AI + EARTH · AN EDITORIAL JOURNAL
THIS WEEK, FOLLOW THE SIGNAL
A question begins as language. Its answer passes through models, hardware, electricity, cooling, and a real place. This week, we follow the route from a sentence to the coast of Finland.
GROUNDWORK 037
The interface feels weightless.
The system beneath it is not.
Groundwork looks past the smooth surface of AI and follows the physical relationships beneath it—without turning uncertainty into a slogan or complexity into a reason to look away.
THE PHYSICAL LAYER
Look past the interface.
The story continues below it.
Photography sets the atmosphere. Evidence tells us what is actually happening.
Editorial landscape · not photography of the Hamina facility
THE EARTH REPORT
This week: seawater cooling, recovered heat, and the physical route beneath an AI response.
The Win
In Hamina, seawater cooling and a district-heating connection make the data center legible as local infrastructure—not an invisible cloud. The project does not erase impact, but it shows how heat can be treated as a resource when a compatible network exists.[7][8][9]
The Tradeoff
Data centers remain a minority of global electricity demand, yet their loads are concentrated. A project can look modest in a worldwide total while materially affecting one grid, water system, neighborhood, or planning process.[5][10]
What We Learned
A prompt becomes tokens, tokens are processed by a model, and the computation runs on hardware. Electricity, cooling, buildings, networks, and geography sit beneath the answer—even when the interface makes them disappear.[1][6]
What We Can Do
Choose a system suited to the task. Ask what is measured, where the computation runs, which grid and cooling system serve it, and whether efficiency gains reduce total demand or simply make more use possible.[5][6]
UNDER THE SURFACE
This is a conceptual path, not a measurement of one product. The exact model, tools, hardware, and facility can change every step.
Prompt
You write or speak a request. The product may add instructions, earlier conversation, retrieved documents, or search results before sending context to the model.
Tokens
A tokenizer represents text as tokens—often words, subwords, punctuation, or other character sequences. The boundaries depend on the tokenizer and language.
Context
The model uses learned parameters and the available context to estimate which token should come next. Modern systems commonly use Transformer-based attention mechanisms.
Response
The selected token is added to the sequence and the process repeats. The language model generates the response; any search or retrieval comes from additional systems around it.
Infrastructure
Servers, accelerators, memory, storage, networking, cooling, and backup equipment support the request. Their contribution varies by system and facility.
Place
The local electricity system, climate, cooling design, water conditions, and community determine what the physical demand means in that place and moment.
NEW TO GROUNDWORK?
These Field Notes stay between issues. Read one now or return when a weekly story raises the question.
THE WIDER JOURNEY
Move through the system slowly—from the words on the screen to the infrastructure and community beneath them.
FOLLOW THE SYSTEMEVIDENCE, METHOD & CREDIT
Citations stay visible in the story. Full context waits here when the reader wants it, rather than interrupting every paragraph.
Google for Developers · Developer education
Introduction to Large Language Models Living educational documentation; checked August 30, 2026Supports the explanation of language models, tokens, context, and next-token prediction. Provider-authored education is used for mechanics, not environmental performance claims.
OpenAI · Model-provider tool
Tokenizer Living documentation; checked August 30, 2026Provides a model-specific way to inspect tokenization. Groundwork does not present one tokenizer's boundaries as universal.
Google for Developers · Developer reference
Machine Learning Glossary Living glossary; checked August 30, 2026Supports the plain-language distinction between training, inference, parameters, serving, and model use.
Vaswani et al. · Primary research paper
Attention Is All You Need June 12, 2017; revised August 2, 2023Introduced the Transformer architecture based on attention mechanisms. It is included as primary technical history, not as a claim that every current language model is identical.
International Energy Agency · Intergovernmental analysis
Energy and AI — Executive summary April 10, 2025; checked August 30, 2026Supports the physical connection between training and deploying AI models, data centers, electricity demand, and concentrated local effects.
International Energy Agency · Intergovernmental analysis
Energy demand from AI April 10, 2025; checked August 30, 2026Supports the description of data-center equipment and the need to consider both efficiency and total demand.
Google Data Centers · Facility-operator source
Hamina, Finland Current facility profile; checked August 30, 2026Supports the facility-specific description of seawater cooling and Google's stated heat-recovery goals. Operator claims remain attributed.
Haminan Energia Oy · Local utility source
Data-center heat for Hamina's district-heating network 2024; checked August 30, 2026Describes the heat-pump project and its stated design potential. Design capacity is not presented as verified annual delivery.
City of Hamina · Local-government source
The impacts of the Hamina data center as a research target June 22, 2026; checked August 30, 2026Provides current local context and reports that heat from data-center cooling is being directed to the district-heating network.
Lawrence Berkeley National Laboratory · U.S. national-laboratory report
2024 United States Data Center Energy Usage Report December 19, 2024Supports the distinction between direct cooling-water use and indirect water associated with electricity generation. U.S. findings are not treated as universal facility averages.
Groundwork uses atmospheric landscapes as editorial context. They are not presented as photography of the Hamina facility.