Local AI is a smaller footprint by design.

Why cloud-AI infrastructure is an environmental story

Modern AI agents — the kind most businesses use today — run on hyperscale data centers operated by a handful of cloud providers. Those facilities are responsible for an increasingly significant share of global electricity use, water consumption, and carbon emissions. The shift to generative AI is accelerating that growth.

1. Electricity demand

Data centers globally consumed an estimated 240–340 TWh of electricity in 2022 — about 1–1.3% of all electricity used worldwide, not counting cryptocurrency mining.^[3] The International Energy Agency's Electricity 2024 outlook expects data-center, AI, and crypto consumption to roughly double by 2026, reaching over 1,000 TWh — comparable to the annual electricity consumption of Japan.^[1]

Generative AI workloads are particularly intense. As MIT researchers note, “a generative AI training cluster might consume seven or eight times more energy than a typical computing workload.”^[2] Power requirements for North American data centers nearly doubled in a single year — from 2,688 MW at the end of 2022 to 5,341 MW at the end of 2023 — driven significantly by generative AI demand.^[2]

The strain is geographically uneven. In Ireland, data centers already account for 18% of national electricity consumption, and could reach 28% by 2031 unless generation capacity expands.^[3] Denmark's data-center electricity use is projected to rise six-fold by 2030, reaching nearly 15% of national demand.^[3]

2. Water consumption

Cooling those servers requires staggering amounts of fresh water. MIT's 2025 analysis estimates that roughly two liters of water are needed for every kilowatt-hour a data center consumes — for evaporative cooling that prevents the hardware from overheating.^[2] When data centers are sited in drought-stressed regions, that draw competes directly with municipal supply and local ecosystems.

“Just because this is called ‘cloud computing’ doesn't mean the hardware lives in the cloud. Data centers are present in our physical world, and because of their water usage they have direct and indirect implications for biodiversity.”
— Noman Bashir, MIT Climate & Sustainability Consortium^[2]

3. Carbon emissions

Data centers and data-transmission networks together account for around 330 Mt CO₂-equivalent per year — roughly 0.9% of energy-related greenhouse gas emissions.^[3] Training a single large model carries a measurable footprint of its own: the training of GPT-3 alone was estimated to consume about 1,287 MWh of electricity and produce ~552 tons of CO₂ — equivalent to powering roughly 120 average U.S. homes for a year.^[2]

And because new models are released every few weeks, much of that training energy is effectively discarded as each generation replaces the last.^[2] Meanwhile, 60–70% of total ML energy use is now attributable to inference — the day-to-day querying of models — meaning the cost compounds with every prompt your team sends to the cloud.^[3]

4. Embodied impact of the hardware itself

The GPUs that make modern AI possible carry their own environmental ledger. Manufacturing a GPU requires more energy than a comparable CPU, and the raw materials involve mining operations and toxic processing chemicals. In 2023, NVIDIA, AMD, and Intel collectively shipped ~3.85 million GPUs to data centers, up from 2.67 million the year before — a number expected to keep climbing.^[2]

Where Sanctum AI fits

We're not going to claim a small startup can rewrite the energy economics of AI. But the design choice we've made — running smaller, focused models on hardware our customers already own or buy locally — has real environmental consequences:

• No hyperscale dependency. Sanctum products don't add load to data centers in drought regions or to grids approaching their limits.
• Right-sized compute. The model your accounting team uses doesn't need to be a 1-trillion-parameter generalist. Smaller, fine-tuned models use a fraction of the energy per inference.
• One-time hardware, long-lived. On-prem inference machines amortize their embodied carbon across years of use, rather than fueling a treadmill of cloud capacity expansion.
• Waste-heat reuse is possible. When the compute is in your building, the heat it produces can offset your heating load instead of being dumped into the atmosphere from a remote facility.

Local AI isn't a silver bullet. It's a meaningful alternative to defaulting every workflow to a cloud model — and for organizations that already have privacy reasons to keep their data home, it's a sustainability win that costs nothing extra.