AI Fantasy Football Could Drink 339,000 Gallons in Two Weeks. Your Flex Is Thirsty.

At some point this weekend, an otherwise responsible adult will open an AI chatbot and type a paragraph that begins, “I’m in a 12-team full-PPR superflex league with two keepers,” continues through the medical histories of six men he has never met, and ends by asking whether a rookie wide receiver has “league-winning upside.”

The chatbot will answer. The adult will ask it to reconsider. The chatbot will reconsider. A podcast host will say the opposite. The adult will return to the chatbot with new evidence, as if conducting appellate litigation before the Supreme Court of Flex.

This year, that ritual is not a niche experiment. The NFL regular season begins on September 9, according to the league’s official 2026 calendar. The two weeks from September 4 through September 18 therefore contain late drafts, waiver claims, injury news, the entirety of Week 1 and the Thursday-night opening of Week 2. They are 14 days of concentrated roster anxiety, which modern technology has helpfully converted into a reason to operate several thousand specialized chips.

So how much AI power, water and tree-equivalent carbon absorption will Americans consume while trying to make good fantasy-football picks during this window?

My central estimate is 229 megawatt-hours of electricity, 66,000 gallons of direct data-center cooling water, about 339,000 gallons when upstream electricity-generation water is included, and roughly 90 metric tons of grid-average carbon dioxide. Under the Environmental Protection Agency’s conversion, that last number equals the carbon absorbed by about 1,500 urban tree seedlings grown for 10 years.

No trees are actually consumed. This is important because the internet has enough trouble with measurement without anyone imagining a cloud provider pushing a maple into the front of an H100 rack.

The range is extremely wide. A conservative case lands at 5.3 MWh and 7,900 gallons of broad water consumption. A nation of frantic agent users could reach 4.17 gigawatt-hours and 6.18 million gallons. The point is not that one of these numbers descended from a carbon-neutral burning bush. The point is to expose every assumption, calculate the plausible order of magnitude and distinguish the footprint of an ordinary prompt from the much hungrier workflows hiding behind a single friendly button.

The Scoreboard Before the Caveats File a Protest

Here is the whole estimate. “Queries” means visible prompt-and-answer exchanges in this model. Some products may secretly turn one visible question into several model calls; we will get to that small computational nesting doll shortly.

The central electricity figure is about 19 years of average U.S. household electricity use, compressed into a fortnight. Spread evenly, it is a continuous 682-kilowatt load. It will not be spread evenly. The queries will arrive in nervous clumps around drafts, injury reports, waiver deadlines, inactive lists and the exact moment somebody realizes his starting running back is questionable with a hamstring.

In the context of the entire AI industry, 229 MWh is lint. The International Energy Agency estimates that data centers used 485 terawatt-hours worldwide in 2025 and could reach 950 TWh by 2030. Our central fantasy-football estimate equals roughly 15 seconds of 2025 global data-center electricity use.

In the context of a leisure activity whose final output is “start Terry,” it is also objectively funny.

First, Count the People Pretending This Is Skill

There is no national meter labeled FANTASY FOOTBALL CHATBOT LOAD. We have to construct the population from the best public survey data available.

In July, the Fantasy Sports & Gaming Association reported that 82.8 million U.S. adults played fantasy sports or bet on sports during the previous 12 months. Fifty-one percent did both. The association also said 75% of fantasy players bet on sports.

Those percentages let us work backward. If 51% of the combined population did both activities, the crossover group was about 42.2 million people. If that group represented 75% of fantasy players, the total U.S. fantasy-sports population was roughly 56.3 million. The figures were rounded before publication, so this is an estimate, not the discovery of a hidden 0.3 fantasy manager wandering Nevada.

The FSGA’s sport breakdown says 79% of fantasy participants play fantasy football. Applying that share gives approximately 44.5 million U.S. fantasy-football players.

There is an awkward vintage mismatch here. The 44.5 million estimate combines the newest 2026 participation and overlap data with the association’s older published sport mix. Fantasy players also participate in multiple sports, which is allowed in the 79% figure. We are not counting unique teams or leagues. We are counting people who plausibly have a football roster and an opinion about target share.

