Environmental Impact of Ubiquitous Generative AI

Author:Murphy | View: 29804 | Time: 2025-03-23 18:06:43

This article ponders the question: What would be the environmental impact of large-scale adoption of generative AI like ChatGPT? That is, what might the environmental impact be if billions of people began to use generative AI extensively on a daily basis?

The reason this question is interesting to contemplate is that we can use its answer to inform how worried we should or shouldn't be about the speedy adoption of this new technology.

As AI models have grown larger and larger [1] and as they have been made widely accessible by companies like OpenAI and Google, the environmental impact of AI models – eg. carbon and water footprint – have become the subject of inquiry and debate. First in academia (e.g. [2] and [3]) and later in mainstream media (e.g. [4] and [5]).

With ChatGPT reportedly having hundreds of millions of users – if not billions [6] – and Google embedding generative AI into several products [7], generative AI is arguably the most widely adopted type of AI at the moment. Combined with the immense size of generative AI models like GPT-4 – rumored to be almost 6 times larger than its predecessor [8] – generative AI is likely also the type of AI to have the largest environmental impact for the foreseeable future.

This article is a thought experiment that contemplates what the environmental impact might be of large-scale adoption of generative AI. Will it lead to environmental catastrophe, will it be a drop in the ocean or somewhere in between? The purpose of this article is to provide a basis for beginning to shed light on that question.

Many assumptions went into making the estimates presented in this article, and if you'd like to play around with your own assumptions, you can do so in this spreadsheet.

If you'd like to get full access to my stories on the environmental impact of AI, become a Medium member through the link below:

Join Medium with my referral link – Kasper Groes Albin Ludvigsen

Stages in an AI system's life cycle

Even though this article does not analyze one specific model, it's informative to distinguish between different stages of an AI model's life cycle. We can think of an AI model's life cycle as consisting of 6 distinct stages [9]:

Raw material extraction
Materials manufacturing
Hardware manufacturing
Model training
Model deployment
End-of-life

In this article, I'll focus on hardware manufacturing (stage 3), model training (stage 4) and model deployment (stage 5) and will therefore briefly describe these stages below.

Hardware manufacturing refers to the environmental impact from manufacturing the hardware on which the AI model runs. Model training is the stage in which the model is developed. Model deployment is the stage in which the model is "deployed" to a "production environment" where it can be used by users. This is also sometimes referred to as the inference stage or the production stage. The life cycle is often depicted as linear although many AI systems require their models to be re-trained or adjusted during the system's lifetime.

To estimate the environmental impact of either of the 3 above-mentioned stages, we need to get an idea of how much hardware large-scale adoption of generative AI would require. This is what we'll consider in the following section.

ChatGPT's Electricity Consumption

How much hardware does large-scale generative AI adoption require?

To assess the potential environmental impact of large-scale adoption of generative AI, we need to know how much hardware would be required to handle billions of daily queries.

To figure out how much hardware is needed, we need to think about how many users the technology will have and how much they'll use it. The more users, the more hardware is needed.

So, what would large-scale generative AI adoption look like in terms of user numbers? Let's assume that 3.5B people start to use ChatGPT or similar technology daily and they make 30 queries per day. That's a total of 105B daily requests. In lieu of ChatGPT's staggering user numbers and efforts by Google and other companies to integrate generative AI into various products, this shouldn't be an unreasonable assumption.

Now we need to get a sense of what kind and how much hardware it takes to handle 100B daily requests.

Patel and Ahmad have previously estimated that it takes around 3,617 Nvidia HGX A100 servers containing 28,936 Nvidia A100 GPUs to handle 195,000,000 daily ChatGPT requests [10]. The A100 GPU is a piece of processing hardware designed for AI workloads. Let's assume that those numbers are in the right ballpark and that they will generalize to other generative AI services. Let's further assume that the number of GPUs increases linearly with the number of daily requests. This means that if 3,617 HGX servers can handle 195,000,000 daily requests, we need 538.46x more compute – ie 1,947,615 Nvidia HGX A100 servers with a total of 15,580,923 A100 GPUs – to handle 105B daily requests.

Now that we have an idea of how much hardware is needed to support large-scale adoption of generative AI, let's look at the environmental impact of manufacturing it.

Environmental impact of large-scale generative AI adoption in the hardware manufacturing stage

In the previous section, we saw that large-scale adoption of generative AI technology may require 1,947,615 Nvidia HGX servers and 15,580,923 Nvidia A100 GPUs. Let's look at the environmental impact from manufacturing this hardware.

