7. Assessment and Calculation

7.1 IT Electricity Usage Calculation

IT Electricity Usage is calculated by mapping each unit of cloud usage to its underlying hardware components and summing the energy contribution of each. The energy for a given unit of usage is built up from the energy of the components identified for that SKU.

Component power figures are grounded in authoritative benchmarks and manufacturer data, as set out in Section 5.2. Server power is aligned with methodologies used by the US EPA, the EU Ecodesign Regulation, and the SPEC Power database, and accounts for the whole server rather than the processor alone. Where no utilisation data is available for a unit of usage, a conservative default utilisation assumption is applied.

7.2 Facility Energy and Emissions Calculation

Total Facility Energy is calculated by multiplying IT Electricity Usage by the applicable PUE factor for the cloud region. Where a cloud provider does not publish a region-specific PUE value, the best available published value is used, which may be a global average. Usage Emissions are then calculated by multiplying Total Facility Energy by the Grid Carbon Intensity for the relevant region and time period.

Usage Emissions are reported on a location-based basis in accordance with the GHG Protocol Scope 2 Guidance. Greenpixie follows the US EPA's guidance on indirect emissions from purchased electricity, giving precedence to regional or sub-national GCI factors where available.

7.3 Embodied Emissions Calculation

Embodied emissions are estimated on a component-by-component basis, using hardware specifications for the instance type, semiconductor die-size data, lifecycle assessment data from Dell's server assessments, NVIDIA's product carbon footprints, the ACT semiconductor model of Gupta et al., Boavizta and Gröger et al., and assumed server operational lifespans based on publicly available cloud provider statements.

Components covered include CPU, RAM, SSD storage, HDD, GPU where applicable, power supply units, chassis and fans, motherboard, expansion cards and assembly.

The silicon die of a CPU or GPU is charged per mm² at a coefficient that depends on its process node, taken from the ACT model of Gupta et al. A newer node takes more lithography energy and more process gas per cm² of wafer, so a cm² of 5 nm die carries about twice the footprint of a cm² of 28 nm die. The coefficient already includes the fab yield, so the raw die area is used.

For CPUs, the methodology takes the total die area of the processor package from the TechPowerUp CPU database, counting both the compute chiplets and the separate IO die that modern AMD and Intel server processors carry, multiplies it by the die coefficient for the CPU's process node, and adds a baseline value for packaging materials from Gröger et al. Where no die area is published, it is estimated from the process node and thread count.

For RAM, emissions are estimated using die density figures for each DDR generation, from Boavizta's crowdsourced table and TechInsights measurements, to derive the effective die size, then multiplying by the per-cm² memory carbon intensity factor from Gröger et al. A baseline is added for PCB, connector strip, and packaging.

For SSDs, the same die-area approach is used with a NAND die density and a die coefficient for each NAND generation. Newer, taller NAND stacks hold more data per cm² and cost more to make per cm², and the generation of an instance's drives is inferred from the launch year of its CPU. For HDDs, a fixed per-device manufacturing footprint is applied from open lifecycle datasets. Power supply units, chassis, fans, motherboard and expansion cards are assigned fixed values per server from Dell's critically reviewed life cycle assessments of its PowerEdge servers.

For GPUs, the methodology sums four terms: the die area multiplied by the die coefficient for the GPU's process node, the onboard memory capacity multiplied by a per-GB factor derived from NVIDIA's HGX H100 and HGX B200 product carbon footprints, a fixed base for the rest of the board derived from the same two footprints, and a thermal solution term that scales with the card's thermal design power. Each GPU model is recorded per memory size, and each instance family is mapped to the size it carries.

Estimated allowances for transport to the customer and end-of-life processing are added to the component manufacturing totals as a percentage uplift, with a separate pair of percentages for servers, disks, GPUs and networking equipment taken from published life cycle assessments of each. End of life follows the recycled content method of the GHG Protocol Product Standard, so no recycling credits are taken. An allocation for data centre shell construction is added on top. Total server embodied emissions are the sum of all component emissions and these allowances. They are then allocated to the client based on the ratio of reserved capacity to total server capacity, amortised over the server's assumed operational lifespan, and pro-rated by usage hours.

Where a server is built to a cloud provider's own design and not sold commercially, its non-chip components are estimated from published assessments of comparable standard servers. Transport allowances assume air freight, and hardware moved by sea has a lower transport footprint, so these allowances are an upper bound.

7.4 Water Consumption Calculation

Total water consumption comprises two components. The first is IT cooling water consumption, the water consumed on site in cooling the IT equipment, estimated as the cloud provider's published WUE multiplied by IT Electricity Usage. The second is electricity generation water consumption, the water consumed during the generation of that electricity, estimated using regional water intensity factors that account for the mix of energy sources powering the local grid, including water withdrawn and consumed for fuel extraction, processing, and power generation.

Where cloud providers publish site-specific or region-specific WUE data, these values are used directly. Where only global averages are available, the global average is applied. Where regional electricity generation water data is not available from the producing country, a global average from the World Resources Institute is used.

The water consumption methodology is developed in collaboration with Dr Shaolei Ren at the University of California, Riverside, and informed by the published work Making AI Less Thirsty by Li et al.