Embodied Carbon & Circularity Across the Lifecycle

For two decades, data-center sustainability was a single-variable problem: drive down PUE, because operational energy dominated lifecycle emissions and the grid was dirty. That framing is now obsolete, and the reason is arithmetic. As fleet PUE flattens around 1.5 industry-wide and best-in-class sites reach ~1.07-1.09, and as operators sign 24/7 carbon-free-energy deals that scrub the operational Scope 2 number toward zero, the operational carbon curve bends down — and the carbon that was always there but never counted stops hiding. Embodied carbon — the emissions locked into concrete, steel, silicon, copper, and the rest of the bill of materials before the facility ever draws a useful watt — does not move when you clean the grid. On a clean grid it becomes the dominant term: studies converge on embodied carbon representing 50-80% of lifecycle emissions for a low-carbon-powered AI facility (Opna/Vertiv; arXiv cradle-to-grave LCA, 2025).

This chapter covers the part of the lifecycle that never appears on the utility bill. We split the embodied problem into its two physically distinct halves — the building shell (concrete, steel, the long-lived structure) and the IT and infrastructure stack (servers, accelerators, switches, copper, the short-lived guts) — because they have opposite time constants and opposite levers. We then confront the fork that defines the AI era: the refresh-cycle dilemma, where the same 18-36 month hardware churn that protects your compute economics (Chapter 1.8) re-pays embodied carbon faster than any other industry. And we close on circularity — reuse, refurbishment, component harvest, recycled-content procurement, certified end-of-life — as the only structural lever that amortizes embodied carbon rather than merely accounting for it.

The shift is not that embodied carbon grew; it is that operational carbon shrank around it. Three forces compress the operational term simultaneously: PUE improvement (less overhead energy per unit of IT work), grid decarbonization (less carbon per kWh), and 24/7 CFE procurement (matching that clean energy to actual consumption hour by hour — Chapter 15.3). Each is a multiplier on operational carbon and none of them touches embodied carbon. The consequence is a crossover: on a coal-heavy grid, a server's manufacturing emissions might be 15-30% of its lifetime total; on a clean grid running a short refresh cycle, embodied carbon can be the majority. An operator who decarbonizes power and declares victory has, on a clean grid, addressed the minority of the problem.

A data-center life-cycle assessment (LCA) is the instrument that makes this legible. Following ISO 14040/14044 and the EN 15978 building stages, an LCA partitions emissions into life-cycle modules: A1-A3 (product/material manufacture — the bulk of embodied carbon), A4-A5 (transport to site and construction), B (use phase — operational energy plus maintenance and component replacement, where the refresh cycle lands), and C/D (end-of-life and the credit for reuse/recycling). 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. The fork an operator faces here is foundational: run a real, project-stage LCA with material EPDs now, or inherit a generic, indefensible number when a disclosure regulation forces it later. Under CSRD/ESRS and the EU EED data-center reporting regime (Chapter 15.7), the embodied number is moving from optional marketing to audited disclosure — and a number you cannot trace to EPDs is a number you cannot defend.

The building shell is the embodied-carbon category with the longest time constant and the smallest number of decision points — which makes it both the highest-leverage 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. Within construction-material carbon, the breakdown is consistent across studies: concrete is ~40% and steel ~10% of the embodied total, with the balance in cabling (copper-intensive), backup-power equipment, and cooling infrastructure (Opna/Vertiv synthesis). Concrete's dominance comes from clinker — the carbon-intensive calcined ingredient in Portland cement — so the building-shell levers are, in priority order, attacks on clinker and on virgin steel.

The lever set is now mature enough to quantify. Low-clinker / supplementary-cementitious concrete (fly ash, slag, calcined clay) and CO2-mineralizing admixtures (CarbonCure-class) can cut concrete's embodied carbon by up to ~50% versus a conventional mix. Electric-arc-furnace (EAF) steel made from scrap, and emerging hydrogen-DRI 'green steel,' displace blast-furnace steel's heavy footprint. Mass timber — 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; ESG Dive; Data Centre Magazine, 2025). Modular and prefabricated construction reduces material waste and on-site (A5) emissions, and right-sized structural design avoids the over-engineering that bloats both concrete and steel tonnage.

The fork, and its downstream cost: low-carbon materials carry a green premium and, for novel options like CLT, structural-loading and fire/permitting questions that interact badly with the density ramp. A liquid-cooled AI hall must bear ~3,000-5,000 lb wet racks (Chapter 1.1); a structural system 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 — a one-way door that strands future capacity. The building-shell decision must be made jointly with the density-ramp basis, never against it.

