District Hunt 2026 was a one-day game played across Capitol Hill in Washington DC on Saturday, October 17. Twenty-four teams and 103 players were sorted into five guilds and competed for control of ten zones through fifteen games and twelve puzzles. The day ran on a web app built for it, and that app was built with AI assistance over three months.
Game Genius runs District Hunt as a nonprofit focused on social impact. Barometer XP, a public benefit corporation, was one of the event's main sponsors. Both organizations tell players that we think about the impact of what we make, including the parts that are hard to measure. The AI work is one of those parts. This page is our accounting of it, and a method anyone can run on their own project.
The AI work did not run on a laptop in Washington. Large AI models run in data centers, and the ones that served this build are in Indiana, Mississippi, Pennsylvania and New York. Three things follow from that, and they shape everything below.
These counts come from the build's own logs rather than an estimate. One continuously resumed working session, July 27 to October 26, 2026, across 14,755 exchanges, split between Claude Opus 5 and Claude Opus 4.8.
| Token type | Count | What it is |
|---|---|---|
| Output | 21,392,843 | Text the model wrote. Produced one piece at a time, which makes it the most energy-intensive per token. |
| Cache writes | 188,507,772 | Context read and stored for reuse. Processed in parallel, so roughly twenty times cheaper per token. |
| Cache reads | 7,642,536,580 | Stored context reused rather than reprocessed. The cheapest tokens by a wide margin. |
| Fresh input | 30,283 | Context read without a cache. A small amount here. |
| Total | 7,852,467,478 | About 7.85 billion. |
About 97% of those tokens were cached reads, which use very little energy. That is why the breakdown matters more than the headline number: a calculation that treated every token the same would come out more than ten times too high.
We give a range rather than a single number because the published measurements vary that much. No energy-per-token figure exists for these specific models. Independent studies of large language models span 0.0001 to 0.002 watt-hours per output token depending on model size, and a frontier model sits at the upper end. We applied that band, discounted cached tokens heavily, then multiplied by 2.4 to cover the parts of a data center that are not the chip itself — idle capacity, processors, memory, redundancy and cooling. That multiplier is Google's published ratio between chip energy and total energy.
Of the water, roughly 90 liters went to cooling the data centers and roughly 190 liters evaporated at the power plants. Published data center water figures usually cover only the first of those two.
Four things are deliberately left out of the figures above. Each is a point where this kind of accounting often goes wrong.
Electricity is an input rather than a pollutant. The carbon and the water are what generating it costs, so addressing those covers the harm. Buying a renewable energy certificate on top would mean two parties claiming the same clean power, because Amazon already reports matching 100% of its electricity with renewables. Counting the same megawatt-hour twice is the most common error in this kind of reporting.
Amazon says it is 75% of the way to returning more water than it uses. That figure is global and self-reported, and we found no independent audit of it. It also excludes the power plant water that makes up about two thirds of the total here. A leaked internal document reported in 2025 indicates the exclusion was deliberate; Amazon called the document obsolete. A complaint filed in Arlington County, Virginia alleges the company's local water claims are misleading, citing public records data. Those allegations have not been tested in court. We did not count the pledge toward our own figures.
Training an AI model uses far more energy than using one, and it is the larger share of AI's overall footprint. No honest per-project number exists for one build's portion of it, because the cost is spread across every user of the model. What is measured here is inference, the part a user controls directly.
The emissions physically happened, on grids in four states. We removed an equivalent amount of carbon somewhere else. Those are not the same thing, and only the second one is something we can stand behind.
100 kilograms of biochar carbon removal through Remove Carbon Today, priced at €198 per tonne, which is about $222 per tonne at late-2026 exchange rates. That came to $23.
We bought 100 kilograms rather than the 48 our central estimate gives. Forty-eight is the midpoint of a range that runs as high as 96, so 100 covers the top of the range rather than the middle of it.
