Downloading vs. Streaming Music

These figures represent the total power draw of the listening system, covering device plus infrastructure where applicable.

Infrastructure overhead (streaming only): When you stream, your device is not the only thing consuming energy. Data centres and network infrastructure also draw power on your behalf. For audio streaming, this component is estimated at approximately 0.002–0.004 kWh per hour (equivalent to 2–4 watts), uncertain but bounded by physical constraints. The figure is derived by applying the Sustainable Web Design operational intensity model (Malmodin et al. 2023, via sustainablewebdesign.org) to audio bitrates, then discounted for the time-based fixed-cost nature of network connections documented by IEA (Kamiya, 2020). The range reflects this methodological uncertainty. For local playback, this component is zero.

Note on device power: The lowest device figure in this table, a modern smartphone streaming audio with the screen off, is anchored by a directly measured test (Steve Paine, UMPC Portal, February 2024): a Google Pixel 6 in battery saver mode, screen off, WiFi on, streaming audio drew 440 mW. This is used as the lower bound for that scenario. Other figures are derived from IEEE device power literature (IEEE Circuits and Systems Magazine, Q1 2023) and standard published ranges for each device class.

What this table shows: For hi-fi listeners, the amplifier consumes so much power that streaming versus local is essentially irrelevant to total energy use. For smartphone listeners, the proportional difference is large, roughly 7–10 times more power for streaming than for local playback when the screen is off, though the absolute wattages remain small.

Using the midpoints of the ranges above (streaming: phone screen off 4W; phone screen on 6.5W; laptop 25W; hi-fi 84W; local: phone screen off 0.5W; phone screen on 2.5W; laptop 18W; hi-fi 80.5W).

Smartphone, screen mostly off (typical background listening in pocket)

Laptop (music playing in background while working)

To put 17.89 kWh in household terms: it is roughly the energy used boiling an electric kettle approximately 179 times (based on 0.1 kWh per boil: a standard 2,200–3,000W UK kettle running around 3 minutes on a reasonably full fill; measured typical use is 0.073 kWh per actual boil), or running a 40W laptop continuously for approximately 447 hours (around 56 eight-hour working days). Source for kettle figure: ecocostsavings.com, direct energy monitor measurement (2024). Source for laptop wattage: Zembl (2026), consistent with Energuide.be and multiple manufacturer specifications.

Stereo amplifier, 80W receiver (dedicated listening sessions)

The most counterintuitive result in this table: The annual energy saving from switching to local playback is almost identical for a smartphone listener (screen off) and a hi-fi listener, approximately 1.3 kWh/year at one hour per day, 8.9 kWh at seven hours per day, because the streaming infrastructure overhead is consistent across both scenarios, while the amplifier simply dominates total consumption. Switching from streaming to local saves the same absolute kWh whether you are using a phone or a high-end stereo system.

To put 8.96 kWh in household terms: it is roughly the energy used boiling an electric kettle approximately 90 times (based on 0.1 kWh per boil), or running a 40W laptop for approximately 224 hours. See the laptop table note above for sources.

These are annual averages; actual intensity varies by hour of day and season, and within large countries by region.

To put these numbers in context: 1 kg of CO₂ is approximately the equivalent of driving an average petrol car 6 km in the UK, 4 km in the US, or 8 km in China. A phone listener doing 7 hours/day of streaming in South Africa saves more CO₂ by switching to local playback (9 kg/year) than the same listener in Sweden achieves from a full year of local playback altogether. The most powerful variable in this table is not the choice between streaming and downloading: it is where you live.

The same streaming-versus-local absolute saving holds across device types. A hi-fi listener doing 7 hours/day saves roughly 1.1 kg CO₂/year on the UK grid, nearly identical to a smartphone listener doing the same. But the hi-fi user's total annual music electricity cost on the US grid (82 kg CO₂/year streaming) dwarfs a smartphone listener's total on the same grid (3.9 kg CO₂/year). The amplifier is the story, not the platform.

To put 9.17 kg CO₂ in context: it is approximately the CO₂ emitted by driving a typical petrol car around 56 km (35 miles) in the UK (165 gCO₂e/km; Statista/GOV.UK 2024), 37 km in the US (249 gCO₂/km; derived from US EPA figure of 400g/mile), or 71 km in China (129 gCO₂/km; ICCT 2021 certified fleet average). Sources: Statista, citing GOV.UK data (2024). URL: https://www.statista.com/statistics/1233337/carbon-footprint-of-travel-per-kilometer-by-mode-of-transport-uk/; US EPA, Greenhouse Gas Emissions from a Typical Passenger Vehicle (2023), URL: https://www.epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle; ICCT, Trends of new passenger cars in China 2012--2021, URL: https://theicct.org/publication/pv-china-trends-report-jan23/

If a listener has solar and/or wind generation with battery storage, and plays music from locally stored files, their ongoing energy cost for playback is effectively zero from a grid perspective. The device draws on stored renewable energy; no server is contacted; no network infrastructure is activated. The only energy event is the initial download, which is itself a one-time cost.

This scenario is only achievable with local files. Streaming requires an active connection to data centre and network infrastructure on every single play, regardless of what is powering the playback device. A listener streaming on a solar-charged phone is still drawing on grid-powered servers and routers for every stream. A listener playing a locally stored file on a solar-charged phone is not.

Data centres also consume water, primarily through evaporative cooling. This is a distinct environmental cost that does not appear in carbon or energy figures.

The industry metric is WUE (Water Usage Effectiveness): litres of water used for cooling per kWh of IT energy consumed. Industry average for all data centres: approximately 1.8--1.9 L/kWh. Leading hyperscalers report substantially lower direct figures: AWS 0.19 L/kWh; Microsoft 0.30 L/kWh (FY24).

These figures cover on-site direct water only. They exclude the water used to generate the electricity powering the data centre -- the indirect footprint. The Lawrence Berkeley National Laboratory (2024) estimated that US data centres consumed approximately 64 billion litres directly and approximately 800 billion litres indirectly through electricity generation in 2023 -- the indirect figure is roughly 12 times the direct figure at the US average grid mix.

Spotify runs on Google Cloud Platform (GCP). Spotify does not publish a WUE figure; GCP does not disclose per-customer water data. Spotify's water footprint cannot be calculated from public information.

Per-stream water estimate, derived from WUE and the infrastructure energy model in Summary Part 1:

These are estimates applying published WUE figures to the infrastructure energy model in Summary Part 1. They are not direct measurements.

A permanently downloaded file, played locally, incurs a one-time water cost at download (estimated at approximately 120 mL for a 300MB album on GCP-level infrastructure) and zero thereafter. Streaming incurs a continuous water draw on every play.

The local water stress dimension is not captured by WUE. Water is not globally fungible in the way carbon broadly is: a withdrawal from an Arizona aquifer is not offset by a wetland project in Virginia. Two-thirds of US data centres built since 2022 are located in water-stressed areas (WRI, 2026). Documented cases include Google's The Dalles, Oregon data centre consuming 29% of the city's total water in 2021, and Google's Santiago data centre delayed by court order over aquifer concerns. All four major hyperscalers have pledged to be "water positive" by 2030; critics note these pledges cover direct water only, and that replenishment in different watersheds from consumption does not remedy local harm.

The per-stream figures in Summary Parts 1--7 describe individual impact. Multiplied across 670 million monthly active users, the numbers are of a different order.

Energy: Bottom-up modelling from listening hours and Spotify's own reported cloud emissions both converge on approximately 1 TWh per year for Spotify's streaming infrastructure -- roughly the annual residential electricity consumption of Sheffield, or 280,000 average UK homes.

Carbon: Spotify's total reported GHG emissions for 2024 were 195,027 tCO₂e (Spotify Equity & Impact Report 2024) -- a 31% decrease from 2023. This is equivalent to approximately:

• 93,000 UK petrol cars driven for a year (at 2.1 tCO₂e/car)
• 42,000 US cars driven for a year (at 4.6 tCO₂e/car; US EPA)
• 101,000 typical Chinese cars driven for a year (at 1.9 tCO₂e/car; ICCT)

As a fraction of national emissions: approximately 0.057% of UK annual GHGs, 0.004% of US emissions, and 0.0015% of China's.

Water: Approximately 209 million litres direct on-site cooling water per year at the central energy estimate (roughly 84 Olympic swimming pools). Including indirect water from electricity generation: 319 million litres at a clean-grid assumption, up to 3.5 billion litres at US average grid water intensity. Spotify does not publish water data; these are derived estimates.

The individual and the aggregate: Each stream is negligible; 670 million users aggregated are not. Spotify's reported total emissions are small as a fraction of global totals, but the infrastructure replacement of physical distribution is not as complete a dematerialisation as it is sometimes presented.

Published estimates for the energy cost of transmitting one gigabyte of data across the internet range from 0.004 kWh/GB to 136 kWh/GB, a factor of more than 20,000. This reflects radically different system boundary choices. The Sustainable Web Design model (sustainablewebdesign.org), which uses Malmodin et al. 2023 data, produces operational intensity figures totalling approximately 0.194 kWh/GB. These figures are used throughout this document with explicit caveats. No single figure should be treated as authoritative.