This is already more grounded than multiplying an old “60 million fantasy users” headline by a number found behind a sponsored calculator. It is still a model. If the true football share has fallen to 70%, the central result drops about 11%. If it has risen to 85%, the result increases about 8%. Population uncertainty matters, but not nearly as much as the number and complexity of AI queries.

One in Four Managers Has Already Invited the Robot Into the War Room

The same 2026 FSGA study found that 25% of fantasy players and sports bettors use AI tools to inform their decisions. ChatGPT was used by 77% of the AI group and Gemini by 54%; respondents could plainly use more than one tool, because nothing says disciplined decision-making like asking two machines and selecting the answer that validates your prior.

That 25% is the foundation of the central case. Applied to 44.5 million fantasy-football players, it yields about 11.1 million AI-assisted managers.

The survey combines bettors and fantasy players, so it does not prove exactly one-quarter of fantasy-football players will query generative AI during this exact two-week period. “AI tools” may include embedded recommendations, predictive models and other analytics rather than a chatbot conversation. Some surveyed users may touch AI only occasionally. Others may spend opening week conversing with a model like an offensive coordinator who cannot leave the building.

That is why the table uses 10%, 25% and 40% adoption cases. The central scenario treats the current industry survey as the best available proxy. The conservative case assumes most nominal AI users sit out this window. The frantic case assumes opening-week intensity, product promotion and embedded tools pull more managers into the machine.

The direction of travel is obvious. ESPN’s 2025 partnership announcement with IBM said its fantasy-football platform had more than 14 million players and was adding weekly watsonx features including waiver grades, trade grades, boom-and-bust probabilities and a trade analyzer. The NFL now sells an official fantasy assistant. Independent products offer roster chats, draft copilots, trade explanations, lineup optimizers and personalized recaps. AI did not sneak into the league. It received a depth-chart position and a subscription tier.

Twenty Questions Is a Normal Amount of Panic

The central case assumes each AI user produces 20 visible prompt-and-answer exchanges over 14 days. That sounds high until we reconstruct the week.

A manager might ask for a draft plan, then ask for changes based on draft position. He might compare two running backs, request sleepers, paste a roster, explain the league’s scoring, reconsider the roster after a late pick, ask about a trade, inspect the Week 1 matchup, check an injury, request a start-sit recommendation, demand a waiver list and then repeat half the exercise after the opening game makes one previously obscure player look like the second coming of LaDainian Tomlinson.

Twenty exchanges is fewer than one and a half per day. It is not 20 separate hours of monastic research. It is one long conversation with follow-ups, plus the occasional panicked second opinion. The conservative case gives five exchanges to 10% of players. The frantic case gives 60 exchanges to 40%, which is what happens when every decision receives research, rebuttal, counterfactual simulation and a motivational speech.

This modeling choice is more important than the population math. At fixed prompt complexity, doubling the exchanges doubles the energy and water. The environmental footprint of AI is mostly a multiplication problem performed on numbers companies rarely disclose at the same boundary.

Our earlier look at how ordinary people use AI agents found the same behavioral pattern: users happily delegate research and recommendation, then become more cautious at the final action. Fantasy football is almost perfectly designed for this arrangement. The bot may collect injury news, compare usage and suggest a lineup. The human still taps “save,” preserving the ancient right to blame himself.

A Prompt Is Not a Unit of Nature

Now we reach the number most AI-footprint arguments pretend is simple: energy per query.

Google published one of the most comprehensive production measurements in 2025. Its researchers estimated that the median Gemini Apps text prompt used 0.24 watt-hours of electricity and 0.26 milliliters of water. The measurement included accelerator power, host systems, idle capacity and data-center overhead. This is much more useful than looking at a GPU’s maximum rating and assuming it spends every second answering your question alone like a private concierge furnace.

