We're living through what the World Economic Forum calls the Fourth Industrial Revolution - a wave of AI, robotics, and digitalization that is "fundamentally changing the way we live, work and relate to one another."
Like all previous industrial revolutions this one isn't just transforming human life it's reshaping the Earth system itself. Rising global temperatures, accelerating ice loss, sea-level rise, more frequent wildfires, shifting weather patterns, and biodiversity loss are all linked to human-driven greenhouse gas emissions.
AI is now a visible part of that story. Every new model, image generator, video model, coding agent, and AI-powered product ultimately translates into demand for chips, servers, data centres, electricity, cooling, grid connections, and raw materials. The question isn't whether we should use AI - it's how we can do it in a way that doesn't blow through our remaining carbon budget.
SaladCloud's model - using GPU capacity that already exists in computers around the world - represents one possible part of that answer.
There is already a vast pool of under-utilized hardware in the wild - the challenge is orchestrating it safely and efficiently.
Instead of assuming every new unit of AI demand needs a new GPU in a new data centre, we can also ask a simpler question: How much more useful work can we get out of the hardware we have already manufactured?
The Two Sides of AI's Carbon Footprint: Embodied vs. Operational
When people talk about AI's environmental impact, the discussion usually starts with electricity. That is only half the story.
The carbon footprint of computing can broadly be split into two categories: "embodied" - emissions that come from everything that happens before you ever run a model: mining and refining metals, manufacturing chips and servers, shipping them around the world, constructing and eventually decommissioning data centers and "operational" emissions - everything tied to electricity use while the system runs: training and fine-tuning models, inference, cooling, networking, UPS systems, etc.
Recent life-cycle assessments of AI hardware show that both sides are significant. Manufacturing advanced GPUs and building data centres consumes large amounts of energy and materials, with impacts such as mineral depletion and human toxicity, while running those GPUs for AI workloads drives ongoing operational emissions tied to the local grid mix. arXiv
In other words, every time we spin up another rack of high-end GPUs in a new facility, we're effectively paying twice: once in embodied carbon, and then again in operational carbon for as long as we keep that hardware running.
Both of these emission streams are growing as more data centres come online and existing facilities devote a larger share of their capacity to power-hungry AI applications like generative LLMs, image generation, and other GPU-intensive workloads.
How Much Power Does AI Use?
Scientists estimate that the power requirements of data centers in North America nearly doubled in a single year, rising from 2,688 megawatts at the end of 2022 to 5,341 megawatts by the end of 2023, driven in part by the demands of generative AI. According to an analysis by MIT on the environmental impact of generative AI, global data centers consumed around 460 terawatt-hours of electricity in 2022-enough to make them the 11th largest electricity consumer in the world, between Saudi Arabia (371 TWh) and France (463 TWh). (MIT)
More recent data shows that this growth is continuing. According to the IEA's 2026 Key Questions on Energy and AI report, global data-centre electricity consumption reached around 485 TWh in 2025 and is projected to roughly double to 950 TWh by 2030, accounting for around 3% of global electricity demand. Electricity consumption from AI-focused data centres grew even faster - 50% in 2025 alone - and the IEA projects it to triple between 2025 and 2030.
Once a model is trained, the energy demands don't stop. Each time someone uses the model-say, asking a chatbot to summarize an email-the hardware consumes energy to process that request. The IEA estimates that energy use per individual AI task has been falling rapidly as hardware and software become more efficient. But newer use cases such as video generation, reasoning models and AI agents can consume hundreds or even thousands of times more energy per query than simple text generation. So while AI is getting more efficient, the amount and complexity of AI compute we use is growing at the same time.
And that's just the frontier models from a handful of tech giants. The long tail is enormous: startups, enterprises, and researchers now train and use millions of smaller models. Hugging Face alone hosts over 2 million models, and that ecosystem keeps growing.
Greener by Design: How the Industry Is Chasing Greener Compute
The industry isn't ignoring the problem. Cloud providers, chip makers, and data-center operators are all under pressure from customers, investors, and regulators to curb emissions, and they're experimenting with a whole toolbox of efficiency measures.
A 2025 review of green cloud computing finds that techniques like server virtualization, dynamic resource allocation, and more efficient cooling can cut data-center energy use and associated carbon emissions by up to roughly 50% when they're applied systematically.
A lot of this comes down to doing more work on fewer machines:
-
Server consolidation and virtualization let multiple virtual machines share a single physical server, improving utilization and reducing the need for extra hardware.
-
Containerization packages applications and their dependencies into lightweight containers, making it easier to pack workloads tightly and move them to where capacity is available.
-
Dynamic resource allocation and autoscaling spin resources up and down with demand, so clusters don't sit over-provisioned and idle. TechRxiv
On the facilities side, operators are investing heavily in:
-
Advanced cooling (hot-aisle/cold-aisle containment, liquid cooling, free-air cooling) to cut the energy spent on keeping servers within safe temperature ranges. ScienceDirect
-
Energy-efficient hardware and accelerators, tuned for better performance per watt.
