Having shattered assumptions in the tech sector and beyond about the cost of artificial intelligence, Chinese startup DeepSeek's new chatbot is now roiling another industry: energy companies.

The firm says it developed its open-source R1 model using around 2,000 Nvidia chips, just a fraction of the computing power generally thought necessary to train similar programs.

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

The AI revolution has come with assumptions that computing and energy needs will grow exponentially, resulting in massive tech investments in both data centers and the means to power them, bolstering energy stocks.

Data centers house the high-performance servers and other hardware that make AI applications work.

So might DeepSeek represent a less power-hungry way to advance AI?

Investors seemed to think so, fleeing positions in U.S. energy companies on Monday and helping drag down stock markets already battered by mass dumping of tech shares.

Constellation Energy, which is planning to build significant energy capacity for AI, sank more than 20 percent.

"R1 illustrates the threat that computing efficiency gains pose to power generators," wrote Travis Miller, a strategist covering energy and utilities for financial services firm Morningstar.

"We still believe data centers, reshoring, and the electrification theme will remain a tailwind," he added.

But "market expectations went too far."

In 2023 alone, Google, Microsoft and Amazon plowed the equivalent of 0.5 percent of U.S. GDP into data centers, according to the International Energy Agency (IEA).

Data centers already account for around one percent of global electricity use, and a similar amount of energy-related greenhouse gas emissions, the IEA says.

Efficiency improvements have so far moderated consumption despite growth in data centre demand.

But the IEA projects global electricity use by data centers could double from 2022 figures by next year, to around Japan's annual consumption.

That growing demand is unevenly spread.

Data centers accounted for about 4.4 percent of U.S. electricity consumption in 2023, a figure that could reach up to 12 percent by 2028, according to a report commissioned by the U.S. Department of Energy.

Last year, Amazon, Google and Microsoft all made deals for nuclear energy, either from so-called Small Modular Reactors or existing facilities.

Meta meanwhile has signed contracts for renewable energy and announced it is seeking proposals for nuclear energy supplies.

For now though, data centers generally rely on electricity grids that are often heavily dependent on fossil fuels.

Data centers also suck up significant amounts of water, both indirectly due to the water involved in electricity generation, and directly for use in cooling systems.

"Building data centers requires lots of carbon in the production of steel and also lots of carbon-intensive mining and production processes for creating the computing hardware to fill them," said Andrew Lensen, senior lecturer in artificial intelligence at Victoria University of Wellington.

"So if DeepSeek was to replace models like OpenAI's... there would be a net decrease in energy requirements."

However, increasing efficiency in technology often simply results in increased demand -- a proposition known as the Jevons paradox.

"Jevons paradox strikes again!" Microsoft CEO Satya Nadella wrote on X on Monday.

"As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of," he added.

Lensen also pointed out that DeepSeek uses a "chain-of-thought" model that is more energy-intensive than alternatives because it uses multiple steps to answer a query.

These were previously too expensive to run, but could now become more popular because of efficiencies.

Lensen said DeepSeek's impact might be to help U.S. companies learn "how they can use the computational efficiencies to build even larger and more performant models".

"Instead of making their model 10 times smaller and efficient with the same level of performance, I think they'll use the new findings to make their model more capable at the same energy usage."