Nvidia have not released any information about the carbon footprint of their products, so we'll have to use some proxies here, which means that the numbers we arrive at are highly speculative, so take them with a grain of salt and please challenge them.

The embodied emissions of Hewlett-Packard's ProLiant DL345 Gen10 Plus server is 2,500 kgCO2e, according to the company's own estimate [11]. This the only reasonably similar server for which I've been able to find embodied emissions data, so we'll use that as proxy like Luccioni et al have previously done [9].

The ProLiant server does not contain any GPUs, so let's add the embodied emissions of 8 A100 GPUs. Again, Nvidia have not disclosed this, but 150 kgCO2e per GPU has been used by others [9] [12].

We're assuming that the Nvidia HGX with 8 GPU slots is used, so let's add 8 * 150 kgCO2e to the 2,500 kgCO2e. That's a total of 3,700 kgCO2e per Nvidia HGX A100 server.

Recall that we need 1,947,615 of these to handle 105B daily requests. The embodied emissions of the GPU hardware needed to accommodate large-scale adoption of generative AI is thus estimated to be 1,947,615 * 3,7 = 7,206,177 tons CO2e.

Let's spread out those emissions evenly over the life span of the hardware. We'll assume that the hardware has a life span of 5 years after which it is either worn out or replaced by newer technology [13].

Based on this, the carbon footprint of manufacturing the hardware needed for large-scale adoption of generative AI is estimated to be 1,441,235.4 tons CO2e per year.

Get an email whenever Kasper Groes Albin Ludvigsen publishes.

Environmental impact of large-scale generative AI adoption in the training stage

Now, let's consider the environmental impact of training the AI models that will underpin generative AI.

The initial version of ChatGPT was based on the large language model (LLM) called GPT-3.5 which is a version of GPT-3. The newest version of ChatGPT is likely based on OpenAI's newest LLM called GPT-4, but OpenAI have not released any information about GPT-4 that can be used to estimate its training costs. We do however have reliable estimates of GPT-3's energy consumption during training. These come from a paper by Google and UC Berkeley researchers who estimate the energy consumption of training GPT-3 to be 1,287,000 KWh [14]. Let's assume that models trained by other companies are in the same ballpark.

To compute the carbon footprint of consuming 1,287,000 KWh, we need to get an idea of how much carbon is emitted when 1 KWh of electricity is produced. This is called the carbon intensity of electricity and varies between regions, because sources of electricity (wind, coal, solar etc.) vary between regions. For this thought experiment, let's use the average carbon intensity of electricity used by Google's, Amazon's, and Microsoft's data centers. Using data from the ML CO2 Impact Calculator [18], the mean carbon intensity of electricity used by these three cloud providers is 484 gCO2e/KWh.

Now, to obtain an estimate of the annual carbon footprint of training generative models, we'll need to know how many companies offer such models and how often they are trained. Let assume that there will be 9 major players in generative AI: OpenAI/Microsoft, Google, Meta, Anthropic, Inflection, Character, Tencent, ByteDance, Baidu. Let's further assume that they train one model each per year. That's an annual training footprint of 6,229 tons CO2e.

The assumptions that go into this part of the article are the most speculative. However, as we'll see, they won't affect the total picture much because the environmental impact of the training stage pales in comparison to the environmental impact of the hardware manufacturing and deployment stages.

Environmental impact of large-scale generative AI adoption in the deployment stage

Let's now consider the environmental impact of the deployment stage of the generative AI models. In other words, let's look at how much electricity it takes to keep the 1,947,615 HGX servers running.

One way to calculate this is to look at the so-called Thermal Design Power (TDP) of the server. TDP is often used to quantify how much power a piece of processing hardware requires to run. The TDP of the HGX server is unknown, but the TDP of a similar Nvidia server, the DGX, is 6.5 kW [15] which we'll assume also applies to the HGX server. So, if the server runs at full power for one hour, it has used 6.5 kilowatt hours (kWh). However, for this thought experiment, let's assume that all the servers run at 75% of their TDP on average, in which case they'll consume 4.875 KWh per hour. That's 1,947,615 servers each consuming 4.875 KWh per hour. That's a total of 9,494,625 KWh per hour, 227,871,000 KWh per day and 83,172,915,000 KWh per year.

