How DeepSeek Made Jevons Trend Again

Jevons paradox is trending, schooling social media posters in economic theory, particularly “marginal utility theory.”

“Jevons” had not been popular since the early aughts, according to a Google Trends scrape of internet use of the word since 2003.

And relatively speaking, it wasn’t so hot back at the turn of the century either, when Jevons was used to explain why the decline in chip prices and rise in compute power rapidly accelerated growth in digitization rather than diminish it.

In short, we just can’t get enough of it.

A home may have had one or two digital devices in 2003. Today, it could have dozens.

The week after DeepSeek roiled capital markets and sent analysts to revisit their modeling of chip and energy demand, “Jevons” scored 100 in use relative to any past all-time high, which was recorded in 2004 and demoted on Jan. 27 from 100 to 33.

Pre-DeepSeek week, the score was 3.

“Jevons paradox getting its moment in the sun,” wrote a Bluesky poster.

Note: The DeepSeek news is not yet confirmed to be a breakthrough. OpenAI is investigating whether DeepSeek didn’t “learn” what it knows by self-training. Rather, the investigation is of whether its knowledge is “distilled” from OpenAI.

This is different than simply using the OpenAI model, which it did, but it is open source and is allowed with a condition that prohibits distillation.

Also, there are doubts about how few chips DeepSeek used. Not many chips are needed in AI training in comparison with the knowledge quality and processing intensity it needs when anyone is using the AI—the “inference.”

And there are other unresolved concerns, such as whether to interact with China-based AI.

Andrew Munoz, COO of AI-enabled E&P asset valuation firm 4Cast, said at NAPE in February, “I’m not advocating using that model, by the way, because user beware: It’s not sourced from the U.S.”

In my inbox of energy research reports, there were 12 notes containing “Jevons” between 2013 and the week before DeepSeek’s news. In three weeks following, “Jevons” appeared in 10.

The theory was minted in 1865 by economist William Stanley Jevons, who determined that technological advances in energy efficiency—in his research’s case, with burning coal—would result in more demand, not less.

Since the Coal Age, it’s been proven correct for oil, natural gas, electricity and other commodities, including chips and data storage.

Microsoft Corp. CEO Satya Nadella jump-started trending Jevons, posting on X the Wikipedia link to “Jevons paradox” the evening before the world woke to the DeepSeek news.

“Jevons paradox strikes again!” he wrote. “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.”

Amazon celebrated DeepSeek’s news, too. CEO Andy Jassy was asked in an earnings call how it affected cost, uptake acceleration and returns on Amazon investments.

RELATED

Paradoxes and Power: Why DeepSeek AI May Be Good News for Energy
NatGas Prices, E&Ps Take a Hit from DeepSeeking Missile

Jassy replied, “Sometimes people make the assumptions that, if you're able to decrease the cost of any type of technology component—in this case, we’re really talking about inference—that somehow it’s going to lead to less total spend in technology.

“And we have never seen that to be the case.”

When Amazon Web Services was launched in 2006, storage cost $0.15 a gigabyte; compute, $0.10 an hour.

Clients could spend less on their in-house infrastructure. But what AWS has seen is that users “will then get excited about what else they could build that they always thought was cost-prohibitive before,” Jassy said.

“And they usually end up spending a lot more in total on technology once you make the per-unit cost less,” he said.

In AI, the cost of inference is declining and “going to be very positive for customers and for our business.”

Gokul Hariharan, tech analyst for J.P. Morgan Securities, wrote, “Even before DeepSeek R1’s introduction, AI inference costs have been dropping at 85% to 90% per year.”

As a result, there has been an explosion in AI use that didn’t exist or was too costly for wider consumption.

Meanwhile, declining costs for training AI models likely won’t result in lower spend.

“On the contrary, we believe they are likely to trigger bigger budgets, as more innovations are typically made possible within a shorter timeframe,” Hariharan wrote. 

What does this mean to previous models for power demand for AI?

Arun Jayaram, the firm’s E&P analyst, reported no change: 60 GW of additional U.S. power demand by 2028 versus the 2022 level.

Meta Platforms’ plan to spend $65 billion on AI this year is unchanged, it reported. Microsoft’s plans are for $80 billion and unchanged.

Meta’s chief AI scientist Yann LeCun posted on Threads, “Major misunderstanding about AI infrastructure investments: Much of those billions are going into infrastructure for ‘inference,’ not training.

“Running AI assistant services for billions of people requires a lot of compute. Once you put video understanding, reasoning, large-scale memory and other capabilities in AI systems, inference costs are going to increase.

“The only real question is whether users will be willing to pay enough—directly or not—to justify the capex and opex.

“So, the [stock] market reactions to DeepSeek are woefully unjustified.”