The second half of embodied carbon lives in the IT and infrastructure stack, and it behaves nothing like the shell. Where the building is poured once and stands for decades, the IT stack is re-manufactured every 18-36 months as accelerators refresh (Chapter 14.9). Its embodied carbon is therefore not a one-time A1-A3 charge but a recurring B-module replacement charge — and at AI refresh velocity, that recurrence is the single most distinctive feature of an AI data center's carbon profile. Advanced-node silicon is carbon-intensive to fabricate (sub-7nm logic, HBM stacks, advanced packaging all push fab energy and process-gas emissions up), 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 reframes where the IT embodied carbon actually sits: across the full server, the host system (CPUs, DRAM, storage, board, chassis) often dominates manufacturing embodied carbon, while the GPU dominates operational carbon (arXiv cradle-to-grave studies, 2025). The downstream consequence is sharp: optimizing only GPU efficiency misses a large slice of the embodied opportunity. A whole-server, whole-rack accounting — not a GPU-centric one — is required to find the real levers, and 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.

Here is the fork that defines embodied carbon in the AI era, and it is a genuine conflict between two goods. Each new accelerator generation delivers a large operational-efficiency gain — more useful FLOPs per watt, less energy per token. From a pure operational-carbon view, refreshing fast looks green: 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 carbon-intensive manufacturing charge, and the less useful work each unit of embodied carbon amortizes over.

The economic engine that drives fast refresh is exactly the one analyzed in Chapter 1.8 and Chapter 14.9: accelerators have a 2-3 year intense-duty economic life against a 5-6 year book life, and a generation that is ~2x more performant per dollar makes the old hardware uncompetitive long before it is worn out. That same compression that protects compute margins is the engine of recurring embodied carbon. The two pressures point in opposite directions, and there is no universally correct answer — only a fork to be decided per workload:

• Refresh-fast (efficiency-led): rotate to the newest accelerator on a 18-24 month cadence, capture the per-token energy win, and rely on a strong secondary market to keep the displaced hardware in service elsewhere — so the embodied carbon is 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 — viable where the workload is interruption-tolerant batch inference or where grid carbon is already low enough that operational penalty is small.

The decisive variable that collapses 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 — which is why the two cannot be decided separately.

Circularity is the only lever that reduces embodied carbon after the bill of materials is fixed, because it changes the denominator — more useful work per unit of manufacturing carbon — rather than the numerator. 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 — DIMMs, NICs, drives, 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.

The numbers make the case. Refurbishment saves up to ~85% of the energy required to manufacture new equipment and costs 55-80% less than buying new (Human-I-T; DCD circular-economy analyses, 2025). Hyperscalers have operationalized this: Microsoft reported a 90.9% reuse-and-recycle rate for servers and components in FY2024 via its Circular Centers, hitting its 2025 target a year early; Google has resold tens of millions of components into the secondary market. Recycled copper — central to data-center power and interconnect — carries far lower embodied carbon than primary copper, and domestic secondary-refining capacity is growing specifically to feed this demand (Electronics360, 2025).

The end-of-life fork is where circularity either happens or fails. Decommissioning that routes assets through certified ITAD (IT asset disposition) — with NIST 800-88 media sanitization and R2v3 / e-Stewards / ISO 14001-certified processors — recovers 35-50% resale value within ~45 days and keeps the asset in the circular loop (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 — UN Global E-Waste Monitor) and converts retained embodied carbon into a Scope 3 disposal liability. The difference between these two outcomes is a procurement decision made years earlier: whether residual-value, take-back, and certified-ITAD clauses were written into the original purchase.

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

• Decarbonizing power, declaring victory. Signing 24/7 CFE and reaching a near-zero CUE while running a 2-year refresh with no circularity program. On a clean grid this addresses the minority of the lifecycle footprint and leaves the majority — recurring embodied carbon — entirely un-abated.
• Fast refresh without a reuse path. Chasing every accelerator generation for the operational-efficiency win while scrapping the displaced hardware. This is the embodied-carbon worst case: manufacturing re-paid every ~2 years, amortized over almost no useful work, then dumped into the e-waste stream as a Scope 3 liability.
• Embodied carbon discovered at disclosure time. Waiting until a CSRD/ESRS or EU EED filing forces the number, then reconstructing it from generic factors because no EPDs were required at procurement. The result is an indefensible, un-auditable figure and a structure already poured — every high-leverage shell lever spent years ago. → Chapter 15.7.