Source: Aslan, J. et al. (2018). Electricity Intensity of Internet Data Transmission: Untangling the Estimates. Journal of Industrial Ecology. Wholegrain Digital (2025). URL: https://www.wholegraindigital.com/blog/website-energy-consumption/

Source: Sustainable Web Design model methodology (2024). URL: https://sustainablewebdesign.org/estimating-digital-emissions/

A large proportion of the academic and journalistic literature on "streaming" conflates audio and video, or applies video-derived figures to music. A 3-minute song streamed at Spotify's standard 128 kbps quality requires approximately 2.9 MB of data. A comparable 3-minute HD video clip requires 100–200 MB or more. The network and data centre energy costs of audio streaming are therefore approximately one to two orders of magnitude lower per listen than for video streaming. Any figure from a study that does not explicitly control for this distinction cannot be applied to music without large downward adjustment.

The Green Web Foundation, the primary independent verifier of green hosting claims, announced in January 2026 that it will no longer accept carbon offsets as evidence of green status. Going forward, verification requires actual Renewable Energy Certificates (RECs) or Energy Attribution Certificates (EACs) matched to the same time period and geographic region. Many providers currently listed as "green" will be affected.

Source: Green Web Foundation (2026). Request for comment: Updates to our verification criteria. URL: https://www.thegreenwebfoundation.org/news/request-for-comment-updates-to-our-verification-criteria-for-data-centers-and-hosting-providers/

Neither Bandcamp, Epic Games (its 2022 acquirer), nor Songtradr (its 2023 acquirer) has published a sustainability report or infrastructure energy data specific to Bandcamp's operations. The analysis of Bandcamp in this document is therefore structural and inferential.

Of the major platforms used by independent musicians (Squarespace, Wix, WordPress.com, Shopify, Big Cartel, Cargo Collective), only Shopify publishes a dedicated, auditable climate report.

The most substantive peer-reviewed research specifically addressing the environmental cost of recorded music remains the collaboration between Dr. Kyle Devine (University of Oslo) and Dr. Matt Brennan (University of Glasgow), published in 2019 under the title The Cost of Music.

Devine examined the environmental cost of recorded music in the United States across four historical benchmark years, converting both physical plastic use and digital electricity consumption into greenhouse gas equivalent (GHGe) figures:

Source: Devine, K. and Brennan, M. (2019). The Cost of Music. Universities of Glasgow and Oslo. URL: https://www.gla.ac.uk/news/headline_643297_en.html

The crucial insight is the structural inversion: as physical plastic use collapsed by roughly 87% between 2000 and 2016, GHGe emissions rose, possibly by more than 100%, relative to the LP era, because the electricity required to store, transmit, and process digital music more than offset the environmental savings from eliminating physical production.

The central empirical claim of the post-CD era: dematerialisation is not decarbonisation.

METHODOLOGICAL CAVEAT: The Devine-Brennan study is US-focused and uses 2016 as its most recent data point. Its GHGe figures for the digital era are presented as a range (200–350 million kg) rather than a point estimate, reflecting genuine uncertainty. The study does not disaggregate between the energy cost of streaming and the energy cost of downloading; both are included in the 2016 digital figure.

The same collaboration produced parallel data on the economic cost of music that provides essential context. In 1977, a US consumer would spend approximately 4.83% of their average weekly salary on a vinyl album. By the era of subscription streaming, the marginal cost per listen approaches zero.

This is not incidental to the environmental question. The dramatic fall in perceived cost correlates directly with a dramatic rise in consumption volume. An individual in 1977 might listen to one album repeatedly because they had paid substantially for it and owned it physically. A subscriber in 2026 faces no such constraint. The behavioural economics of "unlimited access for $11 per month" produces listening patterns (discovery mode, background play, algorithmic autoplay, half-listened playlists) that would have been impossible and unaffordable in earlier eras. Energy use per unit of music has fallen; total energy use from music has risen sharply, in large part because of volumetric consumption change.

A 2026 analysis published via Hypebot calculated that the 20 most-streamed albums in their respective debut weeks on Spotify collectively generated approximately 10.62 billion streams and used an estimated 29.1 million kWh of electricity, roughly equivalent to the daily electricity use of approximately 1 million US homes. This figure is for a single week of top-tier streaming activity on a single platform.

The full global picture (incorporating streaming volumes in markets with less stringent renewable energy requirements, such as parts of China, India, and sub-Saharan Africa) almost certainly presents a more challenging environmental profile than US- or European-centric analyses suggest. These global figures remain poorly quantified in the published literature.

Source: Hypebot (2026). From Audio to Energy: The Carbon Footprint of Streaming Quantified. URL: https://www.hypebot.com/from-audio-to-energy-the-carbon-footprint-of-streaming-quantified/

Data centres house the servers that store the music catalogue and handle streaming requests. Their energy consumption is measured using Power Usage Effectiveness (PUE). The industry average PUE has improved substantially:

• 2009: industry average PUE approximately 2.5 (Uptime Institute)
• 2011: approximately 1.8 (Uptime Institute)
• 2014: hyperscale cloud operators (Google, AWS) reporting PUEs under 1.2
• 2023–2025: leading hyperscale operators reporting PUEs of 1.1–1.2

Source: Uptime Institute Annual Reports (2009, 2011, 2023). Google data centre efficiency disclosures. URL: https://datacenters.google/efficiency/

Importantly, data centre efficiency improvements do not automatically translate into reduced total energy consumption if usage volumes grow faster than efficiency gains. The Lawrence Berkeley National Laboratory's 2016 US Data Center Energy Usage Report (Shehabi et al.) found that despite massive efficiency improvements between 2010 and 2020, US data centre energy use only grew modestly because demand was rising in parallel. As of 2023, global data centre energy use is estimated at 300–380 TWh annually (Kamiya and Coroamă, 2025, EDNA Platform). The rise of AI workloads since 2022 is creating significant new demand pressure.

CDNs are distributed caching systems that sit between origin data centres and end users, duplicating frequently-accessed content closer to consumer populations. Spotify uses Akamai and similar CDN providers in addition to GCP infrastructure. CDN caching means that a very popular song may be served from a local node rather than a distant primary data centre, reducing transmission distance, but it does not eliminate the per-stream server energy cost, because each stream request activates a server process regardless of where the content is cached.

This is the layer where some of the most egregious errors in popular accounts have occurred. The International Energy Agency analyst George Kamiya published a detailed fact-check in 2020 that demonstrated the French think tank The Shift Project had overstated network energy intensity by approximately 50-fold, using outdated figures (0.9 kWh/GB for mobile, versus more recent peer-reviewed estimates of 0.1–0.2 kWh/GB for 4G mobile as of 2019). The Shift Project's methodology applied data-volume-based intensity figures (kWh per GB) to high-bitrate applications, which experts now recognise as inappropriate for time-based services like streaming, where the connection is open regardless of data volume.

The IEA notes that because streaming services are "always-on" data flows, time-based energy intensity (kWh per hour of listening) is more appropriate than volume-based intensity (kWh per GB) for estimating network energy in streaming contexts. This means that switching from high-quality to low-quality streaming saves energy at the data centre, but may not proportionally reduce network energy consumption, because the connection itself is maintained regardless.

Source: Kamiya, G. (2020). The carbon footprint of streaming video: fact-checking the headlines. IEA Analysis. URL: https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines

Counterintuitively, the end-user device is typically the largest single energy consumer in a streaming session, accounting for approximately 72% of total energy use for video according to IEA analysis. For audio, the device's relative share is likely even higher in percentage terms, because the data centre and network components are proportionally smaller. See Part 1 of the Summary for the full device comparison table with measured figures.

The device type variable is a critically important one that is often averaged out in aggregate studies but which has enormous practical significance. A person listening on a hi-fi system uses 20–100 times more electricity per session than a person listening on a smartphone, regardless of whether they are streaming or playing locally.

Lifecycle analysis methodology requires accounting not only for operational energy but also for the energy embodied in manufacturing the hardware required to deliver a service: servers, network equipment, smartphones, and speakers. The IEA has noted that manufacturing accounts for approximately 80% of the lifecycle carbon emissions of a mobile device.

This is seldom incorporated into streaming/downloading comparisons, for good reason: it is essentially impossible to attribute the manufacturing emissions of a smartphone to any particular use case, because the device would have been purchased regardless. It becomes relevant only when asking whether streaming causes consumers to replace devices more rapidly through increased battery drain, a question addressed in Section 5.

Most large data centres rely on evaporative cooling: water is circulated through cooling towers where it absorbs heat from server rooms and is then allowed to evaporate, dissipating the heat to the atmosphere. Approximately 80% of the water drawn for this process is evaporated and lost; the remaining 20% is discharged as warmer wastewater. In cooler climates, free-air cooling -- drawing cold outdoor air across servers without any water -- is possible and water-free. In hot climates, mechanical cooling requiring water is unavoidable for large-scale facilities.

Newer approaches are reducing water dependency: direct-to-chip liquid cooling (circulating liquid coolant directly against processor surfaces), immersion cooling (servers submerged in dielectric fluid), and closed-loop systems that recirculate water without evaporation. In August 2024, Microsoft launched a data centre design using zero water for cooling, relying on chip-level liquid cooling in a closed loop. These technologies remain a small fraction of global installed capacity; evaporative cooling dominates the existing fleet.