-
Waste-heat reuse, where the heat from servers is captured and used to warm nearby buildings or feed district heating networks. MDPI
-
Carbon- and energy-aware workload scheduling, which shifts flexible jobs to locations or times with cleaner electricity. ACM Digital Library
Major cloud providers have also set net-zero and 100% renewable-energy targets, signing long-term power-purchase agreements and siting new facilities in regions with abundant wind, solar, or hydro. ABI Research
Governments are paying attention too. In the EU, for example, the revised Energy Efficiency Directive introduced reporting requirements for large data centres covering metrics like energy consumption, water use and energy reuse, while the EU Code of Conduct on Data Centre Energy Efficiency provides best-practice guidance for improving efficiency. Similar conversations are happening in the US and other regions as data-centre loads begin to show up in national energy and climate planning. Pew Research Center
All of this is necessary-but it still focuses primarily on new and existing datacentre infrastructure. At the same time, we're surrounded by an absurd amount of unused compute. Around 90% of the world's compute resources - more than 400 million consumer GPUs - sit idle for 20–22 hours a day.
The History of Reusing Idle Compute
Reusing idle compute has long been one of the most elegant ways to get more work out of the hardware we already have.
The concept itself isn't new. The BOINC platform (Berkeley Open Infrastructure for Network Computing) has been doing this for science for more than two decades. BOINC's founder, David Anderson, defines volunteer computing as using consumer digital devices for high-throughput scientific computing, harnessing millions of PCs and smartphones to create a virtual supercomputer at very low cost.
In parallel, grid-computing research popularized the idea of "CPU scavenging": forming a compute grid from the idle cycles of participant machines - typically at night, during breaks, or in the tiny gaps while a device waits on I/O.
The history of these projects really comes down to two key ideas:
-
Idle resources exist at massive scale.
-
With the right middleware and scheduling, those idle resources can perform serious work.
What SaladCloud does is take these proven ideas and bring them into the GPU era.
Why Running on Idle SaladCloud GPUs Can Lower Carbon per Job
The strongest sustainability argument for distributed GPU networks isn't that consumer GPUs somehow use no electricity. They obviously do.
The argument is about utilization and avoided infrastructure.
When you run a workload on SaladCloud's idle GPU network instead of building or expanding a dedicated cluster, you're doing three potentially beneficial things from a sustainability perspective:
-
Reusing existing hardware instead of manufacturing new GPUs
-
Life-cycle assessments show that manufacturing GPUs and building data centers carry heavy impacts in categories like mineral depletion, human toxicity, and construction emissions.
-
Every time a workload runs on an already-manufactured GPU, the embodied environmental cost of that device is spread over more useful work.
-
-
Raising utilization of idle devices
-
Many consumer GPUs sit cold for most of the day. Across hundreds of millions of machines, consumer GPUs can be idle 20–22 hours per day.
-
Putting some of that idle time to work improves the ratio of "useful compute" to "baseline energy draw"-especially when devices would otherwise be on for other reasons.
-
-
Avoiding under-utilized new capacity
-
By shifting workloads to a pool of existing, preemptible GPUs instead of building out a new dedicated cluster, you can reduce the need for additional hardware and the under-utilization that often comes with it.
-
In a distributed network like SaladCloud, flexible workloads can also be scheduled where and when capacity is available, further reducing waste.
-
None of this magically makes idle-GPU compute "carbon-free" - electricity still has a footprint - but it can reduce the carbon per unit of work compared with building new capacity that runs at low utilization or sits idle waiting for peak traffic.
How You Can Make AI Workloads Greener with Salad
There are also several decisions developers can make regardless of where their workloads run:
-
Use the smallest model and GPU that actually meet the requirement. Bigger isn't automatically better. Quantization, optimized inference engines and newer model architectures can often produce the same practical result with dramatically less compute.
-
Reuse and fine-tune instead of training from scratch when possible. Training remains necessary, but many applications can start from an existing model and fine-tune or adapt it instead of repeating expensive foundational training.
-
Batch work and keep GPUs busy. A GPU sitting allocated at 5% utilization still consumes power and money. Batching requests, increasing concurrency and scaling replicas with actual demand improves both economics and energy efficiency.
-
Measure useful work, not just GPU hours. Track throughput, GPU utilization, latency and output per dollar or per GPU-hour. Performance optimization and sustainability often turn out to be the same engineering problem.
Those techniques don't just reduce environmental impact.
They usually save money as well.
Conclusion
Reusing idle GPUs does not automatically mean zero additional carbon.
What the literature and practical experience do support is a more nuanced statement:
If we reuse already-manufactured GPUs, increase utilization of idle devices, and apply energy-aware scheduling, we can reduce the carbon intensity per unit of compute compared with building and running under-utilized dedicated infrastructure - especially for flexible, preemptible workloads.
SaladCloud's model - aggregating idle GPUs worldwide and exposing them as a managed, container-based cloud - is one concrete path toward that "sustainability synergy": using what the world already has instead of endlessly building more.