Check the work

1. The building

The model starts from a reference design rather than from a published per-prompt figure, because a building can be audited and a per-prompt figure cannot. The reference is a Basis of Design for a 100 MW facility, Rev 03-S. Critical IT means the power available to the computing equipment itself, after the facility's own overhead is excluded. Of that, 80% goes to GPU compute, 80 MW; the rest is storage and network.

The compute is entirely GB300 NVL72 racks. A rack draws 145 kW on average across a mixed workload: inference runs at 110 to 150 kW, training at 130 to 180. The rack count follows directly.

The building is about ≈216,000 SF, 600 by 360 ft, with 4 data halls of about ≈24,000 SF each. The site is about ≈140 acres, most of it the chiller yard: 620 by 400 ft, with ≈60 chillers of 630 tons on a 62 ft grid, air-cooled, N+1, with integrated free cooling. Electrical distribution is ≈60 line-ups of 2 MW, distributed redundant in a 5-to-make-4 arrangement. Liquid cooling captures 93% of the heat at the chip; the residual is airside, for network and storage. Staffing is 24 to 30 people.

The building's domestic water, for the people in it, is about ≈2,000 L/day. Divided over the users it serves that is 0.7 µL per user-day, and it is the only water term the narrative does not bother to draw.

2. The power

Power usage effectiveness is total facility energy divided by IT energy. A PUE of 1.20 means that for every kilowatt-hour the computers use, the building uses another 0.20 for cooling, power conversion and lighting. It is a ratio of energies over a period, so the number depends on the period.

The peak design PUE of this building is 1.35 PUE: the W32 design lane, five-year extreme ambient, 100% IT load. That is what the chillers are sized for and it is not an operating point. Annualized, with free cooling available whenever the outdoor air is below the 32 °C facility supply temperature, the design lands at 1.20 PUE. The headline scenario uses 1.20, inside that band and slightly conservative. For comparison, Uptime Institute's industry average is 1.52 PUE, flat for seven years; hyperscale fleets report 1.08 to 1.12 PUE; Google's fleet reports 1.09 PUE; new AI builds land at 1.15 to 1.25 PUE.

This design pays for its zero water in electricity. Dry cooling costs an additional 0.30 to 0.74 kWh/kWh per kWh of IT load against evaporative cooling, and that electricity has its own water footprint at the plant. Section 5 counts it.

3. The tokens

Energy per user depends on how many tokens the fleet actually processes per second, which is much lower than the benchmark number. Start from a benchmark blended throughput of 500,000 tok/s per rack and derate it three times, each factor named.

Reality anchor. OpenAI has stated it processes 15.00 billion tokens/min on the API, which is 21.60 trillion tokens/day. At 200,000 tok/s per rack that is ≈1,250 racks, roughly ≈181 MW of GB300 compute. OpenAI's actual fleet is far larger, for three reasons that do not undermine the model: most of the installed base is Hopper and Ampere, and NVIDIA cites 90 tok/s per GPU on Hopper against 6,000 tok/s per GPU on GB300, a 67× gap; training consumes an enormous share; and they are building for future demand. The model here describes what a modern inference fleet costs per token, which is the question a per-person figure needs answered.

4. The user

A day of AI use is taken as 8,000 tokens processed. Processed tokens include the prompt, the context the model re-reads on every turn, any hidden reasoning tokens, and the output. That is far more than the words on screen, and it is why this figure is a fair one for a person who sends about ≈4 substantial prompts a day with some back-and-forth. It is also why the per-prompt spread in published figures is so wide: a reasoning model can spend thousands of hidden tokens on one answer.

The whole-facility 100 MW is used, not the 80 MW compute block, because the storage and network that yield no tokens still draw power on the user's behalf. Users served by one building: 9.50 trillion tokens ÷ 8,000 = ≈1.19 billion users.

Per prompt, the same day is 0.605 Wh at datacenter scope, and 0.47 mL of water at roughly 700 processed tokens per query, the basis used in the almond correction in Act 7. Published anchors bracket this: Google's own measurement of the median Gemini Apps text prompt, May 2025, is 0.24 Wh at datacenter scope, excluding network, device and training own disclosure; Sam Altman's stated average ChatGPT query is 0.34 Wh own disclosure; EPRI's estimate for a ChatGPT prompt is 2.9 Wh against 0.3 Wh for a traditional search independent estimate; reported GPT-5 averages run to ≈18 Wh with heavy responses to 40 Wh third-party estimate.