When calculating the electricity consumption from data center grade hardware, the electricity consumption of the hardware itself is often multiplied by the so-called Power Usage Effectiveness (PUE) of the data center in which the hardware is running. PUE is a metric used to express the energy efficiency of a data center. The more energy the data center uses on, say, cooling compared to the energy used to power the actual computer hardware, the higher the PUE. Microsoft's global PUE is 1.18 [16] and Google's is 1.10 [17], so let's use the average of these two, i.e. 1.14.

If we multiply the HGX servers' estimated annual electricity consumption of 83,172,915,000 KWh by 1.14, we get an annual electricity consumption of 94,817,123,100 KWh.

Next, to calculate carbon emissions, we'll multiply by the carbon intensity of 484 g/KWh presented in the previous section. Based on this, we can estimate the annual carbon footprint of large-scale generative AI adoption to be 45,891,487 tons in the deployment stage.

Combined environmental impact of ubiquitous adoption of generative AI

Now that we have estimated the carbon footprint of the hardware manufacturing, training and deployment stages of large-scale adoption of generative AI, let's combine those into the total annual carbon footprint of large-scale generative AI adoption.

Recall that we estimated the annual emissions from manufacturing the hardware to be 1,441,235 tons. Then, we estimated the annual CO2e emissions from the training stage to be 6,229 tons. Finally, we estimated the carbon footprint of the deployment stage of large-scale adoption of generative AI to be 45,891,487 tons annually.

So the total carbon footprint of ubiquitous adoption of generative AI can be estimated to be 47,338,952 tons CO2e per year.

Figure 1 below depicts how much larger the carbon emissions from the model deployment stage are compared to the hardware manufacturing and model training stages. Consequently, research should address how to reduce the deployment stage emissions rather than the training stage emissions.

Figure 1: Estimated annual carbon emissions from large-scale adoption of generative AI, by life cycle stage. Chart by Kasper Groes Albin Ludvigsen

Putting the annual carbon footprint of large-scale adoption of generative AI into perspective

Throughout the article, you may have wondered: "What do these numbers actually mean? Is it a lot or not?"

Above, we estimated the total annual carbon footprint of large-scale adoption of generative AI to be 47,338,952 tons CO2e across the hardware manufacturing, model training and model deployment life cycle stages.

That's equivalent to the annual emissions of 4,303,541 Danes. When put like that, it sounds like a lot in my opinion. On the other hand, 4.3M people is just a tiny fraction of the world's population. So, let's also compare the carbon footprint of ubiquitous generative AI to the entire world's annual emissions.

Estimates of global annual CO2e emissions vary between sources, but IAE [19] estimate that in 2021, 40B metric tons of CO2e was emitted globally. Of 40B tons, 47,338,952 tons is 0.12 %. Put differently, if 3.5B people made 30 daily queries to generative AI models like ChatGPT, this article estimates that it could increase global CO2e emissions by 0.12 %. I'll let it be up to the readers to decide if they think that's a lot or not.

Water footprint of ubiquitous generative AI adoption

So far we have analyzed the potential carbon footprint of large-scale adoption of generative AI. But carbon emissions are not the only environmental impact of digital technology. Another important consideration is the water consumption from cooling the data centers which host large AI models. Water consumption refers to water that is lost and cannot be reused.

A recent paper by Pengfei Li et al [20] analyzes the water consumption from large language models like the one underpinning ChatGPT. In their paper, the authors present a methodology for estimating the water consumption from data centers and they estimate that ChatGPT consumes 500 ml of water for every 20–50 queries. Data centers consume water in two primary ways:

direct consumption which is when water evaporates and is flushed as data center hardware is cooled, and
indirect water consumption which accounts for the water being used to produce the electricity that data centers need to run.

Let's assume that the water consumption of ChatGPT will generalize to other AI services and that 50 queries require 500 ml of water, so that's 10 ml of water per query. Recall that we assumed that 105B daily requests will be made to generative AI models. That's 1,050,000,000 liters of water per day, or 383,250,000,000 liters in a year. For comparison, the annual recommended fluid intake for adults is 3.2 liters [21], or an annual intake of 1,168 liters. This means that the water consumption from large scale adoption of generative AI could sustain the annual fluid intake of 328,135,000 adults.

You may wonder why water consumption can be problematic. The main issue is when data centers that consume a lot of water are placed in regions affected by drought. This is an issue because we generally don't have the infrastructure needed to move water across large distances. The paper by Pengfei Li et al presents this map which shows areas of the US affected by drought. Thousands of data centers are located in these areas.

