Tired of Bitcoin's power consumption? Research: AI electricity consumption will surpass BTC mining by the end of 2025.

Bitcoin's worries about excessive power consumption seem to have passed, because the more annoying power-eating behemoth is AI. (Synopsis: Pakistan announces 2,000 megawatts of electricity for "Bitcoin Mining and AI Center" Prime Minister appoints Special Assistant for Blockchain and Cryptocurrencies) (Background supplement: Blue and white joint proposal to amend the "nuclear power extension" cap for 20 years, fight to end non-nuclear homes) Artificial intelligence AI's power consumption is growing at an alarming rate, and is expected to surpass the electricity consumption of Bitcoin mining and nearly half of the total power consumption of data centers worldwide by the end of 2025 at the earliest. This grim warning comes from Dr. Alex de Vries-Gao, the Institute for the Environment at Vrije Universiteit Amsterdam, the Netherlands, in his latest study, Artificial intelligence: Supply chain constraints and energy implications, published in the journal Joule. AI consumes far more energy than expected Research data points out that between 2023 and 2024 alone, the cumulative thermal design power consumption of AI accelerator modules produced by NVIDIA and AMD has reached a total (TDP) of 3.8 million (GW). This figure is equivalent to the electricity consumption of the whole of Ireland for an entire year. If all AI hardware (CoWoS) using TSMC's advanced packaging technology is taken into account, such as Google's TPU, and the total power consumption of the entire AI system composed of auxiliary equipment such as heat dissipation, the power demand of AI systems deployed worldwide is estimated to have reached an astonishing scale of 5.3 to 9.4 GW. Dr. Alex de Vries-Gao further predicted in his study: "If TSMC doubles its CoWoS capacity again in 2025 as scheduled, the cumulative power demand for AI systems produced may reach 23 GW by the end of this year. This will not only surpass the electricity consumption of Bitcoin mining, but will also be close to half of the total electricity consumption of non-cryptocurrency data centers in 2024." The remarks highlight the potential pressure of AI development on global electricity supply. In fact, TSMC's CoWoS packaging technology capacity has been widely regarded as a major bottleneck in the AI accelerator manufacturing process. According to statistics in the study, TSMC's CoWoS production capacity has increased from about 126,500 (300mm wafers in 2023 ) to about 327,400 wafers in 2024, an increase of more than double. Market analysts estimate that NVIDIA and AMD together account for 52% of TSMC's CoWoS capacity (2023 ) and 58% ( )2024, of which NVIDIA alone accounted for 44% respectively ( in )2023 and 48% ( )2024, showing the huge demand in the AI hardware market. The tug-of-war for energy efficiency The reason why AI energy demand has soared is mainly because the industry generally holds the concept of "larger models perform better", which directly leads to the increasing scale and complexity of AI models, and the increase in energy consumption. Although some models have improved efficiency, such as the Chinese DeepSeek R1 model, which claims to reduce dependence on high-end hardware, the continuous expansion of model size often easily offsets these efficiency gains, and may even produce a so-called "rebound effect", so that the overall energy consumption does not decrease but increases. At the same time, tech giants generally maintain a low level of transparency about AI energy consumption data, such as Google, which provided data between 2019 and 2021, showing that machine learning accounted for about 10% to 15% of its total energy use at that time. It then stopped disclosing such data, claiming that it "doesn't make sense" to distinguish the energy consumption of AI from other workloads. Although the EU AI Bill requires disclosure of energy consumption during the training phase of general-purpose AI models, there is no mandatory requirement for the inference (Inference) stage, which accounts for a greater proportion of power consumption, and the relevant regulations will not officially take effect until August 2025 at the earliest. Alex de Vries-Gao emphasized: "The lack of transparency is one of our biggest challenges. Without accurate data, policymakers and researchers cannot fully assess the environmental impact of AI and develop an effective regulatory framework." Grid pressures and environmental concerns This rapid growth in energy demand is posing serious challenges to the grids of every advanced country. Data centers are built at a ( typically take only 1 to 2 years ) much faster than the planning and construction cycle of the grid ( often taking 5 to 10 years ), a time lag that has led to the so-called "power capacity crisis." Even more worrying for environmentalists is that new AI data centers could further deepen reliance on fossil fuels, such as a previous partnership between OpenAI/StarGate and Crusoe Energy to provide up to 4.5 GW of natural gas generation capacity for data centers. The United States is the world's largest consumer of data center electricity, accounting for about 45% of the global total, followed by China (25% ) and Europe (15% ). At present, the goal of AI sprint in the United States seems to put aside the hidden worries caused by power consumption for the time being; Related AI data such as China's electric power is even less clear. Related reports Europe's largest power company ENEL will "tokenize solar panels" to buy coins to offset electricity bills, RWA new application Ethiopian power company "18% of revenue comes from bitcoin mining", low-cost hydro mining enterprise new paradise 〈Think bitcoin consumes electricity? Research: AI power consumption surpasses BTC mining by the end of 2025 at the earliest" This article was first published in BlockTempo's "Dynamic Trend - The Most Influential Blockchain News Media".