Over recent years there has been a tremendous amount written on the economics of the cloud, primarily in our industry looking at the migration of traditional media technology to cloud applications. Much of the complexity comes from trying to accurately model the costs of the traditional approach, such as real estate, power, cooling, infrastructure, to compare with the ongoing subscription of a cloud service. As the market matures, companies are increasingly able to make informed decisions on this direct use of cloud in their businesses, but there is also a trend of increasing indirect cloud usage, which is less well understood.
In many areas of our lives, we are using cloud more and more, often not even aware of how or when we are using it. Take a simple example of the photos and videos on your phone; is this all on local storage, or have you exceeded the capacity of your device? Most phones automatically and seamlessly use cloud storage for media synchronising with the local device as content is accessed. Of course this is initially free, but then charges for storage can follow, and how many people have a clear understanding of this cloud usage?
Given the increased use of cloud in so many areas of our lives and businesses, an important question to ask is just how green is it? The cloud providers will argue that it provides economies of scale and efficiency in terms of the previously mentioned overheads like power and cooling. That is a reasonable argument when we are considering replacing X with Y, but there is one area in particular where cloud usage is soaring and it is incremental usage rather than replacing traditional compute, this is AI.
Much like the examples of photos on a phone, the impact or carbon cost of using an AI service is almost entirely invisible to the user. Typing a query into a web page and getting a text response doesn’t give the appearance of high computing demand in the same way that processing video might, but some of the data being published paints a different picture. One estimate for training the GPT-3 large language model is that it consumed 1,287 megawatt-hours of electricity and generated 502 tons of carbon dioxide emissions. Of course, it’s not just OpenAI, there are many services being rolled out. Google’s emissions have increased 48 per cent in five years, partly due to the energy demands of AI. Before continuing, I should point out that I can’t credit the above-mentioned studies or data because they come from an AI summary of a web search. This was automatically done without me asking for it, incurring additional cloud compute/power consumption, which goes back to the point of the energy/carbon cost not being apparent to the user.
Doing this small amount of research leads me to think that there needs to be much more transparency on the carbon impact of using these services. Placing my used plastic in the recycling bin feels a little futile if I then sit at my keyboard asking AI random questions and burn a gigawatt of electricity! It has also led me to question how any of this can be free. If these AI systems consume so much power, then they cost a lot to run, and I don’t buy into the altruistic corporate mission statements that claim they are doing this to help humanity! Ultimately, we will be paying for these ‘free’ services, with the payments falling into two main categories.
The first is that we are paying with our time and knowledge; every time we submit information to an AI system, we are helping train the model it operates on. There is a lot of debate on what this means for intellectual property, from Hollywood actors’ images and voices being duplicated, through to engineers using AI to check their source code, which potentially then effectively makes that code open source because the AI can reproduce it for another user.
The second model is where a free service is provided to build a dependency on that service over time, as it becomes essential to the user, charges and subscriptions can be introduced. The phone memory model at the start of the article is an innocuous example, but the extrapolation is potentially concerning. In the latest series of Black Mirror on Netflix there is an episode where a woman’s life is saved with brain surgery, but the solution requires a medical device in her brain which is linked to a cloud subscription service. Over time, the costs increase, and advertising is introduced for a cheaper subscription—interesting that Netflix allowed that! Of course an extreme fictional example, but a lesson that we need to be careful about what we subscribe to and what the real cost is, both personally and for the planet.