Water Usage Effectiveness (WUE), developed by The Green Grid, is the standard industry metric:

WUE = total annual water used for cooling and humidification (litres) / total IT equipment energy (kWh)

Expressed in litres per kilowatt-hour. The Green Grid's industry average is approximately 1.8--1.9 L/kWh, representing all data centre types. Leading hyperscalers report substantially lower figures:

Google reports total water consumption rather than a WUE figure. In 2023, Google consumed 24.2 billion litres across its operations with 95% attributable to data centres. In 2024, 22.7 billion litres globally -- an 8% annual increase attributed primarily to AI workloads. One Google data centre in Council Bluffs, Iowa consumed 3.8 billion litres in 2024 alone.

Note on reporting quality: WUE figures are self-reported. Fewer than one-third of data centre operators globally measured water consumption as of 2021 (Mytton, npj Clean Water, 2021). Reporting has improved since but remains voluntary and inconsistent across operators. Google has in the past treated site-specific water use data as commercially sensitive.

WUE measures only on-site direct water consumption. It excludes the water used to generate the electricity powering the data centre -- the indirect water footprint -- which is frequently larger than the direct figure.

The mechanism: thermoelectric power plants (coal, gas, nuclear) cool steam condensers using water. A coal steam turbine with wet recirculating cooling consumes approximately 0.53 litres per kWh of electricity generated. A gas combined-cycle plant consumes approximately 0.25--0.40 L/kWh. Solar photovoltaic and wind turbines require essentially no water for generation.

The Lawrence Berkeley National Laboratory's 2024 US Data Center Energy Usage Report estimated that US data centres consumed approximately 64 billion litres directly through cooling in 2023, and approximately 800 billion litres indirectly through electricity generation -- roughly 12 times the direct figure. This ratio reflects the US grid mix in 2023, which was still majority fossil-fuelled. Operators sourcing substantial clean energy have a much lower indirect footprint.

The major hyperscalers' "water positive" pledges (AWS, Google, Microsoft, and Meta have all pledged water positivity by 2030) refer primarily to direct on-site consumption. Indirect water is generally not included in the pledge accounting. This is not hidden -- the scope is stated -- but it means these pledges address a fraction of total water impact.

There is a further parallel with carbon reporting here: purchasing renewable energy certificates (RECs) reduces Scope 2 carbon reporting but does not guarantee that the specific power plant serving the data centre at the specific hour of operation was running on clean energy. The same logic applies to water: at any given hour, the actual electricity mix at the data centre site determines the indirect water draw, regardless of annual certificate matching.

A further methodological challenge worth naming explicitly: calculating the indirect water footprint of streaming but not of comparable electricity uses is inconsistent. A refrigerator, an electric car, or a subway train also draws grid electricity that was generated using water -- but nobody routinely calculates their per-use water footprint on that basis. Damon Becnel, a policy expert and filmmaker who works professionally on Colorado River water issues, made this point directly in a 2026 Outside Online article in which he tracked his own AI water use for 11 weeks across 100 sessions. His lifecycle total -- including direct cooling, electricity generation, and hardware manufacturing -- was approximately 5 gallons (19 litres), compared to roughly 110 gallons for a single 383-mile van drive. His conclusion: "The water is real. The inconsistency in how we talk about it is also real." The direct on-site WUE-based figures in this section are the more defensible comparison precisely because they do not depend on this contested scope 2 attribution choice. The indirect figures are included for completeness but should be read with this caveat in mind.

COI note on the Becnel article: The author has a strong professional and personal stake in the health of the Colorado River, which might be expected to bias him toward overstating data centre impact. That he argues the opposite strengthens rather than undermines the point. The article is an opinion piece in a general-interest publication, not a peer-reviewed study; the underlying sources he cites (LBNL 2024, Epoch AI, Microsoft and Google operator WUE data, Argonne GREET model) are the same primary sources used in this document. Source: Becnel, D. (2026). The Truth About AI's Water Footprint. Outside Online. URL: https://www.outsideonline.com/outdoor-adventure/environment/ai-water-use-colorado-river-footprint/

Spotify operates entirely on Google Cloud Platform. Spotify does not publish a WUE figure, and GCP's WUE is not disclosed at per-customer resolution. Spotify's annual sustainability reporting quantifies GHG emissions in detail but does not include water use.

What can be inferred:

• Spotify is hosted on infrastructure with direct WUE approximately 0.3--0.5 L/kWh (estimated from Google's overall water consumption and electricity consumption data)
• This is substantially better than the industry average but not zero
• Google's total data centre water consumption is increasing at 8% annually, driven by AI; Spotify's 220+ ML projects and LLM-based features (AI DJ, AI Playlist) contribute to this AI-driven load
• The indirect water footprint via electricity generation is not captured in any available Spotify or Google disclosure at a per-service level

Spotify's water footprint cannot be calculated from public data. This is a genuine and unresolved gap.

Applying the infrastructure energy figures from Section 1 (approximately 0.002--0.004 kWh per hour of audio streaming) to available WUE data:

For a listener using Spotify on a clean-energy-heavy grid (Sweden, Norway, France): direct water approximately 0.6--2.0 mL per hour, total including indirect probably 1--4 mL per hour.

For a listener on a fossil-heavy grid (India, Indonesia, South Africa): total including indirect could be 13--26 mL per hour.

These are estimates derived by applying published WUE figures to the infrastructure energy model in Section 1. They are not direct measurements.

One figure that requires scrutiny: A widely circulated claim from the China Water Risk (CWR) analysis (2022) calculates 1,378 litres of water for two hours per day of music streaming over one month -- implying approximately 23 litres per hour. This cannot be reconciled with any published WUE data and is approximately 1,000 times higher than the bottom-up calculation above at hyperscaler WUE, and still roughly 3,000 times higher at industry average WUE. The CWR figure appears to use a water-per-gigabyte-of-data-transferred methodology using a factor that has not been independently published or verified against WUE data. This figure should not be cited without examination of the underlying calculation.

Water is not a globally fungible resource in the way that carbon broadly is. A tonne of CO2 emitted anywhere raises global atmospheric concentration. A million litres withdrawn from an Arizona aquifer cannot be offset by a wetland project in Virginia. This is the structural problem with water positivity pledges that involve replenishment in different locations from consumption.

Documented cases of data centre water impact on specific communities:

• Google's data centre in The Dalles, Oregon consumed 29% of the city's total water in 2021
• Google was forced to delay a data centre in Santiago, Chile after an environmental court ordered the application revised, citing concerns over the capital's aquifer
• The city of Joliet, Illinois projects aquifer depletion by 2030, partly attributed to data centre development in the Chicago area
• In South Carolina, conservation groups challenged Google's access to 1.9 million litres per day from a local aquifer, while the local utility was simultaneously required to reduce its own withdrawals by 57%
• Two-thirds of all US data centres built or in development since 2022 are located in water-stressed areas (WRI, 2026)
• One former AWS water sustainability manager publicly stated he "disagreed quite a lot" with the offsetting approach on ethical grounds, arguing it did not address the harm at the point of extraction

It is also worth situating these local impacts within the broader picture of water use. All US data centres combined -- not just AI-related facilities -- account for approximately 0.3% of total national water withdrawals (LBNL 2024, consistent with USGS 2021 national water use estimates). Agriculture accounts for approximately 80% of Colorado River water consumption; alfalfa irrigation in California's Imperial Valley alone consumes over 800 billion gallons a year. As Becnel (2026) argues, the local data centre impact and the aggregate Western water crisis are different claims that should not be conflated: a single facility can materially stress a small community's supply while data centres remain a minor factor in national totals. This document makes claims about the former, not the latter.

The relevance to streaming: Spotify's infrastructure runs on GCP data centres distributed globally, including locations with varying water stress. A listener in Stockholm cannot know whether the GCP data centre serving their stream is competing for water with a drought-affected community.

A permanently downloaded file, played locally, severs the ongoing data centre water draw from every subsequent play. The one-time download event incurs a proportional share of data centre water for the transfer; thereafter the file plays locally with zero further data centre water consumption.

Estimated water for a one-time album download (300MB, at GCP-level infrastructure energy of approximately 0.3 kWh total for servers and network, and WUE of approximately 0.4 L/kWh): approximately 120 mL of direct on-site water, once, for the lifetime of the download.

For a streaming listener, the water draw is continuous. For an offline Spotify listener, DRM re-authorisation every 30 days and background platform activity maintain a lower but continuous draw.

• Spotify's total water footprint is not calculable from public data
• Indirect water from electricity generation is not consistently included in hyperscaler disclosures
• The relationship between specific listening events and specific data centre locations is not disclosed
• Whether water-positive pledges are being met at the level of individual sites is not independently verified

What can be established: water is a real and growing environmental cost of streaming infrastructure, the local impacts at specific sites are documented and material, and the existing carbon-focused comparisons between streaming and downloading do not capture this dimension.

