Embodied Carbon & Circularity Across the Lifecycle

For two decades, data-center sustainability meant one variable: drive down PUE, because operational energy dominated lifecycle emissions and the grid was dirty. That framing fails where low-carbon grids and short refresh cycles mean operational energy no longer dominates. Survey-average PUE has flattened around 1.5, best-in-class fleets report ~1.08–1.09, and the best individual site is at 1.04. Operators also sign 24/7 carbon-free-energy deals that push the operational Scope 2 number toward zero. As the operational carbon curve bends down, the carbon that was always there but never counted stops hiding. Embodied carbon is the emissions locked into concrete, steel, silicon, copper and the rest of the bill of materials before the facility draws a useful watt, and it does not move when you clean the grid. On sufficiently low-carbon power and a short replacement cycle it can become the dominant term.

None of this appears on the utility bill. Embodied carbon splits into two physically distinct halves with opposite time constants and opposite levers: the building shell (concrete, steel, the long-lived structure) and the IT and infrastructure stack (servers, accelerators, switches, copper, the short-lived guts). Across both runs the refresh-cycle dilemma. An aggressive 18–36-month first-owner refresh case protects your compute economics (Chapter 1.8) but re-pays embodied carbon each time the fleet turns over. Circularity (reuse, refurbishment, component harvest, recycled-content procurement, certified end of life) is the only structural lever that amortizes embodied carbon rather than merely accounting for it.

The shift is not that embodied carbon grew. Operational carbon shrank around it. Three forces compress the operational term at once: better PUE (less overhead energy per unit of IT work), grid decarbonization (less carbon per kWh), and 24/7 CFE procurement (matching clean energy to actual consumption hour by hour, Chapter 15.3). Each multiplies operational carbon down, and none of them touches embodied carbon.

The result is a crossover. On the US grid, HPE's footprint for a mainstream CPU server (a ProLiant DL380 Gen10 run for four years at 30% utilization) attributes about 30% of the total to manufacturing. The share moves with configuration as well as region: about 15% for a performance build in Japan, about 43% for a base build in Europe. In One Click LCA's generic building framework, embodied carbon on clean power can be 50–80% of whole-life building emissions, with no IT refresh counted yet. On a clean grid with a short refresh cycle, embodied carbon can be the majority. An operator who decarbonizes power must model that crossover before declaring victory.

A data-center life-cycle assessment (LCA) makes the crossover legible. Following ISO 14040/14044 and the EN 15978 building stages, it splits emissions into modules: A1-A3 (product and material manufacture, the bulk of embodied carbon), A4-A5 (transport to site and construction), B (use phase: operational energy plus maintenance and replacement of building services such as chillers, UPS and generators), and C/D (end of life and the credit for reuse and recycling). IT equipment normally sits outside a building LCA. It is assessed with product carbon footprints (ISO 14067) and counted in the corporate Scope 3 inventory, which is where the accelerator refresh lands.

The unit of accounting is kgCO2e, and the discipline that makes a number defensible is the Environmental Product Declaration (EPD), a third-party-verified, product-specific carbon datasheet for a material or component. Run a project-stage LCA with comparable material EPDs where available. Where they are missing, document the generic factor’s geography, vintage and uncertainty before it controls a purchase. Missing EPDs do not invalidate an inventory, because GHG Protocol permits supplier-specific, hybrid, average-data and spend-based approaches for capital goods. A documented data-quality hierarchy (method, representativeness, uncertainty) and a plan to replace the weakest factors with supplier data keep the number defensible. An EPD also settles a comparison only when its product category, functional unit and boundary match yours.

