How SOMA calculates the carbon and water footprint of AI usage

This section decides what the factor covers. Purchased AI inference sits in Scope 3 Category 1, whose minimum boundary is cradle-to-gate, so a factor that counts only serving electricity falls short of it. The central factor therefore has four terms: serving electricity (accelerators, the rest of the server, capacity held idle, facility cooling and power distribution, and the upstream fuel cycle of the electricity), embodied server manufacturing and transport, data-centre construction, and the final training runs of the model amortised over the tokens it serves. Model development beyond those final runs is quantified but reported as a separate line and never added, because it rests on sector-wide ratios rather than on anything disclosed about the model in use. Storage, external networking and hardware end-of-life sit outside with small stated bounds; end-user devices and provider staff are excluded and named. A preparer declares this boundary, its exclusions, and that the figure is a bottom-up estimate, not the provider's own allocation of its inventory.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.1 "Scope and Boundary".

This section decides which data each AI service is counted from. A tier is assigned per service, not per company. Tier 3 uses a supplier's own assured cradle-to-gate figure; no major provider offers one yet, so it is empty today. Tier 2a uses exact billed tokens from an API or usage export, split into uncached input, cached input and output. Tier 2b estimates tokens from message or seat counts using published workload defaults. Tier 1 applies a spend-based factor where the vendor, not the customer, holds the AI relationship. The GHG Protocol sets no strict hierarchy, so the choice follows the data a preparer can actually obtain. Billed input includes the system prompt, retrieved context and the conversation history re-sent each turn. Read a product's own message count first and fall back to the per-seat default only where it cannot be retrieved. A preparer declares the tier of each service and why.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.2 "Activity Data: The Four Tiers".

This section decides how billed tokens become the quantity a factor is applied to. Energy is benchmarked per output token on a benchmark that already contains its own input processing, so input tokens are weighted and the result divided by a normaliser so prefill is not charged twice. Effective tokens equal output tokens, plus uncached input weighted at 0.3, plus cached input weighted at 0.05, all divided by 1.339. Electricity, embodied hardware and construction are charged on effective tokens; training is charged on billed tokens, input and output including cached input. Where only a billed total is known, multiply it by 0.302, which assumes an input share of 0.85 and claims no cache credit, the conservative choice. Where messages are counted, apply the weights to the split defaults directly and never also apply that multiplier. A preparer declares the weights used and any cache credit claimed.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.3 "Token Basis".

This section decides how the four terms are built and combined. Facility energy is benchmarked accelerator energy times a host uplift for the rest of the server, an idle uplift for capacity held ready, and the facility's PUE. Electricity is that energy times the lifecycle grid intensity of the serving region. Embodied hardware spreads server manufacturing and transport over its service life at the fleet's utilisation. Construction is a constant per kWh of server electricity. Training is the disclosed emissions of final training runs divided by the tokens a model serves in its life. The first three are per million effective tokens and form the inference footprint; training is per million billed tokens. They are never summed into one unlabelled total, and a billed count must never be multiplied by an inference footprint. The development line is reported separately and never added. Contracted accelerator-hours use a reserved-capacity variant with no idle uplift. A preparer declares which terms were reported and on which denominator.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.4 "Factor Equation".

This section decides how a model name and a region become a factor. Where the model served has its own benchmark result, its own factor is used. Every closed commercial model is instead assigned to a size class, small, mid or frontier, by its disclosed total parameter count, or by the vendor's pricing tier where no size is disclosed; a disclosed size always overrides the tier name. The class value is the mean of its benchmarked members. The assignment is published as data: 187 identifier patterns, read between 2026-09 and 2026-10, matched longest first. An identifier that matches nothing takes class C, the highest-intensity class, and is flagged. A customer-pinned region takes its own grid row, an EU-only processing commitment the EU average row, and an unknown region the global average. A preparer declares the class or own-model factor of each service, the region rule applied, and the share of activity that fell back.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.5 "Resolution: From a Model Identifier and a Region to a Factor".

This section decides how water is counted from the same tokens and energy, without changing the carbon figure. On-site water is the cooling water at the data centre; off-site water is the water consumed generating its electricity. Water is never added to carbon. Three conventions change the answer and are declared rather than settled. Every on-site value is put on a consumption basis, because adding withdrawal to consumption double counts. Evaporation at hydroelectric reservoirs may be allocated to electricity or excluded from it; both totals are reported side by side, neither is preferred, and they are two conventions for one quantity, not an uncertainty range. Scarcity weighting with AWARE 2.0 is applied to the on-site term alone, because the data centre is the only location known. A preparer wanting one volumetric number reports the hydropower-excluded total and discloses the allocated one beside it. Where a region has no water value, the result is "not available", never zero.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.6 "Water as an Extension of the Same Chain".