Microsoft researchers approached the question from the bottom up. Their 2025 paper modeled optimized, production-scale inference on H100 nodes and estimated 0.31 Wh for a standard frontier-model query with a median output of 300 tokens. For test-time scaling—a reasoning-style query with a median 5,000-token output—the estimate rose to 3.91 Wh. That is about 13 times the standard case, not because the machine becomes emotionally invested, but because it generates far more tokens.

The phrase “one query” therefore covers a spectrum. “Who should I start, Player A or Player B?” may be short. “Analyze my roster under full-PPR scoring, retrieve current injury and weather information, compare betting lines and defensive tendencies, simulate game scripts, then explain the downside case” is a small research assignment wearing a football jersey.

For the central estimate, I use a blend: 80% standard queries at 0.31 Wh and 20% long reasoning queries at 3.91 Wh. The weighted average is 1.03 Wh per visible exchange. The conservative case uses Google’s 0.24 Wh median. The frantic case treats every exchange as a long 3.91 Wh query.

This is not a claim that every provider uses the same model, chip, utilization, cooling system or serving software. They do not. Smaller models can be far cheaper. Poorly batched or lightly used systems can be worse. Current flagship models may run on newer hardware than the paper’s H100 baseline. The inference-chip market exists precisely because efficiency is now a business, not an inspirational poster taped to the data-center door.

The Central Case Is 229 MWh, Not “A Small Country”

The electricity calculation is pleasingly unromantic:

44.48 million fantasy-football players × 25% AI adoption × 20 exchanges × 1.03 Wh = 229.1 MWh.

That is enough electricity to supply the average U.S. home for roughly 19 years using the EPA’s 12,194-kWh annual household figure. It equals a continuous average draw of about 682 kW during the two-week window. The conservative case is 5.3 MWh. The frantic case is 4.17 GWh, equivalent to more than 342 average U.S. home-years and a continuous 12.4-MW load.

Please notice what I did not write: “Fantasy football AI uses as much power as 19 homes.” Power and energy are not interchangeable. A home-year is an energy comparison. The instantaneous load depends on when people ask. If 20 million managers all submit lineup questions at 12:47 p.m. Sunday, the peak matters to the servers and grid even though the season-total energy does not change.

This distinction gets flattened every time a data center is described as “using the power of a city.” A facility has a maximum electrical capacity, a changing instantaneous load and an annual energy total. Those are related measurements, not synonyms wearing safety vests.

SiliconSnark has spent much of 2026 following the physical bill behind the chatbot. Our guide to AI infrastructure explains the full stack. Meta’s Louisiana buildout turned the issue into a $50 billion regional extension cord. Etched made inference hardware look like a $10.3 billion machine with cooling attached. Fantasy advice is not the reason these facilities exist. It is one tiny illustration of how millions of trivial, useful and recreational requests become infrastructure demand.

The Cooling Water Is Only the First Water Bill

Water accounting is where a clean single number goes to become a committee.

Google’s median Gemini prompt used 0.26 milliliters of water at 0.24 Wh. That figure describes operational water consumption within the company’s serving boundary. To estimate our scenarios, I preserve Google’s measured water-to-energy ratio—about 1.08 liters per kWh—and scale it with the energy of each query. This produces 248,000 liters, or 65,600 gallons, of direct water in the central case.

That is the liquid associated with cooling the computation. It does not include the water consumed elsewhere to generate the electricity.

Lawrence Berkeley National Laboratory’s U.S. Data Center Energy Usage Report separates the two ledgers. It estimated average on-site water-use effectiveness at just over 0.36 liters per kWh in 2023, rising toward 0.45–0.48 L/kWh by 2028 as hyperscale and liquid-cooled AI facilities expand. Google’s production ratio is higher than that fleet average, which is one reason mixing corporate and national figures deserves labels.

The same Berkeley Lab report estimated 4.52 liters of indirect water consumption per kWh for electricity used by U.S. data centers in 2023. Power plants can consume water through cooling and other processes. The amount varies dramatically by generation technology, plant design, climate, location and what exactly an analyst counts.

Apply 4.52 L/kWh to the central 229 MWh and upstream electricity adds roughly 1.04 million liters, or 273,500 gallons. Add the Google-derived direct cooling estimate and the broader water footprint becomes about 1.28 million liters, or 339,000 gallons.