5. The water

Water use effectiveness is liters of water consumed on site per kilowatt-hour of IT load. It counts evaporation from cooling towers and adiabatic pre-coolers, and it counts nothing upstream. An air-cooled, closed-loop building has a WUE near 0.0 L/kWh: the coolant loop is sealed and the chillers reject heat to air. The headline scenario uses 0.50 L/kWh, an average for a modern hybrid build, deliberately not the reference design's zero.

The larger term is indirect. Thermoelectric generation evaporates water in its own cooling towers, and hydro reservoirs evaporate water from their surface. The US thermal-fleet average used throughout is 1.8 L/kWh. Alternatives, all adjustable above: 1.2 L/kWh for a gas combined-cycle heavy grid; 3.1 L/kWh from Li, Yang, Islam and Ren; and 4.5 L/kWh for the 2023 US national figure including reservoir evaporation, which is 211.00 billion gallons against 176 TWh.

At the headline assumptions the split is 1.01 mL on site and 4.35 mL at the plant: 19% and 81%. Indirect dominates in every scenario, which is why Google's site-only 0.26 mL per prompt and Sam Altman's site-only 0.32 mL per query are both smaller than the figures here. Shaolei Ren of UC Riverside criticised the Google figure for excluding generation water. The criticism is correct, it applies equally to the Altman figure, and it is why this site counts grid water in every row.

The worst scenario on the table is 16.31 mL, about three times the headline. It is still under a tablespoon. The dry-cooling trade is the honest one: about ≈2.15 L/kWh saved on site per kWh of IT load, for 0.30 to 0.74 kWh/kWh of extra electricity, which at 1.8 L/kWh evaporates roughly 0.5 to 1.3 L per kWh of IT load at the plant. A net saving of water, paid in electricity, and the reason the narrative says air cooling moves water rather than eliminating it.

The landscape: AWS reports 0.12 L/kWh fleet-wide, from 0.02 L/kWh in Stockholm to 2.85 L/kWh in Jakarta, a 140× spread inside one operator. Equinix reports 0.91 L/kWh. Microsoft's regional figures include 0.02 L/kWh in Singapore and 1.52 L/kWh in Arizona; see the notes in the appendix on what could and could not be confirmed against Microsoft's own documents. The legacy industry average is 1.8 to 1.9 L/kWh. Best-in-class runs under < 0.05. Nevada's rule is the Southern Nevada Water Authority's moratorium on evaporative cooling in new commercial and industrial buildings; Singapore's roadmap sets a target of 2 L/kWh or lower.

6. The comparisons

The television

The site's unit is one minute of a 96 W television. That wattage is not chosen here; it is reverse-derived from Google's own framing that a 0.24 Wh prompt is "less than nine seconds of TV" (0.24 Wh ÷ 9 s × 3,600 = 96 W). The site is extending Google's unit, not inventing one. Google's water framing, "about five drops", implies the same site-only scope as its 0.26 mL.

Every consistent pairing lands between 90 and 112 seconds. The mismatched pairing reaches 146 seconds and breaks the two-minute claim, which is exactly the kind of scope mixing that this page criticises in others. The scope field in the dataset prevents the narrative from rendering it; the comparison function refuses the pairing rather than labelling it.

The digital day

The full-chain rows in Act 5 split each activity into device, network and datacenter. Streaming uses the IEA allocation of 72% devices, 23% network and 5% datacenters, anchored on the same 96 W television for three hours: 288 Wh on the set. Network intensity is the weakest input: 0.02 kWh/GB fixed and 0.06 kWh/GB mobile, the 2019 IEA range halved for efficiency gains since. Published kWh/GB estimates span an order of magnitude, and the uncertainty runs in AI's favor: older, higher figures make streaming look worse, not AI. The social media datacenter figure, 13 Wh, is Meta's disclosed datacenter electricity divided by daily active people. Usage is 2 h 21 min h of social media a day globally and about ≈7 h of total US screen time. Device draws implied by those rows, ≈3 W for a phone and 30 W for a laptop, are reused for this page's footprint below.

The objects

Almond. 3.56 L per nut, slightly under a gallon, as attributed in press coverage to a 2019 USGS-affiliated study. The Fulton, Norton and Shilling paper linked in the appendix reports about 12 L per kernel on a different basis; if that basis were used, every almond comparison on the site would move further in AI's favor. The figure used is the smaller one.

Hamburger. The narrative says "a fast-food hamburger" and this page gives the basis: a McDonald's hamburger, 1.6 oz raw patty (45 g) plus bun and condiments. At the global-average beef footprint of 15,400 L/kg (Mekonnen and Hoekstra, 2012) the patty alone is about 700 L and the sandwich about 800 L; at a US feedlot figure it is about 330 L. ≈500 L is the planning central and the range is carried on the figure.

