AI Music vs Touring & Vinyl Footprint Calculator

This page is designed to be readable first and auditable second. It does not argue that AI has no footprint. It compares our unusually AI-heavy music practice, likely top 1% among AI artists, with a touring-and-vinyl practice under adjustable assumptions. The AI side uses three cases because model size, datacenter location, retries, cooling, and hardware use are rarely disclosed.

1. Suno scale and spend. Suno usage and spend are cited as reported pitch-deck figures, not audited company disclosures. Techmeme's archive of Kristin Robinson's Billboard report states Suno had spent $32M on compute and $2,000 on data since January 2024. Other reporting and commentary around the same deck cite roughly 7M tracks/day. Techmeme archive of Billboard report.
2. Hosted generation energy. Suno has not published a per-track electricity figure. This calculator uses 12 Wh/track as the default base assumption, then lets the three cases move around it: low-impact 0.5x, default 1x, and stress-test 4x. For comparison, Google reported a median Gemini text prompt at 0.24 Wh and 0.26 mL water, while Sam Altman gave 0.34 Wh and 0.000085 gal for an average ChatGPT query. Music generation is not a text prompt, so this page uses those figures only as scale references. Google Cloud, 2025; Sam Altman, 2025.
3. Model training energy. GPT-3 training has been estimated at 1,287 MWh, but a Suno-class music model should not be treated as GPT-3 unless we know architecture, run count, and training setup. The calculator uses a 200 MWh training-run placeholder and amortises it across 2.5B generated tracks/year. This is a sensitivity assumption, not a disclosed Suno number. Patterson et al., 2021.
4. Model training water. Li, Yang, Islam, and Ren estimate that GPT-3 training in Microsoft US data centers directly evaporated about 700,000 L of freshwater. This calculator uses 7,000 L as a much smaller music-model training placeholder, then amortises it. Training water barely moves the per-practitioner result because the denominator is huge. Li et al., 2025 revision.
5. The three cases. All three cases start from the same visible practice inputs, which default to our own heavy AI use rather than an average artist's use. Low-impact assumes efficient hardware, low-carbon grid power, and few retries. The default estimate uses the visible slider settings plus normal datacenter overhead and a small hardware manufacturing estimate. Stress-test assumes more retries, less efficient inference, dirtier electricity, extra cooling, and larger hardware overhead.
6. Electricity-to-water conversion. For compute water, the default estimate uses 2 L/kWh as a rough placeholder for direct cooling plus electricity-related water. Low-impact excludes most local compute water. Stress-test uses 5 L/kWh for cooling plus upstream electricity water. These are intentionally rough: data-center water depends heavily on location, cooling design, grid mix, and whether hydro reservoir evaporation is attributed to electricity consumers.
7. Vinyl carbon. The VRMA / Vinyl Alliance 2024 first report put a standard 140g black LP around 1.15 kg CO2e cradle-to-factory-gate, with air freight adding 1.36 kg CO2e per record for Europe-US and 3.46 kg for Europe-Australia. Vinyl Alliance later added a 2025 note saying the first report is a starting point and that some methods need strengthening, especially product-level accounting and bio-attributed PVC claims. This calculator therefore uses the plain 1.1 kg baseline and names the caveat. Vinyl Alliance / VRMA report page.
8. Vinyl water. Public vinyl water data is thinner than carbon data. The calculator keeps 5 L/record as a visible process-water placeholder for PVC production and pressing-plant cooling. Treat this as an order-of-magnitude knob, not a settled LCA result.
9. Flight emissions. Flight factors are based on the UK Government 2025 GHG conversion factors for business travel by air, using kg CO2e/passenger-km with radiative forcing and well-to-tank fuel included. Defaults use approximate economy factors of 0.148 short-haul, 0.126 medium/international, and 0.142 long-haul kg CO2e/passenger-km. UK DESNZ / DEFRA conversion factors 2025.
10. Flight water. Touring water now includes a central lifecycle jet-fuel water estimate of 5 L water / L fuel. MIT CS3 reports conventional middle-distillate transportation fuels, including jet fuel, in a lifecycle blue-water range of about 4.1-7.4 L water / L fuel; the calculator uses 5 as a legible middle value. Fuel burn is approximated at 40 / 130 / 320 L per passenger each way for short / medium / long haul, then scaled by the cabin-class multiplier. MIT CS3 water-footprint study.
11. Gaming PC comparator. Nvidia's RTX 4080 family specifications list a 320 W total graphics power figure for RTX 4080-class cards; adding CPU, monitor, and peripherals brings a high-end gaming system to roughly 0.45 kW under sustained load. Nvidia GeForce RTX 4080 family specifications. The 1.2x overhead multiplier reflects PSU conversion loss plus idle/standby draw, anchored to the public 80 PLUS PSU efficiency certification tiers. 80 PLUS certification program. Hours/yr (2,000) assumes 250 working days at 8 hours, a stated behavioral assumption rather than a disclosed usage statistic. Grid intensity uses the visible grid slider; water uses the same default 2 L/kWh as the AI-side calculation.
12. Deliberately excluded. On the AI side: GPU manufacturing beyond the small hardware estimate, chip fabs, data-center construction, local water stress, and aggregate grid strain. On the touring side: audience travel, hotels, ground transport, venue energy, crew travel, instruments, merch beyond vinyl, and freight beyond the optional vinyl air-freight toggle. Flight fuel water is counted; the water behind venue electricity, catering, hotels, and audience travel is not. Those exclusions are named because they matter.