This is why one headline can claim an AI prompt uses five drops while another claims it uses tablespoons. They may be measuring different models, different workloads and different water boundaries. “Direct water” asks what the data center consumes. “Indirect water” asks what the power system consumes to make the electricity. “Embodied water” could go farther still, into chip fabrication and construction. The answer changes because the question changes, a concept fantasy managers understand whenever a ranking forgets to ask whether the league awards a point per reception.

Water Has a Location, a Season and Very Poor Public Relations

Even the broad 339,000-gallon estimate is not a prediction of water leaving one reservoir. AI requests are routed across facilities and regions. A data center using evaporative cooling in a dry area does not have the same local consequence as one using recycled water in a wet region. Electricity generated by a once-through cooled thermal plant has a different withdrawal and consumption profile from wind, solar or dry-cooled generation.

“Consumed” also does not mean all water vanished from the planet. In environmental accounting, consumption generally means water withdrawn and not returned to the original source because it evaporated, entered a product or otherwise became unavailable locally. Withdrawal can be much larger than consumption when water is returned. People hear “used” and understandably picture a server drinking a municipal swimming pool through a silly straw. The plumbing is less cinematic and more important.

The central direct figure—66,000 gallons—is about one-tenth of an Olympic pool. The broad figure—339,000 gallons—is roughly half of one. The frantic scenario reaches 1.19 million gallons direct and 6.18 million gallons broad, around nine Olympic pools. These comparisons are for scale, not absolution. A pool of water spread across a national two-week activity is tiny. A pool taken from a constrained local watershed during a hot month is a planning question.

This is the same collision between global abstraction and local infrastructure we saw in Maine’s fight over an underwater data center. National percentages make loads look modest. Communities experience pipes, substations, wells, tax agreements and specific bodies of water. The cloud always lands somewhere.

No, Fantasy Football Is Not Eating 1,500 Trees

The user asked about trees, so let us rescue the trees from bad equivalence math.

Servers do not consume trees during inference. Trees are sometimes used as a carbon-sequestration comparison: how many would need to grow for a defined period to absorb the carbon dioxide associated with the electricity?

The EPA’s national factor for electricity used is 0.394 kilograms of CO₂ per delivered kWh, including transmission losses. Applying that grid-average factor to 229 MWh produces approximately 90 metric tons of CO₂.

The EPA also estimates that an urban tree seedling grown for 10 years sequesters about 0.060 metric tons of CO₂ under its survival-weighted assumptions. Divide 90 by 0.060 and the central case equals the carbon absorbed by about 1,500 urban seedlings grown for a decade. The conservative case is 35 seedlings. The frantic case is 27,400.

Call these tree-decade equivalents if you enjoy having friends leave the table. They are not trees cut down, permanently required or purchased as offsets. They are a visualization of grid-average emissions.

The carbon estimate may overstate the emissions reported by major AI providers that operate efficient facilities and procure low-carbon electricity. Google measured only 0.03 grams of carbon dioxide equivalent for its median 0.24-Wh Gemini prompt, far below the result from applying the U.S. grid average. Corporate clean-energy contracts, facility locations and market-based accounting all affect that number.

It may also understate full lifecycle impact because our calculation excludes manufacturing accelerators, constructing data centers and training the models. There is no honest way to transform one fantasy-football question into an exact fraction of a chip fab without choosing an allocation rule. The tree number is a carbon illustration, not ecological box score.

The Frantic Case Begins When One Question Secretly Becomes Six

The visible chat box encourages us to imagine one question entering one model and one answer coming out. Modern AI products increasingly behave like small organizations.

The NFL says its Fantasy AI Assistant, built with the AWS Generative AI Innovation Center, uses a multi-agent architecture. Specialized agents retrieve data, analyze it, generate insights and synthesize the response across Next Gen Stats, fantasy data, news and matchup information. The product can answer start-sit, waiver, injury and rest-of-season questions.