Disclosure rules reach the number through separate routes. Material Scope 3 emissions enter corporate CSRD/ESRS reporting. The EU EED data-center regime (Chapter 15.7) reports operational facility KPIs and does not require project LCA or EPD disclosure, so project-stage LCA and product EPDs remain procurement and audit-readiness tools. For new EU buildings, a whole-life number is on its way to being mandatory. Article 7 of the recast EPBD (Directive (EU) 2024/1275) makes each new building carry its life-cycle global warming potential (GWP) on its energy performance certificate. You calculate it by Annex III, using the Union framework set out in Delegated Regulation (EU) 2026/52. New buildings over 1,000 m² need the figure from 1 January 2028, and every new building from 1 January 2030. National roadmaps toward limit values are due by 2027. Check how your member state's transposition treats data centers before design freeze.

The building shell has the longest time constant and the fewest decision points of any embodied-carbon category. That makes it the highest-leverage place to act and the least forgiving place to get it wrong. A data center's structure stands for 20-40 years. Its embodied carbon is poured once, at construction, and cannot be re-decided without demolition.

Where the carbon sits depends on where you draw the boundary. A shell-only assessment is mostly concrete and steel. Once mechanical and electrical plant is in scope, it dominates over the building's life. In Arup's analysis, MEP accounts for up to 70% of upfront embodied carbon and up to 88% across the whole life. Introba's model of one hypothetical data center puts about 90% of 60-year embodied carbon in MEP equipment, mostly from maintaining and replacing it (Introba, Hidden Emissions of the Cloud, 2023). Within the shell, concrete's carbon comes from clinker, the carbon-intensive calcined ingredient in Portland cement. The shell levers are therefore, in priority order, attacks on clinker and on virgin steel.

The levers can be sized, though some figures are pilot targets. Low-clinker concrete replaces part of the clinker with supplementary materials (fly ash, slag, calcined clay, biogenic limestone), and Microsoft is testing mixes in Washington that it says can cut concrete's embodied carbon by more than 50% versus traditional mixes (Microsoft, December 2024). CO2-mineralizing admixtures (CarbonCure-class) add only a few percent, by cutting the cement needed. Electric-arc-furnace (EAF) steel made from scrap, and emerging hydrogen-DRI 'green steel,' displace blast-furnace steel's heavy footprint. On a brownfield site, reusing the existing frame avoids some new steel altogether if testing shows it can carry the hall's loads: Microsoft's Newport, Wales data center takes about 10% of its structural steel from the original building, saving about 520 tons of CO2 (Microsoft, October 2025). Mass timber, with cross-laminated timber (CLT) substituting for upper-floor steel and concrete, is the most aggressive shell lever in production. Microsoft's hybrid steel-concrete-CLT data centers in Virginia are estimated to cut embodied carbon ~35% versus conventional steel and ~65% versus precast concrete (Microsoft, October 2024). Modular and prefabricated construction reduces material waste and on-site (A5) emissions, and right-sized structural design avoids the over-engineering that bloats concrete and steel tonnage. For MEP plant, put product EPDs or CIBSE TM65 estimates in the specification, and design equipment to be repaired in place and replaced less often.

These levers carry costs of their own: a green premium and, for novel options like CLT, structural-loading and fire and permitting questions that collide with the density ramp. A liquid-cooled GB200/GB300 NVL72 hall must bear ~3,000–3,300 lb wet racks, and a Vera Rubin NVL72 hall ~4,000 lb (Chapter 6.2). A structure optimized for embodied carbon must still carry that load and reserve floor-loading headroom for the next generation. Choose mass timber for the wrong hall and you have traded embodied carbon for a structure that caps your density ramp, and lifting that cap later means costly, disruptive strengthening or replacement. Decide the shell jointly with the density-ramp basis.

The second half of embodied carbon sits in the IT and infrastructure stack, and it behaves nothing like the shell. The building is poured once and stands for decades. An aggressive first-owner/economic case replaces accelerators every 18–36 months (Chapter 14.9). Under that case, embodied carbon stops being a one-time A1-A3 charge and becomes a recurring replacement charge, and at AI refresh velocity that recurrence is the defining feature of an AI data center's carbon profile. Advanced-node silicon is carbon-intensive to fabricate (sub-7nm logic, HBM stacks and advanced packaging all push up fab energy and process-gas emissions), and a GPU server bundles that silicon with copper-dense interconnect, NICs, switches, optics and power-delivery hardware.

