LOCAL-FIRST BY DESIGN · ANIMAFLOW’S ENVIRONMENTAL APPROACH 
 AI doesn’t have to mean cloud. 
 Existing production workstations. Already-trained models. A fair comparison based on the work completed. 
 
 Why ANIMAFLOW chose a local-first approach 
 ANIMAFLOW was conceived as an alternative to relying on multiple cloud AI services. Giving studios control over their tools and data—and addressing concerns about the environmental cost of that reliance—were part of that founding choice. 
 In a local-inference deployment, ANIMAFLOW can run downloaded, already-trained models on the same workstations a studio uses for animation, 2D graphics, modelling and 3D production. Those operations do not require an external AI inference service or a newly purchased, dedicated AI workstation. Studios choose the deployment and services appropriate to their needs. 
 Training, local inference and cloud services are different 
 Downloading and using an existing model does not repeat its original training. Training has already incurred an environmental cost; subsequent use is a separate stage. Research on AI life cycles distinguishes training, inference and hardware-related impacts. [1] 
 We cannot undo the footprint of a model’s original training. We can choose how we use it next: reuse existing models, avoid unnecessary retraining and avoid sending locally executable tasks to external AI services. This is a choice about how to proceed, not a claim that local processing erases emissions or is always more efficient than cloud processing. 
 Using AI does not automatically mean using a remote AI data centre. For locally executed tasks, ANIMAFLOW belongs to the studio’s existing production-software ecosystem. Hosting a website or other services externally is a separate deployment decision. 
 35–50% production-time savings in the activities evaluated 
 The ANIMAFLOW team reports 35–50% production-time savings in the activities evaluated . This is a team-reported productivity result, not a universal performance guarantee or an independently certified environmental result. We do not translate that percentage into an equivalent reduction in electricity use or emissions. The scientific sources below concern AI’s environmental impact generally; they do not independently validate ANIMAFLOW’s productivity figure. 
 Compare the same completed production work 
 The relevant question is: what resources are needed to complete the same work, at equivalent quality, with and without ANIMAFLOW? This is the same practical question a studio can ask about rendering, simulations, render farms and other production tools. 
 Existing, operational workstations mean that each installation need not entail new hardware. Their being switched on does not make additional processing impact-free. Equally, automation may replace repeated steps or reduce overall machine time. Whether the total footprint rises or falls depends on the work actually added or replaced. 
 Research finds substantial differences in energy use across AI tasks, models and execution configurations. An “AI” label alone does not establish the footprint of a particular workflow. [2] [3] 
 Our position: informed choices, supportable claims 
 We do not claim that ANIMAFLOW has zero environmental impact or a measured percentage reduction in emissions. We offer a local-first way to use existing models within production infrastructure studios already operate. Its environmental performance should be assessed from the actual deployment and completed work, rather than automatically assigning it the impact of new model training or cloud AI infrastructure that the local configuration does not require. 
 
 Research & sources 
 These publications support the distinctions between training, inference and workload-dependent energy use. They are not environmental certifications of ANIMAFLOW. 
 
 Luccioni, Viguier & Ligozat. Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model (2022 preprint). An AI life-cycle assessment distinguishing training, deployment and hardware impacts. 
 Luccioni, Jernite & Strubell. Power Hungry Processing: Watts Driving the Cost of AI Deployment? ACM FAccT (2024). Empirical comparison of energy use across inference tasks and models. 
 Fernandez et al. Energy Considerations of Large Language Model Inference and Efficiency Optimizations . ACL (2025). Effects of workload, software and hardware on inference energy use. 
 
 Statement updated 24 October 2026. Product-origin and productivity statements are provided by the ANIMAFLOW team. 
 
  
 
 

Sources:
https://arxiv.org/abs/2211.02001
https://facctconference.org/static/papers24/facct24-6.pdf
https://aclanthology.org/2025.acl-long.1563/