That design is sensible. A general language model should not rely on its frozen memory to know whether a player missed Friday practice. Retrieval brings current facts. Separate analytical steps can enforce scoring context and compare options. A final model translates the output into readable advice. This is how AI becomes more reliable: give it tools, fresh data, checks and narrower jobs.

It is also how one user action becomes several inference events, searches, database calls and intermediate outputs. Microsoft’s energy paper explicitly notes that multi-step agentic workflows can be modeled as multiple queries and that orchestration, tool calling and long-context overhead were outside its per-query estimate.

Our central case does not apply an agent multiplier. It treats one visible exchange as one modeled inference event. That makes 229 MWh conservative for products where every question wakes a depth chart of agents. The frantic case—long reasoning energy applied to 1.07 billion exchanges—is partly a way to bound that hidden complexity. It produces 4.17 GWh, 6.18 million gallons of broad water consumption and the carbon equivalent of 27,400 seedlings grown for a decade.

When we asked whether AI agents actually make money, the answer depended on whether the system could reach useful data and complete a workflow. Fantasy football offers a friendly version of the same rule. A chatbot that knows football generally is entertainment. A chatbot that knows your roster, rules, injuries, waivers and deadline is a product. Context improves the answer. Context also has to be moved and processed.

The Conservative Case Is What Efficiency Looks Like When It Works

The low case assumes 10% adoption, five queries per user and 0.24 Wh per query. The result is only 5.3 MWh, 1,500 gallons of direct water and 7,900 gallons under the broad water boundary. Its carbon equals roughly 35 seedling-decades.

This is plausible if most users ask a few short questions, products route simple work to smaller models, popular advice is cached and providers batch requests efficiently. A platform does not need a frontier reasoning model to return a stored injury status, sort a projection table or explain that the defense ranked 28th against wide receivers last season. Calling the largest model for every step would be like sending the team owner to retrieve the kicking tee.

AI efficiency has improved quickly. Google said the energy use of its median Gemini prompt fell 33-fold in one year. Microsoft found large gains available across models, serving systems and hardware. The IEA’s 2026 update says energy per AI task is declining rapidly even as total demand rises because more people are using AI and agentic workloads are becoming more intensive.

This is the rebound problem in tiny shoulder pads. Each answer becomes cheaper. The product therefore adds answers, summaries, agents, personalization and automatic recaps until aggregate demand keeps climbing. Efficiency does not fail. It makes more consumption economically possible.

SiliconSnark calls the larger industrial version AI capexmaxxing: extraordinary infrastructure spending justified by the promise that intelligence will become cheap enough to place everywhere. A fantasy lineup assistant is not the capex thesis. It is what the thesis looks like after it reaches the couch.

AI Did Not Invent Fantasy Analytics. It Gave the Spreadsheet a Mouth.

Fantasy football was computational long before language models learned to spell “YAC.” Platforms have always processed rosters, scoring rules, schedules and player statistics. Projection systems blend historical performance, opportunity, team context, opponent strength and market information. Optimizers rank lineups under constraints. Machine learning has powered advanced football metrics for years.

The NFL’s Next Gen Stats operation already turns player-tracking data into completion probability, expected rushing yards, win probability and other modeled outputs. IBM and ESPN have applied natural-language processing and predictive techniques to trades and player performance across multiple seasons. The new generative layer does not replace those systems. It sits in front of them and converts a database operation into a conversation.

That interface matters. A table can tell you Player A is projected for 13.8 points and Player B for 13.2. A conversational system can ask whether you need floor or upside, notice that your opponent already started a player who exploded on Thursday, account for your league’s bonuses, retrieve a practice report and explain why the difference is smaller than the decimal places suggest.

The value is synthesis, personalization and accessibility. The risk is that polished language disguises weak inputs. If an assistant cannot access current rosters, scoring rules, injuries and depth charts, its eloquence is a liability. It will confidently explain a football season that exists in its training data rather than the one happening outside your window.

This is a recurring lesson in our field guide to AI models and assistants: the model is only one layer. The product is the model plus retrieval, tools, memory, permissions and distribution. For fantasy football, add weather, practice status and a waiver deadline that waits for neither benchmark nor apology.