This section decides what to do when no usage data exist. Apply the environmentally extended input-output factor for computer and related services of the vendor's producing country to spend excluding VAT, deflated from the spend year to the factor's price basis with the harmonised consumer price index. The deflated figure is the method's result; an undeflated figure is shown only for comparison. It is a screening figure that covers the whole subscription, typically several times a physical estimate, and never an upper bound. Unless the AI share can be evidenced, for example by a separately priced AI add-on, keep the subscription as a spend-based screening line and disclose that its AI share is not separately estimable. Do not relabel a whole subscription as AI. A preparer declares the factor, the deflator and the spend year, and removes any spend already counted by a token-based line so the same purchase is not counted twice.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.7 "Spend-Based Fallback".

This section decides how uncertainty is reported, keeping three things apart. The class range is the spread of the benchmarked models of a class around its mean; it is data, not a confidence interval, and disappears where the model has its own factor. The activity band applies only to estimated tokens and is the interquartile range of observed workloads. Systematic uncertainty of the factor is handled by scenarios: one central value, and one conservative high case that raises the host uplift, the idle uplift with lower utilisation, and PUE together while holding training central. The high case is a scenario reported beside the central value, never a confidence interval and never inside the formula; no low case is reported. Data quality is scored on the method's own scale, not a GHG Protocol scale: exact tokens score 2 for activity and 3 for the factor; estimated tokens score 4 and 3. A preparer reports the central value with its class range and high case.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.8 "Uncertainty and Data Quality".

This section decides what a preparer discloses and how the factors stay current; the rules are the method's proposal and go beyond what ESRS or the GHG Protocol prescribe. Disclose the boundary and its exclusions, the location basis and grid year, each service's class or own-model factor and why, the region rule, the token weights and any cache credit, the operating-point rule, and the global warming potential set. Report the central value with the class range and the high case; where a class rests on a vendor tier rather than a disclosed size, add the adjacent-class values as a scenario. Report the figure gross, before offsets. Pin one dataset version per reporting year, and treat a change of version or grid year as a methodology change. Each input family is refreshed from a named source on a stated cadence, and a provider disclosing energy per token for a model it serves supersedes the class chain for that model. SOMA pins soma-ai-ef-2026.10-dataset-v3.0.0.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §3.9 "Reporting Rules and Maintenance".

This section decides the energy anchor. Accelerator energy per output token comes from the ML.ENERGY leaderboard, the only open benchmark covering the same open-weight models on H100 and B200 under one protocol. For each model the method takes the lowest-energy configuration that still streams at interactive speed, the point a cost-minimising provider would choose; this is a best-case serving point, so it biases electricity low. Taking instead the mean over every configuration meeting the same latency limit would raise the result by nearly a factor of three, and is carried as a named scenario, never merged into the high case. A class value is the arithmetic mean of its benchmarked members; on H100 the mid class draws 0.236 Wh of facility energy per thousand effective tokens. The class range at the global-average grid is -40.4 % to 47.4 % for the small class, -38.5 % to 23.2 % for the mid class and -66.6 % to 54.5 % for the frontier class. A preparer declares the operating-point rule and its named scenario.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.1 "Accelerator Energy per Token".

This section decides the three uplifts that turn accelerator energy into the energy a facility buys. The host uplift for the rest of the server is 1.5, between 1.43 and 1.88, from a metered eight-accelerator node, and is more likely to understate than overstate. The idle uplift for capacity held ready is 1.1, between 1.05 and 1.2, from Google's published serving figures; it is tied to a fleet utilisation of 0.75, so the two always move together. PUE is 1.2, between 1.09 and 1.45; it is an author choice set deliberately above the hyperscalers' disclosed fleet values to cover serving outside hyperscale facilities. Construction is charged on server electricity and does not move with PUE. A preparer declares these values and that no on-premises variant exists, so the same uplifts are applied to self-hosted serving.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.2 "Server and Facility Overheads".

This section decides which grid number turns facility energy into emissions. Grid intensities are annual, location-based and on a lifecycle basis for 2023, so upstream fuel-cycle emissions are already inside them. Countries, the EU average and the world average come from Ember; United States grid subregions come from EPA eGRID2023, scaled up to Ember's lifecycle basis. The global average is 0.483 kg CO2e per kWh, the EU average 0.236, and France 0.0533. The dataset carries 240 grid rows and maps 143 public cloud regions to them with a stated confidence; outside the United States a cloud region takes its national average. Because the values are location-based, a provider's renewable purchases do not lower them, and because they are per kWh generated, transmission losses are outside them. Newer grid data are lower, so these rows lean high. A preparer declares the grid year and the location basis.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.3 "Grid Carbon Intensity".