Delivery. One parcel, from the sorting depot to the door. The drive is a class 4 diesel step van on a suburban route, UPS-derived: 100 miles, 151 packages, 8.7 gallons of diesel, 7.8 MJ per parcel, which this site converts to 2.2 kWh (Stolaroff et al., 2018). The same paper gives the carrier's facilities, from UPS's 2015 company-wide totals divided by packages delivered: 0.35 kWh of electricity and 0.93 MJ of natural gas per package, 0.26 kWh as energy. Together, 2.81 kWh. Only the facility electricity carries power-plant water, 630 mL at the site's grid intensity. The warehouse that stored the item, the packaging and the long-haul leg are not counted; no defensible per-parcel figure exists for them, and each would make the parcel larger. An electric truck on the same route is 0.72 kWh at the plug; a small electric van on a dense city route is 0.13 kWh (Rodrigues et al., 2022); a gasoline van at half a mile between stops is 1.0 to 1.6 kWh well-to-wheel (Li, Kim and Keoleian, 2021). No carrier publishes a per-parcel energy figure.

Grid water for electric things. Anything electric on the story page gets its water the same way the AI figure does: the electricity times 1.8 L/kWh. A phone charge is 23 mL; boiling the water for a cup of coffee, 23.7 Wh, is 43 mL; heating a shower is 4.1 L; an electric-car mile is 540 mL.

Coffee, two ways. The 140 L per cup on the story page is the total footprint, almost all of it rain on the plantation. Water Footprint Network's own split puts the pumped share at about 1.0 L per cup, from 139 cubic meters per tonne of roasted coffee at 7 grams a cup. The story page uses the total, as it does for the hamburger, and says so.

Toilet. 6 L per flush matches the federal maximum of 1.6 gal in force since 1994. WaterSense toilets use 1.28 gallons or less; older toilets up to 6.

Shower. 65 L for eight minutes, and 2,270 Wh to heat it. Coffee, egg, bread, t-shirt. Water Footprint Network product footprints: 140 L, 200 L, 40 L, 2,700 L. Streaming hour. ≈133 Wh, one third of the three-hour row.

Blue versus green. Blue water is withdrawn from rivers, lakes and aquifers and consumed. Green water is rainfall held in the soil and transpired by plants. The comparison objects are total footprints, blue and green together; the AI figure is blue plus generation water. That mismatch is labelled in Act 2 and it runs against AI. The distinction matters most for beef in Section 7.

The almond correction

The widely circulated figure of 38,000 queries per almond does not reconcile with Altman's own 0.32 mL per query: 3.56 L ÷ 0.32 mL = 11,125 queries. Reaching 38,000 needs 0.094 mL per query, 3.4× lower than stated. This site's model at 0.47 mL per query gives about ≈7,500 queries. Directionally the claim is right; the arithmetic as quoted is not; and scoped consistently on site against the Act 4 reference design the ratio is infinite, so the argument is stronger than it was made.

7. The aggregate

US annual blue water, national scale, same metric. All US datacenters in 2023, direct plus indirect: 0.80 trillion L (LBNL). US golf courses, 2024 data: 1.59 trillion L, from GCSAA's 1.63 million acre-feet applied, net of the 21% reclaimed share. California almonds alone: 2.59 trillion L from 2.1 million acre-feet applied, or 4.70 trillion L on the higher footprint estimate. US beef, blue water only: 6.76 trillion L. US beef, total footprint including rainfall on pasture: 62.4 trillion L, of which most is green water and would have fallen regardless. Per US resident on 340.00 million people: datacenters 2,353 L, golf 4,672 L, almonds 7,619 L, beef 19,897 L.

Vintage and verification. LBNL reports about 66 billion liters direct and nearly 800 billion liters indirect for 2023, so the 0.80 figure matches the indirect term and understates the total by roughly 8%. That understatement runs in AI's favor and is recorded on the figure rather than corrected silently, because the brief's number is the one the narrative was built on. LBNL's 2025 Update, published June 2026, puts 2024 at 192 TWh (4.7%), 2028 at 464 TWh, and 2030 at 649 TWh; the original brief's 1,050 TWh for 2030 was BCG's high case and is no longer used. Rystad projects datacenter water consumption could roughly triple by 2030, with 34% of direct consumption in high-water-stress regions. S&P Global finds 43% of datacenters globally already exposed to high water stress.