Sources for this section:

Mytton, D. (2021). Data centre water consumption. npj Clean Water. URL: https://www.nature.com/articles/s41545-021-00101-w

Microsoft (2024). Sustainable by design: Next-generation datacenters consume zero water for cooling. URL: https://www.microsoft.com/en-us/microsoft-cloud/blog/2024/12/09/sustainable-by-design-next-generation-datacenters-consume-zero-water-for-cooling/

Lawrence Berkeley National Laboratory (2024). 2024 United States Data Center Energy Usage Report. URL: https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf

Li, P. et al. (2023). Making AI Less Thirsty. arXiv:2304.03271. URL: https://arxiv.org/pdf/2304.03271

EESI. Data Centers and Water Consumption. URL: https://www.eesi.org/articles/view/data-centers-and-water-consumption

WRI (2026). From Energy Use to Air Quality, the Many Ways Data Centers Affect US Communities. URL: https://www.wri.org/insights/us-data-center-growth-impacts

Data Center Dynamics (2025). Google emissions jump 48% in five years due to AI data center boom. URL: https://www.datacenterdynamics.com/en/news/google-emissions-jump-48-in-five-years-due-to-ai-data-center-boom/

Interface Media (2024). AWS data centres "water positive" pledge is not greenwashing but it is misleading. URL: https://interface.media/blog/2024/09/02/aws-data-centres-water-positive-pledge-isnt-greenwashing-but-it-is-misleading/

GIJN (2025). Tips for Researching Massive Water Consumption by Data Centers. URL: https://gijn.org/stories/researching-water-consumption-data-centers/

Both streaming and downloading a song require a similar amount of data to travel from a server to the user's device for the first listen. The critical difference lies in what happens on the second, fifth, twentieth, and hundredth listen:

• When a song is streamed: every single playback event triggers a new data request from the server, requiring server activation, CDN response, network transmission, and device processing of incoming data.
• When a song is downloaded and played locally: subsequent playbacks require none of these. The file exists on the device. The server is not contacted. The CDN is not involved. Network transmission is zero. The only energy cost is the device running its audio playback function.

This structural asymmetry has a clear implication: for any song listened to more than once, downloading is more energy-efficient. The margin of advantage grows with every additional play. For a song listened to only once, there is no energy benefit to downloading.

A key practical question is: how many times does a song need to be listened to before the one-time download energy cost is recouped relative to repeated streaming? The answer is: the breakeven point is the first listen itself. At the first listen, streaming and downloading are energetically comparable (both require the same data transmission). From the second listen onwards, every additional play of a downloaded file is more energy-efficient than every additional stream. There is no threshold to cross, no number of listens to accumulate.

The often-cited claim that "if you stream an album more than 27 times, you should buy a physical copy" conflates the streaming-vs-physical format question with the streaming-vs-download question. The 27-play figure is derived from comparisons with physical disc manufacturing emissions, not with digital download energy. Do not apply it to the streaming-vs-download comparison.

The figure most commonly cited in this area is that downloading music instead of streaming it would result in an 80% reduction in CO₂ emissions after the first listen. This figure requires careful scrutiny of its provenance.

The primary source is an op-ed published in Rolling Stone on 22 April 2022 by Adam Met, the bassist of AJR, citing Spotify's 2020 Sustainability Report as its underlying data source.

The mechanism: according to Spotify's 2020 Sustainability Report, servers produced in excess of 70,000 tonnes of CO₂e per year, and listener-phase emissions constituted approximately 42% of total emissions. The 2020 report attributed the majority of listener-phase emissions to streaming activity. If all content were downloaded rather than streamed, the repeated-playback activation of servers and CDN infrastructure would be eliminated, yielding the 80% post-first-listen reduction.

CONFLICT OF INTEREST ASSESSMENT: The 80% figure derives from Spotify's own sustainability reporting. Spotify is the company whose business model depends on streaming, not downloading. While it is not in Spotify's interest to understate its emissions (doing so would expose it to regulatory and reputational risk), the claim is nonetheless self-reported data from an interested party. No independent third-party verification of the specific 80% calculation appears in the available literature. The directional logic is airtight; the specific percentage requires acceptance of Spotify's own energy accounting.

Source: Met, A. (2022). Protect the Planet: Stop Streaming Songs. Rolling Stone, 22 April 2022. URL: https://www.rollingstone.com/music/music-features/earth-day-climate-change-streaming-downloading-ajr-1339228/

A separate and more directly testable claim is that streaming uses approximately double the device battery of local playback. This is mechanically sound: streaming requires the device to maintain a data connection, run network radio components, decode incoming data packets in real time, and manage buffer states, all of which consume power, in addition to audio playback. Local playback requires only the latter.

This device-level evidence does not depend on accepting any corporate self-reported data. It is a measurable, reproducible physical phenomenon. The measured figure for a Google Pixel 6 streaming audio with screen off (440 mW, UMPC Portal, February 2024) versus the local playback range (0.4–0.6 W) is consistent with the doubling claim. More battery drain means more charging cycles, which means more electricity drawn from the grid, and accelerated battery degradation over time.

Deriving precise per-listen energy figures for audio streaming is difficult because most academic studies address video streaming. Working from the IEA's 2020 updated analysis of video streaming (approximately 0.077 kWh per hour) and adjusting for the data volume ratio between audio and video (approximately 1:25 to 1:50):

• Standard-quality audio (128 kbps) generates approximately 0.058 GB/hour of data.
• Standard HD video generates approximately 1.5–3 GB/hour.
• Adjusted for audio: approximately 0.002–0.003 kWh per hour of infrastructure energy, excluding the end-user device.
• For a 3-minute song (0.05 hours): roughly 0.0001–0.00015 kWh per stream.

These are very small numbers for a single listen, but they compound across the approximately 620 billion streams Spotify reported in 2022 and billions more across Apple Music, Amazon Music, YouTube Music, and other platforms.

PRECISION CAVEAT: These per-stream energy figures are estimates derived from applying video-streaming research with audio adjustments. No major peer-reviewed study has directly measured per-song audio stream energy consumption with the methodological rigour of the IEA's video analysis. The directional order of magnitude is reliable; the specific decimal places are not.

The streaming/downloading comparison is entirely determined by listening frequency at the margin. The environmental case for downloading is strongest for:

• Favourite albums listened to repeatedly over months or years
• Workout playlists used daily
• Children's music libraries played on high rotation
• Any music listened to in contexts of limited or expensive data connectivity

The environmental case for streaming (or indifference) is strongest for:

• Discovery listening: sampling music once before deciding to engage further
• One-time event playlists
• Music listened to on a low-power device over WiFi from a carbon-efficient grid

Higher audio quality requires more data. The relationship between streaming quality and energy consumption is roughly linear for the network and data centre components.

The rise of hi-res streaming tiers (Apple Music Lossless, Tidal MQA, Amazon Music HD) has received almost no attention in environmental analyses of music streaming, but represents meaningful upward pressure on per-stream energy. A user streaming lossless audio repeatedly is generating substantially more network and data centre energy than a user who has downloaded the lossless file once and plays it locally.

Source: Bitrate and data size estimates based on standard audio encoding specifications.

Streaming over mobile cellular networks (4G/5G) is generally more energy-intensive per unit of data than streaming over fixed broadband via WiFi, though the gap depends on network design, signal strength, and load conditions. For downloaded content played locally, this variable is entirely eliminated from subsequent plays. Downloaded content on a device using only local storage is network-agnostic after the initial download.

As established in Part 3 of the Summary, the geographical carbon intensity of the grid in use has a larger effect on the absolute CO₂ footprint of any digital activity than the choice between streaming and downloading. A listener in Sweden streaming on WiFi may have a lower absolute carbon footprint than a listener in Germany playing a downloaded file on an energy-intensive hi-fi system connected to a coal-heavy grid. This does not invalidate the download advantage: it means the advantage is a relative one that can be amplified or attenuated by other factors.

Some platforms implement adaptive caching that pre-loads content predicted to be listened to next: a form of anticipatory downloading that blurs the practical distinction between streaming and downloading at the infrastructure level. The user experience of "streaming" may in some cases involve data that was pre-cached on the device during idle periods, functionally resembling downloading without the user's explicit instruction. This does not eliminate the ongoing infrastructure relationship discussed in Section 8.

From a pure energy-use perspective, "downloading" within a streaming subscription and purchasing a permanent digital download are energetically equivalent in their subsequent playback. In both cases:

• Data transmission occurs once (during the download)
• Subsequent plays require only local device audio playback
• No server is contacted; no CDN is activated; no network transmission occurs

The 80% post-first-listen energy saving described in Section 5.3 applies equally to offline mode within Spotify as it does to a purchased MP3 file. The energy physics do not distinguish between a DRM-locked downloaded stream and a purchased permanent download.

This creates a practically important implication: the environmental advice and the existing commercial arrangement are not in conflict. Spotify Premium subscribers already have access to the tools required to eliminate most of the per-stream energy cost, without any additional purchase.

The distinction that matters for resource use (though not energy per se) is permanence and infrastructure dependency. An offline-mode download from Spotify is:

• DRM-encrypted and cannot be played outside the Spotify application
• Invalidated if the subscription lapses, the platform ceases to operate, or the user is banned
• Dependent on periodic re-authorisation by the platform's servers (typically every 30 days)

A purchased permanent download (MP3, FLAC, ALAC from Bandcamp, Qobuz, or the iTunes Store) requires none of this infrastructure dependency for subsequent playback. It can be stored on a local drive and played indefinitely without any server contact. This represents a structural resource efficiency advantage that does not appear in per-listen energy accounting but is relevant to lifecycle resource analysis.

There is a theoretical rebound risk in encouraging offline downloading: users who know they have downloaded content may also stream more freely when a connection is available (for discovery, social features, etc.), offsetting some of the energy saved through offline playback. This dynamic has not been quantitatively studied in the music streaming context. It is flagged here as a genuine uncertainty, not a refutation of the offline mode recommendation.

Spotify operates as a global streaming platform with approximately 678 million monthly active users as of early 2026. Since 2019, its content infrastructure has been hosted entirely on Google Cloud Platform (GCP). GCP operates at a global average PUE of approximately 1.1. Google matched 100% of its electricity consumption with renewable energy purchases in 2017 and has committed to operating on 24/7 carbon-free energy at all data centres by 2030.

