New Data Shows Massive Input Costs of AI

New data shows the degree to which the AI boom requires input costs

On 15 May 2024, Bloomberg published the story “Microsoft’s AI Push Imperils Climate Goal as Carbon Emissions Jump 30%.” Now, before I lose you, this report and the data from Microsoft are focused on the climate impact of Artificial Intelligence and Machine Learning (AI/ML). Today, I am interested in what contributes to these carbon emissions and what information we can draw from them. What is contributing to these emissions is a massive increase in energy usage for data centers running AI/ML, which is also contributing to a greater demand for more data centers.

The report shows that Microsoft’s carbon emissions have jumped by 30% since 2020 despite its goal to become carbon negative by 2030. These increases, especially the spike in 2030, can be directly tied to the increased need for energy and data-processing power, which risks having a detrimental impact on the global economy.

From Bloomberg article.

Artificial Intelligence Stripped of the Hype

Amid the ongoing hype of AI, we should remember that true AI does not yet exist, and the vast majority of what exists and is used is better described as ML. Once we boil down AI/ML models to their base elements, it can be crudely understood as linear algebra on a massive scale. With each prompt or use of an AI/ML model, the algebraic formulas are processed based on the inputs provided and/or trained upon. These inputs constitute massive amounts of data. This highlights at least two primary inputs for us:

• Compute Power

• Data

You need lots of energy for the processing power required to run AI/ML, and you need to be able to store the data for the AI/ML. It will mean hosting large amounts of data in data centers, especially considering large language models or other large-general AI/ML models. Previous AI/ML models trained on open-source information on the Internet are increasingly coming under fire for copyright infringements and plagiarism. This contributes to some developers turning to closed data that has been “cleaned” to run it in their model. The need for these also means a greater demand for microchips for computing power (as if the demand for microchips could get any higher) and potentially an increased demand for construction supplies for data centers. It is difficult to foresee the overall need for more energy and data centers due to the ongoing hype and almost yearly new developments in AI/ML.

So What?

This will get much worse before it gets better.

Unless efficiencies in energy use in processing power and data storage are found, there is a risk that the demands of the AI boom will lead to increased costs for energy, processors, data storage, microchips, and much more. Countries already recognize a greater need for more microchip production and computing and data processing efficiencies. Still, the scale at which AI/ML adoption is occurring is likely to outpace efforts for efficiencies. This may increase the costs of these goods overall, which could result from shortages that don’t meet demand. The bottom line to this is that there is a risk of SMEs being costed out of access to AI/ML, putting them at a distinct disadvantage.

Recommendations

• Private Sector: Conduct a cost assessment and market analysis of any costs related to AI/ML inputs and their potential impact on your operating costs. Most importantly, check your contracts with your existing providers, as there is a risk that this will begin to impact the industry in the next year or two. Understanding where you are and your business needs can better prepare you in the event market conditions worsen.

• Government: The Government of Canada has said it is investing in computing power to help fuel AI/ML, which is potentially great for SMEs. This and much more were announced in Securing Canada’s AI Advantage, Canada’s $2.4 billion investment. While the specifics of the investments could be nitpicked, the effort and types of investment were welcomed by industry and should be applauded. What does make me worry is that the government does not have a good history of handling information technology-related programs.

  • The Government should review available computing power and data centers and associated costs across Canada by province. This review should identify gaps in coverage, economic trends in computing and data processing availability, and the impact of availability on the economy.
    

• Military: This will likely affect you the most. When dealing with secret data, this will have to be hosted in secure facilities and servers. I do not think Canada’s Department of National Defence and Canadian Armed Forces have the capacity, willingness, or money to build a data center to handle their needs. Treasury Board policy will likely require some change to potentially authorize third-party facilities which meet security criteria. However, the policy related to this is also closely connected to that which is holding back the procurement of ITI in SP of C2, also known as the Canadian Armed Force’s Secret Cloud environment.

Note: This article does not suggest that environmental impacts are less important than economic ones but that they are intrinsically connected.