As major American tech companies like Microsoft, Google, Amazon, and Meta are heavily investing in artificial intelligence (AI), it’s expected that the pharmaceutical and defense industries will start seeing returns on these investments next year. However, securing the necessary power for AI learning and development is expected to be a key challenge.
According to a report by The Economist, one of the most significant AI innovations by 2025 might emerge from drug development or the defense industry. Drugs that began development using AI after the release of ChatGPT in November 2022 are anticipated to enter phase 3 clinical trials next year. Additionally, AI is expected to be integrated into drones, which are rapidly becoming a core component of future weapon systems.
Currently, AI adoption across industries and companies is generally occurring at the individual employee level rather than through top-down mandates. Only 5% of U.S. companies report using AI in their products or services. However, research indicates that about one-third of employees in U.S. firms use AI at least once a week. For software engineers, this usage is at 78%, a significant increase from 40% last year, while utilization in HR departments has surged from 35% to 75%.
The Economist notes that identifying the best ways to leverage AI technology involves reevaluating processes and retraining workers, which takes time. Meanwhile, the information technology (IT) sector sees one in five businesses already integrating AI, suggesting that as AI technology becomes more sophisticated, its adoption could accelerate.
A major barrier to swift AI investment returns is the need to prioritize AI learning and development to gain market leadership, especially since flagship models from companies like OpenAI, Anthropic, and Google show little performance disparity, with Meta, Mistral, and xAI rapidly catching up.
As AI development progresses, obtaining the necessary data becomes challenging, with predictions that high-quality data from the internet may be exhausted by 2028. Companies are therefore exploring ways to generate synthetic data for AI training.
The most significant variable remains power supply. Training OpenAI’s massive language model GPT-3 required enough electricity to power 100 households for a year, but GPT-4 required 50 times more power, sufficient for 5,000 households for a year. This power demand means additional semiconductor purchases are necessary for AI learning and development.
Big tech companies are pouring significant funds into building data centers for this reason. It is estimated that training next-generation AI models alone could cost around $1 billion. AI data center expenditures from 2024 to 2027 are expected to exceed $1.4 trillion. Already, the capital expenditures of MS, Alphabet, Amazon, and Meta in the first half of this year have reached $106 billion. These companies have hinted at increased capital expenditures in the future.
According to CNBC, the power required to operate data centers might surpass the total consumption of the cities or even states where they are located, making it increasingly difficult to find adequate power supply and suitable land, potentially delaying carbon emission reduction goals.