This section decides the two building and hardware terms. Server manufacturing and transport come from Dell's published product carbon footprints for an eight-accelerator H100-class server, which are cradle-to-gate, with no end-of-life stage and no recycling credit. That footprint is spread by time over a hardware life of 5 years at the fleet's utilisation and divided by each model's benchmarked throughput, giving 0.89, 3.34 and 12.84 g CO2e per million effective tokens for the small, mid and frontier classes. Its upper tail is open. Construction is a separate term on server electricity, independent of the grid, at 7.66 g CO2e per kWh, the upper end of a narrow published range; a lower Carbon Trust reading is cited in the source but not emitted. That gives 0.46, 1.50 and 5.60 g per million effective tokens. A preparer declares both terms as inside the boundary.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.4 "Embodied Hardware, Transport and Construction".

This section decides the training term and the separate development line. Training rests on Meta's location-based disclosures for its Llama family, the only verified per-model training figures spanning all three classes, covering final training runs only. They are divided by the tokens a model serves in a service life of 12 months, between 6 and 18, at the one per-model serving volume ever disclosed. That gives 0.60, 2.93 and 12.84 g CO2e per million billed tokens. The volume rests on one early-life datapoint, so the term leans low. Development beyond the final runs is estimated with a ratio of development to serving energy of 1.5, between 0.67 and 2.5. The ratio is a sector spend reconstruction, not a reading of the model in use, so the development line is reported separately and never added to the factor. A preparer who reports a development-inclusive total labels it and states the ratio.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.5 "Training, Service Life and the Development Ratio".

This section decides the numbers behind effective tokens. An uncached input token is weighted at 0.3 of an output token, between 0.1 and 0.4, read from the benchmark's own results for prefill against decode energy. A cached input token is weighted at 0.05, between 0.02 and 0.1; no energy benchmark of a cache read exists, so it sits between a commercial and a physical bound and leans low for cache-heavy work such as agentic coding. Each benchmark run is restated to a reference mix of 1.13 uncached input tokens per output token, which gives the divisor 1.339, between 1.113 and 1.452. At the default input share of 0.85 with no cache credit, effective tokens are 0.302 of billed tokens, between 0.211 and 0.337. A preparer declares the weights and uses the export's own input, cached and output split wherever it reports one.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.6 "Token Weights".

This section decides how messages and seats become tokens when a product reports no tokens. A short chat exchange is 928 billed input and 282 output tokens, the input including re-sent context. An agentic coding request is about 85,000 tokens, most of the input served from cache; count model requests, not human turns. An active enterprise seat sends 30 messages a month and a heavy-use seat 325. The band on estimated tokens runs from -50 % to 130 %, asymmetric because observed workloads are right-skewed. A benchmarked 5.0 % of conversations use a reasoning mode, and no public source gives a reasoning-token multiplier, so at Tier 2b SOMA multiplies estimated output tokens only by 1 + s(k - 1), with k an author assumption of 25 (range 10 to 50), an uplift of about 2.2. Exact token counts already include reasoning tokens and are not multiplied. A preparer declares which defaults replaced missing data.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.7 "Tier 2b Activity Defaults".

This section decides where the water numbers come from. On-site values are published water use effectiveness figures for named cloud regions, or ratios derived from a provider's own published site water and electricity; where neither exists a continental aggregate or an older site series stands in and is marked unconfirmed. They range from 0.02 litres per kWh in Stockholm to 1.57 in Singapore. Withdrawal figures are converted to consumption with the one documented conversion, and the as-published figures become a sensitivity. Off-site values are the consumption factors of electricity generation, with a hydropower-excluded variant beside them; Singapore has no off-site source. Scarcity weighting uses the AWARE 2.0 factor of the region and is not applied to the global or EU averages, which are not locations. Water is published for 17 regions only; any other region reads "not available". A value is applied whichever provider serves from that region, which is a proxy. A preparer declares the basis, both hydropower conventions and each region's confidence.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.8 "Water Inputs".

This section prints the factors a reader can check by hand. At the global-average grid, the inference footprint is 0.0361, 0.119 and 0.443 kg CO2e per million effective tokens for the small, mid and frontier classes. Where only a billed total is known, the convenience factor folds the default multiplier and the training term into one figure per million billed tokens: 0.0115, 0.0388 and 0.147. For comparing models or regions use the inference footprint, never the convenience factor. The dataset applies the same rules beyond the printed regions: 32 benchmarked model rows on H100 and B200 with their own factors, 240 grid rows and 143 cloud regions, with a reference calculator that names the file, line and column behind every number. Closed commercial models have neither a benchmark result nor a disclosed size, so for them the class remains the unit of estimation. A preparer cites the dataset version used.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §4.9 "The Factor Tables".