Spotify's reported total GHG emissions for 2024: 195,027 metric tonnes CO₂e (Scope 1 + 2 + 3). Approximately 98% are Scope 3.

Source: DitchCarbon (2025). URL: https://ditchcarbon.com/organizations/spotify

Source: CarbonCredits.com (2025). URL: https://carboncredits.com/spotify-strikes-a-chord-big-q1-gains-and-bigger-climate-goals-net-zero-for-2030/

Of particular concern to anyone attempting longitudinal comparison of Spotify's reported emissions is the change in reporting methodology between years. Spotify's 2021 total emissions of 490 million kg CO₂e included end-use emissions (user device electricity consumption) accounting for approximately 23% of the total. Spotify's 2023 report changed the methodology in a way that removed or restructured this component.

The practical result: year-on-year comparisons of Spotify's reported emissions are not easily made. The apparent reduction from 490 million kg CO₂e (2021) to 280.7 million kg CO₂e (2023) to 195 million kg CO₂e (2024) may reflect genuine efficiency improvements, changes in scope definition, changes in Scope 3 calculation methodology, or some combination. The direction of travel cannot be reliably determined without access to the underlying data and a consistent methodology applied across years.

Source: Greenly (2025). Notes that Spotify's 2023 report no longer accounts for end-use device electricity emissions. URL: https://greenly.earth/en-us/leaf-media/data-stories/the-carbon-cost-of-streaming

For a Spotify user, even one using offline mode, the following infrastructure layer remains live as long as the subscription is active:

• GCP servers hosting the catalogue and all associated metadata
• Spotify's recommendation and personalisation AI, running continuously
• CDN infrastructure maintaining cached copies of content globally
• Authentication and DRM servers performing periodic licence checks (every 30 days for offline content)
• Rights management and royalty reporting infrastructure logging all playback events
• Spotify's analytics, social, and business infrastructure sustained by the active subscriber base

For a Spotify user who streams rather than downloads within the platform, every playback event additionally activates a new server request, CDN response, and network data transmission.

Even in the best possible scenario (a Spotify Premium subscriber using offline mode for all playback), the following remain true:

• The downloaded files are DRM-encrypted and cannot be played outside the Spotify application
• The files are invalidated if the subscription lapses, the platform shuts down, or the user is banned
• An internet connection is required every 30 days for re-authorisation
• All Spotify's background infrastructure continues to be sustained by the subscription

This is structurally different from a purchased permanent download. There is no exit from that infrastructure while the subscription continues.

The infrastructure summary in Section 8.3 lists "Spotify's recommendation and personalisation AI, running continuously" as a single bullet point. That description significantly understates what is actually happening.

Spotify's ML and AI stack includes at least the following distinct components:

BaRT (Bandits for Recommendations as Treatments): The algorithmic engine behind Spotify's home screen. Uses a Bayesian bandit approach, evaluating each content "shelf" shown to a user continuously. Runs at every app open for every user.

Collaborative filtering models: Identify patterns in aggregate user behaviour to find listeners with similar tastes. Trained on data from hundreds of millions of users and billions of listening events.

Convolutional Neural Networks (CNNs) for audio analysis: Analyse the raw audio of tracks for tempo, key, loudness, and mood. Run continuously as new content is added to the catalogue.

Near-Real-Time (NRT) user embedding inference: Per Spotify Research (2025): "NRT inference, triggered by user activity events (e.g., streaming a new artist), refreshes embeddings within minutes." Every stream triggers additional ML inference cycles beyond the stream itself.

Discover Weekly and Daily Mixes: Large-scale batch inference runs generating personalised playlists for hundreds of millions of users. Discover Weekly alone involves generating 600+ million unique playlists weekly.

AI DJ (launched February 2023): Combines recommendation with a Large Language Model tier to produce personalised spoken commentary. LLM inference is roughly two to three orders of magnitude more energy-intensive per call than traditional recommendation ML.

AI Playlist (launched 2024): Generates playlists from natural language prompts. Requires LLM inference on every user request.

Spotify's Event Delivery Infrastructure processes 500 billion to 1.4 trillion events per day. At even 0.0001 Wh per traditional ML event, the daily inference total runs to 50–500 MWh. If AI DJ is used by 10% of daily active users with five LLM calls per session at 0.1 Wh per call (assuming a smaller, more efficient model than frontier LLMs), AI DJ alone generates approximately 12.5 MWh/day.

GCP's data centre electricity consumption grew 17% in 2023 and 27% in 2024, with Google attributing this growth explicitly to AI workloads. Spotify's 220+ ML projects and growing LLM tier are part of this picture.

None of this infrastructure is activated by a local file owner.

TRANSPARENCY NOTE: The energy costs estimated here are order-of-magnitude approximations derived from analogous systems. Spotify has not published per-system inference volumes and energy costs, and is not required to do so.

Source: Spotify Research (2025). Generalized user representations for large-scale recommendations. URL: https://research.atspotify.com/2025/9/generalized-user-representations-for-large-scale-recommendations

Source: Data Centre Dynamics (2024). Google emissions jump 48% in five years. URL: https://www.datacenterdynamics.com/en/news/google-emissions-jump-48-in-five-years-due-to-ai-data-center-boom/

Spotify had 670 million monthly active users at the end of 2024 (CEO letter, 2024 Equity & Impact Report). Average listening time is widely reported at 148 minutes per day for engaged users; however, not all monthly active users listen daily. Using 55% as an estimate of the daily active user proportion, a central estimate of approximately 50 minutes per day averaged across all 670 million MAUs is defensible. This gives:

Applying the infrastructure energy range from Section 3 (0.002--0.004 kWh per hour of audio streaming) with a 35% uplift for podcast streaming, storage overhead, and AI recommendation processing:

Central estimate with overhead: approximately 0.8--1.1 TWh per year.

1 TWh is the annual electricity consumption of approximately 278,000 average UK homes. Spotify's central infrastructure energy estimate is equivalent to the residential electricity consumption of a city the size of Sheffield.

Spotify's 2024 Equity & Impact Report discloses cloud infrastructure as 35.8% of Scope 3 Category 1 (which is itself 80.4% of total Scope 3), giving cloud-attributed emissions of approximately 54,860 tCO₂e.

Back-calculating to energy:

The market-based figure (1.1 TWh) is consistent with the bottom-up estimate (0.8--1.1 TWh). The convergence of two independent methods strengthens confidence in the central estimate of approximately 1 TWh.

Note: Cloud Scope 3 likely includes hardware manufacturing as well as operational energy; the energy-only component may be somewhat lower. This does not materially change the central estimate.

Spotify's total reported GHG emissions for 2024 were 195,027 tCO₂e -- a 31% decrease in absolute terms from 2023, attributed partly to cloud optimisation work decoupling MAU growth from emissions growth (Spotify Equity & Impact Report 2024).

Expressed as equivalent car-years on the road:

Sources: UK Government/DEFRA [40]; US EPA [43]; ICCT [44].

As a share of national totals: 195,027 tCO₂e represents approximately 0.057% of UK annual GHG emissions, 0.004% of US emissions, and 0.0015% of China's.

Scope caveat: Spotify's reported emissions cover its own operations and immediate supply chain. They do not include electricity consumed by listeners' playback devices (downstream Scope 3, excluded from Spotify's disclosure) or the embodied carbon of those devices. The 195,027 tCO₂e figure is the cost of running the infrastructure, not the total carbon cost of all music listening via Spotify.

Applying AWS WUE of 0.19 L/kWh to the central infrastructure energy estimate of approximately 1 TWh:

The central direct figure of approximately 209 million litres is equivalent to roughly 84 Olympic swimming pools. The indirect figure is highly sensitive to assumptions about GCP's actual electricity mix, as discussed in Section 4; the direct figure is the more defensible number.

Spotify does not publish water consumption data. These figures are derived from infrastructure energy estimates and AWS WUE. They are estimates, not measurements.

Any individual user's per-stream carbon and water footprint is small -- a fraction of a gram of CO₂, a millilitre of water. The summary tables in this document accurately represent individual impact.

But 670 million users each making a small individually-negligible contribution aggregate to infrastructure demands comparable to a mid-sized city's electricity consumption. Spotify's reported total emissions -- while 31% lower than 2023 -- are equivalent to keeping 93,000 UK cars, 42,000 US cars, or 101,000 typical Chinese cars on the road for a year.

This is not an argument against streaming. It is an argument for taking the aggregate seriously when evaluating whether dematerialisation is as complete as it is often claimed. The plastic of a vinyl record has been replaced by a server farm. The farm uses electricity and water, 24 hours a day, for as long as the service operates.

Bandcamp was founded in Oakland, California in 2008. In March 2022, it was acquired by Epic Games, which had migrated its entire infrastructure to Amazon Web Services (AWS) in 2018. In October 2023, Epic sold Bandcamp to Songtradr, a music licensing company, amid broader Epic layoffs. Approximately 50% of Bandcamp's staff did not receive employment offers from Songtradr.

CRITICAL NOTE ON CURRENT STATUS: As of April 2026, Bandcamp operates under Songtradr. Whether its infrastructure has migrated, been reduced, or continues on its prior hosting arrangements is not publicly documented. The analysis below treats Bandcamp as likely to have operated on AWS during the Epic era, and on infrastructure of uncertain provenance and likely reduced scale under Songtradr.

