The author is a Reuters Breakingviews columnist. The opinions expressed are his own.

By Robert Cyran

NEW YORK, Feb 3 (Reuters Breakingviews) - Bosses of big technology firms may find reassurance in comparing artificial intelligence to a lump of coal. Claimed cost-efficiency breakthroughs made by China’s DeepSeek have rocked markets betting on a boom in spending on the chips and servers powering silicon smarts. Analysts and executives like Microsoft’s MSFT.O Satya Nadella reason that this shouldn’t be an issue by appealing to long-dead economist William Stanley Jevons’ observation that more-efficient coal-fired engines drove higher fuel demand. Scrutinizing other, more recent cases shows why that dynamic doesn’t guarantee success for today’s incumbents.

Jevons, in an 1865 book called “The Coal Question,” laid out the seeming paradox. As steam engines improved, they needed less fuel for tasks like hauling trains. Yet coal demand in the United Kingdom kept rising. The reason is simple: railroad operators consumed less of it per train journey, increasing profits and encouraging them to run more locomotives. The trend is durable. Industrial processes keep getting more efficient, but in 2024, 159 years after the book was written, the world burned a record 8.8 billion tonnes of the stuff.

Like energy, intelligence is a necessary input for economic activity. High-margin products from medicine to software, not to mention the day-to-day running of any big corporation, require mental effort for everything from organizational coordination to cutting-edge research. Even a wildly expensive AI that can perform some of this work will find takers where its output is valuable enough. Lower the price, and the range of tasks to which it makes sense to apply it expands.

So when DeepSeek claimed it could train and run cutting-edge language models at a fraction of the cost of Western versions, it’s no wonder some people found comfort in coal. Sure, investors initially panicked, slicing nearly $600 billion off the market value of Nvidia NVDA.O, the dominant maker of chips used for AI, in one day. But Nadella, in a social media post, declared that “Jevons paradox strikes again,” adding that rising efficiency will cause AI use to “skyrocket.”

Yet Jevons can’t help if demand isn’t there or assumptions underlying investments are wrong. Three more up-to-date industrial examples exemplify the pitfalls.

Take hydraulic fracturing, which involves injecting shale formations with water, sand and chemicals to break open previously inaccessible deposits of natural gas and oil. Technological progress steadily reduced costs, leading to a massive rise in production. The U.S. is now the world’s largest producer of crude oil, with about two-thirds coming from fracking. Yet demand did not rise nearly as much. Companies that took on leverage to finance breakneck expansion couldn’t keep up, and waves of bankruptcies followed. Some 42 fracking companies bearing almost $26 billion in debt went under in 2019, according to the Institute for Energy Economics and Financial Analysis. Gas prices scraped along at all-time lows last year, the U.S. Energy Information Administration reported.

This might be unfair to AI. Unlike oil – the need for which is increasingly being displaced by technologies like electric cars – intelligence seems unlikely to go out of style. But that still might not prop up, say, a mooted valuation of $300 billion for ChatGPT-maker OpenAI if competition swamps the market. Just look at solar panels.

The cost of a U.S. solar system has fallen about 40% in a decade, according to the Solar Energy Industries Association trade group, with declines even steeper overseas. Unlike with gas, demand has exploded in response. Installed global solar power capacity increased 10-fold between 2013 and 2023.

Yet competition is fierce because panels are largely interchangeable. All that matters is price. Operating margins for the top ten manufacturers averaged roughly 0% in the third quarter of 2024, according to the National Renewable Energy Laboratory, despite record shipments.

AI insiders like Dario Amodei, the CEO of Anthropic, however, think that competition will thin out. He claims smaller, newer companies have been able to jump into the race because novel techniques that produce big gains are still emerging so rapidly. Over time, improvements will become harder and more costly, by his reckoning. That could mean few firms – or even just one – will dominate.

Yet limited competition and rapid growth do not guarantee extraordinary profit margins, either. Illumina ILMN.O, for instance, has dominated genetic sequencing for over a decade. In 2019, the Federal Trade Commission estimated it had an over-90% share of the market for advanced machines that identify the nucleotides in DNA samples.

These machines have potential uses from telling people where their ancestors came from to designing personalized cancer treatments. Efficiency improvements have been swift and substantial. The price of sequencing one person’s genome, their entire set of DNA, has fallen from roughly $100 million in 2001 to about $200 today, according to the company.

Yet even with nearly the entire market in its grasp, analysts expect Illumina’s revenue in 2026 to reach only $4.7 billion, according to LSEG data, not much different than in 2021. This is the result of a nasty combination that comes from having to outpace both a demand shortfall and rising competition. Genetic testing still needs to improve to be useful enough to draw in more customers; upstarts, meanwhile, are developing their own tests in an attempt to break through. Illumina must run faster just to stay in place.

Perhaps AI will escape all these traps, proving useful enough, cheap enough, but hard enough to copy that it results in not just an explosion of usage, but profit, too. It’s just a much trickier matter than simply offering up Jevons’ paradox as some kind of proof that less is always more.

Follow @rob_cyran on X

(Editing by Jonathan Guilford and Pranav Kiran)

((For previous columns by the author, Reuters customers can click on CYRAN/robert.cyran@thomsonreuters.com))