Biochar is plant waste heated without oxygen until it becomes a charcoal that does not decompose. Worked into soil, it holds its carbon for centuries. The removal is certified by Puro.earth and retired on its public registry, which means it carries a serial number that nobody else can claim.
We checked Carbonmark as the alternative. It supports purchases under a tonne and issues public retirement records, but its current catalogue covers ocean alkalinity, composting, reforestation and afforestation, with no biochar. One limitation at Remove Carbon Today: credits are retired in batches once purchases reach a full tonne, so the registry record can arrive up to a month after the purchase.
Planted through One Tree Planted, which plants approximately one tree per dollar donated. Their staff describe the breakdown as about 60 cents for the sapling and 20 cents for planting, with the remainder going to administration and costs varying by region, so one dollar per tree is close rather than exact. $24.
A tree absorbs roughly 25 kilograms of carbon dioxide a year once established, and roughly one tonne over a 40-year life. The range across species is wide, from about 0.1 to 2 tonnes. If these 24 trees reach maturity they will take up on the order of 24 tonnes between them across their lifetimes.
We are not counting that as an offset. A donation plants trees but produces no registry entry and no serial number, so there is nothing to verify and nothing to retire. It is recorded here as a gift, separate from the carbon arithmetic, at one tree per team that played.
We also gave $20 to Casey Trees, through the Grow Our Canopy part of their giving page, which covers planting, protecting and caring for trees across the region. Casey Trees works differently from One Tree Planted: donations fund its programs rather than buying a counted number of trees, so there is no per-tree figure to quote. We gave there because the game was played under the canopy they maintain. Other established national options with long records: Arbor Day Foundation and American Forests.
Water cannot be bought back into a particular river the way carbon can be removed anywhere. There is no instrument that returns a gallon to the Pearl River or Lake Michigan. Rather than claim restitution we could not deliver, we gave $20 to each of four organizations, covering the watershed of every place the work touched.
| Organization | Why this one | How we checked |
|---|---|---|
| Anacostia Watershed Society | Eastern Market drains to the Anacostia, so this is the water under the game itself. | Charity Navigator four stars, 100%, FY2024. Also vetted by Spur Local, one of our event partners. |
| Chesapeake Bay Foundation | The regional system, and where Pennsylvania's data centers drain through the Susquehanna. | Registered 501(c)(3), site current to October 2026, runs tree plantings. |
| Alliance for the Great Lakes | Lake Michigan, below the Indiana data center, and Lake Ontario, below the New York site. | Charity Navigator four stars and Guidestar Platinum, 2025. |
| Pearl Riverkeeper | The Pearl River, below the Mississippi data center. | Registered 501(c)(3), licensed Waterkeeper Alliance member, running 2026 events. |
For anyone who wants the measured version, Water Restoration Certificates cost $4.00 per 1,000 gallons and verifiably return water to a specific stream. One covers this build thirteen times over. The streams are mostly in the western United States, so the certificate works but does not match the places involved here. Carbon is treated as global by design and electricity certificates trade regionally, but water is the one resource where location is the whole point, and no matched instrument exists yet.
| Carbon removal, 100 kg biochar, registry-verified | $23 |
| 24 trees, one per team, through One Tree Planted | $24 |
| Casey Trees, Grow Our Canopy | $20 |
| Four watershed organizations, $20 each | $80 |
| Total | $147 |
Six steps. Every factor below is either a published figure or an assumption you state openly.
A sustainability report is a company's own account of itself, so treat a statement like “we are 75% of the way there” as a claim and look for who verified it. The Carbon Disclosure Project scores companies independently. European reporting rules now require outside assurance for large firms operating there. The most practical route is local: data centers buy municipal water, municipal utilities are public bodies, and a public records request to the utility will produce real site-level figures when a company does not publish them. That is how the Virginia plaintiffs obtained theirs.
Every link below was checked and loading on October 26, 2026. Figures from these sources are estimates with stated ranges rather than audited measurements.