What's the environmental impact of Atomi's AI?

It's reasonable to ask about the environmental impact of AI. AI systems use electricity, and the data centres that operate them may also use water for cooling.

No digital service is free of environmental impact. But the scale of that impact is often misunderstood.

Public discussion frequently treats every AI interaction as though it were equivalent to training one of the world's largest AI models. It's not. Training a frontier model and generating a short piece of feedback for a student are radically different computational tasks.

Atomi's AI is deliberately designed to operate at the efficient end of that spectrum.

Training a large AI model can require substantial computing resources. It involves repeatedly processing enormous datasets across large clusters of specialised hardware. Using an already-trained model to produce a response is called inference. A short inference request requires only a tiny fraction of the computation involved in training the original model.

Atomi's AI primarily uses inference to give students brief, targeted feedback on their work. We use relatively small, task-appropriate models running on recent-generation Google TPUs designed for efficient AI inference. Responses are generally concise rather than essay-length.

Together, these choices substantially reduce the computation required for each interaction.

Google has conducted a comprehensive measurement of the median text prompt in its Gemini application. It estimated that a single prompt:

Google compared the electricity required with watching television for less than nine seconds. Its calculation included not only the active AI processor, but also CPUs, memory, idle capacity, cooling and other data-centre overheads.

These figures aren't a direct measurement of Atomi's AI. Different models, prompts, hardware and data centres produce different results, and Google's figures are self-reported estimates for its own systems.

They nevertheless provide a useful indication of scale: an individual text-based AI interaction can have a surprisingly small environmental footprint.

Atomi's interactions are intentionally brief and use smaller, task-appropriate models. We design the system to minimise unnecessary computation, although we don't claim a precise environmental footprint per response without directly measuring the complete Atomi system.

Environmental figures can be difficult to interpret without a familiar comparison.

The Australian Government's Green Vehicle Guide uses 180 grams of carbon dioxide per kilometre as its average combined vehicle benchmark. Actual emissions vary by vehicle and driving conditions.

Using Google's estimate of 0.03 grams for a median text prompt, you'd only have to drive an average vehicle for 17cm to produce more carbon than the average text prompt. That's to say, the choices a student or their family makes about how they get to school each day typically have a much larger impact on their total carbon footprint than their relative use of Atomi's AI.

This isn't a complete lifecycle comparison. The vehicle figure measures tailpipe emissions, while Google reports carbon dioxide equivalent from operating the AI service. Google's measurement is also not a direct measurement of Atomi.

The comparison is intended only to communicate scale. The emissions associated with a short text interaction are orders of magnitude smaller than those associated with many routine activities that society performs without considering them environmentally significant.

The language of water consumption can also be misleading.

Water used by a data centre isn't destroyed. Water that evaporates remains part of the Earth's water cycle and eventually returns as precipitation.

In environmental accounting, however, water is described as "consumed" when it's no longer immediately available to the same local water system. Evaporated water may return somewhere else, at a different time or in a form that isn't immediately available for local use.

This means that the environmental significance of water consumption is principally local.

A litre of water used in a drought-stressed catchment isn't environmentally equivalent to a litre used in a region with an abundant supply. The important questions aren't merely how much water is used, but:

Google states that it evaluates local watershed health when selecting data centre cooling systems. In areas facing high water risk, it considers alternatives such as air cooling and recycled water.

Google has also deployed liquid cooling across the vast majority of its AI servers. These systems circulate coolant through the computing infrastructure efficiently in a closed loop, so no water is consumed.

Again, the individual quantities are extremely small. At Google's reference measurement of 0.26 millilitres per prompt, one litre of water would support approximately 3,800 median text prompts.

To put that into context, Sydney Water estimated network leakage of 132 megalitres per day in 2023–24. At Google's 0.26 mL benchmark, that's numerically equivalent to more than 500 billion prompts.

That doesn't make the notion of water use irrelevant. It does put the volume associated with an individual interaction into perspective.

The timing of electricity use can matter as well as the total amount used.

Most educational use of Atomi occurs during the day. Those hours increasingly coincide with high levels of rooftop and grid-scale solar generation. AEMO reports that growing solar output is producing record levels of renewable generation and driving operational electricity demand to its lowest levels during daylight hours.

This makes the timing profile of educational technology comparatively favourable.

However, data centres draw on wider electricity systems, and renewable-energy procurement is more complex than matching an individual request to the electricity generated at that moment.

Daytime educational use, therefore, has the potential to align well with an abundant renewable supply, but we don't count that as a specific environmental saving without measuring where and how the relevant computation is performed.

Environmental costs shouldn't be dismissed simply because a technology has a worthwhile purpose. We should minimise those costs while also considering what is created in return.

A short interaction with Atomi's AI can give a student immediate, specific feedback at precisely the moment they have misunderstood something. It can help them correct a misconception before it becomes embedded, attempt another problem with greater confidence and receive support when a teacher can't personally respond to every answer in real time.

A piece of feedback may take only seconds to generate. What a student learns from it may remain useful for years.

The broader returns from education are unusually significant. World Bank research estimates that an additional year of education is associated with an average increase in annual earnings of 9–10%, alongside broader benefits not fully captured by earnings alone.

Atomi's AI can't claim responsibility for those outcomes on its own. But they illustrate why environmental impact shouldn't be considered in isolation from educational value.

The relevant question isn't simply 'Does this interaction consume any resources?' as almost every worthwhile human activity does.

The more useful questions are 'Are those resources being used efficiently?' and 'Is the value created proportionate to the cost?'

We seek to minimise the environmental footprint of Atomi's AI by:

Efficiency is also improving rapidly. Google reported that the energy required for its median Gemini text prompt fell by a factor of 33 over a 12-month period while response quality improved.

This is an important part of the broader picture. Today's measurements shouldn't be assumed to represent a permanent environmental cost per interaction. AI models, hardware and data centres are becoming substantially more efficient.

Atomi's AI isn't free of environmental impact. No digital service is.

But the footprint of an individual, short AI interaction is far, far smaller than many people assume. Atomi further reduces that footprint by using smaller models, efficient infrastructure and brief, purposeful responses.

Against that small and increasingly efficient cost is the possibility of helping a student understand something they otherwise may not have understood.

We believe the responsible approach is neither to ignore AI's environmental impact nor to deny students access to useful technology simply because it has some impact.

It's to use AI deliberately, efficiently and where it creates genuine educational value.