Source: Variety (2023). Bandcamp Hit With Layoffs After Songtradr Acquisition. URL: https://variety.com/2023/music/news/bandcamps-layoffs-songtradr-1235758123/

Amazon Web Services matched 100% of electricity consumed globally with renewable energy purchases in 2023 and 2024, and maintained a global PUE of 1.15 in 2024. These are genuine achievements. They are not the complete picture.

Amazon's total GHG emissions rose from 64.38 million MTCO₂e (2023) to 68.25 million MTCO₂e (2024) despite the 100% renewable matching. Renewable energy "matching" through annual certificate purchases does not mean that every kilowatt-hour consumed at every data centre at every hour was actually generated by a clean source. The Green Web Foundation is now tightening its verification criteria specifically to require RECs matched by time and region.

Source: Energy Digital (2025). URL: https://energydigital.com/news/whats-inside-amazons-2024-sustainability-report

A Bandcamp purchase involves: loading the album page, browsing, payment processing, a one-time ZIP file download, and local playback from that point forward.

This is the total energy cost of that album for all future playback combined. Every subsequent play draws only local device audio power.

Once the download is complete, Bandcamp's servers are not contacted again for playback. Files are DRM-free, playable in any software, and permanent. If Bandcamp ceases to exist tomorrow, the files continue to function identically.

The scenario: an independent artist hosts their own website on a Virtual Private Server (VPS) provided by a hosting company that claims carbon-neutral or renewable-energy-powered operations. The website contains no tracking scripts, no third-party analytics, no social media embeds, no advertising. It serves a static page and a downloadable ZIP file. The consumer visits, pays (through Stripe, PayPal, or Gumroad), and downloads.

This is the minimum viable digital music distribution infrastructure.

The carbon-neutral VPS scenario is the best-case configuration: a deliberate, informed choice. In practice, the overwhelming majority of independent artists use one of a small number of commercial platforms:

Squarespace runs its core website hosting on AWS. No standalone sustainability report has been published. AWS environmental credentials and their caveats apply.

Wix operates a multi-cloud infrastructure combining AWS, Google Cloud, Fastly's CDN, and Wix's own proprietary data centres. It has published no sustainability report. The proprietary data centres have entirely unknown environmental credentials.

WordPress.com (Automattic) runs on WP Cloud spanning 28+ data centres. In 2020, Automattic announced it was offsetting data centre power emissions using carbon credits: precisely the kind of claim the Green Web Foundation no longer accepts as evidence of green hosting.

Shopify migrated its data operations to Google Cloud Platform in 2018. This is the most transparent and environmentally credible of the commercial platforms. GCP's credentials apply in full. Shopify also publishes an annual Climate Report and is a founding member of the Frontier carbon removal advance market commitment.

Big Cartel actively moved away from Amazon S3 toward Backblaze B2 cloud storage plus Fastly CDN, in part on values grounds as an artist-founded company.

Cargo Collective has no infrastructure disclosure of any kind from publicly available sources.

TRANSPARENCY NOTE: Of the platforms listed above, only Shopify publishes a dedicated climate report with auditable methodology. Any claim about the specific carbon footprint of downloading from a Squarespace or Wix site should be treated as an estimate at best. The direction of travel (one-time event, no ongoing relationship) is reliable. The precise emissions figure is not calculable from public data.

The page-weight overhead of commercial website builders is substantially higher than a minimal static site. A Squarespace or Wix artist page typically loads 2–5 MB of assets before the user even sees a download link. A minimal static HTML page might load in 100–500 KB. This difference affects the energy cost of the page visit, though not the download itself.

Most importantly, none of this changes the fundamental structural conclusion: any download from any of these platforms is still a one-time energy event. Once the file is on the consumer's device, the platform's infrastructure is no longer involved in playback.

The independent website scenario eliminates all platform overhead: no recommendation engine, no social layer, no analytics, no subscription infrastructure. Once the download is complete, the server is not contacted again. The files are unrestricted. No licence server exists to check.

For streaming on Spotify, every playback is a recurring energy event. For a Bandcamp or artist website download, the energy-significant event occurs once. Thereafter: data centre energy zero, CDN energy zero, network transmission zero, recommendation AI zero, rights reporting zero, platform subscription infrastructure zero.

The break-even point is effectively zero additional plays. From the very first play after the download is complete, the downloaded file is more energy-efficient than streaming the same content.

Under Spotify's model, the royalty per stream is approximately $0.003–$0.005. An artist whose album is streamed 1,000 times earns approximately $3–$5. Under Bandcamp's model, artists receive approximately 82% of the sale price. A $10 album sale returns approximately $8.20. Under the independent website model with Stripe (~3% + 30¢), the artist retains 87–97%.

The environmental recommendation and the economic recommendation for artists point in the same direction. Both flow from the same structural fact: streaming monetises access at minimal cost per event, requiring vast scale and continuous infrastructure, while downloading monetises ownership at a one-time cost requiring only a single transmission event.

Source: Variety (2022). Epic Games Acquires Bandcamp. URL: https://variety.com/2022/digital/news/epic-games-acquires-bandcamp-1235194180/

The scenario in this document assumes a consumer who has already decided what they want to download. It does not account for the energy cost of discovery. If a consumer discovers music through Spotify, spends an hour browsing and streaming before deciding to purchase a download elsewhere, the energy cost of that discovery session belongs in the accounting. This is not an argument against the download model; it is an argument for being precise about what the comparison covers.

The "artist website on carbon-neutral VPS" scenario requires deliberate infrastructure choices by the artist. It is not the default. Many independent artists in 2026 host on Bandcamp, Linktree, Squarespace, or similar platforms without investigating their hosting provider's environmental credentials.

Bandcamp under Songtradr is a smaller, less stable operation than under its pre-2022 independent form. The platform's continued operation is not guaranteed. Files purchased there are DRM-free and permanent, but would need to be backed up independently by the consumer to be safe from platform risk.

The per-download energy estimates in Section 9.3 apply a kWh/GB intensity figure to the download data volume. As noted in Limitation 1, kWh/GB figures represent average allocations across the network, not marginal measurements of what any specific download actually caused. The specific kWh figures are estimates with wide uncertainty bands, not measurements.

A consumer who downloads music accumulates files that must be stored. A 1TB SSD draws approximately 0.5–1W in active use. Across a year, a filled 1TB SSD consumes approximately 4–9 kWh. Storage is not a counter-argument to downloading; it is a component of honest accounting.

As electricity grids in major streaming markets decarbonise, the carbon intensity of all electricity-consuming activities falls. This disproportionately benefits the activity that uses more electricity (streaming), narrowing the GHGe gap between streaming and downloading over time on a per-listen basis. In a hypothetical future of fully decarbonised grids, the GHGe difference between streaming and downloading would approach zero, and the comparison would shift entirely to non-carbon resource considerations: server cooling water, hardware manufacturing, e-waste. As of 2026, the majority of global electricity is still generated from fossil fuels, and the argument for downloading over repeated streaming remains live.

• The shift from physical music formats to digital distribution has dramatically reduced plastic use while substantially increasing electricity-based GHGe from the music sector. Dematerialisation is not decarbonisation.
• Streaming requires data transmission on every playback. Downloading requires data transmission only once. For any song listened to more than once, the energy cost of subsequent plays is lower for downloaded content.
• Streaming uses approximately double the device battery of local playback. This is a direct consequence of network radio and data management operations.
• End-user devices account for the majority of energy in a streaming session. The choice of device type has a larger impact on total session energy than the choice of streaming versus downloading.
• Audio streaming is one to two orders of magnitude less data-intensive per hour than HD video streaming. Studies of video streaming cannot be applied directly to music without large downward adjustment.
• Data centre efficiency has improved dramatically since 2009. Spotify's migration to GCP in 2019 materially improved the efficiency of its server infrastructure.
• DRM-free files purchased and downloaded permanently sever the platform infrastructure relationship. Spotify's offline mode does not.
• The Green Web Foundation's 2026 updated verification criteria tighten the evidentiary standard for "green" hosting claims, rendering many existing claims based on carbon offsets no longer verifiable under its standards.

• The claim that downloading rather than streaming would reduce music-related CO₂ emissions by 80% after the first listen. [Directional logic is sound; specific percentage traces to Spotify self-reporting without independent verification.]
• The claim that an artist website on verified carbon-neutral VPS hosting has a near-zero server-side carbon footprint for the download transaction. [Depends on the quality of the hosting provider's renewable energy verification.]
• That Bandcamp under Songtradr continues to operate on AWS infrastructure. [Inferred from Epic's AWS relationship and Songtradr's lack of public disclosure.]

• The precise per-download energy figure for any specific album at any specific quality setting, given the wide variance in published kWh/GB estimates.
• The comparative carbon footprint of the payment processing component across the three scenarios.
• The total ongoing energy cost attributable to an individual Spotify subscriber's proportional share of Spotify's infrastructure, including AI and rights management overhead.
• Whether the efficiency gains in data centres and networks will outpace consumption growth over the next decade, particularly with AI-driven demand pressure.
• The magnitude of the rebound effect from offline mode recommendations.

The total estimated token count for the conversation is approximately 457,000 tokens (input and output combined). The dominant input cost is web search results -- approximately 65 searches at roughly 3,000 tokens each, plus 15 full-page web fetches, plus the growing document held in context repeatedly. Output tokens are approximately 72,000, of which the two documents account for roughly 32,000.