This section decides which choices move the result and how much. The largest mover is not an input but a choice: which configuration of a benchmarked sweep counts as the serving point, worth nearly a factor of three. The training denominator is the second. The conservative high case adds 63.9 %, 63.7 % and 63.7 % to the inference footprint of the small, mid and frontier classes, with PUE and the host uplift carrying roughly equal weight. The hardware inputs move almost nothing. On the billed denominator, the input-token weight matters more than any embodied or construction input. When the input bands are propagated together, the central value sits near the lower end of what they allow, and the high case lies beyond the likely range, so it is a bound, not a likely value. That propagation excludes the class range, the operating point, the serving region and the activity estimate, which stay separate. Two grid conventions, transmission losses and behind-the-meter generation, are named sensitivities rather than inputs. A preparer discloses the high case and the named operating-point scenario.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §5.5 "Sensitivity Analysis".

This section shows the method applied. A hypothetical European professional-services firm has three AI services, one per tier. An API with exact tokens and no disclosed region uses the global average: effective tokens times the mid-class inference footprint of 0.119 kg per million, plus billed tokens times training of 0.00293 kg per million. A seat product estimates tokens from its exported message count, weights its model classes by the days each default model was in force, and keeps the activity band. A bundled subscription stays a spend-based screening line that is not an AI figure. For a firm like this the AI line is small, a documentation task rather than a magnitude one. A second example, fifty developers using agentic coding tools, shows the opposite: there the class assignment and the development line dominate, and the cached-token weight, which has no energy benchmark, widens the range. The suggested disclosure states boundary, activity data and factors, uncertainty, and how double counting with spend-based lines was avoided, and the preparer retains the exports behind it.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §6 "Application".

This section decides where the method sits among standards. It is written for the CSRD and ESRS; the same factors and activity rules would serve an ISSB climate disclosure or a California disclosure, but the reporting rules are not mapped to them. The Green Software Foundation's Software Carbon Intensity specification is a rate per functional unit and maps onto exact-token accounting; its AI extension suggests one token as the unit, which the factor here is compatible with, while carrying four terms on two denominators rather than one. The ITU-T recommendation on AI life-cycle assessment is in force but supplies no emission factors, activity rules or per-token values. The GHG Protocol is consolidating its standards with ISO, and its working group has AI workloads and energy per token under discussion; the method's reporting rules would need revisiting against that text. The claim is that no standard yet supplies a preparer's factors and activity data, not that none addresses AI assessment. A preparer states that the method is a proposal, not a prescribed standard.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §7.4 "Relation to Existing Standards".

This section collects the rules the tiers leave open. A boundary table maps every lifecycle element to the term carrying it, the separately reported development line, bounded exclusions, and what is out of scope: customer fine-tuning, embeddings and image, audio or video work. AI bought through one's own cloud account replaces the cloud tool's line only where that line can be isolated. A supplier's energy quantity is the hybrid method, not a supplier-specific figure. A supplier figure qualifies only with a stated boundary, allocation and assurance status, at least limited assurance in the sense of ISO 14064. Market-based instruments, the preparer's own or the provider's, never reduce this location-based figure. State one set of global warming potentials, AR5 as the grid data use. Pin one dataset version per year, recalculate the base year only above the company's own significance threshold, and back-cast on spend only after adjusting for falling token prices. Remove spend already counted by a token-based line. Every input family has a named source, cadence and out-of-cycle trigger.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §S5 "Decision Rules and the Refresh Schedule".

This section records every assumption the method uses, grouped by where it is discussed, each with its value, source, the reason that source was chosen and the expected direction of error: low means the factor is more likely an underestimate, high an overestimate. It opens with every value changed between dataset versions: the operating point is now the lowest-energy configuration at interactive speed, and a class value is the mean of its benchmarked members rather than one named proxy model, which raises the frontier class. Most named biases point low: the best-case operating point, the host and idle uplifts, the embodied upper tail and the training denominator. Several values are author choices rather than sourced figures, PUE and the cached-token weight among them, and are named as such. SOMA uses dataset version soma-ai-ef-2026.10-dataset-v3.0.0. The concept DOI 10.5281/zenodo.20443585 resolves to the latest version, the record 10.5281/zenodo.22767475 holds the superseded values, and for this version, v3 record: 10.5281/zenodo.23371447 (dataset version 3.0.0, copied 2026-10-19, CC BY 4.0). A preparer cites the version and the register when asked how a figure was built.

Source: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, Llopis (2026), §S7 "Register of Assumptions".