Creating a robust large language model, such as ChatGPT, demands a deep analysis of vast quantities of human-generated text. Yet, this computational endeavor comes with a significant environmental cost. Data centers, responsible for this massive computation, require substantial electricity and efficient cooling systems, often involving the use of water to mitigate the generated heat.
Intriguingly, water resources from the Raccoon and Des Moines river watersheds in central Iowa played an indispensable role in OpenAI’s quest, supported by Microsoft, to nurture its AI systems to emulate human writing.
The true expense of constructing AI marvels like ChatGPT often remains obscured. However, it’s clear that the race to capitalize on the fervor surrounding generative AI has amplified the ecological footprint of leading tech giants. Microsoft, OpenAI, and Google, among others, have tacitly acknowledged the escalating costs of fulfilling the burgeoning demand for their AI tools, extending from the pricy semiconductor hardware to augmented water consumption.
Despite this, the specifics of these costs have often been shrouded in secrecy. The birthplace of OpenAI’s advanced large language model, GPT-4, was virtually unknown to most until a high-ranking Microsoft executive disclosed its existence, nestled “next to cornfields west of Des Moines.”
Building a substantial language model necessitates extensive pattern analysis of human text, consuming substantial electricity and producing copious amounts of heat. To regulate temperature during sweltering days, data centers frequently employ water cooling systems, often utilizing external cooling towers adjacent to their colossal facilities.
Microsoft recently revealed in its environmental report that its global water consumption surged by a staggering 34% from 2021 to 2022, reaching nearly 1.7 billion gallons, equivalent to more than 2,500 Olympic-sized swimming pools. This abrupt surge is primarily attributed to the company’s AI research endeavors, particularly its collaboration with OpenAI on generative AI.
In an upcoming research paper, scientists estimate that ChatGPT consumes approximately 500 milliliters of water, roughly equivalent to the volume of a 16-ounce water bottle, each time it responds to a series of 5 to 50 prompts or queries. The actual consumption varies based on server locations and seasonal factors, encompassing indirect water usage, like cooling power plants supplying electricity to data centers, which often goes unmeasured.
Regrettably, many remain unaware of the resource intensity underlying ChatGPT’s operation, hindering any meaningful conservation efforts. Google, too, reported a 20% surge in water use, closely linked to its AI ventures. This surge was uneven, with notable increases in certain regions, shedding light on the profound influence of AI on water consumption.
In response to inquiries, Microsoft affirmed its commitment to investigating AI’s energy and carbon footprint while striving to enhance the efficiency of large systems, both in training and application. The company emphasizes its pledge to monitor emissions, expedite progress, incorporate clean energy in data center operations, and pursue sustainability goals, including carbon negativity, water positivity, and zero waste by 2030.
OpenAI echoed similar sentiments, highlighting the energy and water-intensive nature of training large models and its ongoing efforts to bolster efficiency.
Microsoft’s partnership with OpenAI, dating back to a $1 billion investment in 2019, laid the groundwork for the development of ChatGPT. Microsoft supplied the necessary computing power for training AI models, leading to the creation of GPT-4, which now powers premium iterations of ChatGPT and several Microsoft products. The revelation of GPT-4’s origin in Iowa underscored the significance of water resources in nurturing advanced AI systems.
Notably, West Des Moines, Iowa, emerged as a strategic location for AI model training, particularly GPT-4, owing to its relatively efficient resource utilization. Its climate permits the use of external air for cooling during most of the year, reserving water usage for times when the temperature exceeds 29.3 degrees Celsius (approximately 85 degrees Fahrenheit).
However, even this efficient approach results in substantial water consumption. For instance, in July 2022, just prior to GPT-4’s training completion, Microsoft channeled around 11.5 million gallons of water into its Iowa data centers, amounting to roughly 6% of the district’s total water consumption, which also serves the city’s residents.
Acknowledging these concerns, the West Des Moines Water Works and the city government expressed a willingness to consider future data center projects from Microsoft only if they demonstrate technology capable of significantly reducing peak water usage. This precautionary measure aims to safeguard water resources for residential and commercial needs.
Microsoft has affirmed its cooperation with the water works to address these concerns, emphasizing its commitment to sustainability.
In conclusion, while the advancement of generative AI technologies like ChatGPT holds immense promise, it also brings to light the vital need for responsible resource management and environmental stewardship to ensure a sustainable future.
Source AP
