Deep breakthrough raises ENA issues from AI

Having broken the hypotheses in the technology sector and beyond the cost of artificial intelligence, the new chatbot of the Chinese startup Deepseek is now doing another industry: energy companies.The company claims to have developed its Open Source R1 model using approximately 2,000 NVIDIA chips, just a fraction of the calculation power deemed generally necessary to form similar programs.

This has significant implications not only for the cost of developing AI, but also the energy of data centers which are the beating heart of growing industry.

The AI ​​revolution has come with hypotheses that IT and energy needs will exponentially increase, resulting in massive technological investments in the two data centers and the means of feeding them, strengthening energy stocks.

The data centers shelter high performance servers and other equipment that operates AI applications.

Could Deepseek therefore represent a less eager way to be able to advance AI?

Investors seemed to think of it, fleeing positions in American energy companies on Monday and helping to cause stock markets already beaten by mass dumping of technological shares. Constellation Energy, which plans to strengthen significant energy capacity for AI, poured more than 20%.

“R1 illustrates the threat that the gains of computer efficiency apply to electricity generators,” wrote Travis Miller, a strategist covering energy and public services for the Morningstar financial services company.

“We always think that the data centers, the reshaping and the theme of the electrification will remain a rear wind,” he added.

But “market expectations have gone too far”.

Nuclear ambitions

Just in 2023, Google, Microsoft and Amazon plowed the equivalent of 0.5% of American GDP in data centers, according to the International Energy Agency (AIE).

Data centers already represent approximately one percent of world electricity consumption and a similar quantity of greenhouse gas emissions linked to energy, according to the IAI.

Improvements of efficiency so far have moderate consumption despite the growth in the data center.

But global electricity consumption of the IAI by data centers could double compared to 2022 figures by next year, to the annual consumption of Japan.

This growing demand is unequally propagated.

Data centers represented approximately 4.4% of American electricity consumption in 2023, a figure that could reach up to 12% by 2028, according to a report commissioned by the US energy department.

Last year, Amazon, Google and Microsoft all concluded offers for nuclear energy, whether from small modular reactors or existing installations.

Meta signed contracts for renewable energies and announced that it was looking for proposals for nuclear energy supplies.

For the moment, however, data centers generally depend on electricity networks which often depend strongly on fossil fuels.

Jeont Paradox strikes again!

Data centers also aspire large quantities of water, both indirectly due to the water involved in electricity production, and directly for use in cooling systems.

“The construction of data centers requires a lot of carbon in the production of steel and also many mining processes and production with high carbon intensity to create the computer equipment to fill them,” said Andrew Lensen, main lecture in intelligence artificial at Victoria University in Wellington.

“So, if Deepseek had to replace models like Openai … There would be a clear decrease in energy needs.”

However, the increase in the effectiveness of technology often results in increased demand – a proposal known as the Jevons paradox.

“Jevons Paradox strikes again!” Microsoft CEO, Satya Nadella, wrote on X on Monday.

“As the AI ​​becomes more efficient and accessible, we will see its use weathered, by transforming it into a product which we simply cannot have enough,” he added.

Lensen also stressed that Deepseek uses a “chain of thoughts” model which is more with a high energy intensity than the alternatives because he uses several steps to respond to a request.

These were previously too expensive to manage, but could now become more popular due to efficiency.

Lensen said that the impact of Deepseek could be to help American companies learn “how they can use calculation efficiency to build even larger and more efficient models”.

“Instead of making their model 10 times smaller and effective with the same level of performance, I think they will use the new results to make their model more capable for the same energy consumption.”