On a straight line to triple, the datacenter bar passes golf in 2,026 and reaches 2.40 trillion L in 2030, still below almonds and well below beef's blue water. The story page says so in one sentence rather than leaving it to a critic.

Electricity, national

All US datacenters used 176 TWh in 2023, 4.4% of US electricity by LBNL's count, which excludes cryptocurrency mining. Against retail sales of 3,874 TWh the share is 4.5%. The story page sets that beside three household uses nobody proposes to switch off, from the EIA's 2020 Residential Energy Consumption Survey: home air conditioning 254 TWh (6.6% of retail sales), electric water heaters 156 TWh (4.0%), electric clothes dryers 56 TWh (1.4%). The household figures are 2020 survey data against 2023 sales, which the story page states. LBNL's 2025 Update projects 464 TWh for 2028 and 649 TWh for 2030, a range of 521 to 843 TWh.

Household bills: the evidence in both directions

The story page says a datacenter can raise household bills and that it depends on where you live. Four findings stand behind that, chosen because each comes from a government body, a national laboratory or an independent market monitor.

Virginia. The state's legislative auditor projected that a typical Dominion household's generation and transmission charges could be $14 to $37 a month higher by 2040, in constant dollars, depending on how much demand arrives. The same report found that datacenters "are currently paying full cost of service, but growing energy demand is likely to increase costs for other customers." It covers one utility and one part of the bill. As a yearly amount it is $168 to $444 a year; the figure is the increase in the monthly bill by 2040, not a total over the period.

Proportion. The US average residential price rose from 13.15 cents/kWh in 2020 to 17.30 cents/kWh in 2025, 32% before inflation. The Energy Information Administration does not attribute a share of that to datacenters; its own analysis points to capital spending on aging generation and delivery infrastructure. In Virginia, the regulator's November 2025 order raised the typical residential bill by $11.24 a month for 2026 without attributing the increase to datacenters. Datacenters are one driver of bills among several, and the story page says so.

The mid-Atlantic grid. PJM's independent market monitor attributes $29.4 billion, 46.2% of capacity charges across the last four auctions, to datacenter demand, existing and forecast. These are wholesale costs from a counterfactual simulation, and they reach household bills to different degrees by state and utility. The monitor's widely quoted larger percentages are increases over its simulated no-datacenter case, not shares of the total, and are not used here.

The other direction. Lawrence Berkeley National Laboratory and Brattle found that load growth "has tended to depress retail electricity prices in recent years, by spreading fixed costs over greater load": 0.6 cents/kWh lower for a 10% rise in a state's use from 2019 to 2024, a pattern their April 2026 update found still held through 2025. Households benefited less than businesses, and the authors warn the past "need not be prelude," naming PJM as the counter-example.

Rules adopted since 2025. Ohio's commission approved a tariff for new datacenters in July 2025; Indiana's approved large-load terms in February 2025; Georgia's adopted a rule in January 2025; Virginia's created a separate rate class in November 2025. Each requires long contracts and minimum payments. In June 2026 the Federal Energy Regulatory Commission ordered all six regional grid operators to justify or reform how they connect large loads and recover their costs.

Not used, and why. No primary source gives a national dollars-per-household figure, so none appears. A consumer advocate's per-month estimates for Maryland describe the whole capacity price jump, not the datacenter share. Savings claimed by utilities, an industry-funded preprint finding that datacenters lowered rates, a law school paper arguing costs are being shifted, and a Federal Reserve Bank of Dallas working paper estimating a rise in wholesale prices are all informative and all contested; they are named here rather than quoted on the story page.

8. What would change these conclusions

• The per-person figure is small. It would move materially if fleet throughput were below about 50,000 tok/s per rack (a further 4× derate), if a typical day were above about 50,000 processed tokens, or if the median user were on a reasoning model at 18 Wh per prompt. Any of those would put a typical day at 10 to 45 minutes of television. None would put it at an hour of streaming.
• The water is at the power plant. It would be false for a building with WUE above about 2.5 L/kWh on a low-water grid, such as an evaporative build in a hot climate on gas combined-cycle. Those buildings exist and are the ones being banned.
• AI loses to golf on a national scale. It stops being true in 2,026 on the straight line above, and earlier if LBNL's direct term is added. The site says so.
• The two-minute claim. It would fail if the television were materially less than 96 W (a 40 W set makes the day about four minutes of it), if the 8,000-token day were doubled, or if a mismatched pairing were allowed. The first two are honest disagreements about the median; the third is not allowed.
• Siting. Nothing here would change the conclusion that averages do not answer a local question, because the site does not claim they do.