A counter-intuitive finding from 2025 LCA work shows where IT embodied carbon sits. In EcoServe’s modeled configurations, the host system (CPUs, DRAM, storage, board, chassis) can dominate manufacturing embodied carbon, while the GPU dominates operational carbon (EcoServe, arXiv:2502.05043, 2025). Optimizing only GPU efficiency therefore misses a large slice of the embodied opportunity. Finding the real levers takes whole-server, whole-rack accounting. Those levers favor denser, higher-utilization configurations (fewer host systems per unit of useful work) and longer service life for the host stack even when accelerators rotate.

Trace and flip. Keep: 10,000 MWh/year × 3 years × 0.030 tCO2e/MWh = 900 tCO2e. Replace: 7,000 × 3 × 0.030 + 450 ≈ 1,100 tCO2e. Keeping saves about 180 tCO2e over the same service. The electricity-factor crossover is 450/[3 × (10,000 − 7,000)] = 0.050 tCO2e/MWh. Above that factor, replacement saves carbon. At an assumed 0.080, keep emits 2,400 t and replace about 2,100 t, a roughly 270 t saving. If the old fleet cannot pass the service test, keeping it is not an option at any factor. Recompute with matched LCAs and energy-at-fixed-output trials before purchasing.

Close the disposition mass balance. The asset owner records each serial number’s destination, sanitization evidence, custody transfers, repair and transport energy, and proceeds net of fees. Incoming mass should reconcile with whole-unit reuse, component harvest, material recycling, disposal and inventory still held. A processor’s combined reuse-and-recycle percentage does not break out those outcomes. Count avoided manufacturing only with an explicit counterfactual for the displaced product serving the same demand over the same horizon, and keep any beyond-boundary credit separate. Chapter 14.9 covers how to run disposition, and Chapter 1.8 prices the refresh.

The refresh cycle sets two goods against each other. Each new accelerator generation brings a large operational-efficiency gain: more useful FLOPs per watt, less energy per token. On operational carbon alone, refreshing fast looks green, because the new hardware does the same work for less power. But every refresh re-pays the embodied carbon of manufacturing the replacement, and on a clean grid, where embodied carbon already dominates, that re-payment can swamp the operational saving. The faster you refresh, the more often you pay the manufacturing charge, and the less useful work each unit of embodied carbon is spread over.

The economics behind fast refresh are analyzed in Chapter 1.8 and Chapter 14.9. The guide tests a contested 2–3-year intense-duty bear case against published 4–6-year useful-life estimates and 5–6-year book policies. A generation that is ~2x more performant per dollar makes the old hardware uncompetitive long before it is worn out. That same compression protects compute margins and drives recurring embodied carbon. The two pressures point in opposite directions, and no answer is universally correct. The choice is made per workload:

• Refresh-fast (efficiency-led): rotate to the newest accelerator on an aggressive 18–24-month first-owner cadence, capture the per-token energy win, and rely on a strong secondary market to keep the displaced hardware in service elsewhere. The embodied carbon is then amortized across a second owner rather than re-paid as waste.
• Extend-life (amortization-led): run accelerators 4-6 years, accept a higher operational-energy cost per unit of work, and amortize the embodied carbon over far more lifetime work. This is viable where the workload is interruption-tolerant batch inference, or where grid carbon is already low enough that the operational penalty is small.

The variable that settles the dilemma is where the displaced hardware goes. If a refreshed GPU is resold and runs three more years for a budget buyer, its embodied carbon is amortized across ~5-6 total service years and the fast refresh is defensible. If it is scrapped, the embodied carbon was paid for 2 years of work and the fast refresh is a carbon disaster. The refresh-cycle dilemma is, in the end, a circularity question, so the two cannot be decided separately.