Image from https://arxiv.org/pdf/2304.03271.pdf

Discussion

Let's now discuss some of the potential implications of the estimates presented above.

Can we power large-scale generative AI adoption with renewable energy?

So how might we mitigate the environmental impact of large-scale adoption of generative AI?

How to estimate and reduce the carbon footprint of machine learning models

One idea that springs to mind is to power it all with renewable energy. Let's consider that idea.

In the following, I'll assume that all queries made of generative AI will require additional energy. I.e. I assume the queries won't substitute queries made to other existing services. You might argue that some ChatGPT queries currently replaces traditional search engine queries, but given that generative AI is being built into both Bing and Google, I'd venture that near all generative AI queries will be additional.

In order to make large-scale adoption of generative AI sustainable, the energy would therefore have to be generated by additional renewable energy capacity – i.e. we'd have to install additional renewable energy capacity. Recall that ubiquitous generative AI could require just shy of 95 billion KWh of electricity per year. The average wind turbine can generate 6 million KWh in a year [22].

So to produce enough renewable energy with wind turbines, we would need to install around 15,800 new wind turbines. To put that into perspective, Denmark, a leading nation in wind energy, currently has 6,286 active wind turbines [23]

I therefore think it's safe to say that establishing enough additional renewable energy to power ubiquitous generative AI adoption would be a massive and expensive undertaking.

On a side note, even renewable energy is considered to have a carbon footprint because the emissions caused by producing and installing e.g. a wind turbine are spread out over the life time of the energy source. Thus, electricity from off-shore wind power is considered to have a median carbon intensity of 12 gCO2e/KWh [24]. So even all generative AI was powered by wind energy, it would have an annual carbon footprint of 1,114,000 tons CO2e – roughly the same as 104k Danes.

Do the benefits outweigh the costs?

Another interesting aspect to consider in this debate is whether the productivity gains we can achieve with generative AI justifies its environmental impact. In a yet to be peer reviewed article (i.e. take it with a grain of salt), MIT PhD students Shakked Noy and Whitney Zhang show that using ChatGPT improved both productivity and quality of work for a number of tasks in an experimental setting [25]. Noy and Zhang measured productivity on tasks such as writing press releases, short reports, analysis plans, and delicate emails. Quality is assessed by (blinded) experienced professionals working in the same occupations.

Whether productivity gains are worth their environmental costs is in essence a value judgement, but it would be a good start for the debate if we could ascertain whether generative AI actually does make us more productive. More research should therefore be conducted, and companies using generative AI should critically assess the effects on productivity.

Caveats

The estimates put forward in this article should be considered educated guesswork. This is first and foremost because we're trying to guess what will happen in the future. Secondly, any effort at estimating the environmental impact of these types of AI models is hampered by the fact that the providers don't disclose the necessary information, which means we must make assumptions. By writing this, I hope to inspire others to challenge my assumptions or produce their own estimates.

Although generative AI is an umbrella term for AI products than can generate both text and/or images, this article focuses on models that generate text. The estimates made here do therefore not consider the adoption of image generation technology per se.

ChatGPT in it's original version (based on GPT-3.5) was the point of departure for this article because relevant data is unavailable for OpenAI's latest model, GPT-4. As written above, GPT-4 is larger than GPT-3, which could mean that it consumes more energy. However, it's not necessarily the case.

I assume that the environmental impact of one company's generative AI product will be in the same ballpark as competing products offered by other companies, but it could be the case that some companies will offer smaller or more specialized models.

Conclusion

In this thought experiment we looked at what the environmental impact could be if a large part of the world's population begins to use generative AI like ChatGPT on a daily basis. The purpose of this thought experiment is give the reader a basis for assessing the question: Should we worry about the environmental impact of large-scale adoption of generative AI?

We estimated that ubiquitous generative AI might consume 95B KWh of electricity annually, and producing this amount of electricity could cause emission of 47,338,952 tons CO2e. That's 0.12 % of global CO2e emissions. In other words, this article estimates that if 3.5B people make 30 queries per day to generative AI services, it could increase global CO2e emissions by 0.12 %. Another environmental impact to consider is water consumption. This article estimates that ubiquitous generative AI might consume 383,250,000,000 liters of water in a year. This amount of water is the same as the recommended annual fluid intake of 328,125,000 adults.

That's it! I hope you enjoyed the story. Let me know what you think!

Follow me for more on AI and Sustainability and subscribe to get my stories via email when I publish.

I also sometimes write about time series forecasting.

And feel free to connect on LinkedIn.