AI may already use more power than Bitcoin — and it threatens Bitcoin mining
AI
BTC
WHEN
REN
READ
The energy required to run artificial intelligence systems may already be higher than the amount of power used to mine Bitcoin, according to new research.
While that may sound like good news for the Bitcoin mining industry, which is regularly attacked over power consumption, the sting in the tail of the report from the Bitcoin Policy Institute is that AI is shaping up as a fierce competitor to Bitcoin for electricity and equipment.
The sectors deep pockets mean AI companies can afford to outbid miners for the same electricity. With AI offering up to 25 times more revenue than Bitcoin per kilowatt hour (kWh), some miners are adding AI processing to their data centers or even switching entirely from Bitcoin to AI.
We will see this trend so long as the revenue per megawatt-hour is higher for AI than Bitcoin, BPI researcher Margot Paez told Cointelegraph.
The AI industry is still in its early stages, but the energy demands of generative AI models is striking. According to Goldman Sachs, a single ChatGPT query consumes nearly 10 times the energy of a typical Google search. MIT Technology Review reports that generating an AI image can use as much power as fully charging a smartphone.
Bitcoin minings high energy use has led to the threat of bans in Europe and moratoriums in New York. According to the BPI, the annual energy usage of US Bitcoin mining facilities is around 121.13 terawatt hours (TWh), while AI consumed between 20 and 125 TWh in 2023 (AI is housed in data centers used for other tasks, making precise figures more difficult).
But with the huge uptake in generative AI this year, the report estimates that AI will use 169 TWh in 2024 and growth will continue to outpace Bitcoin mining, using an estimated 240 TWh in 2027 to minings 160 TWh.
Data centers housing AI models also require significant amounts of water to cool their machinery to maintain efficiency. Research by Shaolei Ren from the University of California Riverside estimates that every five to 50 questions asked to ChatGPT consume approximately 500 milliliters of water.
In comparison, Bitcoin mining in the US alone is estimated to require between 93 and 120 gigalitres of water annually, with each transaction allegedly using enough water to fill a backyard swimming pool (such estimates are controversial with concerns over their potential inaccuracy).
Deep pockets of AI investors pose risk to miners
Profit margins for AI compute are currently much higher than those for Bitcoin mining. Mining generates revenue of $0.17 to $0.20 per kWh, while revenue from the Nvidia graphics processing units used for AI can range from $3 to $5 per kWh, representing a 1725-fold difference.
So why wouldnt Bitcoin miners repurpose their rigs to run AI to make more money?
Bitcoin miner and crypto assets adviser Anibal Garrido told Cointelegraph that its not that easy to make the jump, as miners use application-specific integrated circuit (ASIC) machinery designed solely to calculate the hashes of the PoW protocol, which cant be repurposed for AI.
But Bitcoin miners also need to constantly be updated and replaced and the facilities themselves can be adapted. Paez said that many Bitcoin miners are already modifying their facilities to accommodate GPUs and that she knows of at least one company that has completely transitioned from Bitcoin mining to AI.
Alex de Vries, a data analyst and researcher at Vrije Universiteit Amsterdam and De Nederlandsche Bank, told Cointelegraph that competition for electricity will only intensify.
The pockets of the AI companies are much deeper than those of the crypto mining industry, he said. De Vries believes that AI companies might already be eyeing the power contracts of the Bitcoin miners.
He explained that the AI industry requires immediate access to power and equipment amid the current AI hype and cannot afford to wait for multi-year construction projects to build new data centers meaning the threat to mining is real.
Flexibility vs. fixed
The growing power consumption of AI may help change the politics around the electricity used for Bitcoin mining, which stacks up well by comparison.
Not all power consumption is created equal, and Bitcoin mining rigs are much more flexible and can be shut down or powered up to take advantage of surplus, wasted or cheap electricity.
AI, by contrast, requires 99.9% uptime for the models to function properly. This demand means they absorb all available energy, often regardless of cost, which can lead to the use of less environmentally friendly or hazardous energy sources.
Peaker plants, which are activated to address unexpected spikes in demand often use fossil fuels, exacerbating the environmental impact.
The flexibility offered by Bitcoin mining also allows miners to make deals with governments to ensure they stop consuming energy in case the grid is saturated. Once the grid stabilizes, miners can resume operations, providing the grid with added flexibility to maintain balance.
The BPI report found that US Bitcoin miners shut down operations between 5% and 31% of the time when electricity prices were too high or when directed by grid operators.
The study, which gathered data from eight US mining facilities between July and September 2023, estimated that these interruptions prevented the emission of 13.6 kilotons of CO2. This reduction is equivalent to removing 2,951 cars from the road.
Paez said that as AI is not as flexible as Bitcoin mining, the only way they can manage their emissions is direct investment into renewable energy.
Location-agnostic versus fixed and energy incentives for the industry
Another crucial distinction between the two technologies is in location requirements. Bitcoin mining is location-agnostic, while AI requires low latency to deliver ultra-fast responses, necessitating that data centers be situated near major metropolitan areas.
That means AI data centers must consume whatever energy is available at those specific locations while Bitcoin miners can shift to locations with a surplus of energy, such as renewable energy facilities in remote locations where hydroelectricity, solar or wind is abundant.
Proponents argue the presence of Bitcoin mining can also support the transition to renewable energy by providing predictable demand during otherwise low demand periods, offering financial stability for new projects.
The location-agnostic nature of Bitcoin mining also allows for the utilization of wasted energy. This includes mining with stranded energy from remote hydroelectric power, capturing excess methane emissions, electrifying heating by reusing wasted heat or harnessing renewable energy from solar and wind sources that might otherwise be stranded due to transmission constraints.
Can AI become more efficient?
Recalling the push to use renewable energy for Bitcoin mining, Juan Calvo, senior data engineer and gen AI engineer at Datatonic, believes that AI is duty-bound to become more sustainable:
We need to evaluate whether having the capability to do something justifies its execution, emphasizing the importance of ethical and sustainable choices in technology development.
The engineer explained that AI developers have various techniques to enhance energy efficiency. These include fine-tuning existing models, employing smaller models for specific tasks, and leveraging cloud solutions, which can significantly reduce overall energy consumption.
Hardware advancements can also play a crucial role. Graphics manufacturers like Nvidia are at the forefront of developing specialized hardware that improves performance while consuming less energy. The synergy between more efficient algorithms and advanced hardware can help address AIs increasing energy demands in a more sustainable manner.
However, de Vries pointed out that the bigger is better dynamic in generative AI could undermine these efficiency gains. In a 2023 study, he highlighted that the increasing energy footprint of AI models is driven by an incentive to develop ever-larger models, which in turn escalates the demand for computational resources and energy.
By increasing the efficiency of models and reducing their energy costs, efforts to further improve these models may become more viable, thereby negating some of the efficiency gains.
De Vries compared this dynamic to efficiency gains in cryptocurrency mining hardware. As mining equipment becomes more efficient, Bitcoin miners simply acquire more.