Significant caveat: The conversation was compacted once during its session. Token counts before the compaction point are reconstructed from document word counts and estimated context loads, not measured directly. The true total could be 30% higher or lower.

Estimated range: 0.37--1.05 kWh, midpoint approximately 0.74 kWh.

The per-token energy estimate is derived from Jegham et al. (arXiv:2505.09598, May 2025), which measured GPT-4o short queries at approximately 0.42 Wh, extrapolated to approximately 0.0014 Wh/token. A range of 0.7x--1.4x is applied to reflect uncertainty about Claude Sonnet 4.6 specifically, since Anthropic does not publish per-token inference energy figures. A further 15% uplift accounts for tool overhead: web search infrastructure, file I/O, and Python code execution during document building.

Long-context inference may cost more per token than the per-query figure implies, because attention over long sequences is computationally more expensive. This effect is unquantified.

Direct (on-site cooling, AWS WUE 0.19 L/kWh): approximately 140 mL at the midpoint energy estimate -- roughly a small glass.

Including the indirect water footprint from electricity generation:

The distinction between these figures reflects the same methodological question the document addresses in Section 4: whether annual renewable energy certificate matching should be credited as actually displacing fossil generation, or whether marginal grid intensity is the honest measure.

Approximately 28g CO2e if AWS's annual REC matching is credited; approximately 184g CO2e on a marginal grid impact basis.

The lower figure is roughly equivalent to driving a typical petrol car approximately 170 metres in the UK, 112 metres in the US, or 217 metres in China. The higher figure is approximately 1.1 km in the UK, 0.74 km in the US, or 1.4 km in China.

For comparison: one vinyl record air-freighted from Europe to Australia has a pre-playback carbon footprint of approximately 4,600g CO2e. On the marginal carbon estimate, producing this report cost less than one quarter of that. On the REC-credited estimate, approximately one-fortieth.

• Energy per token for Claude Sonnet 4.6 is not published by Anthropic. The GPT-4o proxy is the best available figure from peer-reviewed measurement as of April 2026.
• All water and carbon figures are subject to the same caveats applied to per-stream estimates throughout this document: grid intensity varies by hour, REC matching is annual not 24/7 CFE, and indirect water figures depend heavily on methodology.
• The calculation itself consumed energy and water. It is not recursive enough to estimate how much.

Source: Jegham, I. et al. (2025). How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference. arXiv:2505.09598. URL: https://arxiv.org/abs/2505.09598

[1] Devine, K. and Brennan, M. (2019). The Cost of Music. Universities of Glasgow and Oslo. URL: https://www.gla.ac.uk/news/headline_643297_en.html

[2] Devine, K. (2019). Music streaming has a far worse carbon footprint than the heyday of records and CDs. The Conversation. URL: https://theconversation.com/music-streaming-has-a-far-worse-carbon-footprint-than-the-heyday-of-records-and-cds-new-findings-114944

[3] Shehabi, A. et al. (2016). United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory. URL: https://www.osti.gov/servlets/purl/1372902

[4] Kamiya, G. and Coroamă, V. (2025). Global data center energy use estimate: 300–380 TWh in 2023. EDNA Platform. URL: https://www.iea-4e.org/wp-content/uploads/2025/05/Data-Centre-Energy-Use-Critical-Review-of-Models-and-Results.pdf

[5] Malmodin, J. et al. (2023). Operational energy intensities: 0.055 kWh/GB (data centres), 0.059 kWh/GB (networks), 0.080 kWh/GB (devices). Via Sustainable Web Design model. URL: https://sustainablewebdesign.org/estimating-digital-emissions/

[6] Aslan, J. et al. (2018). Electricity Intensity of Internet Data Transmission: Untangling the Estimates. Journal of Industrial Ecology. Discussed in: Wholegrain Digital (2025). URL: https://www.wholegraindigital.com/blog/website-energy-consumption/

[7] Kamiya, G. (IEA) (2020). The carbon footprint of streaming video: fact-checking the headlines. IEA Commentary. URL: https://www.iea.org/commentaries/the-carbon-footprint-of-streaming-video-fact-checking-the-headlines

[8] Ember Global Electricity Review 2025. URL: https://ember-energy.org/latest-insights/global-electricity-review-2025/

[9] Carbon Brief UK grid analysis, 2024 annual average.

[10] Australian Clean Energy Regulator. December Quarter 2024 report. 0.56 tCO₂e/MWh.

[11] SA Dept Forestry, Fisheries & Environment, DGGEF (published Feb 2024, using 2021 data).

[12] IEA Emissions Factors 2025.

[13] Enerdata Yearbook 2024.

[14] Paine, S. (UMPC Portal) (February 2024). Google Pixel 6 audio streaming battery measurement: 440 mW. URL: https://www.umpcportal.com/smartphones/reports/streaming-music-efficiency-36-hours-on-a-pixel-6/

[15] IEEE Circuits and Systems Magazine, Q1 2023. Device power modelling.

[16] DitchCarbon (2025). Spotify Emissions Breakdown. 2024 total: 195,027 MTCO₂e. URL: https://ditchcarbon.com/organizations/spotify

[17] Spotify Technology S.A. (2024). Equity & Impact Report 2024. URL: https://www.lifeatspotify.com/reports/Spotify-Equity-Impact-Report-2024.pdf

[18] Spotify HR Blog (2025). Making an Impact, One Note, One Voice, One Idea at a Time (2024 Equity & Impact overview). URL: https://hrblog.spotify.com/2025/03/12/making-an-impact-one-note-one-voice-one-idea-at-a-time

[19] Amazon Web Services (2025). AWS Cloud sustainability disclosures. PUE 1.15 (2024). URL: https://sustainability.aboutamazon.com/products-services/aws-cloud

[20] Energy Digital (2025). What's Inside Amazon's 2024 Sustainability Report. URL: https://energydigital.com/news/whats-inside-amazons-2024-sustainability-report

[21] Google Data Centres. Efficiency: PUE disclosures. URL: https://datacenters.google/efficiency/

[22] Data Centre Dynamics (2024). Google emissions jump 48% in five years. URL: https://www.datacenterdynamics.com/en/news/google-emissions-jump-48-in-five-years-due-to-ai-data-center-boom/

[23] Met, A. (2022). Protect the Planet: Stop Streaming Songs. Rolling Stone. URL: https://www.rollingstone.com/music/music-features/earth-day-climate-change-streaming-downloading-ajr-1339228/

[24] Hypebot (2026). From Audio to Energy: The Carbon Footprint of Streaming Quantified. URL: https://www.hypebot.com/from-audio-to-energy-the-carbon-footprint-of-streaming-quantified/

[25] Greenly (2025). The Carbon Cost of Streaming. URL: https://greenly.earth/en-us/leaf-media/data-stories/the-carbon-cost-of-streaming

[26] Farag, A. (CBC Music) (2023). The environmental impact of music streaming, explained. URL: https://www.cbc.ca/music/the-environmental-impact-of-music-streaming-explained-1.6843948

[27] Green Web Foundation (2026). Request for comment: Updates to our verification criteria. URL: https://www.thegreenwebfoundation.org/news/request-for-comment-updates-to-our-verification-criteria-for-data-centers-and-hosting-providers/

[28] Green Web Foundation (2026). Summarising Public Feedback on updated verification criteria. URL: https://www.thegreenwebfoundation.org/news/summarising-feedback-from-the-community-on-our-updated-verification-criteria/

[29] Variety (2022). Epic Games Acquires Bandcamp. URL: https://variety.com/2022/digital/news/epic-games-acquires-bandcamp-1235194180/

[30] Variety (2023). Bandcamp Hit With Layoffs After Songtradr Acquisition. URL: https://variety.com/2023/music/news/bandcamps-layoffs-songtradr-1235758123/

[31] Wikipedia (2026). Bandcamp. URL: https://en.wikipedia.org/wiki/Bandcamp

[32] Spotify Research (2025). Generalized user representations for large-scale recommendations. URL: https://research.atspotify.com/2025/9/generalized-user-representations-for-large-scale-recommendations

[33] Spotify Research (2024). Contextualized Recommendations Through Personalized Narratives using LLMs. URL: https://research.atspotify.com/2024/12/contextualized-recommendations-through-personalized-narratives-using-llms

[34] arXiv (2025). Energy Use of AI Inference: Efficiency Pathways and Test-Time Compute. arXiv:2509.20241. URL: https://arxiv.org/abs/2509.20241

[35] Google Cloud Blog (2025). Measuring the environmental impact of AI inference. URL: https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference/

[36] Praella (2024). Shopify Sustainability Practices. URL: https://praella.com/blogs/shopify-insights/shopify-sustainability-practices-building-a-greener-e-commerce-future

[37] Backblaze case study: Big Cartel Designs Multi-Cloud Infrastructure. URL: https://www.backblaze.com/case-studies/big-cartel

[38] Automattic (2020). Toward zero: Reducing and offsetting our data center power emissions. URL: https://wordpress.com/blog/2020/09/21/toward-zero-reducing-and-offsetting-our-data-center-power-emissions/

[39] ecocostsavings.com (2024). Electric Kettle Running Costs. Direct energy monitor measurement: 0.073 kWh per typical use. URL: https://ecocostsavings.com/electric-kettle-running-costs/

[40] Statista, citing GOV.UK (2024). Average petrol car: 165 gCO₂e/km. URL: https://www.statista.com/statistics/1233337/carbon-footprint-of-travel-per-kilometer-by-mode-of-transport-uk/

[41] RAC Foundation, citing UK Dept for Transport WLTP data (2024). New petrol registrations: 143 gCO₂/km. URL: https://www.racfoundation.org/motoring-faqs/environment

[42] Zembl (2026). How much energy does a laptop use? URL: https://www.zembl.com.au/blog/how-much-energy-does-a-laptop-use-save-on-power

[43] US EPA (2023). Greenhouse Gas Emissions from a Typical Passenger Vehicle. URL: https://www.epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle. Key figure: average US passenger vehicle emits 4.6 metric tonnes CO₂ per year (22.2 mpg, 11,500 miles/year).