Circularity is the only lever left once the bill of materials is fixed. It works on the denominator (more useful work per unit of manufacturing carbon), not the numerator. Keep the two accounts apart. Extending or transferring an asset's life lowers carbon per unit of service, but it does not erase emissions already booked, because GHG Protocol capital-goods accounting recognizes cradle-to-gate emissions in the year of acquisition rather than amortizing them across the service life. The circular hierarchy runs, in descending carbon value: reuse (redeploy the asset whole, internally or via resale), refurbish (repair and recertify for a second owner), harvest (recover working components such as DIMMs, NICs, drives and power supplies for spares and repair), and finally recycle (recover materials, principally copper and precious metals). Recycling is last because it destroys the most embodied value. A refurbished server preserves the manufacturing carbon of the whole assembly, whereas recycling recovers only the metal.

Refurbishers report energy savings of up to ~85% versus new manufacture and prices 55-80% below new (Human-I-T, a refurbisher, as relayed in DCD circular-economy analyses, 2025). Treat both as advocacy figures until a product LCA and a quote back them. Hyperscalers have operationalized the idea. Microsoft reported 92% reuse and recycling of decommissioned servers and components in FY2025, above its 90% target for the second consecutive year (first met in FY2024, a year ahead of its 2025 goal). It now runs eight Circular Centers (as of Sept 2026). Google harvested more than 7.5 million components from decommissioned hardware in 2025 for its own builds and spares, and has resold more than 54 million since 2015, 2.6 million of them in 2025, once internal needs are met (Google 2026 Environmental Report). Recycled copper, central to data-center power and interconnect, carries far lower embodied carbon than primary copper, and domestic secondary-refining capacity is growing to feed this demand (Electronics360, 2025).

End of life is where circularity happens or fails. Decommissioning that routes assets through certified ITAD (IT asset disposition), with NIST SP 800-88 Rev. 2 media sanitization and R2v3 / e-Stewards / ISO 14001-certified processors, keeps the asset in the circular loop. The provider-reported 35-50% of value within ~45 days is a vendor benchmark with no denominator, asset cohort, condition or fee treatment given. Underwrite residual value against a specific transaction sample, separate gross sale price from net proceeds after transport, sanitization and broker fees, and carry a zero-recovery downside case (Chapter 14.9; facility-side decommissioning in Chapter 14.10). Decommissioning that treats hardware as waste lands it in the 62-million-tonne global e-waste stream (heading to 82 Mt by 2030, only ~22% formally recycled, per the UN Global E-Waste Monitor). Scrapping adds waste-treatment emissions (Scope 3 Category 5) and throws away hardware whose manufacturing carbon was already booked. The difference between the two outcomes is a procurement decision made years earlier: whether residual-value, take-back and certified-ITAD clauses were written into the original purchase.

Resale also has a legal gate. H100-class and later accelerators stay under US export control after their first sale: some destinations and end users require a license, and if the buyer diverts the hardware, the seller remains liable. Before accepting a bid, get the buyer's end-use and end-user statement, screen both the buyer and the final destination, and write a no-re-export clause into the sale (Chapter 14.9).

Three recurring mistakes follow from treating embodied carbon as an afterthought to the operational story:

• Decarbonizing power, declaring victory. The operator signs 24/7 CFE and reaches a near-zero CUE while running a 2-year refresh with no circularity program. Operational emissions are addressed, and recurring embodied carbon is untouched.
• Fast refresh without a reuse path. The operator chases every accelerator generation for the operational-efficiency win and scraps the displaced hardware. This is the embodied-carbon worst case: manufacturing re-paid every ~2 years, spread over almost no useful work, then scrapped with waste-treatment emissions landing in Scope 3.
• Embodied carbon discovered at disclosure time. The operator waits until a material Scope 3 obligation under CSRD/ESRS forces the number, then rebuilds it from generic factors because no EPDs were required at procurement. The inventory is less precise and needs documented data quality, and the structure is already poured, with every high-leverage shell lever spent years ago. Under Omnibus I, mandatory CSRD reporting now covers only companies above 1,000 employees and €450M turnover. Smaller developers and colos are more likely to meet the number through a customer's Scope 3 data request (see Chapter 15.7).