AI done wrong could land us in a climate catastrophe. So far, regulation such as the EU AI Act has focused on ethics and privacy, ignoring the glaring impact on the planet. Sustainable AI needs low-carbon, water-efficient infrastructure and clear policy.
With traditional, centralised AI, data is sent to data centres, which drive enormous energy, water and e-waste demands. AI system power is already approaching that of a country the size of the UK. This is estimated to soar to nearly triple the combined annual electricity use of Pakistan, Bangladesh and Nigeria annually by 2030.
In a blog last autumn, we predicted that if AI destroys humanity, it will sooner be from overheating the planet than from robots outsmarting us. The future of AI’s impact on the environment is largely a political decision: there’s still time to build AI that benefits humanity without causing irreparable damage to the planet. For this, alternative methods are necessary.
Distributed inference is the greener, more scalable way forward. Instead of shipping raw data across long-distance networks to data centres, AI moves closer to where data is generated. Edge nodes process it locally on distributed hardware. Couple this with federated learning for the training phase, and we have a promising solution to AI’s part in the climate crisis.
In this blog we dig into why centralised AI could be antithetical to net zero, then explain how FLock.io is helping governments, firms and organisations to navigate this.
Centralised AI is an impending climate catastrophe
AI adds unprecedented demand for electricity at a time when fossil fuels still provide over 60% of total global electricity generation .
Data centres are springing up around the world. The US has approximately 4,400 to 4,500 active data centres – mainly in Virginia – representing the largest data centre market in the world. The UK is the runner up with 500 to 560 active data centres, followed by Germany, France and China. They are run by cloud providers and tech giants like Amazon or Microsoft, but are used by AI startups too.
ChatGPT requires a staggering amount of computational power to operate, consuming vast amounts of energy. Every time you prompt ChatGPT to generate an image or draft an email, the host company’s servers run thousands of calculations. A single AI search consumes about 10 times the energy of a standard Google search.
Emissions are from two main processes: training and inference. Training an LLM requires massive computing power over several weeks or months. Inference is the everyday processing of use prompts and generating responses. Day-to-day usage (inference) accounts for roughly 80–90% of total energy demand.
However, energy estimates are complicated by data centre operators not publicly disclosing the required inputs. This lack of transparency poses a significant challenge for accurately assessing the carbon footprint of proprietary models (unlike open source models, of which the carbon footprint can easily be calculated ).
To keep the servers from overheating, water systems are often used to cool them. But England faces a shortfall of five billion litres of water daily by 2050, even without allowing for data centre growth according to the Guardian.
According to analysis by the LSE, data centres are a growing target of global climate-related legal cases. It found a growing number of cases challenging the energy sources, water consumption and air pollution of datacentres, all of which have related climate implications.
Distributed inference and FL have lower carbon emissions
A more sustainable AI ecosystem can be built through distributed inference and federated learning on edge computing. Federated learning (FL) uses 80% less training energy per model update. By moving AI training away from centralised, energy-intensive data centres and toward distributed inference, we can continue to innovate while also protecting our planet.
It has begun – for example, Orange is carrying out a working proof-of-concept using an MQTT server as a federated learning coordinator, vehicles training local PyTorch models and syncing to a global model, framed around reducing energy costs.
Edge computing started in the 1990s with networks of distributed servers to move any compute task closer to the data source, to reduce latency and bandwidth use. Applying the concept to AI, we have distributed inference: splitting a single AI model across multiple smaller devices or local nodes to share the processing load, rather than relying on a single distant cloud. This lowers bandwidth use and cuts out network travel time for faster, real-time responses.
Back in 2021, Cambridge University researchers carried out the first ever systematic study of the carbon footprint of FL. They discovered that it had a significantly greener impact, with lower carbon emissions than traditional machine learning. They also made available a first-of-its-kind “Federated Learning Carbon Calculator” so that the public and other researchers can estimate how much CO2 is produced by any given pool of devices.
One of the researchers said: “Although smartphones have much less processing power than the hardware accelerators used in data centres, they don’t require as much cooling power as the accelerators do. That’s the benefit of distributing the training of models across a wide pool of devices.”
FLock.io is the key technical partner in a government initiative to develop sovereign AI for Sarawak, the largest state in Malaysia, including for healthcare and the civil service. The initial project showed that FL can collaboratively train a model without sharing the data, and also that distributed inference enables a large model to run efficiently on smaller GPUs. FLock.io won an award from the Ministry of Energy & Environmental Sustainability (MEESty) of Sarawak.
Federated learning is also more scalable and better for data privacy
As well as being greener, federated learning protects data privacy by keeping sensitive raw data local on devices or private networks. Data privacy regulations and security concerns restrict the use of AI by regulated industries holding sensitive data, including hospitals, banks and governmental agencies. It forces organisations to either forgo AI adoption or rely on generic models that lack domain accuracy or introduce compliance risk.
Conventional approaches – such as centralised cloud-based AI training and on-premises model deployment – typically require significant computational resources. They cannot guarantee robust privacy protection or protection against model poisoning attacks and data leaks, and can compromise model accuracy.
FLock.io’s solution allows collaborative AI model training without sharing raw data. Each participant trains the model locally and securely on-premises or on edge devices. They share only encrypted model updates, which are then aggregated to improve the model’s performance, enabling real-time inference.
More about FLock
FLock is an AI research and infrastructure company pioneering enterprise-grade federated learning and distributed AI solutions that prioritise data privacy. Its decentralised federated learning architecture and production-ready platforms (AI Arena, FL Alliance, and FLock API Platform) enable organisations to train and deploy their own custom AI models on local hardware while maintaining full data privacy, model ownership, and regulatory alignment by design.
FLock effectively combines FL and blockchain-based verification for a 37% improvement in model accuracy, a 44% reduction in total cost ownership, a reduced risk of data breaches or model poisoning attacks and a 63% shorter deployment time.
FLock was recently spotlighted by the World Economic Forum (WEF)’s MINDS programme for its privacy-preserving AI training for the NHS. It showcases use cases in eye disease detection and diabetes management that keep patient data secure, and Moorfields Eye Hospital and UCLH using FLock’s federated learning platform to train clinical AI models while maintaining 100% data sovereignty.
Follow FLock on LinkedIn and X and email the team at hello@flock.io.