Can 229 MWh Actually Beat Kyle From Accounting?

AI can make a fantasy manager better informed. It cannot make football deterministic.

A useful system can compare projections, identify a workload change, summarize injuries, flag a favorable matchup, translate scoring rules and expose the assumptions behind a recommendation. It can search more material than a casual player has time to read. It can remember that your league gives six points for passing touchdowns and penalizes missed field goals, rules that generic rankings sometimes treat as decorative.

The machine remains trapped inside the same uncertainty as everyone else. Coaches change plans. Players aggravate injuries. A defensive back falls down. A quarterback launches a beautiful pass into 25 miles per hour of weather. A touchdown goes to the blocking tight end because the universe enjoys content.

Fantasy outcomes are particularly noisy because touchdowns and turnovers have enormous scoring effects and relatively small sample sizes. A correct projection is a distribution, not a prophecy. If one player has a projected mean of 13.8 and another 13.2, the honest conclusion is often “close decision,” not a seven-paragraph declaration that one is the optimal asset.

Generative AI can make this worse by turning uncertainty into prose. The answer grows more confident because the sentence is complete. A good assistant should surface ranges, assumptions and freshness. It should say when the choice is marginal. It should identify which new fact would change the recommendation. It should not call a 52% preference “the data-driven answer” and then add three fire emojis.

The highest-value AI use may therefore be compression rather than prediction: take five injury reports, three matchup notes, your scoring system and two plausible game scripts, then produce the decision case. That saves time. Whether it wins the week still depends on professional athletes making orderly contributions to your private spreadsheet.

The Right Comparison Is Not Always Zero

Our estimate calculates incremental AI inference. It does not prove that every watt-hour is new consumption caused by fantasy football.

Without a chatbot, the manager may open a dozen websites, stream rankings videos, listen to podcasts, refresh social feeds and leave three glowing monitors displaying target-share charts. AI could replace some of that activity. It could also encourage more activity by making research easier, producing unlimited follow-ups and inserting advice into moments when the user would previously have accepted ignorance.

The counterfactual is unknowable without behavioral data. A one-minute AI answer might substitute for 20 minutes of browsing. A 45-minute AI spiral might substitute for making a decision and walking the dog. The environmental question is not merely how much the model uses. It is how the product changes the whole information routine.

Our table excludes phones, laptops, home networks, displays, ordinary fantasy-app servers, advertising systems, video streaming and the electricity used to watch the games. It also excludes the existing analytical pipelines that generate projections before anyone asks a chatbot. Those systems may be substantial, but allocating them to AI-assisted picks would require usage and infrastructure data the platforms do not publish.

This boundary keeps the estimate legible: the added generative-inference layer. It also means the result is not “the environmental footprint of fantasy football.” The full footprint would include media production, broadcast networks, stadiums, travel, devices and enough chicken wings to trigger its own watershed assessment.

Training Does Not Happen Again Because You Asked About a Tight End

Another common mistake is to attach an entire model-training run to current prompts without a defensible denominator.

Training large models consumes significant electricity and water. Manufacturing accelerators, constructing facilities and fabricating chips have embodied impacts. Those costs belong in lifecycle accounting. They do not recur every time a manager asks whether to start the Buffalo defense.

To allocate training energy per fantasy question, we would need the model’s training footprint, expected useful lifetime, total lifetime inference volume and a rule for assigning shared costs across commercial, educational, medical, coding and recreational use. Change the lifetime query estimate and the allocated training cost changes. Guess badly and the result becomes arithmetic theater.

So this article does not amortize training or hardware manufacturing. It measures modeled inference electricity, direct cooling water, upstream generation water and grid-average operational carbon. The exclusions are not proof those impacts do not exist. They are an admission that precision requires a boundary.

The industry frequently prefers one magic sustainability number because magic numbers fit dashboards. Reality arrives as scopes, locations, percentiles and allocation choices. The plumbing is the point, and the plumbing would like you to read the footnote.

