CONNECT WITH US

Generative AI faces bottlenecks from limited data, rising resource demands, and copper shortages

Chiang, Jen-Chieh, Taipei; Jingyue Hsiao, DIGITIMES Asia 0

Credit: AFP

Generative AI requires ample resources, yet challenges such as limited high-quality data, rising freshwater and electricity consumption in data centers, and a widening copper supply gap may emerge as critical bottlenecks to AI progress.

Decline in high-quality data fuels synthetic data demand

A CB Insights report reveals that by 2026, the availability of high-quality data for training large language models may fall short, pushing developers to rely on costly licensed data or synthetic data—such as artificial text and images. Since 2022, nearly 50 vendors have entered the synthetic training data market, with around 30 companies securing equity funding.

However, the entrance of major tech players like Microsoft and Meta into the synthetic data sector has slowed funding for startups. Only five funding rounds took place in 2024, a sharp decline from 21 in 2022, and some companies faced workforce cuts. Israel-based Datagen, founded in 2018, declared bankruptcy in 2024.

Despite this, certain startups are expanding. Italy-based Aindo generates synthetic data for finance, healthcare, and administration, while UK-based Synthesized offers table and text synthetic data for training AI models, including datasets tailored for fraud detection in financial institutions.

How much power does generative AI require?

The growing adoption of AI applications is significantly raising electricity consumption in data centers, with AI-powered Google searches reportedly requiring up to ten times more energy than standard searches.

According to the International Energy Agency (IEA), electricity demand from data centers, cryptocurrency, and AI-related activities reached 460TWh in 2022 and is expected to increase to 620–1,050TWh by 2026. The IEA estimates that by 2026, AI-driven industries may consume up to ten times the power used in 2023.

Technological innovations could help curb this energy demand. The Central Research Institute of Electric Power Industry (CRIEPI) notes that with advancements in energy-saving technologies, data center power demand growth might be kept under two times current levels, limiting the overall rise in electricity demand to around 1%. However, the commercial feasibility of these power-saving technologies remains uncertain.

Intensive water consumption poses risks for AI development

The rising water demands of data centers, driven largely by cooling requirements, underscore growing concerns about water scarcity. Cooling systems in data centers typically involve either cooling towers, which use water evaporation to dissipate heat, or chillers, which rely on refrigerant compression and electricity. Some data centers incorporate ambient air for natural cooling when possible.

Microsoft reported a 23% increase in total water consumption in 2023, totaling 7.844 million cubic meters, primarily due to data center expansion. For instance, training OpenAI's GPT-3 model (with 175 billion parameters) required 700,000 liters of water for server cooling at Microsoft's latest US data center. Factoring in water used for electricity generation and hardware production, the total demand for GPT-3's training reached 5.4 million liters.

As GPT-4 and subsequent models emerge, water usage is expected to climb further. GPT-3 is estimated to use 500 ml of water every 10–50 queries. By 2027, water demand for cooling and power generation in data centers could reach 4.2–6.6 billion cubic meters annually, equivalent to half of the UK's yearly water consumption.

Power demand drives copper needs

The copper supply shortage poses a growing threat to AI development, as demand for copper has soared with the expansion of renewable energy, electric vehicles, and generative AI.

JPMorgan Chase & Co. projects that by 2030, AI data centers alone could require approximately 2.6 million tons of copper annually. Each additional megawatt of data center power capacity may demand 20–40 tons of copper, given the IEA's forecast of a 15% annual increase in power consumption.

The copper supply gap is expected to reach 4 million tons per year by 2030, with AI demand adding another 2.6 million tons to this shortfall. Yet, efficiency gains might help manage copper demand.

According to a May 2024 report from Macquarie Investment Bank, annual copper demand growth could be limited to 200,000 tons through 2030, significantly lower than other forecasts. This outlook assumes a copper usage rate of 27 tons per megawatt for data centers, consistent with JPMorgan's estimates.

Macquarie notes that advancements in energy-saving technology, such as Google's reduction in data center power usage to one-fifth of 2009 levels, could temper the rise in power and copper demand for AI infrastructure.