[44] ICCT (International Council on Clean Transportation) (2023). Trends of New Passenger Cars in China: Air Pollutant and CO₂ Emissions and Technologies, 2012--2021. URL: https://theicct.org/publication/pv-china-trends-report-jan23/. Key figure: certified CO₂ emission rate of new passenger car fleet 129 g/km (NEDC) in 2021.

[45] IFPI (2025). Global Music Report 2025. URL: https://www.ifpi.org/our-industry/global-music-report/. Key figure: 752 million users of paid music streaming subscription accounts globally at end of 2024, up 10.6% year-on-year.

[46] US Energy Information Administration (EIA) (2023). Use of Energy Explained: Energy Use in Homes. URL: https://www.eia.gov/energyexplained/use-of-energy/homes.php. Key figure: average US household annual electricity consumption approximately 10,500 kWh in 2022.

[47] IEA / National Bureau of Statistics of China. Chinese residential electricity consumption approximately 1,800 kWh per household per year (central estimate; range 1,500--2,100 kWh/household/year). Used to derive the 484,000 Chinese homes equivalence in the Summary.

24/7 CFE (24/7 Carbon-Free Energy)

A more stringent standard than simple annual renewable energy matching. A company committed to 24/7 CFE aims to match every hour of its electricity consumption with carbon-free generation from the same grid region during that same hour. Google has committed to achieving this across all its data centres by 2030.

AAC (Advanced Audio Coding)

A compressed audio format similar in purpose to MP3 but using a different compression method. Apple's preferred format; plays natively on Apple devices.

ALAC (Apple Lossless Audio Codec)

Apple's lossless audio format. Like FLAC, it compresses audio files without discarding any data. Plays natively on Apple devices and in the Apple Music app.

AWS (Amazon Web Services)

Amazon's cloud computing division, the world's largest cloud infrastructure provider by market share. Associated with Bandcamp during its Epic Games ownership period.

BaRT (Bandits for Recommendations as Treatments)

Spotify's primary homepage recommendation algorithm. Uses a Bayesian bandit approach, continuously testing which content to show each user and learning from engagement. Context-aware for device, time of day, and listening history. Runs on every app open for every Spotify user.

Carbon offset

A financial mechanism by which an organisation pays for activities elsewhere that reduce or remove greenhouse gas emissions. As of January 2026, the Green Web Foundation no longer accepts carbon offsets as evidence that a hosting provider operates on clean energy.

CDN (Content Delivery Network)

A geographically distributed network of servers that cache copies of content close to end users. Part of the infrastructure cost of streaming; not activated by local file playback.

CFE (Carbon-Free Energy)

Electricity generated from sources producing no direct carbon emissions during operation: wind, solar, hydro, nuclear, geothermal. Distinct from "renewable energy," which typically excludes nuclear.

CNN (Convolutional Neural Network)

A type of machine learning model used for pattern recognition. Spotify uses CNNs to analyse the acoustic properties of tracks for use in recommendations.

CO₂e (Carbon Dioxide Equivalent)

A standardised unit expressing the warming effect of different greenhouse gases on a common scale over 100 years.

Collaborative filtering

A recommendation technique identifying patterns across many users' behaviour. If users A and B have both enjoyed tracks X, Y, and Z, and user A also enjoys W, collaborative filtering predicts user B might too.

DRM (Digital Rights Management)

Technical controls restricting how digital content can be played, copied, or transferred. Spotify's downloaded tracks are DRM-encrypted and tied to the Spotify app; they cannot be played in any other software and are invalidated if the subscription lapses. Files purchased from Bandcamp or an artist's website are DRM-free.

EAC (Energy Attribution Certificate)

A certificate documenting the generation of 1 MWh of renewable energy at a specific time and location. The Green Web Foundation's updated 2026 criteria require time- and region-matched EACs as evidence of green hosting.

FLAC (Free Lossless Audio Codec)

A lossless audio format. Files are mathematically identical to the source recording. Does not play natively in Apple's Music app; requires conversion to ALAC for Apple devices.

GCP (Google Cloud Platform)

Google's cloud computing infrastructure, used by Spotify for all content hosting since 2019. Global average PUE approximately 1.1.

gCO₂/kWh (grams of CO₂ per kilowatt-hour)

The standard measure of an electricity grid's carbon intensity. Ranges from approximately 15 gCO₂/kWh (Sweden) to approximately 1,013 gCO₂/kWh (South Africa).

GHG (Greenhouse Gas)

Gases that trap heat in the atmosphere, driving climate change: primarily CO₂, methane, and nitrous oxide. Emissions in this document are expressed as CO₂ equivalent (CO₂e).

Green Web Foundation

An independent non-profit verifying whether websites and hosting providers are powered by renewable energy. Updated its criteria in January 2026 to require time- and region-matched energy certificates, no longer accepting carbon offsets.

Inference (ML context)

The process of running a trained machine learning model on new data to produce a result such as a recommendation. Inference is distinct from training. For streaming platforms, inference happens continuously at scale.

kWh (kilowatt-hour)

The standard unit of electrical energy. One kWh = a 1-kilowatt device running for one hour.

kWh/GB (kilowatt-hours per gigabyte)

A measure of the energy intensity of internet data transmission. Published estimates vary enormously (0.004 to 136 kWh/GB). This document uses the Sustainable Web Design model figure of approximately 0.194 kWh/GB, with explicit caveats.

LLM (Large Language Model)

AI trained on large text datasets, capable of generating language. Powers Spotify's AI DJ and AI Playlist features. LLM inference is roughly 100–1,000 times more energy-intensive per call than traditional recommendation ML.

ML (Machine Learning)

A branch of AI in which systems learn patterns from data rather than being explicitly programmed. Spotify runs 220+ active ML projects.

MP3

The most widely used compressed audio format. At 320 kbps, essentially indistinguishable from lossless for most listeners. Plays on virtually every device and software player.

MTCO₂e (metric tonnes of CO₂ equivalent)

One million grams of CO₂ equivalent. Spotify's total reported 2024 emissions: 195,027 MTCO₂e.

MWh / GWh / TWh

Larger units of electrical energy. 1 MWh = 1,000 kWh. 1 GWh = 1,000 MWh. 1 TWh = 1,000 GWh.

NEM (National Electricity Market)

Australia's main electricity grid, covering Queensland, New South Wales, Victoria, South Australia, and Tasmania. Carbon intensity 560 gCO₂/kWh in 2024.

NRT (Near-Real-Time)

Spotify's ML inference process that updates each user's taste profile embeddings within minutes of any listening event.

PUE (Power Usage Effectiveness)

A measure of data centre energy efficiency. Calculated as total facility energy divided by IT equipment energy alone. Lower is better. Industry average approximately 1.5; GCP fleet approximately 1.1; AWS fleet approximately 1.15.

REC (Renewable Energy Certificate)

A certificate representing the generation of 1 MWh from a renewable source. The Green Web Foundation now requires RECs to be matched by time period and geographic region.

Scope 1, 2, and 3 emissions

A standardised framework for categorising GHG emissions. Scope 1: direct emissions from owned sources. Scope 2: indirect emissions from purchased electricity. Scope 3: all other indirect emissions across the value chain. Spotify's reported emissions are approximately 98% Scope 3.

SSD (Solid-State Drive)

A storage device using flash memory rather than spinning magnetic disks. Approximately 0.5–1W in active use.

Sustainable Web Design model

A methodology for estimating the energy and carbon footprint of websites and internet data transfer, developed by Wholegrain Digital and others, drawing on Malmodin et al. (2023). Produces operational energy intensity figures of approximately 0.194 kWh/GB.

VPS (Virtual Private Server)

A partitioned virtual machine running on a shared physical server at a hosting provider's data centre. For independent artists, a VPS from a verified green host offers the lowest overhead and most carbon-controllable distribution option.

WUE (Water Usage Effectiveness)

The standard metric for data centre water efficiency, developed by The Green Grid. Calculated as total annual water used for cooling and humidification (litres) divided by total IT equipment energy (kWh). Expressed in L/kWh. Industry average approximately 1.8--1.9 L/kWh; leading hyperscalers report 0.19--0.30 L/kWh. Measures direct on-site water only; does not include the indirect water footprint from electricity generation.

W / Wh (Watt / Watt-hour)

The watt is the base unit of electrical power. A smartphone streaming audio with screen off draws approximately 0.5–1W; an 80W stereo amplifier draws 80W. One Wh = 0.001 kWh.

WAV (Waveform Audio File Format)

Uncompressed audio. Files are typically 400–600 MB per album. Offers no quality advantage over lossless compressed formats like FLAC or ALAC for listening purposes.