Power Usage By AI: DeepSeek The 'Green' Alternative To OpenAI, Meta AI, Grok?

The rapid rise of artificial intelligence (AI) has been a game-changer across industries, but with it comes a pressing concern: the escalating energy demands of training and running AI models. As AI models like OpenAI's ChatGPT and DeepSeek-V3 from China gain in prominence, the energy consumption required to support their operations is starting to create significant strain on the global energy sector.

With the US and China both intensifying their AI ambitions, the question is, can the energy sector sustain this growth, and does DeepSeek offer a lower-energy alternative?

There is also a pertaining question of whether DeepSeek's latest free chatbot app, which was based on a new model called DeepSeek-V3, is on par with the multibillion-dollar platforms like that of ChatGPT, Google's Gemini, Elon Musk's Grok or Meta AI.

AI's insatiable hunger for energy is becoming increasingly evident. As AI models continue to evolve, their data centre requirements have surged. "If you look at data centres for cloud workloads versus data centres for AI workloads, data centers for AI workloads require ultra-high density. They need a concentrated power source, and so intermittent power like wind and solar doesn’t fit the bill," said Jared Cohen, President of Goldman Sachs Global Affairs, to CNBC.

According to Cohen, AI demands base load power—power sources that are continuous and stable, like nuclear, coal, and natural gas.

This poses a significant challenge for the US, which is expected to consume between 220 to 275 TWh of energy by 2026 to power its AI operations, as per the World Economic Forum.

To put this in perspective, 275 TWh is enough to power 2.75 crore homes in India for an entire year assuming the annual energy consumption of an Indian household is 1,000 kWh.

The 24/7 nature of AI systems drives this massive energy requirement. It implies that when you put an AI model in function it never switches itself off, it is always learning, always moving.

The environmental toll of AI, due to the high power consumption, is now clear. "Microsoft, which has invested in ChatGPT maker OpenAI and has positioned generative AI tools at the heart of its product offering, recently announced its CO2 emissions had risen nearly 30% since 2020 due to data centre expansion," the World Economic Forum reports.

Google’s greenhouse gas emissions in 2023 were also significantly higher, up nearly 50% from 2019, driven largely by the energy consumption of data centres.

Amid the growing energy concerns in the western economies, China's DeepSeek could possibly emerge as a game-changing AI model that uses significantly less energy compared to Western counterparts like ChatGPT.

"Across the industry, the increasing energy demand, primarily from building and running the data centres used to train and operate AI models, is contributing to global greenhouse gas (GHG) emissions", says The World Economic Forum.

One of the critical factors behind DeepSeek’s energy efficiency is its use of less-advanced chips, which bypass the US-imposed restrictions on semiconductor exports to China.

As per its own research paper, DeepSeek-V3 required 2,788,000 GPU hours for training, translating into 836,400 kWh of energy.

"The world’s top companies typically train their chatbots with supercomputers that use as many as 16,000 chips or more. DeepSeek’s engineers said they needed only about 2,000 Nvidia chips," reported The New York Times.

If DeepSeek has found a way to use fewer chips to make a model as efficient as existing platforms, it would likely mean less energy is needed to build and run these models. This could make DeepSeek more compatible with green energy sources.

As energy consumption in the AI sector skyrockets, the push for renewable energy sources is gaining momentum. The dilemma, however, is whether AI companies can truly rely on intermittent power sources like wind and solar. According to Cohen, for AI workloads to be efficient and sustainable, "nuclear, coal, and natural gas" will likely remain crucial.