, the water table, and the siting is an ethics-of-AI governance question. The footprint forces the second — Siau & Wang (2020).
Two halves, one conversation: the bill nobody itemized — then the rules nobody agreed on.
A frontier model has a body. Silicon, electricity, water, copper, land. What an honest accounting of the environmental cost looks like — and where it lands.
Training a frontier model and serving its inference both draw megawatts continuously. Datacenter demand is now a material driver of grid planning in several U.S. regions.
Evaporative cooling consumes potable or industrial water — sited disproportionately in arid regions. The Southwest is a relevant case in this room.
Accelerator manufacturing, server build, building shell, fiber, copper. The carbon cost is paid before the first prompt — and is rarely on the accounting.
Hyperscale datacenter buildout reshapes local economies, tax bases, transmission corridors, and noise envelopes. The externalities are not evenly distributed.
Frame: AI's environmental cost is not one number. It is at least four numbers, on at least three timescales (training, inference, buildout).
IEA's Base Case: data-center use roughly doubles to ~945 TWh by 2030 — just under 3% of global electricity — growing ~15%/yr, four times faster than demand overall. AI "accelerated servers" grow ~30%/yr. — IEA, Energy and AI (2025).
An early estimate put a chatbot query near 2.9 Wh — roughly 10× a web search (EPRI / IEA, 2024). Newer medians run lower (~0.3 Wh), long reasoning and multimodal far higher. The exact figure is contested; the direction is not. (Bars illustrative.)
GPT-3's ~700,000 L (2020) is the last training figure anyone published. GPT-4 is estimated near 70× that — tens of millions of liters, much of it a single month's cooling draw at one Iowa cluster.
For today's frontier models — GPT-5-class, Claude Opus 4.x — the training-water cost is not published at all. The footprint grows; the transparency doesn't.
— GPT-3: Li et al. (2023); GPT-4: estimates from reporting on Microsoft's Iowa training cluster (2023).
Microsoft's Goodyear campus drew ~56M gal/yr of potable water (~670 homes) and is now switching to zero-water air cooling amid local limits; Phoenix-area direct cooling runs ~385M gal/yr; the most water-intensive Google site hit ~1 billion gal in 2024. — APM Research Lab; Microsoft; Google (2024). (Bars illustrative.)
Electricity to train GPT-3 (2020) — about 552 tonnes CO₂e, roughly the annual tailpipe emissions of ~120 U.S. cars. Frontier models since are substantially larger.
— Patterson et al. (2021). Car comparison illustrative.
Accelerator fabrication, server build, the building shell, fiber and copper carry carbon spent up-front. For modern hardware it increasingly rivals operational emissions — yet routinely never reaches the ledger. — Gupta et al., "Chasing Carbon" (2021). Split illustrative.
Per-call energy and embodied compute scale with model size. Most tasks fit inside a smaller model — defaulting to the frontier is a habit, not a requirement. (Figure is illustrative, not measured.)
Agentic loops fire speculative calls, retries, and tool calls behind one user action. Cumulative compute, not the visible turn count, is the cost. Cache. Cap retries. Watch the loop.
Collective, diffuse, opaque harms that older rights frameworks address only partially.
Datacenter siting, grid load, and water draw produce harms that are collective, diffuse, and opaque — the same shape as algorithmic harm. Marshall's social-rights tier is the same conceptual address.
Laudato Si' (2015, on care for creation) and Magnifica Humanitas (2026, on AI) share a single underlying claim: the environment and the economy and the technology are one moral question.
Corporate AI constitutions and public AI bills of rights are doing very different work. Alondra Nelson argues one of them is structurally insufficient.
The model developer's vision of how its own model should behave. Canonical example: a frontier-lab "constitution" document used in training.
Developed through public process. Canonical example: the White House OSTP Blueprint for an AI Bill of Rights, October 2022.
Source: Alondra Nelson, A civic grammar for AI rights, Science (2026). DOI 10.1126/science.aeh7153
Non-discrimination, transparency, data privacy, notice, human alternatives — a civic vocabulary traveling across partisan lines by institutional diffusion (Strang & Meyer), mapping onto Marshall's third wave of citizenship rights:
Human dignity affirmed as infinite. Transhumanism (engineering away finitude) and posthumanism (blurring the human/machine line) are named an "anti-human vision" — measured against dignity, common good, justice.
Disarm AI from military and purely economic ends; subject companies to stricter state and international regulation — explicitly not industry self-governance.
| Document | Issuing body | Force | What it does |
|---|---|---|---|
| Framework Convention on AI & Human Rights | Council of Europe | Treaty | Binding international convention tying AI systems to human-rights, democracy, and rule-of-law obligations. |
| Recommendation on the Ethics of AI | UNESCO | Recommendation | First global standard-setting instrument on AI ethics, adopted by 193 member states. |
| OECD AI Principles · G20 endorsement | OECD · endorsed by G20 | Principles | Trustworthy-AI principles widely adopted as a starting point for national strategies; signal of cross-bloc consensus. |
| Political Declaration on Responsible Military Use of AI & Autonomy | U.S. State Dept. + co-signatories | Declaration | Norms for military AI: human accountability, legal review, controlled escalation. |
| Toronto Declaration | Amnesty Intl. + Access Now | Declaration | Machine learning, equality, and non-discrimination — a civil-society anchor for the Nelson framework. |
Prompt for the room: which of these has any practical effect on what a public health researcher in Arizona actually does?
Seed prompts, not a script — the framing above is shared substrate, and disagreement with it is fair game. The mic is yours.
Compare to "use less plastic."
The hardest question. Save it for last.
Not principles. Mechanism.
Nelson asks both. Pick one to defend.
You can dislike the framing and still answer.
Be specific. Name a system.
The civic grammar gets traction because publics use it. The footprint shrinks because engineers and researchers refuse the default. Both are unfinished work — and both belong to the room that just spent three days in it.