How to Stop the Robot Coach From Calling Four Audibles

Individual users should not be made to feel guilty for asking a useful question. The central two-week estimate is tiny beside data-center demand, air conditioning, transportation and the broader entertainment footprint of professional football. The large decisions belong to model providers, cloud companies, utilities and regulators.

Still, efficiency and answer quality happen to agree on several practical habits:

• Batch the roster context. Give the assistant your format, lineup, bench and decision set once instead of reconstructing the league in six separate conversations.
• Ask for current sources. A concise answer grounded in fresh injury and depth-chart data is more useful than a long answer generated from stale model memory.
• Request a decision rule. Ask what assumption would flip the recommendation. This exposes uncertainty and reduces the ritual of regenerating until the machine agrees with you.
• Use the smallest adequate tool. A projection lookup does not require a 5,000-token reasoning monologue. Providers should route routine retrieval and calculation to efficient systems.
• Do not ask for five versions unless five versions change the decision. “Make it more savage” is a defensible request for a league recap. It is not medical care.

Platform design matters much more than user etiquette. Caching a widely requested explanation, sharing retrieved data across requests, using structured calculations, routing by task complexity and keeping outputs concise can reduce load without turning every fantasy player into a junior data-center engineer.

The best system will not use a giant generative model to calculate arithmetic, a search agent to rediscover stored facts and a second giant model to congratulate the first. This sounds obvious. So did not putting a touchscreen on a refrigerator.

Fantasy Platforms Need an Environmental Box Score

The FSGA found that 85% of fantasy players and sports bettors want AI-generated content labeled. That is a request for provenance: tell users when the analysis came from a machine.

Platforms should add a quieter second label for the people trying to measure the machine. They need not display a guilt meter beside every waiver claim. They should publish enough aggregate information for credible accounting:

• median and high-percentile energy per visible user request;
• the number of model calls triggered by an average request;
• the share of requests routed to small, standard and reasoning models;
• average input and output tokens;
• cache and retrieval behavior;
• direct water consumption and the geographic basis of the estimate;
• location-based and market-based carbon figures;
• whether upstream electricity water and embodied hardware impacts are included.

Without these disclosures, researchers are forced to combine one company’s prompt measurement, another laboratory’s reasoning model, a national grid factor and a fantasy-industry survey. That is what I have done here. It is transparent and directionally useful. It is not a substitute for operators measuring their own products.

AI companies already optimize latency and cost at exquisite resolution. They know tokens, accelerators, utilization and routing because invoices have achieved observability. Environmental measurement is technically possible. It has merely not received the same product urgency as making the assistant write a limerick about your opponent’s 0–2 start.

The Final Score: Small Footprint, Enormous Group Chat

Between September 4 and 18, the defensible central estimate for U.S. AI-assisted fantasy-football picking is:

• 44.5 million fantasy-football players;
• 11.1 million using AI in the central adoption case;
• 222.4 million visible prompt-and-answer exchanges;
• 229 MWh of inference electricity;
• 65,600 gallons of direct cooling water;
• 339,000 gallons when estimated upstream power-generation water is added;
• 90 metric tons of CO₂ using the U.S. grid-average factor;
• and the carbon-absorption equivalent of roughly 1,500 urban tree seedlings grown for 10 years.

The conservative case is radically smaller. The frantic agentic case is radically larger. Both are plausible because the decisive variable is not whether a chatbot answer requires “a bottle of water.” It is whether millions of people ask five short questions or dozens of long, tool-using ones—and whether one visible request wakes one model or an entire virtual coaching staff.

None of this means AI fantasy advice is an environmental emergency. It means ordinary digital conveniences have physical denominators. One prompt is negligible. A national habit is measurable. A national habit repeated across shopping, coding, school, work, health, entertainment and every other decision becomes the infrastructure boom we keep describing as if it materialized spontaneously behind a tasteful gradient.

Use the robot coach. Demand current data. Ask it to show uncertainty. Prefer one good, roster-aware analysis to 12 regenerations of the same hunch. Then make your decision and accept the part no model can optimize away: the player you bench will score two touchdowns.

The trees had nothing to do with it.