ENVIRONMENTAL IMPACT
As an AI-powered application, we believe you have a right to know what our infrastructure costs the planet. This page is our attempt at an honest answer.
OUR INFRASTRUCTURE
Civitas runs entirely on a single Raspberry Pi 5 — a credit-card-sized computer with a 4-core ARM Cortex-A76 processor drawing 3–11 watts depending on load. There is no cloud provider, no managed Kubernetes cluster, no CDN replication across dozens of edge nodes.
ENERGY USE
Measured with a USB-C inline power meter, the Pi 5 draws approximately 3 W at idle and peaks around 10–11 W during the nightly pipeline. We use 7 W as our stated operating average — conservative but defensible across idle periods, web serving, and inference runs.
For context: a typical US refrigerator uses 400–500 kWh/year. Civitas consumes roughly one-seventh the energy of keeping your food cold.
CARBON FOOTPRINT
We calculate our carbon footprint using the EPA eGRID 2024 national average grid intensity of 0.384 kg CO₂e per kWh. We do not know the exact energy mix of our ISP or the local grid — we use the national average as the honest, conservative estimate.
WHY LOCAL AI MATTERS
Every AI application makes an infrastructure choice. We chose to run a small language model locally on low-power ARM hardware rather than sending queries to cloud AI APIs. Here is what that means in practice.
| METRIC | CLOUD AI (GPT-4o) | CIVITAS (LOCAL) |
|---|---|---|
| Hardware power draw | ~700 W per H100 GPU | ~10 W (Pi 5) |
| Energy per inference | ~0.30 Wh | ~0.15 Wh |
| Data center overhead (PUE) | 1.15× multiplier | None |
| Network round-trip to GPU | Yes (internet) | No (localhost) |
| Query data leaves device | Yes | No |
| Pipeline AI energy / year | ~55 kWh (est.) | included in 61 kWh total |
ChatGPT serves ~700 million users weekly and is estimated to consume 340+ MWh of electricity per day for inference alone. Civitas's entire annual energy budget — 61 kWh — is consumed by ChatGPT in roughly 11 seconds.
DESIGN CHOICES THAT REDUCE IMPACT
NO PERSISTENT USER DATA
Civitas stores no user accounts, behavioral profiles, or tracking data. No analytics databases, no session replication, no data pipelines churning user signals 24/7.
NO CDN EDGE REPLICATION
A typical production web app replicates assets across 30–100+ CDN edge nodes globally. Civitas serves content from a single machine. If you're in Seoul, you're hitting the Pi. That's a deliberate trade-off.
ARM EFFICIENCY OVER X86 THROUGHPUT
The Cortex-A76 core in the Pi 5 runs at a TDP of ~3W per core. A comparable Intel Xeon core at the same process node consumes 5–10× more. ARM's RISC architecture is inherently more efficient for the steady-state serving workloads Civitas runs.
SINGLE-DEVICE CONSOLIDATION
Web server, API, database, and LLM inference all run on one physical device drawing 7W total. A cloud-equivalent architecture typically involves 3–5 separate service instances plus managed databases, with a combined idle baseline of 50–100W before data center overhead.
LOCAL INFERENCE = NO DATA CENTER TRANSIT
Each cloud AI query traverses public internet routing, BGP peering, CDN edges, and data center networking before reaching a GPU. Local llama.cpp inference happens entirely on-device — the energy cost of network transit is zero.
WHAT WE CANNOT CLAIM
Transparency requires acknowledging what we don't know and what we haven't measured.
Our ISP's energy mix is unknown. The 7W figure covers the Pi only. Modem, router, upstream fiber infrastructure, and internet backbone all consume power we cannot attribute or measure. Networking infrastructure adds an estimated 0.06 kWh per GB transferred (IEA), but that cost is diffuse and shared.
Accessories are not counted. The real system idle — including router, modem, and power supply losses — is likely 12–20W, not 7W.
We don't hold renewable energy certificates. Without a PPA or REC, we use the US national average grid intensity. If the Pi is on a cleaner regional grid (e.g., Pacific Northwest hydro at ~0.08 kg CO₂e/kWh), the real footprint could be 4× lower — but we don't claim that.
Embodied carbon is not included. Manufacturing a Raspberry Pi 5 has a carbon cost. Raspberry Pi Ltd. does not publish lifecycle assessment data. We estimate 10–30 kg CO₂e for manufacturing based on comparable devices, amortized over a 5-year lifespan.
Inference energy varies by model and query length. Longer prompts, larger models, and unquantized weights all increase energy per inference. Our 0.15 Wh estimate assumes a quantized sub-7B model on typical short civic queries.
DATA SOURCES
EPA eGRID 2024 — US grid average carbon intensity (0.384 kg CO₂e/kWh)
UK DESNZ 2025 — UK grid emissions factor (0.177 kg CO₂e/kWh)
Epoch AI (2024) — GPT-4o energy per inference (~0.30 Wh)
raspberry.tips (2026) — Pi 5 measured power consumption (USB-C inline meter)
Arxiv 2511.07425 — LLM inference on single-board computers (Pi 5 ~10W under load)
Amazon Sustainability Report 2024 — AWS global data center PUE of 1.15
IEA (2023) — Network energy intensity estimate (0.06 kWh/GB)
EPA (2024) — Passenger vehicle CO₂ per mile (~0.25 kg CO₂/mile average)
Energy and carbon figures are estimates based on published data. We update this page when better measurements become available. Last reviewed: July 2026.