Don’t Forget Embodied Carbon of Your AI Gear

The 2025 Green AI Institute white paper warns that embodied carbon from servers and GPUs can rival operational emissions. A Kenyan telco extended server life to five-plus years, bought refurbished edge boxes for non-critical inference, and tracked embodied CO2 alongside kWh. Fewer new imports meant lower capex and a cleaner ESG story without slowing rollouts. They also kept a “reuse first” policy for lab hardware, moving boxes from AI experiments into test environments instead of buying new. Ask vendors for lifecycle data, prefer modular gear you can upgrade, and only import accelerators where workloads truly need them. Track embodied carbon per rack and per project so finance sees the full cost—not just the electricity line. When you retire gear, donate or resell locally to extend its life. Every reuse delays manufacturing a new unit, which is where much of the carbon sits, and that is a line you can proudly explain in board meetings.