Climate impacts of AI hardware manufacturing rival those of data centers

Technical Summary: Climate impacts of AI hardware manufacturing rival those of data centers

Problem Statement
Current discourse regarding the climate impact of Artificial Intelligence (AI) predominantly focuses on the electricity consumption and operational greenhouse gas (GHG) emissions of data centers. However, this focus overlooks the upstream emissions embedded in the manufacturing of advanced AI hardware. As AI accelerators require increasingly complex components—such as high-bandwidth memory (HBM), advanced-node logic, and advanced packaging—the production processes have become highly energy- and emissions-intensive. Furthermore, the reliance on potent fluorinated process gases and the concentration of manufacturing in regions with high grid carbon intensity suggest that the manufacturing supply chain may constitute a significant, yet under-quantified, portion of AI's total climate footprint. Existing assessments often fail to distinguish between the physical electricity burden of the grid and the procurement-adjusted emissions reported by corporations, leading to an incomplete understanding of the system-level emissions.

Methodology
The authors compiled a facility- and corporate-level annual GHG emissions inventory for 2024, covering two distinct subsystems: AI hardware manufacturing and AI data center operations. The study employs a dual-accounting framework to compare emissions under two distinct methodologies:

1. Location-based accounting: Uses the average emission intensity of the local grid where the facility is physically located.
2. Market-based accounting: Incorporates contractual instruments such as renewable electricity procurement, power-purchase agreements (PPAs), and electricity attribute certificates (EACs) to reflect reported Scope 2 emissions.

For AI data center operations, the study utilizes a facility-level modeling framework linking 100 facilities (15.0 GW of power capacity) to local grid factors and company-disclosed market-based factors.
For AI hardware manufacturing, the study maps AI-relevant hardware modules (based on product carbon footprints of the Dell XE9680L and NVIDIA H100) to 61 manufacturing sites (approximately 4.2 million 300-mm-equivalent wafer starts per month) across five major entities: TSMC, Intel, Samsung Electronics DS, Micron, and SK hynix. The manufacturing emissions are allocated based on module shares, supplier-evidence mapping, and coverage modifiers derived from corporate disclosures.

Key Contributions

• System-Level Inventory: The paper provides the first comprehensive, side-by-side comparison of annual GHG emissions from global AI hardware manufacturing and AI data center operations for the year 2024.
• Dual-Accounting Framework: It extends AI emissions accounting beyond simple electricity use by explicitly separating physical grid exposure from procurement-adjusted reported emissions across the entire infrastructure system.
• Module-Level Decomposition: The study decomposes manufacturing emissions by hardware module, identifying High-Bandwidth Memory (HBM) and accelerator cores as the dominant contributors to the manufacturing carbon footprint.
• Supply Chain Transparency: It highlights the disparity in renewable electricity adoption between data center operators (who often utilize aggressive procurement strategies) and hardware manufacturers (who remain concentrated in GHG-intensive grids with lower renewable coverage).

Results

• Emission Hierarchy Reversal: Under location-based accounting, AI data center operations emitted 19.6–21.3 Mt CO2e yr⁻¹, significantly exceeding the 6.5–8.2 Mt CO2e yr⁻¹ from hardware manufacturing. However, under market-based accounting, operational emissions dropped to 5.5–6.0 Mt CO2e yr⁻¹ (a 72% reduction), while manufacturing emissions only declined to 5.2–6.6 Mt CO2e yr⁻¹ (a 20% reduction). Consequently, under market-based accounting, hardware manufacturing emissions are comparable to, and in the upper-bound case, exceed, operational emissions.
• Procurement Gap: The divergence is driven by asymmetric reductions in disclosed emissions. Major data center operators have aggressively adopted onsite renewables and market-based instruments, reducing their reported intensity to 93 g CO2e kWh⁻¹. In contrast, hardware manufacturing hubs (primarily in Taiwan, South Korea, and China) retain a high market-based intensity of 363 g CO2e kWh⁻¹, nearly four times that of operations.
• Module Drivers: HBM is identified as the single largest module-level contributor, accounting for 32–33% of total manufacturing emissions. The concentration of HBM production among a few suppliers (Samsung, SK hynix, Micron) with varying levels of renewable electricity coverage (ranging from ~8% to 30%) creates a structural bottleneck for decarbonization.
• Corporate Disparities: Among data center operators, emissions rankings shift significantly under market-based accounting, with facilities lacking explicit 2030 decarbonization pathways remaining prominent. In manufacturing, the geographic distribution of emissions remains largely unchanged under market-based accounting, as supplier procurement adjustments do not significantly redistribute the physical manufacturing geography.

Significance
The paper argues that understanding AI's climate footprint requires looking beyond data center electricity use to include hardware manufacturing and corporate decarbonization commitments. The findings indicate that while data center operators have successfully reduced their reported operational emissions through procurement strategies, the physical climate burden of the AI system is increasingly embedded in the upstream supply chain. As frontier AI hardware becomes more carbon-intensive (e.g., the shift from H100 to B200 architectures increases embodied carbon by 73%), and as HBM capacity expands, the manufacturing sector represents a persistent and growing source of emissions. The study concludes that effective AI emissions governance must treat data center operations and hardware manufacturing as coupled components, necessitating a system-level disclosure framework that links facility location, electricity procurement, and hardware supply chains to distinguish physical grid impacts from procurement-adjusted claims.